NVIDIA and OpenAI ink massive 10GW AI compute pact

Back in 2016, Jensen Huang personally hand-delivered the very first NVIDIA DGX supercomputer to a tiny non-profit called OpenAI. Nine years later, that spark has turned into the most aggressive infrastructure deal in tech history.

OpenAI and NVIDIA have signed a letter of intent for a landmark strategic partnership to deploy at least 10 gigawatts of NVIDIA systems. To power this rollout, NVIDIA plans to progressively invest up to $100 billion into OpenAI as each gigawatt comes online.

The initial phase is targeted to go live in the second half of 2026, delivering the first gigawatt on NVIDIA’s upcoming Vera Rubin platform.

The 10-gigawatt AI factory

To put 10 gigawatts in perspective, that is roughly the power output of several commercial nuclear reactors, now earmarked entirely for millions of cutting-edge GPUs.

OpenAI president Greg Brockman noted that this planned setup represents a staggering one billion times more computational power than that first DGX box delivered in 2016.

Under the terms, NVIDIA becomes OpenAI’s preferred compute and networking partner. The two will co-optimize OpenAI’s models and software stack directly with NVIDIA’s silicon and system architectures.

The companies also made clear this setup will complement existing mega-builds with collaborators like Microsoft, Oracle, SoftBank, and the Stargate initiative.

A 700-million user compute wall

OpenAI’s appetite for electricity and silicon is not theoretical. The platform now boasts over 700 million weekly active users, fueled by rising demand for agentic workflows, complex reasoning models, and massive multimodal processing.

Speaking to CNBC, Sam Altman argued that while the cost per unit of intelligence continues to plummet, frontier model capabilities are rising so fast that aggregate demand is exploding.

Without an unprecedented buildout, Altman warned, society might face an absurd trade-off between using compute to cure cancer or providing free education. His answer: build enough capacity so no one has to choose.

Huang echoed the sentiment, calling this project the biggest AI infrastructure effort in human history, adding that this first 10 gigawatts is just the beginning.

Why it matters

This agreement resets the bar for what it costs to compete at the frontier of artificial intelligence.

For everyday builders and enterprises, you will not see instant changes this week. But starting late 2026, when the Vera Rubin racks begin churning out tokens, the bottleneck on long-horizon reasoning and real-time agent execution should dramatically loosen.

For the industry, it is a formidable moat. By pledging up to $100 billion tied directly to gigawatt milestones, NVIDIA essentially finances the very customer buying its chips. It creates a self-reinforcing flywheel that locks OpenAI into NVIDIA’s hardware ecosystem for the next decade.

We think this maneuver is brilliant financial and industrial engineering, but it carries glaring physical risks. Securing 10 gigawatts of grid interconnects, transformers, and clean power is a massive civil engineering headache, not just a software roadmap. Whether regulators, local grids, and existing partners like Microsoft let this flywheel spin unchecked remains the trillion-dollar question.