Nvidia rose to become the world's most valuable company driven by relentless demand for its GPUs, the primary chip for building and running AI. However, Nvidia has begun shipping its own central processing units (CPUs), giving cloud providers an alternative to chips from AMD and Intel and creating a fresh competitive front in AI servers.
On Tuesday, Nvidia disclosed fresh details about its Vera data center CPU, such as specifications and benchmarks that potential buyers require to thoroughly assess the chip.
Nvidia is widely regarded as the trendsetter for the IT industry. In the CPU arena, however, Nvidia is again an underdog, facing entrenched rivals Intel and AMD that maintain strong relationships with hyperscalers and cloud providers. Nvidia's CPU push is another instance of its vertical integration strategy, aiming to manufacture more of its own chips and components each year. Instead of selling chips individually, it intends to offer complete rack-scale computing systems.
Before the AI boom, CPUs were the most critical server component. The first AI servers in 2022 used one CPU per eight GPUs, signaling a shift toward Nvidia. However, the emergence of agentic AI has refocused interest on CPUs, as they must supply data to agents.
Get the market news that matters in a five-minute read with Market Briefs, our free daily newsletter
Ian Buck, Nvidia's vice president of hyperscale, stated during a presentation last week that agents have made CPUs "much more integral," adding, "Particularly how fast a CPU can answer one question." Nvidia projects the entire server CPU market may reach $200 billion.
Last week, Nvidia announced that Vera is its first server CPU built from the ground up, rather than relying on a ready-made Arm design. Nvidia explained that earlier cloud CPUs from Intel and AMD prioritized core count, whereas its Vera chip and Olympus core emphasize single-core speed.
In a presentation, Nvidia product marketer Hannah Coutand said the chip "emphasized per-core speed, high memory bandwidth, and latency," adding, "so that agents can return to their GPUs as quickly as possible and keep those GPUs, which are a very expensive and a highly valuable asset in the AI factory, as highly utilized as possible."
Nvidia will sell Vera as a standalone chip and also bundle it with its GPUs, including in a liquid-cooled rack of 256 chips, a dual-chip server, or the Vera Rubin system. The chip supports up to 1.5 TB of low-power memory per chip.
Karl Freund of Cambrian AI Research noted that Nvidia's CPU is not for traditional server tasks but for intense AI workloads. However, convincing cloud providers to buy may remain difficult, even for Nvidia. Coutand said Vera is in "early innings." Nvidia's partner list included only Oracle among major cloud providers.
"The CPU is something they've done to kind of unhook their customers from using Intel or AMD CPUs, and they covet that revenue," said Freund. "What they've done is they decided to focus on a unique CPU that isn't available in the market from anyone right now."
Adoption is early; Nvidia only named Oracle as a cloud partner. But Nvidia has direct ties to OpenAI and Anthropic, making them natural customers for the full package.
The bottom line: Nvidia is creating a new CPU category that incumbents cannot yet replicate. If it works, a huge revenue stream opens; if not, incumbents hold a big lead. The expanding AI server market may accommodate both.
Join Market Briefs, our free daily newsletter, for a quick daily rundown of the markets
