Nvidia Is Pushing Hard to Own Every Chip Inside AI Data Centers

Key points

• Nvidia's new Vera Rubin chip system claims to process ten times as many AI tasks per watt as its previous Grace Blackwell system.
• OpenAI already has one Vera Rubin rack running in its facilities, Nvidia confirmed at a technical briefing in June 2025.
• Nvidia is selling the Vera CPU, the processing brain inside the new system, as a standalone product and told Chinese customers it could ship as early as August 2025.
• AMD revealed its competing Helios AI chip rack on Sunday, the same week Nvidia staged its Vera Rubin briefing, in a clear race for multi-year chip contracts with companies like Meta, Amazon, Anthropic and OpenAI.

Nvidia built its fortune selling GPUs, the specialised chips that do the heavy number-crunching AI needs. Now it wants to sell the other kind of chip inside AI data centers too.

At a briefing for journalists at its Santa Clara headquarters last week, first reported in detail by Wired AI, Nvidia executives laid out the case for Vera Rubin, its next-generation chip system. The headline claim: Vera Rubin processes ten times as many tokens per watt as its predecessor. A token is a small chunk of text or data, and tokens per watt measures how much AI work a system can do for a given amount of electricity.

That matters to companies running AI at scale. Electricity is one of the biggest costs in a data center, and efficiency gains translate directly into lower bills.

Why is Nvidia selling CPUs now?

Because AI is getting more complicated. GPUs crunch numbers fast. CPUs, the general-purpose processors that have run computers for decades, are better at coordinating tasks and managing data flows. As companies build AI agents, software that can carry out multi-step tasks on its own like booking a meeting or analysing a spreadsheet without a human clicking through each step, they need more CPU firepower alongside their GPUs.

Vera Rubin pairs one CPU with every two GPUs. A full Vera Rubin NVL72 rack, a stack of chips packed into a single liquid-cooled unit, holds 36 Vera CPUs and 72 Rubin GPUs.

Nvidia also claims the new system is far easier to install than earlier products. It describes Vera Rubin as "cable-free compute" and says setup time can drop from a couple of hours per rack to a few minutes. Liquid cooling, which uses less energy than blowing air over hot chips, runs throughout the whole system.

The company does benchmark the Vera CPU as faster than rival chips from AMD and Intel. The comparison tests apparently used slightly older generations of those competitors' products, which is the kind of fine print that deserves attention when a company grades its own homework. Ian Buck, Nvidia's vice president of accelerated computing, framed the pace of change plainly: "We're on a roadmap to crank out new architectures, not just GPUs but CPUs."

Early customers named by Nvidia include Microsoft and Oracle alongside OpenAI. Vera Rubin is expected to ship broadly in the second half of 2025.

We've been tracking Nvidia's chip ambitions closely. Our 17 July story on H200 exports to China is a useful reminder that hardware availability and geopolitics are never fully separate conversations.

For anyone whose job touches AI tools, the practical upshot is simpler than the chip specs. Faster, more efficient hardware means the AI services you already use, from writing assistants to customer-service bots, get cheaper to run. That tends to make them more widely deployed. Whether that's a threat or a gift depends almost entirely on how you use them.

Takeaway: The efficiency story is real, but watch the benchmark fine print. If your organisation is evaluating AI tools this year, ask vendors which hardware generation their product runs on. Efficiency improvements at the chip level often show up as lower per-use costs within 12 to 18 months.