Nvidia became the most valuable company on earth by selling one kind of chip, the graphics processor, to a world desperate to train AI. So it says something about where the technology is heading that the company spent this week talking up a different chip entirely. On Tuesday Nvidia published the full specifications and benchmarks for Vera, its first data-center CPU built around a core it designed itself, called Olympus. In the market it is trying to enter, the GPU king is, for once, the challenger.
The reason is the same force reshaping every other AI story right now: agents. When ChatGPT launched, an AI server might pair eight power-hungry GPUs with a single CPU, and the CPU was an afterthought. But an autonomous agent does not spend its life doing matrix math. It runs code, queries repositories, waits on tools, and feeds data back and forth, all work that lands on the CPU. As one Nvidia executive put it, agents have made the CPU "much more integral," and the ratio of GPUs to CPUs has been sliding toward one-to-one. Vera is built explicitly for this: rather than piling on cores, it dedicates enormous die area to a fabric moving 3.4 terabytes per second between them, chasing low latency so that "agents can return to their GPUs as quickly as possible." Nvidia claims it edges out AMD's flagship server chip on a key benchmark despite having fewer threads.
What makes the moment striking is the scoreboard. Nvidia's stock is up a mere 8 percent this year, and roughly a trillion dollars has been shaved off its peak valuation. The two best-performing chip stocks of 2026 are not Nvidia at all, but Intel and AMD, up 149 and 128 percent, as investors bet that the CPU incumbents will ride the agent wave. AMD already holds around a third of the server-CPU market and deep relationships with the cloud giants, and it launches its next-generation Venice chips this week, which are expected to leapfrog Vera on memory bandwidth. Nvidia, the disruptor of the last three years, now finds itself pitching against entrenched rivals on their home turf.
Nvidia's answer is not really to win that turf. Its executives are candid that Vera is a specialist, tuned for data-rich agentic work and traded off against the "legacy" tasks that still make up most of what servers do. The pitch is vertical integration: sell the whole rack, the Vera Rubin system that binds its CPUs, GPUs and networking into one machine, and quietly unhook customers from Intel and AMD in the process. Early buyers, including OpenAI, Anthropic and SpaceX, received the first chips in June, and Nvidia says it wants to grab share in a market it insists is expanding fast, not fight over a fixed pie.
Strip away the specifications and there is a tidy lesson here about how booms mature. The first phase of the AI build-out rewarded one component so completely that it minted the most valuable company in history. The second phase, the one where models stop answering single questions and start doing long chains of work, is spreading the value back out across the rest of the machine. The humble CPU, written off as plumbing, turns out to be the thing that keeps the expensive silicon busy. And the company that owns the expensive silicon has just conceded, by shipping a CPU of its own, that the plumbing was worth fighting for all along.