At the AI Infra Summit in Santa Clara, Nvidia's Ian Buck and Intel's Lip-Bu Tan spoke one after the other, and both kept returning to the same component: the CPU. For three years the AI infrastructure conversation has been a GPU conversation. Now the company that created the GPU boom and the largest maker of general-purpose processors are describing the same shift, in the same week, from the same stage — AI agents are creating enormous new demand for the chip the industry had stopped talking about.
Buck made the case with one comparison. A typical AI chat in 2023 ran about 1,000 tokens across three exchanges, with a person stopping after each one to read the answer and type the next prompt. In the agentic workload he showed, the average request carried 142,000 tokens and 65 sequential steps. The person waits for the result; the agent keeps going on its own.
That is a 142-fold jump in tokens per request, and it is worth holding next to the forecast both companies are leaning on: Goldman Sachs Research expects token consumption to grow 24 times by 2030 as agentic AI spreads among users and companies. The two figures measure different things — one request against total industry volume — but the smaller number is the one describing six years of adoption, and the larger one describes a change that has already happened inside a single task.
The reason this lands on the CPU is structural. A GPU is good at enormous quantities of parallel math, which is why it suits model training and token generation. The CPU does the varied work around the model: it runs secure sandboxes, compiles code, works with Git, searches files, retrieves context, calls tools, checks results and decides what to do next. Each of those steps can be small. A single agent task can involve dozens of them.
Nvidia's own framing for this is "a new CPU moment in the AI factory." By the company's explanation, sandboxes, tool calls, orchestration and context retrieval from long inputs all add load to the CPU.
Nvidia's new Vera CPU is built for exactly that work. In the results Buck presented, Vera was 1.6 times faster across more than 80 agentic tasks, with gains reaching two times in compile and Git cycles. Perplexity separately reported that secure agent sandboxes started 1.9 times faster. That last number matters more than it looks: for short jobs, creating the sandbox can take longer than the work done inside it.
Intel is sending the same signal from the other side of the market. Tan said the company is allocating CPU shipments quarter by quarter and fielding calls from other companies' executives asking for more volume. He tied the demand to reinforcement learning, agent orchestration and a changing ratio between CPUs and GPUs. Earlier this year he described the broader transition he expects: the next wave of AI moving intelligence closer to the end user, from foundation models through inference to agentic systems. In Intel's first quarterly results, the company said that transition had noticeably increased demand for its CPUs.
There is a genuine inversion in this. Nvidia, which made the GPU a household term, is shipping its first CPU of its own and selling separate systems built around it. Intel, routinely written off as the company of the previous computing era, is getting a new lease from the AI buildout.
OpenAI's Jalapeño chip adds a third angle. It is a custom inference accelerator — an ASIC — developed with Broadcom. Its advantage comes from OpenAI being able to tie hardware development tightly to its own models, kernels, compiler, memory, networking and serving software. Because the company controls all of those layers, its engineers can see exactly where models slow down, burn too much power or move more data than they need to, then design silicon around those specifics. OpenAI says Jalapeño will be able to run current and future large language models from different developers, so the chip is not locked to one model. The extra advantage is alignment: one customer, one workload, one product roadmap, with chip and models evolving together.
Here is where I would push back on the neatness of the story. Two of these three pieces of silicon are being sold by the people presenting the evidence for needing them, and the evidence is presented without a stated baseline. Vera is 1.6 times faster — than what? The Russian and English coverage alike carries the multiple without naming the processor it beat, and a multiple without a denominator is a marketing number, not a benchmark. Perplexity's 1.9x on sandbox startup is the more useful datapoint precisely because it comes from a customer describing its own workload. And Jalapeño sits oddly in a story about general-purpose compute: an inference ASIC is the least general-purpose chip in the building. It belongs to the same shift only in the sense that everyone has concluded the GPU alone is not the whole machine.
The economics are what make the argument serious regardless of who is making it. Money depends on utilization. Nvidia argues the governing infrastructure metric is becoming agentic tokens per megawatt. A fast GPU creates no value while it waits for the CPU to launch a tool, move data or finish an orchestration step; during that wait the expensive accelerator draws power and does nothing useful. That reframes five practical questions at once. Infrastructure buyers should be sizing the CPU-to-GPU ratio, not just counting GPUs. Teams running agents should measure end-to-end task time, including CPU and tool work. CFOs should be asking how often expensive accelerators sit idle waiting for the rest of the system. Developers should profile sandbox startup, compilation, retrieval and tool calls. Investors should watch CPUs, memory and orchestration software alongside accelerators.
The more interesting question is what happens to the current accounting if this holds. Datacenter capex has been justified with GPU counts and GPU depreciation schedules, on the assumption that the accelerator is the thing doing the work. If a meaningful share of an agent's wall-clock time is spent compiling, retrieving and orchestrating, then some portion of those GPU-hours was never compute at all — it was waiting, billed at accelerator prices. Nobody on that stage put a number on the idle fraction, and that is the number that would settle the argument.
GPUs remain critical. But an AI factory is only as productive as its slowest component, and as agents move from answering questions to performing tasks, the chip that keeps them running may be the one the industry spent three years treating as overhead.