Inside the $100B OpenAI–NVIDIA Pact: Chips, Compute, and the New Economics of Model Building
NVIDIA’s pledge to invest up to $100 billion in OpenAI, tied to a 10-gigawatt buildout of AI supercomputing, is not just another mega-deal—it is the capital markets’ clearest signal yet that compute is the strategic high ground of artificial intelligence. The architecture is unusually explicit: money arrives in $10 billion tranches, capacity arrives in gigawatts, and the first phase targets the second half of 2026 on NVIDIA’s next-generation Vera Rubin systems. OpenAI positions NVIDIA as a preferred, not exclusive, supplier across chips and networking, preserving leverage with other partners while concentrating on the stack that currently defines frontier AI performance. The stakes extend well beyond a bilateral relationship. A 10 GW program equates to roughly 4–5 million GPUs—about NVIDIA’s total expected shipments this year—and it forces hard choices about energy, siting, and financing. The pact reverberated immediately in markets, with NVIDIA shares rallying on the announcement and broader indices hitting fresh highs. Behind the pop is a recalibration of AI’s cost structure: concentrated access to compute becomes a moat, training throughput becomes the new velocity metric, and the economics of inference compress toward power, density, and interconnect performance. This article dissects the capital stack, engineering constraints, chip and cloud implications, and policy risks that will determine whether this bet on scale earns the returns its size implies.