AMD argues that agentic AI doesn’t reduce GPU sales; it also boosts CPU demand

AMD has aimed to settle one of the most repeated debates in artificial intelligence infrastructure: whether the growth of agent-based AI and inference can shift some spending from GPUs to CPUs. Lisa Su, the company’s CEO, stated during the Q1 2026 earnings presentation that this additional processor demand is “largely additive” to the accelerator market, not a direct replacement.

The underlying explanation is simple. Large models still require GPUs and accelerators for training, heavy inference, and efficient large-scale execution. But AI agents add another layer of work: coordination, data movement, parallel task execution, tool calls, memory management, context setup, and flow control. All of this increases the role of CPUs within data centers.

CPUs Return to the Center of AI Industry Debate

Over the past two years, the conversation about AI infrastructure has been dominated by GPUs. NVIDIA has set the market’s pace, and hyperscalers have competed to secure accelerators, HBM memory, high-speed networks, and power capacity. But inference and agents are changing the composition of those systems.

Lisa Su explained that agent-based AI is increasing the need for CPU compute in servers because these workloads require more processing for orchestration, data movement, and parallel execution, in addition to using CPUs as control nodes for GPUs and accelerators. AMD is already reflecting this trend in its results: the data center segment reached $5.8 billion in Q1, a 57% year-over-year increase, driven by demand for EPYC processors and higher GPU Instinct shipments.

This perspective helps explain why AMD has revised upward its CPU market outlook for servers. According to Reuters, the company now expects the addressable CPU server market to grow more than 35% annually, surpassing $120 billion by 2030, up from an 18% annual growth estimate made in November.

The thesis is not that data centers will buy fewer accelerators. AMD’s view is that they will buy more of everything, but in different proportions. In traditional AI setups, CPUs often act primarily as host nodes, with ratios around one CPU per four or eight GPUs. In scenarios with many agents, Su suggested that this ratio could approach one-to-one, or even tilt toward more CPUs than GPUs in certain deployments.

This is significant because it positions AMD as less dependent solely on winning the accelerator race. While competing with NVIDIA in GPUs via its Instinct line, AMD also has a strong foothold in servers with EPYC. If agent-based AI prompts a rethink of the CPU-GPU balance, AMD could benefit on both fronts.

MI450, Helios, and the Push for Large-Scale Inference

The GPU side remains essential. AMD confirmed it has started sending samples of its Instinct MI450 to leading clients and maintains its schedule to commence production shipments of Helios in the second half of 2026. The company reports that customer forecasts already surpass initial plans, with more ongoing discussions for large deployments involving several gigawatts.

MI450 and Helios are key components in AMD’s strategy to compete with NVIDIA in full-rack AI infrastructure. Merely selling chips isn’t enough anymore: large clients want integrated solutions with accelerators, CPUs, memory, networking, cooling, and software. AMD aims to provide an open, standards-based alternative, with EPYC, Instinct, Pensando, and ROCm as core components.

AMD had already detailed that each MI450 GPU can reach up to 432 GB of HBM4 memory and 19.6 TB/s memory bandwidth. At rack scale, a Helios system with 72 MI450 GPUs could deliver up to 1.4 exaFLOPS in FP8 and 2.9 exaFLOPS in FP4, with 31 TB of HBM4 memory integrated.

Inference appears to be the most promising application field. Based on insights from the call, AMD sees deployments of MI450 for both training and inference, but larger projects are likely geared toward inference. This makes sense: once models are trained, serving responses to millions of users, agents, and enterprise applications becomes a persistent and growing load. Memory, token cost, energy consumption, networking, and stable scalability are critical factors.

The next step is already in sight. Su also mentioned that many clients are discussing AMD’s upcoming MI500 series, indicating that accelerator roadmaps are being planned well in advance. In AI, big buyers don’t plan quarter-to-quarter; they reserve power capacity, racks, memory, packaging, and manufacturing capacity years ahead.

Strong Results and an Outlook Above Expectations

The financial results provide context for this optimism. AMD reported revenue of $10.253 billion in Q1 2026, a 38% increase year-over-year. GAAP net income was $1.383 billion, and diluted earnings per share (EPS) were $0.84. On a non-GAAP basis, EPS was $1.37, with a gross margin of 55%.

Free cash flow reached $2.566 billion, up from $727 million in the same quarter last year, with a free cash flow margin of 25%. The company ended the quarter with $12.347 billion in cash, equivalents, and short-term investments.

For the second quarter, AMD expects revenue around $11.2 billion, with a ±$300 million variance. The midpoint indicates nearly 46% YoY growth and approximately 9% sequential growth. The company also forecasts a non-GAAP gross margin of about 56%.

The market responded positively to the outlook. Reuters reported a roughly 12% increase in extended trading after the earnings, supported by strong data center chip demand and higher-than-expected forecasts.

The New Balance: More GPUs, but Also More CPUs

The most interesting discussion is shifting from quarterly results to the architecture of future data centers. Agent-based AI doesn’t operate as a simple query to a model. An agent can split a task, query a database, call an API, generate code, verify results, ask for context again, execute other tools, or coordinate with other agents. Each action adds workload around the accelerator.

In this environment, while GPUs remain central for intensive compute, CPUs are gaining importance as coordinators and general processors. Networking, memory, storage, and orchestration software requirements also increase. Infrastructure is becoming more like a factory of distributed tasks rather than a single inference engine.

Thus, Lisa Su emphasizes that increased CPU demand doesn’t cannibalize the GPU market. If agents multiply useful tasks that can be automated, overall infrastructure spending might rise. This is a pragmatic version of Jevons’ paradox applied to AI: lowering the cost of a technology leads to more uses and potentially higher total consumption.

For AMD, this narrative is highly favorable. It positions AMD not only as an alternative to NVIDIA in accelerators but as a comprehensive AI infrastructure provider: EPYC CPUs for servers and control nodes, Instinct GPUs for training and inference, Pensando networks, ROCm software, and rack platforms like Helios.

The challenge lies in execution. NVIDIA still holds a clear lead in software, ecosystem, solution availability, and adoption. AMD needs to deliver MI450 on time, secure HBM4 memory, improve ROCm, expand its customer base, and turn initial interest into actual deployments. It also depends on TSMC and a supply chain strained by the same AI demand that is boosting its numbers.

Intel adds another layer of competition. The recent revaluation of CPUs in AI also benefits its longtime rival, which aims to leverage its own manufacturing capacity to regain ground in servers. Reuters notes that Intel is accelerating its internal production while AMD continues to rely on external foundries like TSMC—a significant difference given how critical manufacturing capacity is in this market alongside chip design.

Nevertheless, AMD’s message is clear: the next phase of AI will not be just a GPU race. It will be a race of complete systems. Inference and agents require balancing accelerators, CPUs, memory, networking, and software. If this viewpoint is confirmed, AMD could find more avenues for growth than the market traditionally associated with its data center business.

Frequently Asked Questions

What did Lisa Su say about agent-based AI?
Lisa Su argued that agent-based AI increases CPU demand for orchestration, data movement, parallel execution, and control nodes, but this demand is largely additional to the GPU market.

Can CPU demand reduce GPU sales?
AMD believes not significantly. According to their view, accelerators are still necessary for models, while agents create more additional workload for CPUs and supporting systems.

What are AMD MI450 and Helios?
MI450 is the upcoming generation of Instinct GPUs for AI, and Helios is AMD’s rack platform integrating GPUs, EPYC CPUs, networking, and software for large training and inference deployments.

When will the MI450 production shipments begin?
AMD states it has already started sending samples to key clients, with production shipments of Helios scheduled for the second half of 2026.