$120 billion → $220 billion! AMD Revalues CPU in the AI Era

The competition for AI infrastructure continues to intensify, but the market's focus is no longer limited to GPU performance; rather, it has shifted to a re-evaluation of the entire computing system.

A research report released by Barclays on July 23 indicates that the most significant signal sent by AMD at the Advancing AI conference was not about the new generation of GPUs or AI servers, but rather a substantial upward revision of its long-term outlook for the server CPU market.

The company expects that by 2030, the global server CPU market will reach $220 billion, nearly doubling from the prior forecast of $120 billion. This reflects that, driven by Agentic AI, CPUs are transitioning from a supporting role in AI servers to becoming a core computing resource.

On this basis, AMD further raised its forecast for the overall computing market. The company expects that by 2030, the data center AI accelerator market will reach $1.4 trillion, while the serviceable compute market (Compute TAM), which includes server CPUs, AI accelerators, and other sectors, will expand from $365 billion in 2025 to about $2 trillion.

Barclays believes this signifies that competition around AI infrastructure is evolving from single GPU performance to a full-stack competition that encompasses CPUs, GPUs, networking, software, and system architecture.

Server CPU Market Size Significantly Revised Upwards

The most noteworthy aspect of this event for the market was not the Helios rack or the new generation GPU, but AMD’s redefinition of the future scale of the server CPU market.

The company projects that by 2030, the server CPU market will reach $220 billion, with a corresponding compound annual growth rate now significantly revised upward from over 35% to more than 50%. Barclays considers this aggressive revision as reflecting a major shift in AMD’s outlook on future AI server architecture.

The core logic supporting this forecast is the rapid development of Agentic AI. AMD anticipates that in the future, "Agentic CPUs" aimed at agent workloads will account for more than half the overall server CPU market.

As more AI applications shift from training to real-world deployment, CPUs will undertake tasks such as model scheduling, agent operations, task collaboration, and data management. They are no longer merely auxiliary components in GPU servers but will become an important and independently growing market.

Inference Becomes New Focus for AI Infrastructure Competition

Beyond the re-evaluation of market space, AMD's outlook on AI workload structure has also drawn attention from Barclays.

The company projects that by 2026, inference will become the largest AI compute scenario for the first time, accounting for about 60% of all AI workloads. This compares to 50% and 40% in 2025 and 2024, respectively. This means that the industry’s competition focus is shifting from "who can train the largest models" to "who can run models more efficiently and cost-effectively".

Barclays believes this is also the key reason why AMD simultaneously introduced the Helios system platform, MI455X GPU, Ethernet interconnect, and the ROCm.AI software platform. The future competition for AI infrastructure will no longer be limited to GPU compute power but will revolve around system-wide cost, memory capacity, bandwidth, network interconnect, and software ecosystem.

The report notes that the industry as a whole is still in a state of tight compute supply. As major cloud vendors continue to expand and existing customers enter the volume deployment phase, growth in AMD’s data center business is expected to outpace the overall market. However, the conference did not announce any new major customers, with AMD’s cooperation with Anthropic previously disclosed.

From Chips to Platforms: AMD Refines Its Product Roadmap Through 2030

In line with these judgments, AMD further refined its product plan for the coming years.

On the hardware front, the company officially launched the Helios AI rack, integrating the MI455X GPU, Venice CPU, Salina DPU, and Vulcano AI NIC into a complete system solution. The MI455X is configured with 432GB HBM4 and 23.3TB/s memory bandwidth, supporting up to 40PFLOPS of FP4 performance. AMD expects that compared to its previous generation, token throughput can increase by up to 34 times, and token cost can be reduced by up to 18 times.

AMD also focuses its competitive edge on system efficiency. The company expects that compared with Nvidia's Vera Rubin platform, Helios offers not only a larger HBM capacity and higher horizontal interconnect bandwidth, but also higher token output per dollar, demonstrating that the competition is shifting from single GPU performance to the total cost of ownership (TCO) of the entire system.

Meanwhile, AMD extended the CPU and GPU roadmap through 2030. The Venice processor is now in production and set for deployment on cloud platforms and OEM servers starting in Q4; the Florence and Ravenna CPU generations are scheduled for release in 2028 and 2030, respectively. For GPUs, the MI500 is expected in 2027 with MI600 following in 2028. Notably, the MI500 will be the first to employ both copper and optical interconnect technology.

Beyond hardware, AMD also introduced the ROCm.AI AI development platform and continues to expand its product portfolio around inference, networking, and robotics ecosystems, including an uncoupled inference solution developed in collaboration with Cerebras, the professional inference card MI350P, and the KRIA AI System-on-Module robotics platform, aiming to further strengthen its full-stack capabilities for AI infrastructure.

Barclays believes that, compared to specific product performance, what truly changed market expectations at this Advancing AI conference was AMD's redefinition of future compute demand scale and its latest assessment of the evolution direction for AI infrastructure. If inference demand accelerates as expected, the importance of server CPUs and system-level solutions may exceed previous market expectations.