Two million GPUs is not a purchase order — it’s a fence. When the biggest cloud on earth pre-books the next two generations of NVIDIA silicon through 2028, the practical question for everyone else is what capacity you can actually get a date on.
AWS and NVIDIA announced an expansion of their partnership that adds roughly two million additional GPUs to AWS’s global infrastructure across 2027 and 2028, spanning Blackwell Ultra, Rubin and Rubin Ultra. That comes on top of the million-plus GPUs AWS committed at GTC in March. The deal also brings NVIDIA’s Vera CPUs to AWS for agentic workloads, integrates NVLink Fusion and custom NVHBM memory with Trainium, and carves out 100,000 GPUs for US government and national-security workloads on secure AWS infrastructure.
The subtext is allocation. NVIDIA’s Q2 FY27 results the same day showed $89 billion of data center revenue in a single quarter, and supply of leading-edge HBM, CoWoS packaging and rack-scale liquid-cooled systems is still the constraint. Every million GPUs a hyperscaler locks in for 2027–28 is a million that neoclouds, enterprises and sovereign buyers will not see in the first wave.
AWS CEO Matt Garman framed it as customers wanting “freedom to choose the best tools” — which is true for AWS customers renting by the hour. For buyers who want committed, dedicated capacity at a fixed rate, the freedom is in what’s left over.
What it means if you’re buying: Rubin is now a 2028 conversation for anyone outside the top five clouds. The compute you can contract today, stand up this year and run at a fixed rate through the Rubin ramp is Blackwell-class — B300 clusters with a real RFS date. A 36-month B300 term signed this fall runs almost entirely inside the window before Rubin reaches general availability at scale, and Blackwell holds its value longer precisely because the next generation is fenced off.
What to do about it: Live on the marketplace now: ORION 1024 (1,024× B300 liquid-cooled, East Coast US, December RFS, $4.60/GPU-hr on a 36-month term), MAGNETAR (1,024× B300, Ohio, December 1 RFS) and BLAZAR (1,024× B300 build-to-order, Washington, 10-week delivery). If you need Hopper for inference today, LYNX (2K× H200, US) is available immediately. Capacity with a date on it beats a roadmap with a queue on it.
Original announcement · nvidianews.nvidia.com ↗ · Summary and analysis by Ax3
