Nvidia Announces Mass Production of AI-focused CPU Vera, Server CPU Demand Surges

For the past three years, AI hardware anxiety has centered almost entirely on GPUs, with CPUs relegated to booting tasks, moving data, and managing servers—rarely taking the spotlight. But AI agents have changed the nature of tasks: a typical chatbot receives text, runs a model, and generates text, whereas an agent must also read files, search the web, query databases, call APIs, and execute code. A single task might involve a dozen or more cycles of 'reasoning, acting, and observing,' each switching between GPU and CPU.

The GPU still handles the heaviest model computations, but parsing model outputs, determining tool types, making requests, running programs, managing files, collecting results, and scheduling the next round of tasks are typically handled by the CPU. Madhu Rangarajan, AMD's corporate vice president of compute and enterprise AI, stated that in a real agent pipeline tested by the company, 7 out of 8 stages run entirely on the CPU. This ratio comes from AMD's own testing and may not represent all systems, but it underscores that model inference is just one part of an agent's task; all other coordination falls on the CPU.

The overhead for a single agent may be small, but it becomes significant at scale. When enterprises deploy agents to tens of thousands of employees, concurrent tasks expand rapidly. Each agent parses outputs, executes code, and reads/writes data, accumulating CPU overhead into capacity demands. If the CPU is slow, the GPU must wait for the inputs needed for the next round of inference. Security checks add further burden: the more permissions an agent has, the more the system needs to inspect files, interfaces, and sensitive data. Rule matching, log analysis, and small model reviews typically stay on the CPU because they are distributed and latency-sensitive, making it inefficient to offload them to the GPU.

Another easily overlooked burden is tokenization. Before processing text, large language models must convert it into tokens. This process involves extensive string manipulation and conditional branching, which is not as suited to GPUs as matrix operations. In ordinary conversations, inputs are short, so tokenization latency is negligible. But agents continuously accumulate tool results, code, logs, and historical actions, and inputs can grow to tens or even hundreds of thousands of tokens.

Euijun Chung, a PhD student at Georgia Tech, noted in an interview with IEEE Spectrum that in some implementations, when a context of 100,000 tokens already exists and 1,000 tokens of tool results are added, the tokenizer may still reprocess the entire input. Prefix caching and incremental tokenization can reduce redundant work, so this is not necessarily how all systems operate, but the CPU pressure from long texts remains.

Chung's team tested models like Llama 3.1 and Qwen3 in their latest paper. In long-sequence, large-batch configurations, tokenization accounted for up to about 80% of first-token latency. When CPU cores are insufficient, tokenization threads compete with task scheduling and GPU kernel launches for resources, leading to CPU saturation while the GPU is underutilized. Adding CPU cores improved first-token latency by 1.47-5.15x across different configurations, without needing additional GPUs.

Industry data has begun to reflect this shift. In January 2026, KeyBanc, based on supply chain surveys, concluded that Intel had sold out most of its server CPU capacity for the year, and AMD's server CPU supply was also near full capacity, giving both companies room to raise prices. This is an analyst's channel assessment, not disclosed order data from the manufacturers.

Intel's subsequent financial filings provided partial confirmation. The company stated that server CPU demand exceeded available supply in the first quarter of 2026, with internal capacity constraints expected to persist through year-end. Server product average selling prices rose 27% year-over-year, while shipments fell 5%. Intel attributed the price increase to a higher mix of premium products, demand-based pricing, and increased costs.

AMD's data center business revenue reached $5.8 billion in the same period, up 57% year-over-year, with growth coming from both EPYC server CPUs and Instinct GPUs. AMD did not separately disclose server CPU revenue growth, so it is not possible to attribute all of the 57% to agent demand. The company's future outlook is more aggressive: AMD previously expected the server CPU market to grow about 18% annually; in May 2026, it raised that forecast to over 35%, projecting a market size exceeding $120 billion by 2030.

Data centers have traditionally used one CPU to manage four to eight GPUs. AMD has observed that agent systems are pushing this ratio toward 1:1, with some workloads using even more CPUs. This is still a supplier's assessment based on customer demand; actual configurations vary widely depending on the workload. The direction of the change is clear, however: adding a few more CPUs is no longer enough, and cloud providers are beginning to consider building dedicated CPU racks for agents.

In March 2026, Arm made its first move from selling processor designs to directly selling data center CPUs. Its AGI processor has up to 136 cores and was developed with Meta. Arm estimates that, at the same power consumption, agent data centers may require up to four times more CPU capacity than existing data centers. Qualcomm followed with the Dragonfly C1000, featuring over 250 cores, and signed a multi-generational server CPU cooperation agreement with Meta.

Within a few months, Intel, AMD, Arm, Qualcomm, and Nvidia have all incorporated 'agent CPUs' into their product roadmaps, indicating that this competition has moved beyond the traditional x86 market to include Arm servers, custom cores, and full rack designs. Nvidia holds a unique position: it sells both the GPUs that create CPU pressure and the Vera CPUs that alleviate it. From processors to networking to storage, Nvidia can provide the entire stack. Vera is thus not just a product line expansion; it also protects GPU utilization.

When chipmakers follow the money, someone has to bear the cost. As with GPUs and memory chips, this capacity shift is likely to be paid for by consumers. In its Q4 2025 earnings call, Intel acknowledged that it is trying to shift internal wafer supply toward data centers while increasing external foundry use for client products. Management later added that Intel cannot completely exit the client market.

By Q1 2026, Intel's client CPU average selling prices rose 16% year-over-year, while unit sales fell 13%. The company expects supply constraints to last at least through the first half of the year. Changes in client products cannot be entirely attributed to agents; Intel's own process node transitions, internal and external capacity, and memory supply also affected pricing and shipments. But with limited capacity, server chips have more cores and higher prices, and cloud providers are willing to sign long-term orders, so production resources naturally shift toward data centers.

Over the past three years, consumers have already experienced GPU price hikes and shortages. Next, the same may happen with CPUs.