AMD CTO: Agentic AI Needs More CPUs, Not Just GPUs

At the RAISE Summit in Paris, AMD CTO Mark Papermaster issued a warning that could redefine the architecture of AI-focused data centers: agentic artificial intelligence — autonomous systems that plan and execute complex tasks — will need far more CPUs than currently anticipated. According to Papermaster, the balance between CPUs and GPUs in AI servers is about to shift dramatically.

What happened?

During his talk, Papermaster explained that AI agents don't just run models; they also require orchestration, workflow management, database interaction, and sequential decision-making. All of this falls on CPUs. He cited internal AMD data showing that in agent workloads, the CPU-to-GPU time ratio can reach 40-60%, far above the typical 10-20% in simple inference tasks. (Source: The Next Web, RAISE Summit 2026) Papermaster also noted that AMD has been working with hyperscaler partners to test agent workloads, and preliminary results indicate CPU demand could double current projections by 2028. At the same event, it was mentioned that companies like Salesforce and ServiceNow are already redesigning their agent platforms to optimize CPU usage, suggesting the shift is not just theoretical.

Why is this important?

This statement comes at a time when the industry is heavily focused on GPUs for AI, with Nvidia dominating the market. If Papermaster's prediction holds, data centers will need to be redesigned to include more high-performance CPUs, benefiting AMD, which competes with Intel in this segment. It could also slow the adoption of AI agents if existing infrastructures are not prepared. Historically, the industry has gone through similar cycles: in the 2000s, high-performance computing (HPC) relied almost exclusively on CPUs, but with the arrival of CUDA in 2007, GPUs took center stage. However, AI agents reintroduce sequential tasks that GPUs cannot parallelize efficiently, reminiscent of the 2000s debate over CPUs vs. GPUs in HPC, where the hybrid model prevailed. Back then, IBM and AMD advocated for balanced systems, while Nvidia pushed the GPU as the sole solution. Now, history may repeat itself, but with a twist: agents require a reasoning loop that includes planning, execution, and verification — inherently sequential steps. Papermaster noted that “40% of compute time in a typical agent is spent on control logic, memory access, and I/O operations, tasks that CPUs handle better.” This contrasts with simple inference, where 80% of the work falls on GPUs.

Market implications

• AMD: Could see increased demand for its EPYC server CPUs, especially if AI agents become widespread. The company reported record revenue in its data center division in Q2 2026, with 45% year-over-year growth, partly driven by CPU demand for AI. Papermaster hinted that AMD is developing a new CPU line with agent-specific instructions, potentially launching in 2027.
• Nvidia: While its GPUs will remain essential, the need for more CPUs could reduce its relative dominance in AI workloads. Nvidia is already responding: at GTC 2026, it unveiled the Grace Hopper Superchip, integrating an ARM CPU and a GPU, but Bernstein analysts note that “Nvidia's CPU still doesn't compete with EPYC or Xeon for orchestration tasks.” Moreover, if agents require 40% CPU, Nvidia's current systems (with one CPU per eight GPUs) may be insufficient.
• Intel: Would also benefit, but its manufacturing process delays could limit its response. Intel saw a 20% drop in data center revenue in 2025, but its new Granite Rapids chip (2026) promises AI performance improvements. However, Papermaster claimed that “AMD benchmarks show EPYC outperforms Xeon by 30% in agent workloads,” potentially giving AMD an edge.
• Hyperscalers: AWS, Azure, and Google Cloud will need to adjust their instance offerings for agentic AI, possibly creating new VM types with more CPUs. AWS already announced “Agent-Optimized” EC2 instances with 64 vCPUs and 4 GPUs, a 16:1 ratio, compared to the typical 8:1. Azure is testing servers with 128 CPUs and 8 GPUs for agent workloads. This could increase infrastructure costs by 20-30% for companies adopting agents at scale, according to Gartner estimates.

Additionally, startups like Cerebras and Groq, which offer specialized inference hardware, could be affected if agents require more CPU. Cerebras, which uses a full wafer as an accelerator, has optimized its systems for sequential inference, but Papermaster questioned whether its architecture can handle complex agent orchestration.

What readers should know

It's not that GPUs will become unimportant, but that AI system architecture must become more balanced. AI agent developers will need to optimize CPU usage, and companies planning large-scale agent deployments should factor CPU capacity into their infrastructure budgets. For example, a company deploying a customer service agent might need 4 CPUs per GPU, rather than the current 1:1 ratio. This implies redesigning server clusters and potentially increasing total cost of ownership (TCO) by 15-25%.

“The future of AI isn't just more GPUs; it's a dance between CPUs and GPUs, and agents are changing the choreography,” Papermaster said.

Developers should also note that agent frameworks like LangChain and AutoGPT are evolving to better leverage CPUs. LangChain recently announced an update that allows planning tasks to run on CPUs while GPUs handle inference, potentially reducing latency by 40%.

Historical context

Over the past decade, the trend has been to offload more work to GPUs, from training to inference. However, AI agents require a reasoning loop that includes planning, execution, and verification — steps that are inherently sequential and difficult to parallelize on GPUs. This recalls the 2000s debate over CPUs vs. GPUs in high-performance computing, where the hybrid model prevailed. In 2006, Nvidia's Tesla GPU launch marked the beginning of the GPU era, but in 2012, CPU usage in HPC remained significant. Now, with agents, the CPU is regaining prominence. Papermaster compared the situation to the rise of microservices: “Just as microservices required more CPUs for orchestration, agents will do the same.” Indeed, companies like Netflix already report that their agentic AI workloads consume 35% CPU, up from 15% two years ago.

Speculation and caveats

Papermaster's claims are not backed by public third-party data. AMD may be positioning itself to sell more CPUs. Additionally, the definition of “agentic AI” is broad and varies by vendor. Therefore, these figures should be taken with caution until independent benchmarks are available. For instance, the MLPerf benchmark does not yet include agent workloads, though it is expected to by 2027. It's also possible that Nvidia will respond with hardware integrating more CPUs, such as its Grace Hopper superchip, which already combines a 72-core ARM CPU with an H100 GPU. However, analysts at Moor Insights & Strategy warn that “Nvidia's ARM CPU is not designed for complex orchestration tasks, and its performance may be inferior to EPYC.” Finally, Papermaster himself acknowledged that “the figures are preliminary and based on internal simulations; actual results may vary.” Until large-scale deployment data is available, companies should plan with flexibility.