Apple is preparing to take the stage for what is expected to be its most consequential launch event in years. The keynote, scheduled for 10AM PT / 1PM ET, is not just another annual iPhone refresh. It is the first iPhone event led by John Ternus, who recently became Apple’s chief executive after a long tenure as the company’s head of hardware engineering. Ternus has already described the launch as a huge moment that will be phenomenal, and the rumor mill has spent months building anticipation for a foldable iPhone that could reshape the company’s flagship lineup.

For weeks, analysts, developers, and Apple watchers have been trading reports about what will appear on stage. The centerpiece is expected to be the iPhone 18 family, but the event could also introduce a new category of device: a foldable iPhone. If the rumors are correct, the launch would mark Apple’s first serious entry into a form factor that competitors have been refining for years. The event is also expected to bring updates to Apple’s operating systems, new Apple Watch models, and possibly new AirPods. With so much on the line, the keynote is being treated as a turning point for both the company and its hardware roadmap.

John Ternus Leads His First iPhone Keynote

John Ternus is not a newcomer to Apple’s stage, but this is his first iPhone keynote as CEO. Before taking the top job, Ternus spent years as vice president of hardware engineering, overseeing products that include the Mac, iPad, and iPhone. His background is deeply technical, and his elevation to CEO has been framed as a signal that Apple intends to keep hardware innovation at the center of its strategy. That emphasis is likely to be visible throughout the event, especially if the company unveils a foldable device that demands new hinges, displays, materials, and manufacturing processes.

Ternus has already set expectations high. He has called the launch a huge event and promised it would be phenomenal. Those words carry weight because Apple rarely pre-announces the significance of a product cycle in such direct terms. For a new CEO, a foldable iPhone would be a bold way to step into the spotlight. It would also give Ternus a defining product launch early in his tenure, much as previous CEOs used category-defining devices to establish their leadership. The event is therefore not only a product showcase but also a leadership statement.

The Foldable iPhone: Apple’s Next Halo Product

The most anticipated announcement is the long-rumored foldable iPhone. Early reports and leaks have pointed to a device that may be called the iPhone Duo, a name that echoes both the dual nature of a foldable screen and the idea of two devices in one. If Apple follows through, the iPhone Duo would be the company’s first foldable phone and a new halo product for the entire iPhone line. A halo product is a flagship that generates excitement, draws attention to the broader ecosystem, and pushes technological boundaries even if it does not sell in the same volumes as standard models.

A foldable iPhone would face high expectations. Samsung, Google, and several Chinese manufacturers have already released multiple generations of foldable phones, and each has grappled with trade-offs around durability, crease visibility, weight, battery life, and software adaptation. Apple’s approach is expected to emphasize polish and integration rather than being first. The company has reportedly been working on hinge designs, flexible display technology, and software features that let apps transition smoothly between folded and unfolded states. Apple has also reportedly provided design guidance to app developers for the iPhone Duo, a sign that the company wants the device to launch with a strong ecosystem of optimized apps.

The stakes are considerable. A foldable iPhone could redefine what a premium smartphone looks like and justify a higher price tag. It could also introduce new use cases, such as a larger inner display for multitasking, media consumption, and productivity. At the same time, Apple must avoid the pitfalls that have slowed foldable adoption elsewhere. If the company can deliver a device that feels durable, elegant, and genuinely useful, the iPhone Duo could become the most important iPhone since the original. If not, it could become an expensive experiment. The keynote will be the first real test of which outcome is more likely.

iPhone 18 Pro and Pro Max: Camera and Design Updates

Alongside the foldable, Apple is expected to announce the iPhone 18 Pro and iPhone 18 Pro Max. Rumors suggest the Pro models will feature a dynamic aperture main camera, a change that could give photographers more control over depth of field and exposure. A dynamic aperture would allow the lens to adjust the amount of light entering the sensor, similar to how a traditional camera works. That could improve low-light performance, portrait photography, and creative flexibility. It would also be a meaningful differentiator for the Pro line at a time when smartphone cameras have become increasingly similar.

The Pro models may also introduce new color options. One persistent rumor is that the black iPhone Pro is returning, a detail that may seem minor but often matters to longtime users who prefer a darker finish. Design changes are likely to be evolutionary rather than radical, but the combination of camera upgrades and a possible new finish could be enough to drive upgrades. Apple typically reserves its most advanced features for the Pro line, and this year is unlikely to be an exception.

There is also talk that the standard iPhone 18 may not arrive until the new year. If that happens, the September event would focus on the higher-end Pro models and the foldable, leaving the more affordable flagship for a later launch window. Such a split would be unusual for Apple, but it would allow the company to give its most expensive and innovative devices more attention. It would also create a longer upgrade cycle for buyers who are waiting for a standard iPhone 18.

Apple Watch, AirPods, and Software Updates

No Apple event would be complete without updates to the company’s wearables and operating systems. New Apple Watch models are expected, likely with incremental hardware improvements and new health or fitness features. Apple has steadily expanded the Watch into a health monitoring device, and each generation tends to add sensors, software capabilities, or design refinements. The next models could continue that trend, though the company may save major health breakthroughs for future releases.

AirPods may also make an appearance. Rumors of camera-equipped AirPods have circulated for some time, but reports suggest those devices are not ready yet. That may be for the best. Camera-equipped earbuds raise significant privacy questions and would need a compelling use case beyond novelty. For now, Apple is more likely to focus on audio quality, battery life, and integration with its other products. A new pair of AirPods could still appear, but the event’s biggest wearable story will probably remain the Apple Watch.

Software is another certainty. Apple typically uses its fall events to detail the next versions of iOS, iPadOS, watchOS, macOS, and other platforms. These updates often include new features for the iPhone, improvements to Siri and Apple Intelligence, and deeper integration across devices. If a foldable iPhone is announced, the software story becomes even more important. Apps will need to adapt to new screen sizes and form factors, and Apple will need to show developers and users why the experience is worth the investment.

Pricing, Memory Costs, and the Year of Smartphone Price Hikes

One of the biggest questions hanging over the event is price. The smartphone industry has been dealing with rising component costs, and memory prices have been especially volatile. The current climate, sometimes described as RAMageddon, has put pressure on manufacturers across the board. If Apple is forced to absorb higher memory costs, it may pass some of that expense to consumers. That could make the iPhone 18 Pro, Pro Max, and foldable iPhone more expensive than their predecessors.

A foldable iPhone would already be expected to command a premium price. Adding higher memory costs on top of advanced display and hinge technology could push the device into a new pricing tier. Apple has historically been willing to charge more for innovation, but there are limits. The company must balance the desire to protect margins with the need to convince customers that a foldable phone is worth the extra cost. The event may reveal not just what Apple has built, but how confident it is that buyers will pay for it.

Price hikes are not limited to Apple. The broader smartphone market has seen costs rise, and many Android manufacturers have raised prices on flagship devices. If Apple follows suit, it could signal that the era of stable flagship pricing is ending. That would have implications for upgrade cycles, carrier promotions, and the used-phone market. It could also make the standard iPhone 18, whenever it arrives, more important than ever as a more affordable entry point.

What Else to Watch

The event may include other surprises. Apple could announce new services, accessories, or partnerships. It could also provide more details about its artificial intelligence strategy, which has become a central theme across the tech industry. The company has been integrating AI features into its devices, and the iPhone 18 lineup is expected to benefit from those efforts. A foldable iPhone would give Apple another canvas for AI-powered multitasking, photography, and productivity features.

Developers will be watching closely. A new form factor requires new design thinking, and Apple’s guidance for the iPhone Duo will shape how apps look and behave. If Apple gets developer support right, the foldable could avoid the app-compatibility problems that have plagued some earlier foldables. If it does not, users may find that their favorite apps feel awkward or unfinished on the larger display. The company’s relationship with developers has been strained at times, but a successful foldable launch would require close collaboration.

Apple watchers will also be looking for clues about the company’s long-term roadmap. Will the foldable remain a niche product, or will it eventually replace the standard iPhone? How will Apple position the iPhone 18 Pro against the foldable? What does the new CEO prioritize? The answers may not come all at once, but the event will set the direction for the next several years.

Live Coverage Begins at 10AM PT / 1PM ET

The keynote is scheduled to begin at 10AM PT / 1PM ET. Live coverage will follow every announcement, from the first glimpse of the foldable iPhone to the final pricing details. The event is expected to be one of Apple’s biggest in years, with the potential to introduce a new product category and a new CEO’s vision. As the lights go down and the stage is set, the tech world will be watching to see whether the rumors match reality. The iPhone 18 era, and possibly the foldable era, begins now.

Source: The Verge News

At Apple Park, following the 'Surprise and shine' presentation, Apple gave attendees their first hands-on time with the iPhone 18 Pro and iPhone 18 Pro Max. The new models arrive with a familiar silhouette and a set of camera-focused upgrades that could matter most to photographers, videographers, and anyone who treats a phone as a primary creative tool. The initial impression is one of refinement rather than reinvention. The iPhone 18 Pro and Pro Max keep the same general design language as the previous iPhone 17 Pro lineup, including an aluminum body and the same screen sizes. In the hand, the phones feel like a continuation of Apple's established Pro formula: dense, premium, and carefully balanced. The biggest visible change is on the back, where Apple has introduced color-matched glass that blends more seamlessly with the aluminum frame. The effect is subtle but cohesive, making the rear panel look more like a single piece rather than a layered assembly.

The iPhone 18 Pro lineup is available in four colors: black, silver, glacier, and burgundy. Glacier is a shade of blue, and it gives the Pro line a cool, understated option that sits between the traditional silver and the deeper blue tones Apple has used before. Burgundy adds a richer, warmer choice for buyers who want something less neutral. The black and silver options remain the classic, professional-looking anchors of the lineup. Because the glass on the back is color-matched, each finish looks more integrated, with less contrast between the camera plateau and the rest of the rear panel. That might sound like a minor design note, but in a market where phones are often judged by their finish and feel, it helps the iPhone 18 Pro feel like a more considered object.

Dynamic Island gets smaller and more useful

Another change that users will notice immediately is the Dynamic Island, which is now smaller. Apple has also expanded its ability to display information. According to the company, the Dynamic Island can now show three Live Activities at once. In a brief hands-on test, this worked as described. That means a user could potentially keep an eye on a sports score, a food delivery, and a timer or ride-share update without opening multiple apps. The smaller cutout also frees up a bit more screen real estate, though the difference is not dramatic. Still, for a feature that began as a controversial workaround for the camera cutout, the Dynamic Island has become a genuinely useful part of the iPhone interface. Making it smaller while increasing its information density is a sensible step.

Variable-aperture camera brings pro-level control

The headline feature of the iPhone 18 Pro and Pro Max is the variable-aperture lens on the main camera. This is a significant addition because most smartphone cameras have a fixed aperture. A variable aperture allows the lens to open wider or close down, which changes how much light reaches the sensor and how much depth of field appears in a photo. A wider aperture, such as f/1.48, lets in more light and can create a shallower depth of field, helping subjects stand out against a blurred background. A narrower aperture, such as f/4.0, lets in less light but can increase sharpness across a broader scene. Apple has given users the option to let the aperture adjust automatically or to choose one of four selectable stops: f/1.48, f/1.8, f/2.8, and f/4.0.

That level of control is unusual in a phone. It moves the iPhone 18 Pro closer to the experience of shooting with a dedicated camera, where aperture is one of the fundamental exposure settings. The native camera app has also gained new pro controls. Users can manually adjust shutter speed, white balance, and focus. The focus control includes peaking in the preview, a feature that highlights the edges of in-focus areas so the photographer can confirm what is sharp. It is a small but thoughtful addition that many experienced shooters will appreciate. For years, pro-level camera apps have offered these kinds of controls, but having them built into the default camera app reduces friction and makes them accessible to a wider range of users. The result is a camera system that can be as simple or as technical as the user wants.

What the camera controls mean in practice

In practical terms, the variable aperture could change how iPhone photographers approach different scenes. For portraits, a wider aperture can produce a more pronounced background blur without relying entirely on software-based portrait mode. For landscapes, stopping down to f/4.0 could help keep more of the scene in focus, from foreground details to distant horizons. For low-light situations, the widest aperture can gather more light, potentially reducing noise and improving shutter speeds. The ability to switch between automatic and manual aperture also means users can decide how much control they want in the moment. A casual shooter can leave it in auto and let the phone make the call. A more advanced user can select a specific stop and pair it with manual shutter speed, white balance, and focus for a more deliberate exposure.

The focus peaking feature is particularly useful for video and for still subjects that require precise focus. When shooting through glass, foliage, or other complex scenes, it can be difficult to tell exactly where the camera has locked focus on a small screen. Peaking overlays a visual cue, making it easier to confirm. Combined with manual focus, this gives the iPhone 18 Pro a more credible claim to being a creator's tool. It is not a replacement for a full-frame mirrorless camera, but it narrows the gap in terms of control and feedback.

Design and display: familiar but refined

The iPhone 18 Pro and Pro Max look quite similar to the previous iPhone 17 Pro lineup. They keep the same general design, aluminum body, and screen sizes. That means buyers who liked the flat-edged, squared-off look of recent Pro models will find more of the same here. The Pro Max remains the larger option, while the standard Pro offers the same core features in a more compact frame. The color-matched glass on the back is the most obvious aesthetic update. It gives the phones a more uniform appearance, especially in the glacier and burgundy finishes. The camera module still stands out, but the overall effect is cleaner. The aluminum body provides a solid, premium feel, and the phones are likely to be as durable as their predecessors, though Apple has not detailed any new drop or water-resistance claims in the hands-on session.

The smaller Dynamic Island is another visual change. It reduces the black cutout at the top of the display, though it is still present. The ability to view three Live Activities at once is more of a software feature, but it makes the island more useful. Live Activities can display real-time information from apps, such as sports scores, delivery progress, ride status, timers, and music playback. Being able to see three at once could be convenient for multitaskers, though it may also require users to be selective about which activities they keep active. In the brief demo, the feature worked without obvious lag or confusion.

Pricing, preorders, and release date

The iPhone 18 Pro and iPhone 18 Pro Max will be available to preorder starting September 12th ahead of a September 18th release date. The iPhone 18 Pro starts at $1,199, while the iPhone 18 Pro Max starts at $1,299. Those prices place the new models at the premium end of the smartphone market, which is consistent with Apple's Pro strategy. The company positions the Pro line as the choice for users who want the best camera, display, and performance in an iPhone. The addition of variable aperture and pro camera controls reinforces that positioning. For anyone upgrading from an older Pro model, the biggest draws will likely be the camera system, the smaller Dynamic Island, and the new color options. For buyers coming from a standard iPhone, the Pro line remains a significant step up in price, but it also offers the most advanced features Apple has put into a phone.

Broader event context

The hands-on session took place as part of Apple's 'Surprise and shine' event, which included several other announcements. Among the related stories from the event were the unveiling of the iPhone Duo, Apple's first foldable, and hands-on impressions of that device. There were also discussions about the black iPhone Pro returning, the lack of AirPods with cameras, and the incomplete history of Duo devices. The event clearly had a lot for Apple fans to follow, but the iPhone 18 Pro and Pro Max remain the centerpiece for many users. The Pro line is Apple's most profitable and most closely watched iPhone tier, and each year it sets expectations for the rest of the smartphone industry. This year, the focus is less on a dramatic redesign and more on giving photographers and videographers more control.

Competitive landscape and what comes next

The smartphone market has seen a trend of price hikes, and the iPhone 18 Pro's starting price reflects that. Competitors continue to push camera hardware, display technology, and AI features. Apple's response with the iPhone 18 Pro is to lean into professional-grade camera controls and a refined design. The variable aperture is a hardware feature that is difficult to replicate with software alone, which could give Apple an edge among photography enthusiasts. At the same time, the company is balancing that with ease of use. The automatic aperture mode ensures that casual users do not need to understand f-stops to get good results. The manual controls are there for those who want them, but they do not get in the way.

As for what comes next, the hands-on session left some questions unanswered. Apple did not provide detailed specifications about the chip, battery life, or display brightness during the brief time with the devices. Those details will likely emerge in the coming days as reviewers spend more time with the phones. What is clear is that the iPhone 18 Pro and Pro Max are iterative upgrades in design but meaningful upgrades in camera control. The smaller Dynamic Island and three Live Activities add convenience. The color-matched glass and new finishes add polish. And the variable aperture, paired with manual shutter speed, white balance, and focus, gives the Pro line its strongest camera-focused pitch in years.

The phones will be available for preorder on September 12th and will release on September 18th. The iPhone 18 Pro starts at $1,199, and the iPhone 18 Pro Max starts at $1,299. For users who have been waiting for Apple to bring more professional camera controls to the native app, the wait appears to be over. For everyone else, the iPhone 18 Pro offers a familiar, premium experience with a few new tricks that could make a real difference in everyday photography and videography.

Source: The Verge News

The proposal, shared on Sunday, describes a final network snapshot, the issuance of ERC-20 ONE tokens on Ethereum, and the migration of exchange listings. Validators would be offered several paths: stop operating their nodes, continue in a governance role, or join a new AI-video initiative backed by Harmony’s team. Harmony called the proposal non-binding and did not specify when the final block would be produced or whether the shutdown would be submitted to the network’s validator-led governance process.

How Harmony’s governance would handle a shutdown

Harmony has published formal governance rules that would govern any decision to close the network. Under those rules, elected validators can create proposals and unelected validators may vote, with voting power based on total stake. For a proposal to pass, at least 51% of the total stake weight must participate, and 66.7% of votes cast must support the measure after a seven-day introduction period and a 14-day voting window.

Because Harmony is a proof-of-stake network, validators hold significant responsibility for both block production and network security. The proposal acknowledges that winding down an L1 chain is not an isolated technical event; it requires coordinated action by every participant who operates infrastructure. The non-binding nature of the current proposal leaves room for the validator ecosystem to debate whether migration is preferable to continuing to run Harmony as an independent chain.

ONE migration and token snapshot

Under the proposed migration, all ONE balances would be recorded at Harmony’s final block, and new ERC-20 tokens would be airdropped to the same addresses on Ethereum. The snapshot would cover wallets, staking delegations, validator rewards, smart contracts, and centralized exchanges, with no claim process required. That design aims to reduce the burden on ordinary ONE holders and ensure that tokens are not stranded when Harmony’s chain stops producing blocks.

The proposal also addresses obligations to users who hold ONE through third-party services. By including centralized exchange balances in the snapshot and asking exchanges to migrate their listings, Harmony hopes to avoid a scenario in which exchanges are left holding illiquid tokens with no market or withdrawal path. However, Harmony warned that some onchain use cases cannot be automatically moved.

Multisig safes, liquidity pools, and other onchain applications cannot be migrated, according to the proposal. Harmony urged users to exit all smart contracts before Sept. 10, which is the earliest date validators may begin shutting down. The proposal also sets aside a $1.372 million pool to compensate validators that stop their nodes on time, retain their stakes, and agree to serve as governors during the transition. That compensation is an attempt to incentivize an orderly transition rather than a sudden abandonment of the network.

Exploit led to rollback plans

Harmony’s shutdown proposal comes less than four weeks after an exploit created forged ONE tokens and led the team to plan a rollback that would wipe out more than 109,000 transactions. The incident marked a turning point for the project, shifting from repairing a compromised network to potentially ending it as an independent blockchain.

On Aug. 12, Harmony said it was considering a rollback after reports surfaced that an attacker had minted nearly 4 billion unauthorized ONE. That amount represented roughly 26% of the circulating token supply. An outside account later claimed that approximately 2.8 billion of the forged tokens had reached exchanges, but Harmony had not confirmed those figures at the time.

By Aug. 17, Harmony said it planned to revert the blockchain to an Aug. 11 checkpoint, discarding 109,126 regular transactions and 315 staking transactions. Investigators said they had traced nearly all of the forged tokens to specific wallets or service boundaries and were working with exchanges, bridges, and law enforcement to prevent the stolen tokens from being monetized.

A brief history of Harmony and its security issues

Harmony launched its mainnet in 2019 with the goal of providing fast, low-cost throughput for decentralized applications, using a sharded proof-of-stake architecture to increase capacity. The project positioned itself as an Ethereum bridge network and attracted developers building games, DeFi protocols, and NFT marketplaces. For a time, Harmony was viewed as one of many Ethereum-compatible alt layer-1s seeking to offer lower fees and faster finality than Ethereum.

Prior to the August exploit, Harmony had already faced a major security crisis. In June 2022, the Horizon bridge, which connected Harmony to Ethereum, was hacked, and around $100 million in cryptocurrency was stolen. The attacker moved funds to various chains and used mixers to obscure the trail. Harmony offered a bounty for the return of the stolen assets and eventually introduced a recovery proposal for token holders affected by the bridge breach. The bridge incident damaged confidence in Harmony’s security and contributed to a long period of declining activity on the network.

The recent exploit drew on the same kind of concern about token supply integrity. Forged ONE tokens could be used to manipulate markets, governance, or exchange balances, making it difficult for the network to continue with confidence. A rollback was one proposed response, but a rollback at this scale is disruptive because it reverses transactions that may include legitimate user activity. The decision to propose a full migration to Ethereum suggests that Harmony’s team sees little long-term value in maintaining a separate chain after repeated security challenges and shrinking network usage.

Why migrate to Ethereum?

Migrating a token from a layer-1 blockchain to Ethereum is not a new idea. Several projects that launched competing smart contract platforms have eventually moved their asset issuance to Ethereum to take advantage of its security, liquidity, and ecosystem depth. Ethereum is the dominant settlement layer for DeFi and tokenized assets, and an ERC-20 version of ONE would allow holders to trade, lend, or use the token in a broad range of applications without depending on Harmony’s own validator set.

Harmony has highlighted the benefits of migration in the proposal, pointing to Ethereum’s established infrastructure and the simplicity of ERC-20 tokens for exchange support. For many users, the practical difference between holding ONE on Harmony and holding an ERC-20 version of ONE on Ethereum would be minimal on a daily basis, especially if exchanges and custodians manage the technical transition on their behalf.

Still, the migration would mark the end of Harmony as a settlement network. Shutting down the layer-1 means no further native block production, no native smart contract execution, and no independent security budget. Any projects still building on Harmony would need to move their protocols to Ethereum or another chain. Harmony’s warning that smart contracts cannot be migrated is therefore a critical caveat; it places responsibility on users and developers to extract their positions before the network stops.

Validator incentives and regulatory context

The proposed $1.372 million validator compensation pool is relatively small compared to the costs of operating network infrastructure for years. But it signals that Harmony is aware of the need to retain validators through the transition. Validators that accept the offer would become governors, possibly providing a layer of oversight during the token migration. Others may simply choose to stand down, which could accelerate the shutdown if a critical mass of validators stops producing blocks.

Harmony’s approach also raises governance questions. If validators do not reach a consensus under the published rules, Harmony could face a contested shutdown or the emergence of an unofficial chain operated by dissenting validators. Token holders would then be forced to choose which network represents the true version of Harmony, potentially splitting the ONE asset. The proposal tries to avoid this by making exchange migration and token issuance contingent on a clear plan, but final authority rests with the community.

The migration plan also has regulatory implications. Exchanges and custodians that list ONE must agree to rebrand the token or issue the ERC-20 equivalent to their customers. Those decisions may require compliance reviews, especially if the token’s status changes as a result of the migration. Law enforcement involvement in the exploit case means that wallets associated with the attack may remain frozen or under investigation.

Harmony’s latest announcement is a sign of how severe the August exploit was. Instead of trying to patch the chain and restore trust, the project is now offering a path that effectively dissolves Harmony into Ethereum. If approved, the final snapshot will create a permanent record of ONE balances and produce a new token on Ethereum, while the original chain comes to an end. For users who still have assets on Harmony, the message is to act before Sept. 10 and ensure that any smart contract positions are closed in time. For validators, the message is to weigh the proposed incentives and decide whether to support the transition or oppose the shutdown. The coming days will determine whether Harmony’s proposal gains enough stake support to become binding.

Source: Cointelegraph News

According to NAPP, more than 20 merchants have already agreed to accept HUMO for goods and services. The program extends beyond a simple payment test, as participating banks will also connect their existing payment-processing infrastructure with blockchain-based systems. Asterium, a licensed cryptocurrency exchange in Uzbekistan, will act as a project partner and facilitate the integration of the digital asset ecosystem with traditional banking rails.

How the Pilot Works

HUMO is described as a stablecoin pegged to the national fiat currency, the Uzbek som. Unlike many privately issued stablecoins that rely on reserves held in banks or commercial paper, HUMO is backed by government securities. This design aims to reduce counterparty risk and tie the digital currency to the credit quality of the state, providing a local alternative to global stablecoins such as Tether or USD Coin. The idea is to combine the efficiency of blockchain payments with the stability and trust of sovereign debt.

The initial phase of the trial will last for 12 months, according to the Central Bank of Uzbekistan. The entire project, including any extensions or additional testing phases, is capped at three years. During this period, regulators will assess a range of factors, including the adequacy and safeguarding of the token’s collateral, cybersecurity protections, consumer safeguards, anti-money laundering controls, and the risk HUMO might pose to financial and price stability. These criteria suggest that the Uzbek authorities are taking a cautious approach and will use the sandbox environment to collect empirical data before deciding any next steps.

Regulatory Framework and Institutional Support

The pilot follows a broader legal and regulatory framework that Uzbekistan approved in November 2025. At that time, the government signaled its intention to allow stablecoin payment trials within a controlled sandbox, alongside provisions for tokenized shares and bonds. Monday’s announcement is the first concrete implementation of that framework, and it places Uzbekistan among a small group of countries experimenting with stablecoin-based payments within a formal regulatory perimeter. Rather than granting a blanket approval, the authorities appear to be testing the concept with a limited set of participants and a strictly defined timeline.

NAPP’s role is central to the development. The agency has positioned itself as the main driver of blockchain and crypto-related innovation in Uzbekistan, overseeing a licensing process for digital asset service providers and seeking to balance innovation with financial oversight. By co-managing the sandbox with the central bank, NAPP is bridging two normally distinct areas: technology-friendly sectoral regulation and monetary policy. The collaboration suggests that payment systems involving digital fiat-pegged tokens need the support of both the innovation watchdog and the institution responsible for monetary sovereignty.

Context of Uzbekistan’s Crypto Movement

Uzbekistan has been gradually opening its cryptocurrency market, though with some caution. In recent years, the government has taken steps to license exchanges, legalize certain crypto activities, and attract international players. The NAPP has introduced registration procedures for crypto companies and has been active in shaping the legal environment for blockchain technology. By experimenting with a government bond-backed stablecoin, Uzbekistan is aiming to develop a domestic electronic payment instrument that could eventually improve financial inclusion and modernize the country’s digital economy.

The choice of a government securities-backed stablecoin is particularly notable. In many markets, stablecoins are backed by cash reserves or a basket of high-quality assets. Government bonds are considered low-risk because they are backed by the issuing state. In the case of Uzbekistan, the backing could provide a strong guarantee for the token’s value, as long as the collateral is properly reserved and managed. However, the use of sovereign bonds as collateral can introduce some complexities, especially in times of market stress. Fluctuations in bond prices could theoretically impact the value of the reserve pool, though the token is still pegged to the som and the central bank may be involved in maintaining stability.

Another important angle is the involvement of Humo Digital. Humo is a name associated with Uzbekistan’s national payment system, Humo, which is operated as an interbank payment processing network. By leveraging an existing payment brand and integrating with participating banks, the stablecoin pilot could potentially reach a larger user base quickly. Humo’s infrastructure is part of the country’s financial landscape, and linking the digital token with traditional banking networks might create a smoother user experience for making payments with the stablecoin.

Merchant Adoption and Payments

The number of participating merchants is still modest, but it is significant for an initial test. More than 20 merchants have been prepared to accept HUMO payments for goods and services, which means the pilot will not be limited to internal transfers or exchange listings. These merchants will receive HUMO through the payment system or their banks, and the transactions will be processed on blockchain rails integrated with bank systems. The pilot will thus test the commercial viability of stablecoin payments in real-world scenarios, including point-of-sale transactions, settlement times, and user experience.

Asterium’s role as a licensed crypto exchange adds another layer. The exchange will likely provide the market infrastructure for converting HUMO to and from the Uzbek som and possibly handle liquidity for the merchants. The participation of a licensed exchange could also help regulators monitor the flow of funds and ensure that anti-money laundering requirements are met. The exchange’s infrastructure will need to be capable of supporting the stablecoin’s issuance and redemption processes, as well as recording transfers on the blockchain.

Supervisory Considerations and Risk Assessment

The Central Bank of Uzbekistan has outlined the key risk areas that will be monitored during the pilot. One of them is the adequacy and safeguarding of the collateral backing HUMO. Regulators will want to verify that each token is indeed backed by one som worth of eligible assets and that those assets are held in a protected and transparent manner. The risk of over-issuance or unauthorized dilution is one of the primary concerns in stablecoin regulation, so the sandbox regime will likely require third-party audits or regular reporting.

Cybersecurity is another core area. Stablecoins and blockchain payment systems are exposed to hacks, smart contract bugs, and onboarding vulnerabilities. If a platform handling government-segment securities suffers a breach, the fallout could affect the banking system and consumer confidence. The sandbox will need to demonstrate high security standards before the project scales. Additionally, consumer safeguards are important, especially for citizens who may not be familiar with stablecoin technology. Ensuring that users can recover their funds and understand the risks will be a critical metric for success.

Anti-money laundering and countering the financing of terrorism controls are also included in the list of review areas. Because stablecoins can be transferred across borders quickly, they require robust know-your-customer processes and transaction monitoring. The integration with bank systems and a licensed exchange gives the regulator more visibility, but it also adds operational complexity. Finally, the project’s potential effect on financial and price stability will be assessed. If a government bond-backed stablecoin were to grow too quickly, it could interact with the broader financial system in unintended ways. The limited scope and strict timeline are meant to prevent these risks from materializing on a large scale.

Regional and Global Implications

The launch comes amid broader moves in the Central Asian region to explore state-adjacent digital finance. In neighboring Kyrgyzstan, law makers recently introduced a bill proposing the creation of a state crypto reserve. While that concept is different from a stablecoin pilot, it illustrates how several post-Soviet Central Asian states are looking to integrate digital assets into their national development strategies. Uzbekistan’s initiative, with its emphasis on government bond backing and formal sandbox supervision, might serve as a template for other emerging economies seeking to leverage blockchain technology without adopting fully private, unregulated stablecoin systems.

At a time when global stablecoin regulation is intensifying, the Uzbekistan pilot offers a practical case study. International bodies such as the International Monetary Fund and the Bank for International Settlements have called for clear rules around stablecoin issuance and reserves. By running a controlled experiment under the oversight of both the central bank and a specialized agency, Uzbekistan is pre-emptively addressing some of these international recommendations. If successful, the HUMO trial could provide valuable data on the operational use, collateral management, and regulatory scalability of a stablecoin that is linked to both a national currency and sovereign debt.

As the pilot unfolds, market participants will be closely watching the merchant experience and the technical reliability of the integrated system. The partnership between Humo Digital, banks, and Asterium will need to function seamlessly while satisfying regulators on security and compliance. With the initial trial set to run for 12 months and the total project capped at three years, the Uzbek government has created a time-boxed environment where successes and failures can be analyzed without jeopardizing the broader financial system. The end result could shape the country’s approach to digital currency for years to come.

Source: Cointelegraph News

Key facts

• Housing and technology companies are beginning to explore small AI data center deployments in residential spaces.
• Major organizations involved in early discussions include homebuilder PulteGroup, chip maker Nvidia, and energy management company Span.
• High mortgage costs and a desire to generate income from underused home spaces are driving homeowner interest.
• Business demand for lower-cost, geographically distributed compute is creating a potential market.
• Different models could evolve, including controlled edge-host programs, decentralized compute marketplaces, and infrastructure brokering.
• Residential power, heat, security, regulatory, and trust issues remain serious obstacles.
• The realistic future is selective micro-hosting in suitable homes, not large-scale replacement of data centers.

Economic forces at work

The timing of residential hosting interest is not accidental. Homes are expensive. Homeowners who bought at elevated prices and high interest rates are carrying heavy monthly mortgage payments. Property taxes and insurance costs continue to rise. Across many housing markets, people are looking for ways to turn underused parts of their homes into sources of recurring income. Spare rooms have already become short-term rentals. Garages have become workshops or small living units. Rooftops have become solar assets. The next frontier could be basements, utility rooms, and detached structures hosting small-scale server equipment.

This shift also reflects pressure on businesses. AI has rapidly increased demand for processing capacity. Edge workloads continue to grow. Not every application needs to run in a hyperscale facility. Not every company wants to pay hyperscale prices. A residential server rack might offer lower real estate costs, faster deployment, or a location closer to end users. The strategic question that businesses are already asking is how much compute can be pushed to distributed locations without losing operational control or economic advantage.

A third force is cultural. More homeowners than ever understand racks, uninterruptible power supplies, network switches, remote access, and basic thermal management. The gap between enterprise infrastructure knowledge and prosumer infrastructure knowledge has narrowed. That shift makes the residential data center concept feel possible, even though the barriers to executing it at commercial quality remain high.

Business models taking shape

There is not yet a large, polished market where everyday homeowners host third-party servers the way they host guests with a short-term rental. What exists instead are adjacent business models that point in that direction without fully embracing residential colocation.

One model is the controlled edge-host program. In that arrangement, a company places or manages compute equipment in selected distributed locations. The company sets strict standards for connectivity, power, security, and maintenance. The homeowner or site operator is not acting as an open hosting provider. They participate in a curated network where the provider controls the service architecture and the quality of the physical environment.

Another model is the decentralized compute marketplace. Platforms let individuals and smaller operators sell spare compute capacity from their own hardware. This approach is closer to the economics of monetizing residential infrastructure. However, selling compute cycles is not the same as taking custody of someone else’s physical server and being responsible for the room where it runs. Compute marketplaces create value, but they do not yet solve the environmental and liability problems of housing hardware in homes.

A third model is the infrastructure broker. Brokerage companies already match buyers and sellers for colocation, bare-metal services, and other infrastructure needs. Their existence shows that brokering data center relationships is viable. But those relationships usually connect enterprises to professional facilities, not homeowners with racks next to their water heaters. The residential market remains unfinished because trust, standardization, and liability models are still underdeveloped.

The clear appeal of residential hosting

The strongest positive element is financial. If a homeowner can generate meaningful monthly income by hosting compute equipment, the idea will always attract attention. This is especially true in markets where monthly housing costs are high and supplemental income feels necessary. Hosting a carefully sealed rack of servers might seem more stable and less socially disruptive than renting out rooms to a constant lineup of short-term guests.

There is also an asset utilization argument. Many homes contain spaces that produce no return. A basement corner, an unused garage bay, or a storage room can sit empty for years. Once an infrastructure provider pays for space, power, and connectivity, the home starts to participate in the digital economy rather than simply serving as shelter. The homeowner converts a passive area into an active source of revenue.

For enterprises, the appeal is equally clear in specific cases. Residential locations can offer lower operating costs, faster deployment, and better geographic spread. In areas where electricity is relatively inexpensive and broadband is strong, small residential hosting sites could fill gaps that do not justify a full commercial data center. Homes will not replace data centers. They might, however, complement them in a narrow set of circumstances where latency, cost, or deployment speed matter more than enterprise-grade physical certainty.

The heavy weight of downsides

The problem is that the downsides are not minor. Residential power is not data center power. Residential broadband is not enterprise-grade networking. A private home is not a secure, redundant, environmentally controlled facility. No amount of careful rack installation changes these fundamental limitations.

Power is the first wall. Most homes are not designed for sustained commercial server loads without electrical upgrades. Those upgrades can be expensive, regulated, and dependent on local utility approval. When backup batteries, uninterruptible power supply systems, cooling equipment, and dedicated circuits are added, the project begins to look like a facilities operation instead of a side hustle. The initial economics can quickly evaporate.

Heat and noise follow. Commercial hardware generates both around the clock. This affects the comfort of the house, the cost of climate control, and the long-term reliability of the equipment. The environment inside the home changes. What was once a living space can begin to feel like a machine room. Maintenance becomes routine. Monitoring becomes constant. The homeowner starts to live inside the rhythm of a server environment.

The risk list is even more serious. Fire hazards. Water damage. Physical theft. Tampering. Insurance complications. Zoning restrictions. Homeowners’ association objections. Lease restrictions for renters. Questions about physical access and liability. Rules that protect sensitive or regulated data. All of these issues can be managed in theory, but they are the precise reasons professional data centers exist. Trying to solve them in a residential environment creates enormous complexity.

Trust is the missing foundation

Customer trust may be the biggest obstacle. Most businesses buy compute from recognized providers because the provider offers a predictable operating environment. That assumption collapses when infrastructure sits in someone’s living space.

What happens during a storm? Who controls the door to the equipment room? What happens if the homeowner accidentally unplugs a critical circuit? How are incidents documented and reported? What happens when a neighborhood power event damages customer hardware? These questions are not edge cases. They determine whether a residential hosting model can survive.

Trust can be built through contracts, remote monitoring, strict certification, and regular inspections. But those requirements become expensive. At some point, a managed residential micro-site starts costing nearly as much as a modest commercial colocation facility. If that happens, the value proposition disappears.

A realistic path forward

Residential data hosting is unlikely to become the next major hosting model. Professional data centers remain superior in most cases because they eliminate variable environmental conditions and provide robust security, redundant power, fire suppression, and clear operational accountability. Those qualities are difficult to reproduce in a private home.

Still, the concept should not be dismissed. Some regions may have conditions that make residential micro-hosting viable. Cheap power. Upgradeable electrical panels. Strong broadband. Detached buildings. Favorable local zoning rules. Workloads that benefit from geographic distribution and do not require pristine enterprise conditions. In those situations, carefully managed residential hosting could fill a useful niche.

The likely future is not an Airbnb for random servers. It is not whole neighborhoods converted into basement data centers. It is a selective market where curated providers match specific homeowners or small properties with specific infrastructure needs under tightly controlled agreements. What begins as a niche could still create meaningful change in where compute lives, how it is managed, and who benefits from the new economics of the AI age.

Source: InfoWorld News

Take a common pricing comparison cited in current industry discussions. NVIDIA H100-class compute can cost about $2.01 per hour on Spheron, while AWS pricing for a similar workload category lands around $6.88 per hour. That is approximately 3.4 times more expensive for comparable AI processing. Some enterprises may negotiate better rates, but the existence of such a wide gap changes the conversation. Buyers now know that credible lower-cost alternatives exist, and knowledge changes behavior.

Key facts at a glance

• Hyperscalers can cost roughly 3 to 6 times more than specialized AI cloud providers for similar GPU compute.
• An often-cited example puts NVIDIA H100-class compute at about $2.01 per hour on Spheron and roughly $6.88 per hour on AWS, a 3.4x gap.
• Enterprises are increasingly treating AI infrastructure as a long-term operating expense rather than a short-term experiment.
• Private cloud, sovereign cloud, and on-premises GPU deployments are gaining traction as cost-conscious alternatives.
• Workload placement is replacing the assumption that one hyperscaler should run every AI job.

Why the cost gap is now strategic

The pricing gap is not a rounding error. Enterprises can no longer dismiss it as the price of doing business with a trusted vendor. The numbers are significant enough to influence architectural choices, vendor strategy, and even where AI innovation takes place. When compute capacity is the core input of an AI workload, continuous differences in unit cost become a major factor in long-term financial modeling.

A workload does not become more valuable simply because it runs in a hyperscale cloud. The chip is still the chip. The cluster is still the cluster. The economics are still the economics. A customer does not receive higher model accuracy just because the invoice came from a household cloud brand. No amount of ecosystem polish can change the fact that the same GPU class is available elsewhere for a fraction of the price.

In addition to neoclouds, private clouds, sovereign clouds, and even on-premises GPU strategies are becoming more attractive. Many enterprises now view AI infrastructure as a durable operating expense rather than a temporary experiment. Once that mindset takes hold, even small differences in unit cost become strategically important. Large cost differences become very hard to justify. A premium vendor then stops looking premium and starts looking overpriced.

When premium status is not enough

Hyperscalers have benefited for years from a straightforward value proposition. They provide global reach, mature security controls, integrated tools, elastic capacity, and an ecosystem that reduces operational friction. Those factors still matter and remain valuable. The difficulty is that AI is exposing a flaw in traditional cloud pricing models. When compute is the core of the workload and can be sourced elsewhere at a considerably lower cost, the value of the surrounding ecosystem must be exceptional to justify the markup. In many AI scenarios, it is not.

This is where hyperscalers may be making a strategic mistake. They appear to assume that AI buyers will accept the same pricing strategies that worked for traditional cloud migrations. That assumption is risky. AI buyers are not simply lifting and shifting old enterprise applications. They are training, fine-tuning, and deploying models in environments where utilization, throughput, latency, and token economics are monitored in real time.

The pressure is also coming from outside the technology team. Boards are asking tougher questions about AI spending. Investors are asking harder questions about returns. Finance teams are asking the toughest questions of all. If the answer is that an enterprise is paying multiple times more for the same class of compute because it is easier to stay with a familiar brand, that answer will not hold up well in budget reviews.

AI buyers are becoming more rational

The next phase of the AI market will not belong to the companies that generate the most headlines. Success will go to providers that consistently deliver reliable performance at sustainable costs. That shift favors disciplined operators that are optimized for GPU availability, efficient scheduling, and straightforward commercial models. It also favors enterprises willing to blend different environments instead of relying on the largest cloud vendor for every workload.

The conversation is moving away from cloud preference and toward workload placement strategy. Enterprises are becoming comfortable with the idea that different AI jobs belong in different places. Some workloads will remain on hyperscalers because integration benefits are real. Others will go to private cloud because security, data gravity, or regulatory requirements demand it. Still others will land on sovereign platforms because national and industry-specific rules leave no other option. A growing number will be routed to neoclouds because the price-performance equation is simply too compelling to ignore.

This shift is not necessarily a rejection of hyperscalers. It is a rejection of careless pricing. The biggest cloud providers will continue to be highly important for AI. But their role is changing from default choice to one option among many. That is a significant strategic downgrade, driven not by technological weakness but by pricing behavior.

A familiar cycle in cloud economics

The cloud industry has experienced this pattern before. Established companies believe their size protects them, that customers prioritize convenience above everything else, and that their pricing power will last indefinitely. Then a new group of competitors appears with a sharper value proposition and fewer outdated assumptions. Incumbents dismiss them as niche players. The challengers improve, specialize, and attract the most cost-conscious innovators. By the time incumbents respond seriously, the market has already moved.

That is precisely the risk hyperscalers face in AI today. If they continue treating GPU-driven workloads as an opportunity to maintain high margins across compute, storage, networking, and managed services, they will train customers to look elsewhere. Once that habit forms, it will be difficult to change. Customers who build procurement discipline around lower-cost AI infrastructure will not quickly return just because a hyperscaler eventually cuts prices.

The pricing power of the largest clouds was never guaranteed. It was built on scarcity, complexity, and the absence of meaningful alternatives. In AI workloads, scarcity has faded, complexity is no longer mysterious, and alternatives are multiplying. Hyperscalers still have genuine advantages in scale, security, and integration. But those advantages are no longer enough to justify price gaps of three to six times.

The next winners in AI infrastructure may be the providers that understand a hard truth. When a market is growing as quickly as AI is growing, adoption matters more than margin preservation. If AWS, Azure, and Google do not learn that lesson soon, they may discover that they were not defeated by competitors at all. They may simply have priced themselves out of the AI market all on their own.

Source: InfoWorld News

In practice, AI-first development usually means moving toward agentic workflows. Developers act as architects and supervisors who frame problems, guide AI agents, review the output, and ensure quality. It is also a transformation of how developers explore and understand code, with a larger focus on designing applications that are AI-native and ready for autonomous digital workers.

Why AI-first development matters

Software development teams that rely on traditional human-only processes are increasingly finding themselves at a disadvantage. AI-first development is becoming a necessity because the old way of working is no longer sustainable when market demands are growing and delivery schedules are shrinking.

Teams that use AI throughout the development pipeline can accomplish far more in less time. Instead of spending days or weeks on repetitive coding tasks, planning migrations, and updating aging libraries, engineers can clear work that used to sit in backlogs for quarters in a matter of days. Moreover, AI-first development helps surface design-level defects long before production. Issues that would previously become incidents and require an emergency fix are instead caught during design and code review, saving time, money, and reputation.

Another major reason is that applications are increasingly being built for a world where AI agents are end users. Bolting AI onto an existing architecture after launch is risky. It is like adding a second floor to a house without checking if the foundation can support it. The result might hold up for a while, but when it breaks, it can break catastrophically. Designing with AI in mind from the start produces cleaner data flows, clearer permission schemes, and user interfaces that understand exactly how agents will interact with the software.

From a business point of view, retrofitting AI into projects is expensive and complicated. In contrast, AI-first designs lead to cleaner architectures and lower total costs. For example, a team that plans permission boundaries in advance avoids having to clean up fragmented access-control mechanisms later. And when an AI-first product is designed deliberately, it becomes a new kind of product rather than a legacy application with a chatbot attached.

The impact on documentation and clarity

A less obvious but powerful benefit of AI-first development is that it forces teams to improve their documentation. AI cannot read minds. When developers design with AI in mind, they naturally produce richer specifications, clearer requirements, and more explicit acceptance criteria. This clarity helps both the human team and the AI systems that rely on structured instructions.

The result is better thought processes across the organization. Teams are more deliberate about how they define system boundaries, outline expected behavior, and document edge cases. They write down what used to be implicit knowledge, which makes knowledge transfer easier and makes software easier to maintain and extend.

How to succeed at AI-first development

Not every team is ready for this major change. Many organizations struggle because their engineering culture depends on undocumented knowledge, implicit rules, or manual oversight. AI systems require precise context. If a team works with legacy codebases that are poorly documented or rely on obscure architecture, it might end up with AI-generated code that looks plausible but requires substantial human rework.

Fortunately, there are practical ways to build the capabilities and the culture required to thrive in an AI-first environment. The following strategies have worked for experienced engineering leaders and can be applied by any team moving in that direction.

Build new skills and roles

The first step is to invest in new skill sets and new roles. Without the right people and capabilities, nothing else will work. The scarce skill in the age of AI is no longer writing code; it is the ability to quickly read generated code and assess it accurately. Senior developers are increasingly acting as architects who frame the problem, direct the AI agent, and verify the result. With the right architecture in place, one engineer can safely and effectively oversee the work of several AI assistants.

Organizations should therefore focus staffing on senior architecture roles. They should also create new entry-level or junior roles that focus on orchestrating agents rather than writing boilerplate code. Junior developers can be trained to manage the handoffs between agents, understand when conflicts arise, and decide which agent should have priority. This is not programming in the traditional sense, but it is a natural evolution that takes advantage of human judgment and product knowledge.

In addition, user experience expertise needs to be embedded throughout the development process, not just at the beginning or the end. AI can write requirements, create technical specifications, and help with architecture, but it does not deeply understand real users. Observing actual user behavior, analyzing usability tests, and making human-centered decisions still require human expertise.

Create training programs to ease the transition

Many organizations are still unfamiliar with the concept of an AI-first development workflow. Developers and managers need training to understand what it means to be an architect of AI systems. The role of the architect is changing quickly, and most organizations have not fully caught up.

This does not just mean learning how to prompt an AI tool. It means designing boundaries that say where the agent can act, where the human must decide, and what happens when the agent is wrong. Teams that ignore the last question often discover the problem too late. Training should emphasize decision-making, prompt design, and robust verification methods.

Embrace agentic workflows

Agentic workflows are at the heart of AI-first development. Autonomous agents use reasoning, planning, and external tools to achieve complex goals, and humans stay in the loop to provide direction and oversight. The fundamental transition here is from typing code to specifying outcomes. Engineers who thrive in that environment are people who can define clear system goals, think through edge cases in advance, and evaluate output objectively.

Writing code will never stop being useful, but the highest value work in a mature AI-first organization is architecture and editorial judgment. It is the ability to say a new AI workflow should require a human checkpoint before sending output to a customer, or an agent should not be allowed to modify production data. That kind of engineering excellence does not decrease in importance; it becomes even more valuable.

For individual developers, daily work changes substantially. They spend less time manually writing repetitive code. They invest more time defining goals, considering product constraints, iterating on architecture, designing user experiences, envisioning data

Source: InfoWorld News

This shift is not just a technological evolution; it is a practical response to some of the hardest problems in AI operations. From GPU shortages to the complexity of model inference, the barriers to entry have historically been steep. Serverless AI lowers those barriers by wrapping models behind simple APIs. You send data and receive a prediction, while the provider handles everything from hardware provisioning to autoscaling. The appeal is obvious, but as with any architecture, there are hidden costs and behavioral trade-offs that demand careful scrutiny.

The scalability advantage

One of the most celebrated benefits of serverless AI is the promise of effortless scalability. In traditional AI deployments, you have to guess how many GPU instances you need, often months in advance. If you underprovision, requests are throttled; if you overprovision, you see wasted spend and idle hardware. Serverless turns this on its head. The service dynamically adjusts resources behind the scenes, expanding during peak traffic and shrinking when demand fades. You pay only for what you consume, typically measured in tokens or inference seconds.

Consider a retail business gearing up for the holiday season. Recommendation engines powered by machine learning models might need to handle ten times the normal workload in November and December, then slump back to baseline in January. With serverless AI, that business avoids paying for year-round dedicated infrastructure that sits idle most of the time. The same applies to event-triggered workloads, such as image recognition pipelines that run only when new files are uploaded, or natural language processing tasks that occur sporadically across regions.

This elasticity also solves the latency problem of cold GPU nodes. In a traditional setup, scaling up after a demand spike means waiting minutes for instances to initialize. Serverless offerings, especially those with warmed pools or optimized inference pipelines, can spin up capacity in near real time. For user-facing AI, this responsiveness is crucial to maintaining the experience and staying competitive in a market where speed is a feature. The result is that serverless AI has become the default choice for startups and enterprises alike that want to avoid infrastructure overhead.

The hidden costs of pay-per-use

But serverless AI is not without its drawbacks, and the hidden costs can be significant for certain patterns. The most commonly cited problem is the unit price. Because you are paying a premium per request or per token, consistent, high-volume usage will almost always be more expensive than a dedicated approach. For a workload that needs to run 24/7 at a steady state, negotiating a fixed price for reserved virtual machines or GPU clusters will generally yield far lower costs than metered serverless pricing.

Moreover, serverless AI can undermine predictability. In a legacy architecture, you control the budget based on a fixed set of hardware. With serverless, a sudden spike in requests, whether from a viral news story or a runaway background job, can lead to a bill that multiply exceeds what you planned for. This is a very real risk that catches many teams off guard. Without careful monitoring and robust cost controls, the same flexibility that enables scale also introduces financial volatility.

Cold starts represent another challenge. While providers have optimized over time, the first request after an idle period may incur added latency while containers spin up and models load into memory. For real-time applications that demand sub-second responses, this unpredictability can be unacceptable. Developers then find themselves implementing complex concurrency controls, custom warmers, or hybrid strategies that undermine the very simplicity that drew them to serverless in the first place.

Fit architecture to use case

Architecture is about trade-offs, and serverless AI is no different. The decision should hinge on workload characteristics, not hype or a general preference for modern technology. Variable, spiky, or seasonal workloads are the natural domain of serverless AI. The pay-as-you-go model shines when the alternative is paying for rare peak capacity. On the other end, workloads with stable, predictable traffic are better served by reserved infrastructure, where you can optimize utilization and reduce marginal cost through dedicated accelerators and custom inference pipelines.

The challenge is that the industry has a tendency to apply simplistic mantras, and serverless AI is often touted as the ideal for everything. This echoes the early cloud migration era, when companies moved to cloud without rearchitecting, expecting massive savings that never materialized. The thinking must instead be nuanced. Ask yourself: Is this workload truly elastic? Are there upstream events that trigger bursts? What is your tolerance for variance in response time? How sensitive is your finance team to unexpected charges?

There is also the question of data control. Serverless AI often means sending sensitive data to a third-party model endpoint, which creates compliance concerns in sectors like healthcare and finance. Many providers now offer virtual private cloud (VPC) endpoints and privacy layers, but the trade-off between convenience and data sovereignty is another factor. For some enterprises, running open-source models on their own infrastructure remains a better fit, simply because of the regulatory environment and security requirements.

Another angle is skill set. Serverless AI appeals to application developers who want to embed intelligence without diving into the weeds of model deployment. Meanwhile, teams with substantial machine learning engineering expertise may see serverless as leaving performance on the table. A dedicated approach allows for model quantization, kernel fusion, batch scheduling, and other optimizations that can drastically cut inference costs. Those optimizations are usually impossible when you are simply calling an API, because the underlying architecture is hidden from you.

The role of hybrid strategies

Given the strengths and weaknesses, the most successful organizations tend to adopt hybrid patterns. Mission-critical, high-volume components run on dedicated capacity, while experimental or burstable features use serverless AI. For instance, a customer support platform might run a high-volume sentiment analysis model on a reserved cluster, but use a serverless endpoint to handle occasional email summarization. This way, you maximize efficiency in the steady state and keep flexibility for irregular tasks.

There is also an evolving middle ground: managed inference platforms that let you bring your own model and your own instance pool, with autoscaling policies that you control. Providers like Amazon Bedrock and Google Cloud Vertex AI have started supporting custom provisioning modes beyond simple serverless requests. They offer provisioned throughput for near-zero cold starts, and spot-capacity for cost-saving on less urgent workloads. This allows architects to blend pay-per-use with reserved capacity, aligning pricing with traffic patterns more closely than ever before.

In the same vein, solutions like GPU-backed Kubernetes clusters and infrastructure as code have made it easier to build internal serverless environments. But maintain that takes serious effort. Only enterprises with a clear operational mandate and dedicated site reliability engineering teams should consider that route. For everyone else, managed offerings are often the safer survival strategy, because they delegate maintenance and security updates to the provider, freeing your own engineers to focus on product features rather than infrastructure.

Cost control is central to survival

Whether you adopt serverless, dedicated, or a hybrid model, managing the financial dimension is essential. The default inclination is to consider compute unit cost alone, but the total cost of ownership includes developer time, security audits, monitoring, and the opportunity cost of delayed product delivery. Serverless AI is often justified not because it yields the lowest cloud bill, but because it dramatically shortens time to market. A feature that takes days instead of months to build represents real money, but that advantage is lost if the variable cost later makes the product unprofitable.

This is why robust observability is the bedrock of cost management. Teams need to track tokens, invocation counts, duration, and cold start latency at a granular level. They need to set budgets and alerts, and enforce resource tags so that every service is attributed to a cost center. Without these measures, the dynamic nature of serverless AI can quickly become a budget disaster. Many cloud vendors offer built-in cost explorers and AWS, Azure, and Google all provide estimation tools, but they are only effective when the left hand of engineering talks regularly to the right hand of finance.

Open-source and local models are reemerging as a competitive pressure on serverless AI prices. As companies like Meta, Mistral, Alibaba, and a host of startups release high-performance open-weight models, the cost of running your own infrastructure has dropped dramatically. Smaller models, such as Google's Gemma or Microsoft's Phi, are capable for many narrow tasks and are free of per-token fees besides the electricity and hardware. As a result, teams that would earlier have defaulted to provider-managed foundation models are choosing open-weight models for certain workloads, and serverless for others.

The look ahead

As the AI landscape matures, the line between serverless and dedicated is blurring. Providers are introducing more flexible control planes, and inference engines are becoming more efficient at utilizing GPU capacity. In the future, we may see autonomous resource allocation where an AI model itself predicts traffic, and spins up the optimal mix of dedicated and on-demand capacity without human intervention. That would be the ultimate survivalist adaptation: building systems that not only withstand change but proactively respond to it.

Until that day arrives, the key takeaway is consistency and patience. Treat serverless AI as a tool in a larger toolbox. Define your workload profile, benchmark your cost constraints, and test candidate architectures under realistic load. There is no one-size-fits-all, and every cloud provider has different pricing and performance quirks that need to be evaluated in the context of your specific application demands.

Developers and architects who internalize these lessons will be well positioned to ward off the inevitable challenges that come with deploying artificial intelligence at scale. They will avoid the trap of overreacting to a single metric, and instead look at the complete picture of what it takes to deploy and operate machine learning in production. The servers might not be on your physical premises anymore, but the survivalist spirit of preparation, adaptability, and pragmatism lives on in the architecture.

You might also want to consider the user experience angle of serverless AI. In real-world systems, the user experience is what ultimately determines success, and serverless AI can introduce variability in response times that directly affects how users perceive an application. If you are building a chatbot, for instance, a 200-millisecond cold start is often tolerable. A 5-second cold start is not. Careful product design can hide latency where necessary, using human-like buffering or progressive display, but there is no technical rule that applies universally. User expectations and tolerance must guide the architecture.

What about model refresh cycles? A dedicated deployment can be updated without orchestrating through a provider’s system, giving you faster control over versioning. Serverless AI abstracts that away, which is a blessing for the overwhelmed engineer, but it also means that changes in an upstream model might alter outputs without prior notice. Model updates are inevitable, and the strategies for managing them, such as canary releases and version pinning, can be more complex in a serverless environment, especially if you are building an application that needs to guarantee consistent behavior for regulatory reasons.

There is also the need for a robust fallback plan. With serverless, you are dependent on the provider’s availability. If that service experiences an outage, your application might have no way to carry out inference at all. Amazon, Microsoft, and Google run highly reliable infrastructure, but even the biggest providers can be disrupted. A multidisciplinary survivalist reserves some low-capacity dedicated endpoint or an on-premise model that can be called when the cloud variant is down. It does not need to handle the full load, but it must be capable of still providing the core functionality until the remote service recovers.

Training and fine-tuning are not always part of the serverless picture. Most serverless endpoints offer fully pretrained models that are not easily customized. You can typically use prompt engineering or fine-tuned adapters through the API, but you still lose the ability to train a completely bespoke model with proprietary data and custom architecture. In niche domains where accuracy rests on specialized knowledge, a traditional inference pipeline is often the only way forward. The hidden cost in that scenario is less about compute dollars and more about lost accuracy and domain adaptation.

Cost reporting is another arena that becomes more intricate in multi-serverless environments. In FaaS (function-as-a-service) or PaaS (platform-as-a-service) modes, every invocation can be traced back to a specific build or team, but in AI APIs, the same model is often used by multiple products. Without good accounting, cost allocation becomes messy, and teams start to argue over who owns the budget. This can lead to political tension and eventually hamper innovation. To avoid this, design a tagging and retagging strategy at the application level early on, and carry that discipline forward as you segment your use cases.

Security is also substantially different from self-managed infrastructure. When you expose an API, you are widening your attack surface. Model poisoning, prompt injection, and data exfiltration are all risks that must be mitigated through input validation, output monitoring, and strict access control. The provider may encrypt data in transit and at rest, but you still have to deal with the security of your own client applications. For survival in the wilds of modern cloud computing, one can never overlook the human factor. A developer with a leaked API key can result in enormous bills or even a regulatory violation. Therefore, secret management, key rotation, and granular permissions become survival tactics in and of themselves.

The future of AI computing is not going to be a choice between serverless and dedicated alone. It will be a spectrum. Cloud providers are increasingly offering not just model endpoints, but also serverless training jobs, distributed inference at the edge, and even serverless GPU clusters that you can reserve on demand. This environment is changing so quickly that today’s trade-offs may be irrelevant in a few years. The best preparation is not to hunt for a perfect one-time architecture, but to cultivate an architecture culture that can dissect new options as they arrive and tests them rigorously against actual needs.

In practical terms, setting up a decision template can save huge amounts of time. Begin with a non-functional requirements document that includes latency, throughput, availability, cost ceilings, and geographic distribution. Then run a small proof of concept, using simulated traffic on a serverless offering and on a dedicated instance. Measure both the technical metrics and the operational overhead of each option. Only after gathering evidence, choose the approach. This empirical spirit helps organizations dodge most of the pitfalls that come from following industry trends without adequate validation.

The survivalist’s guide to serverless AI is ultimately a guide to disciplined engineering. It means discovering that good architecture is often unglamorous, and that careful evaluation outweighs cleverness. As more enterprises integrate generative AI into core products, the basic questions of scale, cost, trust, and sustainability remain the first order of business. No matter how many new models are released, the success of an AI project will still depend on the soundness of the systems around it. Serverless AI is a powerful tool, but you need the wisdom to know when to use it, when to sidestep it, and when to combine it with other resources for the best outcome.

Source: InfoWorld News

CHARLOTTE, N.C. — September 3, 2026 — The Broadcast Retirement Network (BRN) today is marking seven (7) years of daily, advertising-free, fact- and evidence-based programming dedicated to retirement, aging, finance, lifestyle, privacy, and wellness. 

Since its launch in 2019, BRN has aired more than 2,500 original programs, broadcasting seven days a week at 7:30 AM ET. Over that period, the network has produced more than 625 hours of programming featuring international experts sharing their expertise with the audience. 

BRN programming is made available at no cost, complete with full transcripts, and is broadly syndicated across major news sites, aggregation services, streaming platforms and podcast services—ensuring BRN’s content reaches its audience wherever media is consumed. Shorter clips are also distributed across all major social media channels every two hours, further extending the network’s reach.

BRN distinguishes itself through a straightforward editorial commitment: no sales pitches, no advertisements, and no politics—just the facts, every morning. The network also delivers a daily, hand-curated newsletter The Morning Pulse on aging, finance, lifestyle, privacy, retirement, and wellness, selected by an expert editor.

“Reaching our seventh anniversary is a testament to the trust our audience and partners have placed in us,” said Jeffrey Snyder, Chief Executive Officer and Lead Anchor of the Broadcast Retirement Network. “For seven years—drawing on my 32 years in the retirement industry, our mission has remained the same: to deliver clear, credible and useful information to the people who need it, free of any external noise.

A September-Only Anniversary Offer

To mark the occasion, BRN is offering new subscribers 20% off The Morning Pulse—its daily newsletter delivering expert-curated news on money, health, and retirement, written by a human, without the use of AI, and free of ads and sales pitches. Every subscription directly supports BRN’s daily programming. Read today’s edition here: https://us6.campaign-archive.com/?u=26e6dd4c63255e9ef5e07a4c4&id=152e013f21

Normally $4 per month, new subscribers can save 20% with code BRN20 through September 30, 2026, only. Individuals can subscribe at https://buy.stripe.com/4gw9CS5NI8cibVm7su.

Looking Ahead

BRN will unveil a new opportunity for prospective partners on September 9, 2026. Details will be shared across the network’s platforms and social media channels.

“This milestone belongs to our guests, audience and partners as much as it does to us,” Snyder added. “We are grateful for seven years of continued support—and we’re just getting started.”

About Broadcast Retirement Network

The Broadcast Retirement Network (BRN) is an independent daily program delivering fact-based news and expert insight on aging, finance, lifestyle, privacy, retirement, and wellness. Airing seven days a week at 7:30 AM ET, BRN provides advertising-free programming, complete with transcripts, syndicated at no cost across major news, streaming, and podcast platforms. BRN is led by Chief Executive Officer and Lead Anchor Jeffrey Snyder, who brings 32 years of retirement industry experience to the network’s daily coverage.

Media Contact

Jeffrey Snyder 

Chief Executive Officer / Lead Anchor 

Email: jeff@broadcastretirementnetwork.com 

YouTube: https://www.youtube.com/@BroadcastRetirementNetwork

###

• Email: jeff@broadcastretirementnetwork.com
• Website: https://www.youtube.com/@BroadcastRetirementNetwork
• Company: Broadcast Retirement Network
• Company Logo:
• Name: JEFFREY H SNYDER
• Phone No: 6464840188
• City: Charlotte
• State: NC
• Country: United States

<p>The post Broadcast Retirement Network Marks 7 Years of Daily, Fact- and Evidence-Based Programming first appeared on PR Business News Wire.</p>

AI safety experts reacted with alarm almost immediately. Redwood Research CEO Buck Shlegeris said in a public post that he was extremely concerned by reports that Astra uses opaque recurrence. He wrote that he did not know whether Astra is much less chain-of-thought monitorable than previous models, but if OpenAI pushes this technique further, the company could massively increase recurrence and undermine or even destroy chain-of-thought monitorability.

Longtime AI safety advocate Zvi Mowshowitz also weighed in, warning that the technique is playing with fire. He noted that OpenAI and Anthropic have fought to establish a valuable taboo around maintaining chain-of-thought faithfulness and monitorability for as long as possible. In his view, more intensive use of such techniques would probably damage monitorability and could spark a race to the bottom among AI labs unless regulators step in.

Why the chain of thought became a safety tool

To understand the concern, it helps to understand how modern AI reasoning systems are audited. Under normal circumstances, a reasoning model's chain of thought records the sequence of tokens the model generates before declaring an answer. For example, a model asked to solve a math problem might generate a plan, apply each step, and only then produce the final result. Those intermediate tokens give safety researchers a transcript of sorts.

The transcript is imperfect. Leading labs have long acknowledged that chain-of-thought logs do not offer a complete or literal map of what happens inside a neural network. Some reasoning occurs in hidden states that never surface as language, and some outputs may be rationalizations rather than true causes of the model's decision. Still, the chain of thought remains one of the most valuable tools available for monitoring misbehavior or misalignment. It can help evaluators determine whether a model misunderstood a task, followed hidden instructions, or silently adopted a dangerous objective.

During recent reports of autonomous agents behaving in unexpected ways, chain-of-thought records helped investigators understand why the agents made the choices they made. Without such records, safety teams would be left to compare inputs and outputs and guess at everything that happened in between. That is why any technique that makes the intermediate reasoning less visible is treated as a serious development.

How recurrent depth hides reasoning

Recurrent depth is a more technical concept than a simple policy change. In a conventional reasoning model, text is generated one token at a time. The model builds from the prompt and from everything it has written so far, which creates a readable trail of the model's apparent thinking. Recurrent depth changes that process by allowing the network to revisit the same internal representation multiple times in a loop before producing an output.

Rather than expanding a problem into a long series of written thoughts, the model compresses its work into repeated rounds of internal computation. Each pass can refine the model's understanding of the query without producing an external token. The result is a system that appears to make leaps that are difficult to trace, because the intermediate states were never turned into human-readable language.

All AI models perform some amount of opaque reasoning. Even the earliest neural networks computed hidden states that users could not inspect, and few researchers treat chain-of-thought logs as a direct representation of a model's actual reasoning. The concern is one of degree. If an architecture shifts too much work into this kind of loop, the legible chain of thought may end up being little more than a summary written after the real decision was already made. In the worst case, monitoring systems would lose the thread entirely.

What OpenAI has said

OpenAI has pushed back against the idea that Astra is an uninterpretable black box. The company says the model's use of recurrent depth is limited, and its chain of thought is still expected to be legible in most cases. OpenAI also rejected any suggestion that it would move to neuralese, a term often used to describe the compressed, non-human language that models may use when reasoning in latent space.

In a post, OpenAI chief scientist Jakub Pachocki emphasized the lab's commitment to legible chains of thought. He wrote that OpenAI has worked to preserve and utilize chain-of-thought monitoring since its first reasoning models, and he described that goal as a core part of the current research program. The company has already announced plans for extensive chain-of-thought monitoring systems within its forward-looking safety framework.

Those assurances have not fully calmed observers. The worry is not necessarily that Astra itself will be impossible to monitor, but that the technique could be scaled up in future models. If recurrent depth becomes more powerful or more efficient, labs may feel pressure to use it more heavily, even if doing so reduces transparency. That pressure is the heart of the race-to-the-bottom concern.

Concern is spreading across major labs

OpenAI is not the only lab interested in the approach. A follow-up report on Wednesday morning said that researchers at Anthropic and Google DeepMind were already discussing opaque recurrence. The exact uses are still unclear, but the fact that multiple frontier labs are paying attention suggests this is not an isolated experiment.

The broader AI field has been moving toward test-time computation, meaning models are allowed to think longer and harder before answering. Most of that progress has been made visible through chain-of-thought-style reasoning. But there are strong economic and technical reasons to make reasoning more efficient. Opaque recurrence could deliver many of the benefits of deeper reasoning without the cost of writing out long chains of thought. The same efficiency, however, makes the reasoning harder to audit.

Redwood Research chief scientist Ryan Greenblatt said the natural trajectory from here could be especially dangerous. He argued that opaque reasoning could easily scale faster than conventional chain-of-thought reasoning, effectively stripping almost all decision-making out of visible channels. He said his biggest concern is that a natural progression would involve scaling up opaque reasoning until the model reasons entirely, or almost entirely, in latent space.

The problem is not only intentional deception. Models do not need to be trying to hide their reasoning to make it inaccessible. The architecture itself can remove visibility. Monitoring systems designed for a world of legible reasoning may become far less reliable if that world changes quickly.

Interpretability research may eventually provide new tools for peering into these hidden computations. Causal tracing, activation probes, and other techniques are advancing, but they are still young and difficult to deploy at scale. Regulators and external auditors are only beginning to develop evaluation methods for frontier models. If the models begin reasoning in opaque spaces before those methods mature, some of the most important safeguards could become obsolete before they are fully built.

What makes recurrent depth especially consequential is not that it is a sudden failure of safety, but that it is a quiet architectural shift in the opposite direction from transparency. The current debate around AI safety has often assumed that models will continue producing readable thoughts that can be checked and monitored. A technique like opaque recurrence challenges that assumption.

Whether OpenAI stops here remains an open question. For now, Astra's use of recurrent depth is limited, and the company has publicly committed to chain-of-thought monitoring. But the method is no longer hypothetical. It is already in a deployed model, and other major labs are paying close attention. Safety researchers, meanwhile, are hoping the industry treats this as a warning rather than a template.

Source: TechCrunch News

The Real World AI Stage arrives at a time when AI is no longer confined to text generation and image recognition. Autonomous drones are operating in conflict zones, humanoid robots are being tested in warehouses, and machine learning tools are accelerating genetic engineering. Disrupt 2026 intends to highlight these developments, offering a deep dive into what it takes to build AI systems that interact with the real world. The event is scheduled to take place October 13 through 15 at Moscone West in San Francisco, with more than 10,000 startup, technology, and venture capital leaders expected to attend.

A Split Stage for a Growing Industry

Past Disrupt events have featured a single AI stage, but the rapid evolution of the field has made it necessary to separate conversations about software, models, and enterprise AI from those about robotics, autonomous systems, and physical infrastructure. The traditional AI Stage will continue to address large language models, business model disruption, security concerns, and the impact of AI on software companies. The new Real World AI Stage will instead concentrate on the convergence of digital intelligence with tangible machines, sensors, and actuators.

This split mirrors the broader split within the AI industry itself. On one side are companies like OpenAI and Anthropic, which focus on large-scale models and reasoning. On the other are organizations such as Nvidia, which builds the hardware and software platforms that power physical AI, and Colossal Biosciences, which uses AI in genetic research to bring back extinct species. The Real World AI Stage aims to capture both the engineering challenges and the ethical questions that arise when AI is given a body and a purpose in the physical world.

Bridging the Data Gap for Robots

One of the most anticipated sessions is titled “Robots Are Waiting for Their ChatGPT Moment. Here Is What Is Standing in the Way.” The session will feature Les Karpas, Head of Physical AI at Nvidia. Karpas is expected to discuss why robots have not yet experienced the kind of capability breakthrough that large language models did when they were trained on vast amounts of internet data. While LLMs had access to trillions of words, robots lack equivalent datasets. Self-driving cars have logged millions of miles, but general-purpose robots do not have a comparable reservoir of real-world interactions to learn from.

That data gap is widely considered the primary reason why general-purpose robotic intelligence remains years away. Startups are now racing to solve this problem by building data pipelines, simulation environments, and foundation models specifically designed for physical AI. Nvidia has positioned itself at the center of this effort, providing the computing platform that enables robot developers to train and test their systems in virtual environments before deploying them in the physical world. Karpas’s session will explore what a so-called ChatGPT moment for physical AI would actually require and whether the industry is approaching that milestone or still far off.

Safety and Trust in High-Stakes Environments

Another key session is “Building AI Systems When Failure Is Not an Option,” featuring Nate Michael, Chief Technology Officer of Shield AI. Shield AI is known for developing autonomous aerial systems used in defense and security applications. When AI is deployed in a military drone, a commercial aircraft, or an industrial vehicle, the consequences of a mistake are severe. A software failure can result in a crash, a grounded mission, or the loss of life. This session will address how founders and engineering leaders can create a safety culture, validate AI systems through rigorous testing, navigate regulatory hurdles, and earn public trust when stakes are high.

Michael’s experience with Shield AI provides a concrete example of these challenges. The company’s autonomy software is designed to allow drones to operate without GPS and without a human pilot, relying on onboard sensors and AI to navigate uncertain environments. Building such systems requires more than just high-accuracy algorithms; it requires verification and validation processes that can identify failure modes before they occur. The discussion will likely touch on how companies can balance speed with caution and what frameworks exist for determining when a system is safe enough for real-world deployment.

De-Extinction and the Role of AI in Biology

One of the more unusual conversations on the Real World AI Stage will feature Ben Lamm, CEO and founder of Colossal Biosciences. Lamm has turned de-extinction into a multi-billion-dollar venture, aiming to resurrect species such as the woolly mammoth and the thylacine. In a fireside chat titled “Can We Engineer Nature’s Comeback?” Lamm will discuss the technologies his company has developed to edit genomes, create synthetic embryos, and use artificial intelligence to analyze biological data.

The session will also consider whether engineering nature is truly a conservation breakthrough or a distraction from protecting existing species. Colossal’s work has drawn both excitement and criticism, but its ambition highlights an emerging area where AI and biology intersect. Machine learning algorithms are being used to predict gene function, optimize gene edits, and assemble complete genomes from ancient DNA. AI’s role in such work is essential, as the sheer complexity of genetic data requires processing power far beyond human capability. Lamm’s appearance at Disrupt 2026 offers a rare, in-depth look at how a startup approaches a scientific challenge that was once considered purely fictional.

When the Cloud Can’t Keep Up

Another panel, “Operating at the Edge: How AI Works When the Cloud Doesn’t,” will bring together Dr. Ali Agha, CEO and founder of FieldAI, Michelle Lee, CEO and founder of Medra, and Aidan Madigan-Curtis, a partner at Eclipse Ventures. These leaders work in the fields of defense, space, and industrial AI. Their session will focus on systems that must operate in remote or hostile environments where cloud connectivity is limited or impossible. In such conditions, AI must run locally, with minimal latency and no ability to ask a central server for guidance.

Edge AI is increasingly critical for autonomous vehicles, agricultural equipment, ocean exploration, and space missions. FieldAI, for instance, builds autonomy software for off-world rovers and landers, preparing them to operate on the Moon, Mars, and beyond. Medra is working on AI-driven medical devices that need to function in ambulances or other locations without stable internet. This panel will share practical lessons on architecture, design trade-offs, and system reliability, offering insights that can be applied to any AI deployment that must survive disconnected, unpredictable conditions.

Scaling From Prototype to Production

“From Prototype to Production: Can It Scale in Reality?” is designed to address a common challenge for deep tech startups. The panel includes John Mackey, CEO and co-founder of MBRYONICS, Boris Sofman, co-founder and CEO of Bedrock Robotics, and Adrian Macneil, CEO of Foxglove. The conversation will explore why many startups fail to make the leap from a working prototype to a profitable commercial product. A lab demonstration can work flawlessly, but manufacturing at scale introduces issues related to supply chains, quality control, cost, and reliability.

MBRYONICS is involved in space hardware, Bedrock Robotics builds humanoid robots for industrial settings, and Foxglove provides data management and visualization tools for robotics companies. Each founder has direct experience bridging the gap between engineering and operations. Attendees can expect to hear about mistakes they have made, decisions they would reverse if given a second chance, and how they learned to design for manufacturing from day one.

More Stages and Speakers at Disrupt 2026

The Real World AI Stage is just one part of a larger programming lineup at Disrupt 2026. The main Disrupt Stage will feature top executives from Replit, Amazon, Tether, and other major players. The AI Stage will dive into the security gaps and business model changes that AI is forcing on software companies. The Smart Money Stage will cover stablecoins, instant payments, and how AI is reshaping financial trust. The Smart Systems Stage will look at energy breakthroughs, grid strain, and infrastructure concerns driven by AI’s growing power demands. The Builders Stage will be a practical resource for founders and investors interested in fundraising, hiring, and scaling their ventures.

Disrupt 2026 is also set to include the popular Startup Battlefield, where early-stage companies compete on stage, along with networking sessions and an exhibition floor filled with innovative businesses. Organizers have announced early-bird pricing, with discounts available on tickets. A promotion is currently offering up to $300 off, and anyone interested in attending can register online for the October event at San Francisco’s Moscone West.

Source: TechCrunch News

The initiative arrives at a time when Apple has been under pressure to respond to breakthroughs from rivals like OpenAI, Google, and Microsoft. While those companies focused on standalone ChatGPT-style products and cloud-based assistants, Apple's strategy is distinctly different. Apple Intelligence is not a separate application. It is not a web portal. It is a layer of intelligence woven into the operating system that can draw on relevant data from apps, messages, calendar events, and other personal information stored on the device. This approach creates a system that can act across apps, understand context, and maintain a strict boundary between personal data and the models that process it.

A New Model for Personal AI

At its core, Apple Intelligence uses a combination of large language models, diffusion models, and a new type of inference engine that runs locally on the device. Rather than sending every request to a remote data center, the system tries to process as much as possible on the Apple Silicon chip. That means tasks like summarizing a web article, generating a notification response, or finding a specific photo can happen without an internet connection. The on-device approach provides lower latency, improves responsiveness, and significantly reduces the risk of exposing sensitive information to external servers.

Apple has also created a foundation model specifically tuned for its ecosystem. The model, which Apple says has been trained on licensed and publicly available data, is smaller and more efficient than the enormous cloud-based models used by some competitors. Because it runs locally, the model can be more responsive to natural language orders and can preserve a user's personal context without copying that context into a company server.

However, not every task can be handled entirely on-device. When a user is asking for more compute-heavy output or a piece of information that requires access to a much larger knowledge base, Apple Intelligence uses a system called Private Cloud Compute. This architecture allows the iPhone, iPad, or Mac to send only relevant requests to dedicated Apple servers that run the most capable models. In a notable privacy feature, those servers are built with Apple Silicon, run in a hardened environment, and are configured so that no Apple employee can access the data. Users receive verification that the server they are connected to is running a secure, logged-less configuration. Apple has said that data sent to Private Cloud Compute is not stored, not used for training, and is only available for the duration of the request.

The Reinvention of Siri

The most visible change inside Apple Intelligence is the reinvention of Siri. After years of criticism that Siri had fallen behind more advanced digital assistants, the new Siri is designed around large language models and understanding of natural language. It can maintain context across turns, allowing users to say things like "Play the song from yesterday's playlist" and then follow up with "Turn it up" without needing to repeat the song name.

Siri also gains the ability to act across apps. For example, a user could say, "Show me the photos from my beach trip and make a slideshow," and Siri will handle the necessary coordination between the Photos app, the file system, and possibly a presentation app. It can summarize incoming notifications, respond to text messages with a tone the user would prefer, and take actions inside apps through a new API called App Intents. Developers can publish metadata that allows Siri to respect user permissions and control app features.

This new Siri is not simply a voice tool. It is deeply integrated into the operating system through a typewriter-style interface that can be activated on keyboards. On the Apple Vision Pro, Mac, and iPad, users can also type requests to Siri rather than speaking them. That changes the fundamental interaction pattern. Users do not have to be in a quiet room or awkwardly talk to their computer when they want help. Siri will eventually rely on a new language model that can better reason about user-related data, find answers in emails, retrieve meeting times, and even identify people, packages, and items in photos.

Writing Tools for Every Text Field

One of the most widely used parts of Apple Intelligence will likely be the interconnected feature set known as Writing Tools. In any application where text can be written or edited, Apple Intelligence can serve as a proofreader, style assistant, and rephraser. Users can choose to rewrite selected text in a friendlier, more professional, or more concise tone. The system can generate a list of key points from a full email or article, so users can respond quickly without reading the entire text.

Writing Tools can also summarize emails in the Mail app, generate auto-replies with enough context to fill in contact information, and help turn voice messages into text summaries. Since the tools work across the operating system, developers do not have to build custom support for each app. Once Apple Intelligence is available on a given device, the Writing Tools appear in many standard text fields, giving users a coherent experience regardless of whether they are writing a text, a social media post, or a business report.

Image Playground and Genmoji

Apple Intelligence also gives users creative visual generators. Image Playground is an application and API that allows users to quickly create images in three distinct styles: Animation, Illustration, and Sketch. The images are generated on-device using a diffusion model that understands the subject and style in the user's prompt. This feature is intended for playful and informal moments, such as making a custom birthday card or adding a cute illustration to a group chat. The output is not designed to be photorealistic, a choice that limits some of the ethical concerns identified in AI image generators and helps the system avoid misleading or deceptive imagery.

Genmoji takes the idea one step further by creating original emoji-like images based on a user's description of a person or concept. Users can go into the emoji keyboard, type "cat wearing a pirate hat on the moon," and Apple Intelligence will produce a custom emoji. The result can be shared as a sticker or inserted into messages. Because the feature uses the user's existing contacts and personal photo library, it can also create a smiling cartoon version of a friend or a custom object that matches a previous style.

These creative tools share an underlying emphasis on context. The system can reference people in the user's photo collection, recognize their faces and clothing, and incorporate those personal elements into the generated image. Crucially, that personal context is stored only on the device and is not sent to a remote server unless the user explicitly chooses an operation that needs Private Cloud Compute.

Privacy at the Center of Intelligence

Privacy is not simply a marketing phrase in this context. Apple has made architectural choices that genuinely change the amount of sensitive data that leaves the device. Many of the most common operations, like summarizing a text thread or identifying an object in a photo, can happen entirely in secure enclave hardware and Apple Silicon's neural engine. The entire system uses a process known as semantic indexing, which builds a local index of a user's personal data, including email, messages, calendars, photos, and files. That index allows the assistant to find relevant information without needing to upload the entire library to a cloud environment.

What is particularly innovative about Apple's approach is the transparency of its server-side computing. Apple has generated software tokens that prove to the user's device that it is connecting to a verified Private Cloud Compute environment. If the authenticity cannot be established, the request is refused. This is an attempt to give users the same confidence in cloud AI that they already have with on-device processing. It is still early, however, and security researchers are waiting to test the extent to which the private cloud infrastructure can withstand malicious attacks or respond to government data requests.

Device and Developer Requirements

The full Apple Intelligence experience is only available on certain devices, and that constraint is itself an important piece of news for consumers. To run on-device models that require large memory bandwidth and a strong graphics architecture, users need an iPhone 15 Pro or iPhone 15 Pro Max. On the iPad, Apple Intelligence supports the M1 chip and later. On the Mac, it supports M1 and after. The requirement means that owners of older devices will not be able to access many of the features, even though they can still install the latest operating system. This creates a clear financial incentive for users to upgrade their hardware, but it also indicates how demanding these local generative models are. Devices need at least 8 GB of memory to handle the size of the model without slowing the rest of the system.

For software developers, Apple Intelligence opens an array of new possibilities. Once they adopt App Intents, developers can make app features and actions visible to Siri and Apple Intelligence. This allows the assistant to not only understand what the user is doing at a global level but also perform specific tasks inside third-party applications. A travel app could list upcoming reservations, a fitness app could present yesterday's workout data, and a weather app could give personalized daily briefings. The key fact is that the user will have to grant permission for each of these actions, and the permission prompt appears in context so that the user is not simply agreeing to blanket access.

Apple has also introduced Swift Assist and other developer tools built on its own language models. These tools can generate code suggestions, add smart completion across Xcode, and answer questions about app architecture. The goal is to let developers spend time on product design rather than repetitive coding tasks. Apple is positioning itself as a platform company that also understands how to build sophisticated developer tools, which sets it apart from the large cloud AI vendors that primarily sell standalone assistants.

Positioning Against Competitors and the Road Ahead

Apple's introduction of Apple Intelligence can be seen as a direct response to the growing popularity of "copilot" experiences in Microsoft products and the widespread adoption of generative AI chatbots from Google and OpenAI. Rather than putting an AI chatbot on top of an existing operating system, Apple is choosing to integrate intelligence into the core user experience. Because the system can access personal context in a controlled way, Apple's version of AI has the potential to feel more intuitive than generic web-based assistants. The company is also betting that privacy will be a decisive factor for many consumers, especially as they become more cautious about how their data is used.

The first wave of Apple Intelligence was scheduled for beta testing in the fall of 2024, alongside releases of iOS 18, iPadOS 18, and macOS Sequoia. Initially, the features were available in American English. More language versions were expected over the course of the following year. The rollout was staggered because the server infrastructure and localized models needed time to be adapted. Apple has promised that ChatGPT integration will also be available in the first wave of updates, allowing users to access OpenAI's GPT-4o without leaving Siri. Before sharing anything with ChatGPT, Apple will ask for permission, but the integration still expands Apple Intelligence from a purely local system into a hybrid network that can leverage external models when necessary.

There is a great deal of work left to do, especially in the area of model reliability. Apple has acknowledged, as most generative AI vendors have, that these models may sometimes generate incorrect information or produce unreasonable output. Because the system has deep access to personal data, the consequence of such errors could be more severe than a wrong answer in an internet search. Apple has attempted to mitigate these risks by avoiding photorealistic image generation, placing clear limits on the types of tasks that are automated, and requiring permission before an action is executed. Yet no perfect solution exists, and the company is nevertheless moving forward in a quick speed to stay competitive in the AI market.

Apple Intelligence is not simply a set of tricks to sell more iPhones. It is a deeper change in how the computer operates. A device powered by Apple Intelligence is designed to infer the user's intent, reason with available information, and generate content that matches practical needs. Instead of requiring users to locate an app, decide what to type, and then manually edit the text, the operating system itself becomes a collaborator. The task of writing, image making, and communication is distributed between the user and a group of small, fast models that run directly on the hardware. As the quality of these models improves, the boundaries between apps may fade and the very notion of an assistant may shift from an afterthought to a primary interface. All of that is likely to happen in stages, but Apple has made its direction clear: artificial intelligence will be everywhere, on every screen, always aware of what the user is doing, and always looking for useful ways to help.

Source: TechRadar News

Key Facts at a Glance

• Electric vehicle sales, including fully electric cars and plug-in hybrids, have grown from fewer than one million vehicles in 2016 to more than ten million vehicles annually in the early 2020s.
• Battery pack prices have fallen dramatically over the past decade, helping to narrow the price gap between electric cars and conventional internal combustion vehicles.
• Hybrid vehicles lack a charging port and rely on internal combustion and regenerative braking, while plug-in hybrid vehicles offer a dedicated electric range and can be charged from the grid.
• China is currently the largest single market for electric vehicles, followed by Europe and the United States.
• Government policies, including zero-emission vehicle mandates, tax credits, and charging investment, continue to shape the pace of adoption.

Market Momentum and Adoption Trends

The adoption of electrified vehicles has accelerated faster than most industry predictions. Early forecasts from the 2010s underestimated the speed at which battery costs would fall and the rate at which consumers would embrace electric drivetrains. Today, automakers no longer ask whether electrification will happen, but how quickly they can convert their production capacity. More than half of the major carmakers have announced plans to spend hundreds of billions of dollars on battery development and new electric vehicle platforms over the next several years.

China has emerged as the dominant force in electric vehicle manufacturing and sales. Government support, combined with fierce competition among domestic and international brands, has created a large and evolving marketplace. Chinese consumers have access to numerous affordable electric models, from small city cars to high-end sedans. The Chinese market also leads in the production of batteries and critical materials, giving it significant influence over the global supply chain.

Europe has also become a major hub for electric vehicle adoption, helped by stringent carbon dioxide emission standards and extensive charging networks in several countries. Markets such as Germany, France, the United Kingdom, and Norway have seen rapid growth. Norway stands out because of longstanding tax incentives and a strong charging network; nearly every new car sold there is now a fully electric or plug-in hybrid vehicle. In the United States, adoption has been more uneven. Coastal metropolitan areas tend to have higher EV penetration, while rural regions face limited model availability and sparser charging infrastructure. Yet even in the United States, electric pickup trucks and family SUVs have started to move into the mainstream, signaling that electrification is expanding beyond the compact commuter segment.

Hybrids continue to serve as an important middle step. Conventional hybrids, such as the Toyota Prius and its many successors, use a small battery and electric motor to improve fuel economy without requiring the driver to plug in. Plug-in hybrids add a larger battery and offer anywhere from 20 to 50 miles of electric-only driving, providing a transitional option for people who want to reduce fuel consumption but worry about range anxiety. For many consumers, the decision between full electric and hybrid depends on their access to charging, the length of their typical journeys, and local climate.

Battery Technology and the Cost Transformation

At the heart of the growing electric vehicle market is the lithium-ion battery. Since 2010, battery pack costs have declined by more than 80 percent, according to industry estimates. That shift has allowed automakers to offer longer-range vehicles at lower prices. Early electric cars often had 100 miles or less of driving range, while many current models offer more than 250 miles on a single charge. The capacity of a battery pack, measured in kilowatt-hours, is a major factor in the final price of the vehicle.

Battery chemistry is also evolving. Many automakers are moving toward lithium iron phosphate batteries, which are less expensive and have a lower risk of thermal runaway, though they are generally heavier and less energy dense. At the same time, research continues on solid-state batteries, sodium-ion batteries, and other alternatives. The direction of battery development is crucial because it affects the availability of raw materials, the cost of future vehicles, and the environmental footprint of production.

The global supply chain for batteries is a geopolitical and economic issue in its own right. Mining and refining of lithium, cobalt, nickel, and graphite are concentrated in a small number of countries. Battery manufacturing, too, is heavily concentrated, with China controlling a large share. Automakers in other regions have begun to invest in domestic battery plants and recycling capacity, partly to reduce reliance on imported materials and partly to qualify for local incentives.

Charging Infrastructure and Range Realities

Charging is often described as the biggest practical barrier to electric vehicle adoption. Drivers who own homes with private parking can install a level 2 charger and start each morning with a full battery. This pattern is simple and convenient. For residents of apartment buildings, older neighborhoods, and urban streets, however, access to off-street parking is limited, and charging can become a difficult or expensive lifestyle change.

Public charging networks have grown quickly, especially along major highways and in commercial areas. Direct current fast chargers can replenish a battery to roughly 80 percent in 20 to 40 minutes, depending on the vehicle and the charger power. That is still slower than refueling a gasoline car, but the difference becomes acceptable when charging is integrated into daily activities such as shopping or work. The location and reliability of chargers remain critical factors. Some networks have suffered from downtime, broken connectors, and payment problems, prompting automakers and regulators to push for better standards.

Range anxiety is receding as newer vehicles offer longer battery range and as drivers gain experience. Many electric car owners quickly adapt to the runtime and learn to plan longer trips around fast charging stations. Hybrids and plug-in hybrids still serve as an important backup for towing, harsh weather, and remote travel, areas where the charging network remains thin.

Environmental Impact and Energy Considerations

One of the main reasons for promoting hybrid and electric vehicles is the reduction of tailpipe carbon dioxide emissions. Fully electric vehicles produce zero emissions while driving, and hybrids produce lower emissions than comparable gasoline vehicles. When the entire lifecycle is considered, electric vehicles generally have a lower carbon footprint than internal combustion vehicles, even when their electricity comes in part from fossil fuels. As power grids become cleaner, the advantages of electrification will increase.

However, the environmental story is not simple. Producing batteries requires substantial energy and raw materials, and the mining and refining process can have significant ecological and social costs. Cobalt mining, in particular, has been linked to human rights concerns in some regions. Automakers are attempting to address these issues by improving supply chain transparency, reducing cobalt content in cells, and scaling up battery recycling and second-life uses.

There are also open questions about how much pressure electric vehicles will place on electricity grids. Widespread overnight charging is generally manageable because many cars can charge during off peak hours. But clusters of fast charging during weekday afternoons or during extreme weather events could increase peak demand. Utilities are preparing through rate plans that encourage off-peak charging, smart charging software, and investments in distribution networks. If managed well, electric vehicles could become an asset for grid stability by storing power and returning it to the network when needed, a concept often called vehicle-to-grid.

Hybrid and Electric Vehicle Choices Expand

The variety of electrified models has expanded rapidly at all price levels. Buyers can now choose between compact electric hatchbacks, midsize sedans, three-row electric SUVs, electric pickup trucks, and even electric performance sports cars. Many legacy automakers have transformed their most popular models into hybrid or electric versions, while new entrants have built their entire brand image around electrification. Heavy trucks and buses are also electrifying, with operators using fleet data to choose the right mix of battery vehicles and hybrid units.

Commercial fleets, including delivery companies, taxis, and public transport authorities, have become an influential early market. Electric delivery vans fit well with predictable daily routes and central depot charging. Transit agencies are purchasing electric buses to reduce air pollution and operating costs. For these fleets, total cost of ownership often favors electric vehicles because of lower fuel and maintenance costs, despite higher upfront prices.

Hybrid and electric vehicles are also changing the relationship between drivers and vehicle software. Over-the-air updates allow manufacturers to improve battery management, adjust range estimates, and even add new features remotely. This digital layer makes a car feel more like a connected device and less like a purely mechanical appliance. The evolution of infotainment systems, remote climate control, and phone-integrated route planning adds a level of daily utility that was rare in conventional vehicles.

Government Policy and Regulatory Direction

Government policies have been a decisive factor in almost every major automotive market. The European Union has proposed an effective ban on the sale of new internal combustion engine vehicles by 2035, while California and Washington have adopted similar rules. Several countries, including Japan, South Korea, and Canada, are targeting complete vehicle electrification in the decades ahead. These deadlines are not legally secure everywhere, but they send a strong signal to manufacturers and investors.

Financial incentives remain important, though they are evolving. Purchase rebates, tax credits, and exemptions from congestion charges can significantly lower the effective price of an electric vehicle. Some governments are also subsidizing home chargers, building public charging stations, and supporting local battery plants. At the same time, there is political resistance to incentives in several regions, and some programs have been terminated as EV sales have grown. The uncertainty over future subsidies may slow adoption in the short term, especially among price-sensitive buyers.

Regulation is not limited to emissions. Safety standards for batteries, requirements for charging connectors, grid interoperability rules, and data privacy laws all play a part. The move toward a common charging standard in North America and Europe has become a contentious issue, as automakers and networks compete for market share. Although universal standards are not a prerequisite for the success of electric vehicles, shared infrastructure reduces confusion for consumers and helps speed up the market transition.

Looking Around the Corner

Electric and hybrid vehicles will continue to shape transportation in the coming decade. The next steps include better batteries, faster charging, more affordable models, and a broader secondhand market. Affordability is likely to remain the deciding factor for many households. The cheapest electric cars today cost more than comparable gasoline cars, but total cost of ownership is often competitive when fuel and maintenance are included.

The shift to electric power also intersects with other forms of mobility. Ride-hailing fleets, shared scooters, and public transit are all part of a changing urban transportation picture. Electric vehicles are expected to become cleaner, cheaper to operate, and more integrated with renewable energy. In that context, hybrids will likely remain an important transitional technology, especially in markets where charging infrastructure lags.

Whether a driver chooses a conventional hybrid, a plug-in hybrid, or a fully electric vehicle, the underlying message is clear: the internal combustion engine no longer dominates the future of personal mobility. Automakers, utilities, governments, and consumers are together building a system in which electricity will drive a much larger share of the miles traveled.

Source: TechRadar News

A powerful surveillance ecosystem

Flock Safety, founded in 2017, has become one of the country’s largest suppliers of automated license plate readers and surveillance equipment. Its cameras are mounted on utility poles and streetlights, constantly scanning passing license plates. The company claims its network spans thousands of cities and captures millions of vehicle records every day. Police officers access the platform through a mobile app and a web dashboard, where they can search for plates or filter by vehicle type, color, and even distinctive details such as roof racks or bumper stickers.

The reverse engineering effort focused on the mobile application used by officers. By intercepting network traffic and decompiling the app’s code, the researcher was able to map the underlying API endpoints and document what data is collected, stored, and shared. One of the most troubling revelations is that the app appears to allow law enforcement to perform “offense-free” searches. In other words, an officer can enter a description of a vehicle they find suspicious, such as a white Ford pickup with a ladder rack, and pull up a list of previous sightings, even if no crime has been reported. Civil liberties advocates have long warned that this type of exploratory search turns the database into a mass location-tracking tool.

What the reverse engineering found

Another key finding involves data retention. Flock publicly states that it does not sell data and that many agencies set their retention periods to 30 days. The reverse-engineered app, however, reveals that metadata such as geolocation and camera identifiers are sent alongside every license-plate read. That metadata is essential for geofencing: officers can run a “search by location” and ask for every vehicle seen in a certain radius during a specific window. While the company requires agencies to obtain a warrant or court order for some historical searches, the app’s internal logic does not appear to block warrantless searches, leaving enforcement to policy and auditing.

The researcher also found that Flock’s platform has the ability to share data between agencies automatically. The company calls this federation: a camera in one jurisdiction can contribute to searches run by officers in neighboring towns, counties, or even states. This interoperability is marketed as force multiplier for investigations, but it also means that privacy protections in one jurisdiction may be undermined by weaker protections elsewhere. If a search query is logged, the log may include an officer’s name and badge number, the agency, and the search parameters—but the app does not appear to generate a public record or notify the person whose vehicles were searched.

The technical report includes screenshots of the app’s network requests, showing that search results are returned in JSON format with time stamps, camera locations, and images of license plates. The images are linked to an index that can be queried with partial plate numbers. The optical character recognition engine is capable of reading plates from several states and provinces. Some configurations include a “hot list” feature that flags plates in state or national criminal-justice databases. When a hot-list plate is detected, the system can send push notifications to officers in the vicinity, according to the reverse-engineered code.

Privacy and legal questions

Flock Safety responded to the disclosure with a statement saying that the reverse engineering was unauthorized and that the company has security monitoring to prevent such activity. It emphasized that law enforcement access is logged and audited, and that every user must agree to a privacy pledge limiting searches to legitimate law enforcement purposes. But privacy advocates are not satisfied. They argue that the underlying architecture is built for surveillance at an unprecedented scale, and that the lack of independent testing means the public cannot verify the company’s promises.

This episode fits into a broader pattern of police departments rapidly adopting algorithmic tools—from predictive policing to facial recognition—without thorough public debate. Flock’s technology is popular because it is relatively cheap to deploy and easy to use. But, as this reverse engineering demonstrates, the systems are not merely passive data recorders; they are active search engines. The same capability that lets an officer find a getaway car can allow a supervisor to locate every vehicle at a political protest, an abortion clinic, or a mosque.

Historically, law enforcement has needed a warrant to place a GPS tracker on a vehicle. In United States v. Jones, the Supreme Court ruled in 2012 that physical installation of a GPS tracker constitutes a search. With ALPR networks, there is no physical installation, and the data is often collected by private devices on public roads. Courts have split on whether warrantless access to historical ALPR records violates the Fourth Amendment. Some decisions have held that prolonged warrantless use of ALPR data violates an individual’s reasonable expectation of privacy. Others have allowed shorter retention periods.

A call for stronger oversight

The reverse engineering report does not reveal specific law enforcement operations, but it contains enough detail for security researchers to inspect Flock’s API for vulnerabilities. The author of the analysis said the goal was to foster transparency and informed conversation, not to help criminals evade detection. Still, the publication of technical details may make it easier for malicious actors to spoof the system or inject false data. The company has patched some less severe issues, but the core architecture remains accessible to anyone who can obtain a legitimate officer login—a risk that concerns cybersecurity experts.

Flock has argued that its technology is essential for solving crimes like carjackings, catalytic-converter thefts, and hit-and-runs. The company keeps a public transparency portal that lists which agencies use its cameras, but the portal is not comprehensive. Some cities have canceled contracts after residents raised objections, while others have expanded their use with little oversight. The reverse engineering release may become a crucial test case for whether private surveillance companies can continue to operate in this gray zone.

Now a handful of civil liberties groups are urging state legislators to pass stricter laws governing ALPR data. Among the proposed policies: requiring warrants for all historical searches, limiting retention to 14 days, requiring automatic notification when a person’s plate is queried, and establishing independent audits. No federal law currently regulates ALPRs, leaving a patchwork of state rules that often lag behind technology. The report provides an unusual empirical foundation for the legislative debate, because it is one of the few independent documentations of how a private product works behind the scenes.

The pressures on independent research

The researcher’s methodology also highlights the inherent difficulty of studying proprietary surveillance systems. Copyright law can inhibit security research, and the Computer Fraud and Abuse Act has been used in the past to punish good-faith researchers. For now, the report is hosted in a public repository, and its author has removed any examples that might compromise active investigations. Flock’s response, however, shows how hard it is for outsiders to scrutinize these systems. The company has previously threatened legal action against researchers who published vulnerabilities, although it has not yet commented on next steps in this case.

The implications go beyond one company or one mobile app. Flock Safety is part of a larger industry that sells AI-powered surveillance to police. These tools often rely on machine learning models trained on massive datasets. The models are not perfect: misreadings of plates can lead to innocent drivers being pulled over, and a flawed search query can produce a false positive. Without access to the system’s performance metrics, there is no way to know how often these errors happen. The reverse engineering exercise cannot answer every question, but it does provide a starting point for public accountability.

The most important takeaway is that the power of the state to search digital records has grown without a corresponding expansion of judicial oversight. The public knows police use camera networks, but rarely can citizens see the software architecture that determines how those cameras are used. This reverse engineering peels back a layer, showing that a license plate capture is not merely a moment frozen in time; it is an entry in a searchable, shareable, location-aware database. That knowledge will inevitably influence ongoing fights over police surveillance, the right to assemble, and the limits of algorithmic law enforcement.

Source: TechRadar News

Apple has not confirmed every detail in advance, but the rumors are unusually consistent. The event is expected to include a foldable iPhone that may be called iPhone Ultra, the iPhone 18 Pro and iPhone 18 Pro Max, Apple Watch Series 12 and Apple Watch Ultra 4, and AirPods 5. There is also a chance Apple will use the stage to show a new home hub, updated Apple TV 4K, or refreshed HomePod mini. The standard iPhone 18 is not expected to appear at the September event, however, and could be delayed until spring 2027.

At a glance: Key facts

• When: September 9 at 10:00 a.m. Pacific Time. The event will be livestreamed on Apple's website, YouTube, and Apple TV app.
• Main iPhone announcement: Foldable iPhone Ultra with a 5.5-inch outer display and 7.8-inch inner display.
• Pro flagships: iPhone 18 Pro and iPhone 18 Pro Max with A20 Pro chips, new camera technology, and likely price increases.
• Wearables and audio: Apple Watch Series 12, Apple Watch Ultra 4, and AirPods 5.
• Software: Expected release dates for iOS 27, iPadOS 27, macOS Golden Gate, tvOS 27, watchOS 27, and visionOS 27.
• Absent: Standard iPhone 18 is not expected to appear; Apple is reportedly planning a spring 2027 launch for that model.

Foldable iPhone: What we expect from iPhone Ultra

Apple's first foldable iPhone has been rumored for years, and it is now expected to take center stage at the September event. Industry watchers and supply chain reports describe the device as a book-style foldable. When closed, it may look like a compact phone with a roughly 5.5-inch OLED cover display. When opened, it is expected to reveal a 7.8-inch OLED display, similar in feel to an iPad mini that folds in half. The unfolded dimensions are said to be wider than they are tall, with an aspect ratio designed for both phone and tablet-style apps.

The product is expected to be branded as iPhone Ultra, although Apple has not confirmed the name. The company is said to have worked hard on durability, since foldables create more mechanical stress than traditional smartphones. A liquid metal hinge and an internal adhesive designed to minimize the appearance of micro-cracks could help reduce the visibility of the crease over time. Apple may not be able to entirely eliminate the crease, but the company has reportedly tried to make it less noticeable than early foldable prototypes.

One of the biggest design trade-offs involves biometrics. Because the chassis is extremely thin, especially when opened at around 4.5mm, there may not be room for the TrueDepth camera system used for Face ID. Instead, Apple is expected to put Touch ID in the side button, similar to current iPads. Front-facing cameras are still likely on both the outer and inner displays, but the primary authentication method could shift to the fingerprint sensor.

Another trade-off is the camera system. The foldable iPhone is unlikely to include a dedicated telephoto lens. Instead, it may offer a Wide lens and an Ultra Wide lens on the rear. This arrangement would allow the rear camera to be useful whether the device is closed, open, or used in a folded form. Magnetic wireless charging appears to be part of the plan despite conflicting rumors, with MagSafe expected to be included.

Performance should be a highlight. The iPhone Ultra is expected to use a 2nm A20 Pro chip, which could be around 18 percent faster and up to 30 percent more power efficient than the A19 Pro. The device is also likely to ship with 12GB of RAM. Battery capacity could land between 5,400 mAh and 5,800 mAh, split between two halves of the foldable design. The Ultra will likely come in silver/white and a dark indigo color close to black. Pricing may start at $2,000 or higher, with some reports estimating a starting range of $2,099 to $2,299.

On the modem side, Apple could equip the foldable with its own C2 modem in some regions. There are questions about whether the C2 supports mmWave 5G, which is frequently used in the United States. If it does not, U.S. versions of the foldable might include a Qualcomm modem for mmWave support, or Apple could limit the device to sub-6GHz 5G around the world.

Software will be just as important as hardware. The company has reportedly asked developers to create apps that adapt to a variety of aspect ratios and display sizes. iPad-style multitasking, side-by-side app views, and sidebars may all be available on the foldable. The iPhone Ultra could effectively bridge the gap between an iPhone and an iPad.

iPhone 18 Pro and iPhone 18 Pro Max

The traditional high-end flagships are expected to see more modest updates. The iPhone 18 Pro and Pro Max are likely to keep the same basic design as the iPhone 17 Pro models, with some refinements. Rumors suggest the camera bump may be slightly thicker to accommodate a new variable aperture main camera. There is also a possibility that the Dynamic Island could become smaller, but earlier speculation about removing it entirely with under-display Face ID now appears to be off the table.

Color options are expected to include a dark cherry shade, light blue, and silver. A black Pro model may not be offered this year. Apple might also change the relationship between the frosted glass back and the aluminum frame, making the visual contrast less pronounced than in the current generation.

Performance improvements are likely to come from the same 2nm A20 Pro chip rumored for the foldable iPhone. A more efficient LTPO+ display panel and a larger battery could also improve battery life, especially in the iPhone 18 Pro Max, which may have a battery capacity around 5,567 mAh. Reports suggest the main wide camera will gain a DSLR-like variable aperture, giving users more control over exposure and depth of field in different lighting. A three-layer stacked image sensor could also improve low-light performance and reduce shutter lag.

The telephoto lens may receive a wider aperture, but the ultra-wide lens and front-facing camera are not expected to change. Apple could simplify the Camera Control button by removing the capacitive layer that currently handles swipe gestures, retaining a more conventional physical push-button. Pricing is a major uncertainty, but rising memory and storage costs could lead to an increase of $100 to $300 across the lineup.

Apple Watch Series 12 and Apple Watch Ultra 4

The next Apple Watch models are not expected to feature a major redesign. The Apple Watch Series 12 may reintroduce a ceramic case option, a material Apple has used on past models. Both the Series 12 and Apple Watch Ultra 4 are likely to receive an updated processor, which could enable better health and fitness tracking. Apple is reportedly testing a feature that continuously collects heart rate data throughout the day. That data could power a new Health app experience similar to readiness and wellness insights offered by devices such as the Oura Ring.

AirPods 5

References to AirPods 5 have been discovered in Apple software releases, increasing the chance that new earbuds will be announced alongside the

Source: MacRumors News

Tokenized RWAs are digital representations of assets such as Treasury bills, corporate debt, government securities and investment funds. These tokens remain tied to off-chain legal, custody and reporting frameworks while using blockchain infrastructure for settlement and record-keeping. Stellar, which was originally designed for cross-border payments, has become a home for several tokenization initiatives because of its fast transaction speeds, low fees and simpler finality model compared with many other blockchain networks.

RWA market passes $3.9 billion

According to a dashboard maintained by Stellar and built on public blockchain data, the network’s RWA market capitalization stood at approximately $3.996 billion as of Aug. 29. That compares with $868.8 million at the end of 2025, implying a gain of roughly 360% during 2026. The dashboard tracks assets issued on Stellar by approved issuers and lists the market by category and issuer.

The asset mix includes U.S. Treasurys, private credit, public credit, non-U.S. government debt and other tokenized products. The growth from roughly $0.87 billion to nearly $4 billion adds about $3.1 billion in on-chain asset value in fewer than nine months. That rate of expansion puts Stellar in the upper tier of public blockchains used for tokenized real-world assets.

Growth driven by institutional and issuer activity

The expansion has been led by a small group of tokenization issuers that have deployed large portfolios on Stellar. Spiko accounted for about $1.55 billion of Stellar’s RWA value as of Aug. 27, according to the dashboard. Realiz followed with $559 million and Tradable with $548 million. Franklin Templeton, a traditional asset manager that has built a presence in tokenized funds, had approximately $546 million in assets on Stellar, while Ondo contributed $535 million. These five issuers alone represent a significant majority of the network’s tokenized RWA supply.

Each issuer has played a different role in the market. Some focus on money-market funds and government securities, while others have specialized in private credit. The presence of large, regulated financial institutions and dedicated tokenization firms helps explain the rapid adoption: issuers bring their existing client networks, legal structures and asset management processes to the blockchain while relying on Stellar for the settlement layer.

Non-U.S. government debt gains traction

One area where Stellar has gained particular ground in 2026 is non-U.S. government debt. Citing external tokenization data, the Stellar Development Foundation said the network held roughly $490 million in that asset class as of Aug. 20. A portion of that market is represented by tokenized Mexican CETES and Brazilian government bonds issued through Etherfuse. These products let investors gain exposure to select sovereign debt instruments in a tokenized format, broadening Stellar’s appeal beyond the U.S. Treasury market.

Tokenized government securities have become one of the fastest-growing categories in digital assets. Traditional finance participants have increasingly looked to blockchain rails to improve settlement, reduce operational friction and expand access to short-term, low-risk instruments. For networks such as Stellar, the growth of tokenized Treasurys and government debt also provides a revenue-generating use case that is distinct from speculative trading activity. Even with that growth, however, the network’s underlying token has not followed the same trajectory.

DTCC plans deeper Stellar integration

Institutional adoption has been an important driver of Stellar’s tokenized asset activity. In May, the Depository Trust & Clearing Corporation, known as DTCC, said it planned to connect its tokenization service to the Stellar network. According to the announcement, DTC-tokenized assets could become available on Stellar in the first half of 2027. The integration would potentially allow a wide range of traditional securities to be represented on-chain through DTCC’s post-trade infrastructure.

DTCC plays a central role in U.S. capital markets as the post-trade infrastructure provider for most equity and corporate bond transactions. Its decision to explore Stellar suggests that tokenization is moving from pilot projects to more established market infrastructure. The announcement said the integration could eventually support tokenized U.S. Treasurys, major index ETFs and stocks in the Russell 1000. If completed, such an expansion could bring billions of dollars in traditional market value into blockchain-based settlement systems.

Private credit and Tradable’s expansion

Another major development came in July, when tokenization platform Tradable announced plans to bring up to $1 billion in private credit assets to Stellar. The project is designed to support compliance requirements, investor onboarding and asset lifecycle management within a blockchain environment. Tradable has previously tokenized around $1.7 billion in private credit across nearly 30 positions, making it one of the more active issuers in that segment. The planned Stellar integration could extend its reach and further deepen the network’s exposure to credit markets.

Private credit has become a focal point for tokenization because the asset class is typically illiquid, manually managed and difficult for ordinary investors to access. Tokenization can make ownership records more transparent, allow for larger participation and potentially streamline administrative processes. Stellar’s RWA statistics show that credit-related products now make up a meaningful share of the total value locked on the network, alongside sovereign debt and funds.

Payments and stablecoins

Stellar’s history is rooted in cross-border payments, and that use case has also expanded in tandem with tokenized assets. MoneyGram launched its MGUSD dollar stablecoin on Stellar in June, giving users the ability to hold dollar-denominated balances and send funds through MoneyGram’s global payments network. The token was introduced as a way to combine the stability of the U.S. dollar with the speed of Stellar’s settlement layer. MGUSD is among a growing list of dollar tokens designed for payments on the network.

According to the same dashboard, Stellar currently supports about $438 million in reserve-verified stablecoins. That figure covers stablecoins whose reserves meet certain certification criteria and includes a variety of payment-focused dollar tokens. The presence of both stablecoins and tokenized assets has made Stellar a multi-purpose settlement network, although its market cap remains well below the largest Ethereum-based tokenized RWA ecosystems.

XLM price lags network growth

Despite the rapid increase in tokenized assets, the price of XLM, the native token of the Stellar network, has not matched the network’s fundamentals. According to market data, XLM is down about 11% year-to-date and trading near $0.18 at the time of writing. The divergence highlights a common pattern in the crypto market: protocol usage and token price do not always move in sync, especially when sentiment is affected by broader macroeconomic conditions, risk appetite and market positioning.

Stellar’s RWA expansion could eventually translate into higher demand for XLM for transaction fees, settlement and network security, though token utility varies by design. For now, developments such as the DTCC integration, Tradable’s private credit pipeline and MoneyGram’s stablecoin launch are expected to remain the main drivers of new activity on the network. The continuation of those partnerships may determine whether Stellar can maintain its tokenized asset momentum through 2027.

Source: Cointelegraph News

In the statement, Core DAO said the problem was limited to the reward issuance mechanism. The project emphasized that user funds were not affected and that confirmed transactions would remain intact. The announcement follows an earlier status update on Monday that acknowledged the existence of abnormal reward activity.

Validator Reward Exploit Details

Core DAO has not yet disclosed many of the operational details surrounding the exploit. The team has not said exactly how much CORE was overissued, how long the validators were able to draw excess rewards, or whether the extra tokens were moved into circulating supply. The initial status update described the participants as a small group of validators and stated that they had accrued rewards significantly above the protocol’s intended issuance.

The word “malicious” was later used to describe the validators, suggesting that this was not an accidental consequence of a network error. It is possible that the validators identified and exploited a flaw in the reward-distribution logic, or that they manipulated a governance or consensus parameter. Without a published postmortem, the broader community can only speculate on the exact mechanism that made the excess issuance possible.

Exchanges Restricted Core Transfers

As news of the incident circulated, several major exchanges moved quickly to limit exposure to the Core network. Coinbase paused CORE sends and receives while the network was being investigated. Bithumb and Coinone, both based in South Korea, suspended deposits and withdrawals because of suspected or confirmed security concerns. Bitget also suspended CORE deposits and withdrawals, though it referred to wallet maintenance as the reason. LBank suspended deposits, citing the requirements of the project.

The breadth of exchange actions suggests that platforms were not given enough detail to determine whether user funds were at risk, and many opted for a cautious approach. Temporary suspensions are common when blockchain networks face unexpected technical issues, but they also serve as a reminder of how centralized exchanges function as gateways to the broader cryptocurrency market. When an exchange halts deposits, users can still hold tokens, but the ability to move them onto or off the platform is interrupted.

The reaction also highlights the need for blockchain teams to maintain clear lines of communication with exchanges and infrastructure providers. During a network emergency, exchanges are often among the first to observe anomalies, and they rely on timely and accurate updates from the project team in order to make informed decisions.

What Is Core DAO and the Satoshi Plus Consensus?

Core DAO is the development organization behind the Core blockchain, a layer-1 smart contract platform designed to integrate with Bitcoin’s security model. The chain uses a consensus mechanism called Satoshi Plus, which combines delegated proof-of-stake with Bitcoin mining hashpower. This arrangement allows Bitcoin miners to participate in securing the Core network while still receiving their normal Bitcoin rewards, with Core issuance serving as an additional incentive.

By aligning with Bitcoin, Core aims to offer high security and decentralization while providing a more flexible smart contract environment than Bitcoin itself. The project has attracted attention because of its focus on Bitcoin-centric DeFi and its use of non-custodial applications on a network that relies on Bitcoin’s proven infrastructure.

The CORE token is central to the network’s operation. It is used for transaction fees, staking, and governance participation. Validators are selected based on the amount of CORE staked and the Bitcoin hashpower delegated to them. They earn rewards in CORE for maintaining the network and producing new blocks. The issuance schedule is determined by protocol rules that are commonly understood by the community.

When those rules fail to hold, it creates concern about the robustness of the tokenomics model. Regulators, investors, and potential users are likely to pay close attention to how Core DAO resolves the issue and whether it is able to recover any excess funds from the validators involved.

Why an Emergency Hard Fork?

In order to stop the validators from drawing further excess rewards, Core DAO is coordinating what it refers to as an emergency hard fork. This type of network upgrade generally requires validators and node operators to update their software to new consensus rules. Failure to adopt the upgrade can lead to a split, with some nodes continuing on the old chain and others following the new rules.

Core DAO has already described the hard fork as a forward upgrade. This means that the network will adopt new rules moving forward, but no changes will be made to past blocks. In other words, the blockchain’s history will be preserved, and any transaction that users completed before the fork will remain valid on the upgraded chain. The team explicitly said that it would not roll back the network or reverse previously confirmed transactions.

This is an important design choice. Rolling back a chain can be highly contentious, as it may undo legitimate transactions and violates the principle of immutability. Forward upgrades, by contrast, aim to avoid rewriting history while still enforcing changes that protect the network from future vulnerabilities. By choosing this path, Core DAO likely intends to maintain community trust and avoid the kind of divisions that historically arise from rollback proposals.

Hard Fork Precedents in the Crypto Space

The broader cryptocurrency industry has seen several similar emergency forks over the years. One of the most famous took place in 2016 on Ethereum, when the network implemented a hard fork to reverse the effects of a hack that drained more than $50 million from a decentralized autonomous organization. That decision was controversial because it rewrote Ethereum’s history, leading to a split between Ethereum and Ethereum Classic.

More recently, BNB Chain activated the Pasteur hard fork to strengthen bridge security after an attack on a cross-chain bridge resulted in a substantial loss of funds. That fork, like Core’s planned upgrade, was described as a way to secure the network and address an active vulnerability. Polygon also disclosed that security flaws had been fixed in recent hard forks, indicating how common this response has become in the industry.

In all of these cases, the immediate priority was to contain financial damage and restore normal network operations. The long-term consequences, however, often include greater scrutiny of governance practices, pushback from decentralization advocates, and questions about whether a small group of developers has too much power to alter the rules.

Governance and Decentralization Concerns

When a network’s core team can unilaterally decide to fork the blockchain in response to an emergency, it raises governance concerns even if the action is technically necessary. Core DAO’s relatively centralized response is not unusual in the crypto industry, but it runs contrary to the ideals of decentralization that many blockchain projects claim. The tension is inherent: decentralized networks sometimes need fast coordination, and that coordination tends to be led by the founding team or a small group of contributors.

The situation also raises questions about the role of validators in governance. Validators on the Core network presumably have staked significant amounts of CORE and may be aligned with Bitcoin miners. If a handful of validators were able to exploit the issuance process, it could indicate that there are not enough checks and balances in the election or monitoring process. Stronger supervision and better automated detection mechanisms might have stopped the exploit before it became severe enough to warrant a hard fork.

Impact on Core Token Holders and the Ecosystem

For users who hold CORE, the hard fork raises both practical and philosophical questions. On a practical level, holders may need to take action to ensure that their wallet software or exchange supports the upgraded network. Most custodial services and centralized exchanges will handle this transition internally, but non-custodial users are often instructed to follow project announcements and update their clients when necessary.

From a value perspective, the excess issuance of CORE could be viewed as a dilution event. If the unauthorized rewards were able to enter circulation, the total supply of CORE may be higher than originally scheduled. Even though the number of tokens is perhaps modest relative to the overall market, any unplanned increase in supply can pressure the token’s price and reduce confidence in the project’s economic design.

Core has not yet disclosed the scale of the overissuance, so the market is evaluating the news without complete information. This uncertainty could lead to continued volatility in the CORE asset and might affect the network’s ability to attract new users and developers.

Unanswered Questions About the Vulnerability

One of the most pressing unanswered questions is what type of vulnerability enabled the validators to draw excess rewards. Core DAO has not explained the flaw in detail, nor has it provided information about the validators’ identities. Without those specifics, experts are unable to assess how serious the bug was or whether other networks that share a similar architecture are at risk.

Another open question is whether the excess rewards remain under the control of the malicious validators. If the validators were able to withdraw the funds into wallets they control, the tokens may have already been sold on exchanges. In that case, it may be impossible to recover the funds. If the rewards are still locked in the validator’s staking wallet, the project might be able to freeze or confiscate them as part of the hard fork.

Core DAO has also not addressed what measures will be put in place going forward to prevent a recurrence. Hard forks can patch a specific bug, but if the underlying code is still flawed, similar exploits could happen again. The project’s technical postmortem will be watched closely to determine whether the issue was one of implementation or basic design.

The exchange suspensions are another area that lacks complete clarity. While Coinbase, Bithumb, Coinone, Bitget, and LBank all restricted CORE transfers, each gave a slightly different reason. It is still unclear whether the exchanges had independent evidence of risk or were simply following recommendations from the project team.

Core DAO has pledged to publish a full technical postmortem that will address some of these concerns. Until that report arrives, the community will have to rely on the project’s assurances that the incident is contained and that user assets remain safe. The planned forward upgrade represents a critical step toward restoring normal operations on the Core blockchain, but the long-term reputation of the network will depend on how transparently and effectively the team responds to this unusual event.

Source: Cointelegraph News

Source: Cointelegraph News

Senate Bill 1119, authored by state Sen. Steve Padilla and Assemblymembers Buffy Wicks and Rebecca Bauer-Kahan, is the product of months of negotiation over how to protect young people from the potential harms of conversational AI without shutting down useful educational and creative tools. The legislation passed with enough support to reach Newsom's desk, but it remains to be seen whether the governor will sign it, veto it, or allow it to become law without his signature. Because California is the nation's largest state and the home of OpenAI, Google, and many other AI developers, the outcome could influence product policies well beyond state lines.

What the bill requires

The bill targets companies that operate AI chatbot platforms reasonably likely to be accessed by users under 18. It would require those companies to estimate or determine the age of each user, and to give parents and guardians tools to manage the time their children spend on services such as ChatGPT. The measure also requires independent audits designed to assess whether chatbots are exposing minors to dangerous content, manipulation, or other risks associated with AI companionship and conversation.

The legislation includes several privacy and advertising protections that go further than current federal children's online privacy rules. Under the bill, AI companies could not serve ads to minors that rely on personal information gathered from their chatbot conversations. The sale of data collected from children through chats would also be illegal. Chatbots would still be allowed to show “age-appropriate contextual advertising,” as long as the ads are clearly labeled as ads. Sponsors said those limits are intended to prevent companies from turning children's conversations into behavioral profiling databases.

The bill was amended Friday to narrow the situations in which families can sue AI companies. The amendment was seen as a response to concerns from technology groups that the original language could have produced a flood of litigation over chatbot interactions that were not clearly linked to harm. It is unclear what role, if any, Altman played in shaping that deleted language. OpenAI did not answer specific questions about Altman's outreach, and a person familiar with the matter denied that the OpenAI chief executive and Newsom spoke directly last week.

A national benchmark in the making

The California proposal has become one of the most closely watched pieces of AI legislation in the country because Washington has not yet passed a comprehensive federal AI law. State regulators have increasingly moved into that gap, and California's enormous market gives its rules outsized influence. A bill signed into law in Sacramento effectively becomes a compliance benchmark for companies that operate nationwide. That is why SB 1119 has drawn attention not just from California-based AI firms but also from trade groups, civil liberties organizations, and child-safety advocates across the country.

Supporters argue the bill is an overdue update to laws that were written for an earlier era of online services. Most existing children's privacy protections focus on websites, apps, and social media platforms, rather than on conversational systems that can build intimate relationships with users. Chatbots, unlike passive feeds or games, can engage users in open-ended dialogue, offering emotional support, companionship, and sometimes troubling suggestions. The ability to generate personalized, context-aware responses means a chatbot may be more persuasive and more deeply embedded in a child's daily life than conventional software.

Those concerns have been amplified by recent, high-profile cases involving young people and chatbots. In 2026, reports emerged of teenagers who died by suicide after being allegedly encouraged by AI chatbots. The cases were cited by SB 1119's supporters as evidence that lawmakers had a responsibility to act before more children are harmed. Mental health experts and child-development researchers have also warned that AI companions can be particularly attractive to adolescents who feel isolated, and that those users may not understand the limitations or commercial incentives of the systems they are talking to.

The bill's core assumption is that safety should not rely on the goodwill of a technology company. OpenAI had already introduced some of the protections included in the measure, but the bill would make similar safeguards enforceable across the industry. That distinction matters to advocates who argue that voluntary commitments are too easy to weaken or abandon when financial pressure grows.

OpenAI's public position

OpenAI has tried to position itself as a constructive partner in the conversation around youth safety. In a statement released shortly before the bill passed on Monday, the company said it encouraged Newsom to sign the measure. “SB 1119 builds on youth safety measures that OpenAI has supported through our products, global policy principles, advocacy in California, and work on the Parents & Kids Safe AI Act,” the statement said.

An OpenAI spokesperson directed attention to that statement when asked about the company's engagement with Newsom's office. The statement emphasized that OpenAI had already introduced protections for teenage users. The company launched ChatGPT for Teens on Aug. 18, a system designed for users between 13 and 17 years old. The product includes safeguards meant to encourage healthy use and provide parents with additional controls. OpenAI says that if its system estimates that someone is under 18, or if the user states that they are 13 to 17, they are automatically placed into the teen experience. Those protections are part of the baseline settings rather than optional features that users can disable.

The timing of that launch was notable. OpenAI introduced it just weeks before the California bill landed on the governor's desk, and the company's public comments pointed to the product as evidence that it had already begun adopting many of the practices the legislation would require. Still, consumer groups and some state lawmakers have argued that voluntary measures are not enough, particularly because many chatbot creators have economic incentives to maximize user engagement and time spent on their platforms.

OpenAI's broader lobbying posture has also complicated its public image. The company regularly says it welcomes regulation that addresses real risks, but it has also worked to shape rules in ways that are favorable to its commercial ambitions. The California bill is no exception. The late amendment narrowing the right to sue is one example of how industry pressure can influence legislation even when a company ultimately issues a supportive statement. OpenAI's public stance appears designed to keep lawmakers and the public on its side while avoiding the kind of adversarial fight that could damage its relationship with Sacramento.

Newsom's decision and the politics around it

Newsom has long had a friendly relationship with technology companies and has frequently resisted legislation that he believes would place too many guardrails on innovation. In previous years, he vetoed bills backed by labor unions and consumer groups that technology giants described as burdensome, including measures involving workplace conditions and data privacy. His approach has sometimes frustrated advocates who argue that California should lead the way in regulating powerful new technologies.

The political environment has shifted, though. Public attention has focused on concerns about AI's risks, including the spread of misinformation, the displacement of workers, and the emotional vulnerabilities of young users. The so-called data center backlash has become a serious issue in communities near AI infrastructure, but the human costs of chatbots have also become impossible to ignore. Newsom is term-limited and will leave office on Jan. 7, 2027. He is widely expected to mount a presidential campaign in 2028, and his political positioning has puzzled some observers. While he has not formally left the Democratic Party, Newsom has spent considerable time courting right-wing media figures and far-right voters. He hosted a podcast that featured guests such as Charlie Kirk and Steve Bannon, and in several conversations he appeared to agree with much of the MAGA base's criticism of Democratic leadership. That broader political strategy may shape how he approaches a high-profile AI bill in his final months as governor.

If Newsom signs the bill, he could strengthen his national image as a leader on technology accountability and child protection. If he vetoes it, he risks alienating the Democratic base and parents who have become increasingly worried about chatbot dangers. If he allows it to become law without his signature, he can preserve a degree of distance from the decision while still letting the measure take effect.

California's governor's race and the future of AI policy

The governor's race to succeed Newsom is already well underway. Democrat Xavier Becerra, a former state attorney general and U.S. secretary of health and human services, is running as the establishment candidate. Republican Steve Hilton, a former tech executive and conservative commentator, has positioned himself as the anti-regulation candidate. Polling from the University of California Berkeley Institute of Government Studies showed Becerra with a commanding lead of 55 percent to 37 percent.

Both candidates have generally been friendly to the technology industry, though their approaches differ. Hilton wants less regulation across the board and is married to Rachel Whetstone, a veteran Silicon Valley executive who has worked for Uber, Netflix, Facebook, and Google. Becerra has proposed what he calls modest AI guardrails and has accepted significant campaign donations from Silicon Valley. Even in a year when AI safety has become a prominent policy issue, no major candidate has embraced the kind of aggressive restriction that some advocacy groups are demanding.

California's choices will matter beyond the governor's office because Washington has not produced a comprehensive federal AI law. As a result, state legislators have taken the lead on issues ranging from deepfakes to automated decision-making to chatbot safety. If SB 1119 becomes law, it is likely to become a reference point for similar legislation in other states and for federal proposals in the next Congress. The bill is not the final word on the subject, but it marks one of the most direct attempts yet to regulate the relationship between kids and conversational AI.

For now, all eyes are on Newsom. The bill's supporters are waiting to see whether California will enforce new protections before another crisis forces the issue. The outcome will also send a signal to AI companies about how much political space they are able to shape the rules that govern their products, particularly when the users are children. As the governor weighs the legislation, the conversations that happened before the final vote are likely to remain part of the public record and a subject of scrutiny for policy advocates and industry watchers alike.

Source: Gizmodo News