Nvidia’s Vera Rubin Debuts Alongside a $100B OpenAI Pact — and a Loop Critics Can’t Ignore

*Dek: The new Rubin platform promises a 10x cut in inference cost, but the headline-grabbing 10GW deal with OpenAI puts a spotlight on how tightly the AI economy now revolves around a single chipmaker — and on money that flows in a circle.*

Nvidia arrived at the turn of 2026 with two announcements that, read together, define the shape of the AI buildout. The Vera Rubin platform was unveiled at CES in January. The OpenAI deal predates it: a letter of intent signed in September 2025 to deploy at least 10 gigawatts of Nvidia systems, backed by an Nvidia investment of up to $100 billion in OpenAI.

The hardware story is straightforward enough. Vera Rubin is a six-chip platform — pairing the new Vera CPU (88 custom Armv9.2-compatible Olympus cores) with the Rubin GPU, plus NVLink 6, ConnectX-9, BlueField-4 and Spectrum-6 networking parts. Nvidia claims each Rubin GPU delivers 50 petaflops of NVFP4 compute, 288 GB of HBM4 memory at 22 TB/s of memory bandwidth, and 3.6 TB/s of NVLink 6 GPU-to-GPU interconnect bandwidth. The platform can cut inference token cost by up to 10x versus Blackwell (on mixture-of-experts workloads at long sequence lengths) while needing 4x fewer GPUs to train MoE models. The flagship NVL72 rack packs 72 Rubin GPUs and 36 Vera CPUs. Nvidia confirmed chips were in production by January 2026, with partner products due in the second half of the year.

That second-half timing matters, because it lines up with the OpenAI deal. Under the announced letter of intent, OpenAI's first gigawatt of Vera Rubin systems is targeted for the back half of 2026, with Nvidia investing progressively "as each gigawatt is deployed." Sam Altman framed compute as "the basis for the economy of the future"; Jensen Huang called the partnership "the next leap forward."

Where the concentration risk lives

Look past the superlatives and a pattern emerges. The same list of early Rubin deployers — AWS, Google Cloud, Microsoft, Oracle Cloud, CoreWeave, Lambda and others — is effectively the entire frontier of AI compute. When one vendor's roadmap dictates when the world's largest model builders can scale, that is concentration, not just market leadership. A delay, a yield problem, or a pricing shift at Nvidia now propagates straight into every hyperscaler's capex plan.

The OpenAI arrangement adds a sharper concern that analysts have flagged across the sector: circularity. Nvidia invests up to $100 billion into OpenAI; OpenAI then spends heavily on Nvidia systems. The supplier is helping finance its own largest customer's purchases. That can accelerate deployment, but it also blurs the line between genuine end demand and demand that a vendor is, in part, underwriting. Revenue that loops back to its source is harder to read as a clean signal of market health.

It is worth stressing what is — and isn't — locked in. The OpenAI agreement is a letter of intent, described as non-binding, with terms to be finalized "in coming weeks." The $100 billion figure is a ceiling tied to staged deployment, not a wire transfer. Those caveats deserve to travel with the numbers.

None of this dims Rubin's engineering. But the more the AI economy's growth runs through one company's silicon — and, increasingly, its balance sheet — the more its risk concentrates there too.

Fontes

• https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
• https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems