The Long Beach News

collapse
Home / Daily News Analysis / AMD CEO Lisa Su sees AI inference boom fueling $1.5T demand by 2030: 'About 60% of the market going forward'

AMD CEO Lisa Su sees AI inference boom fueling $1.5T demand by 2030: 'About 60% of the market going forward'

Oct 17, 2026  Twila Rosenbaum  110 views
AMD CEO Lisa Su sees AI inference boom fueling $1.5T demand by 2030: 'About 60% of the market going forward'

Advanced Micro Devices CEO Lisa Su is betting that the next phase of artificial intelligence spending will be defined less by model training and more by inference, the process of running trained models in production. Speaking at the OCP Global Summit, Su said inference and agentic workloads are becoming the primary demand drivers for AI infrastructure, and she projected that the market for data center AI accelerators could approach $1.5 trillion by 2030.

Su’s comments point to a shift in how the AI buildout is being framed. The first wave of generative AI investment focused heavily on training large models with massive GPU clusters. Now, as those models are deployed in consumer and enterprise applications, demand is moving toward systems that can handle high volumes of inference queries reliably and efficiently. Su said inference could represent about 60% of the AI market going forward, with agentic workloads—AI systems that can carry out multi-step tasks with limited human involvement—adding a new layer of computing demand.

AMD sees inference and agents changing the AI infrastructure mix

“This year is really the year of inference,” Su said. “This is the year where we believe inference is now about 60% of the market going forward. And it’s not just inference. It’s really the agentic workloads that have really taken off.” The distinction matters for chip designers because inference has different performance priorities from training. Training clusters are built for maximum throughput on very large workloads, while inference systems often need to balance latency, power efficiency, memory bandwidth, and cost per token across many simultaneous requests. Agentic workloads add another dimension: they can chain together multiple model calls, tools, and data lookups, which can increase CPU and networking activity alongside GPU compute.

Su said AMD expects the market for data center AI accelerators to approach $1.5 trillion by 2030. She also pointed to a sharp increase in demand for central processing units, driven in part by AI agents that can carry out tasks with limited human involvement. That combination—GPUs for model execution and CPUs for orchestration, data movement, and general-purpose computing—is central to AMD’s pitch as it tries to sell complete rack-scale systems rather than individual chips.

“We have actually raised the TAM on this three times in the last nine months because we’ve seen that much growth in the overall CPU demand,” Su said. “We now believe [CPU demand] will grow at approximately 50% CAGR over the next four or five years to over a $200 billion TAM by the time you get to 2030.” The comments suggest that AMD sees the AI infrastructure opportunity extending beyond accelerators. As agents become more capable, they may require more host processors to manage workflows, schedule tasks, and move data between systems. That could lift demand for server CPUs at a time when the market has been focused on GPU supply.

Helios rack shipments begin as AMD targets large-scale AI deployments

AMD is expanding its AI infrastructure lineup as demand shifts toward systems that can run AI applications at scale. Su highlighted the company’s Helios and Helion rack systems, growing demand for processors, and the development of a new EPYC chip. The company’s Helios rack-scale system, introduced in July, began production shipments in September. Su said initial systems have reached major customers, including Meta Platforms, Microsoft, OpenAI, Oracle, and Anthropic. Production is expected to increase through the fourth quarter and into the first half of 2027.

Developed with Meta, Helios follows Open Compute Project standards and combines 72 GPUs into a single accelerator system. The rack contains more than 18,000 GPU compute units and 4,600 Zen 6 CPU cores. AMD estimates that the platform can deliver 30% more tokens per dollar, a measure of AI processing value relative to cost. That metric is important for cloud providers and large enterprises that are trying to control the cost of running AI models at scale. “Helios is absolutely the most sophisticated and complex system that we’ve built at AMD together with our partners,” Su said.

The customer list is notable because it spans hyperscale cloud providers, a leading AI model developer, and an enterprise software and cloud infrastructure company. Meta and Microsoft have been investing heavily in AI data centers, while OpenAI has become one of the most closely watched buyers of AI compute. Oracle has been expanding its cloud AI capacity, and Anthropic is a major developer of frontier models. Shipping to those customers gives AMD a foothold in deployments that have traditionally been dominated by other accelerator vendors.

AMD’s Helios system is designed around rack-scale architecture, meaning it treats multiple servers and accelerators as a combined unit rather than a collection of separate boxes. That approach can simplify deployment and improve efficiency for large AI workloads. Open Compute Project standards are intended to make it easier for data center operators to integrate hardware into their existing facilities and supply chains. The next production ramp through the fourth quarter and into the first half of 2027 will be an important test of AMD’s ability to convert design wins into volume revenue.

Helion targets agentic AI and enterprise workloads

AMD also introduced Helion, a rack system designed for agentic AI workloads. The company said the system can support more than 32,000 Zen 6 cores and up to 32 Instinct MI350P GPUs. It supports liquid or air cooling and uses a standard 21-inch rack design. The system is intended to help data center operators match computing resources to different tasks, from everyday personal assistants to enterprise applications and demanding technical workloads.

The Helion design reflects the growing diversity of AI workloads. Not every application needs the same balance of GPU acceleration and CPU capacity. A personal assistant might require low-latency inference and efficient CPU orchestration, while an enterprise application might need more memory, security features, and integration with existing software stacks. A technical workload could require high-throughput compute for simulation or data processing alongside AI models. By offering rack-level configurations with different ratios of CPU and GPU resources, AMD is trying to address those varied requirements without forcing customers into a single architecture.

Su also presented EPYC Verano, a server processor based on AMD’s Venice CPU core family. Designed to reduce system power consumption, the chip features six core-complex dies and up to 192 CPU cores, along with improved bandwidth between CPUs and GPUs. The processor is aimed at data center workloads that increasingly mix general-purpose computing with AI acceleration. Better CPU-to-GPU bandwidth can help reduce bottlenecks when data must move between host processors and accelerators, a common issue in AI inference and agentic systems.

The emphasis on power consumption is significant. Data center operators are facing constraints on power availability and cooling capacity, especially in dense AI clusters. Chips that can deliver more compute within the same power envelope can improve total cost of ownership and make it easier to deploy AI infrastructure in existing facilities. AMD’s focus on core count, bandwidth, and efficiency shows that it is designing CPUs not just as standalone server processors but as part of a broader AI system architecture.

Market reaction and retail sentiment

Despite the long-term growth outlook, AMD stock traded more than 1% lower in Tuesday’s premarket. The shares have gained more than 177% year-to-date, reflecting strong investor enthusiasm for the company’s AI positioning. On social media, retail sentiment around the stock slipped to bearish from neutral the previous day, according to a measure of investor chatter. The mixed reaction suggests that expectations for AMD’s AI business are already high, and investors may be looking for proof that the company can convert its product announcements into sustained revenue growth.

AMD’s challenge is not only technical but also commercial. The AI accelerator market is highly competitive, and customers often qualify multiple vendors to avoid dependency on a single supplier. Large cloud providers design their own infrastructure and work closely with chip vendors on roadmaps, so winning a production shipment is an important milestone but not a guarantee of long-term share. AMD’s ability to ramp Helios and Helion, support customers through deployment, and deliver on token-per-dollar claims will shape how much of the projected market it can capture.

The CPU side of the opportunity may be equally important. If AI agents drive demand for more general-purpose computing, AMD’s EPYC franchise could benefit alongside its Instinct accelerators. The company has raised its CPU total addressable market estimate three times in nine months, a sign that it sees faster growth than previously expected. A 50% compound annual growth rate over four or five years would be unusually strong for the server CPU market, but AMD argues that AI-related workloads are changing the demand curve.

Su’s remarks also highlight how the AI industry is moving from an experimental phase to an operational one. Training a frontier model is a discrete project, while inference and agentic applications run continuously and must be cost-effective at scale. That shift favors vendors that can offer complete systems, software support, and supply chain reliability. AMD is positioning Helios, Helion, and EPYC Verano as pieces of that broader infrastructure story, with the goal of selling more value per rack rather than competing only on individual chip specifications.

As production of Helios increases and Helion moves toward customer deployments, AMD will face scrutiny on execution. The company has set ambitious targets for market size and growth, and it has begun shipping to some of the most demanding customers in the industry. The next several quarters will show whether those early shipments translate into repeat orders and broader adoption. For now, Su’s message is clear: AMD believes inference and agentic AI will be the dominant workload drivers of the next infrastructure cycle, and it is building its product roadmap around that expectation.


Source: MSN News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy