AWS and Nvidia's 2-Million-GPU Bet, and the Bottleneck Neither Can Buy Its Way Past

AWS and Nvidia announced plans to deliver an additional 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs across AWS's global infrastructure through 2027-2028. This is a raise, not a first commitment: AWS had already announced more than 1 million Nvidia GPUs at GTC earlier this year, and this expansion exists specifically because demand ran ahead of that original number.

The number worth sitting with is what Nvidia itself is now quoting as the economics of this build-out: revenue per gigawatt of deployed infrastructure rising from $18B on the Hopper generation to $40B on Vera Rubin -- more than doubling the revenue density of the same physical footprint in roughly one hardware generation. That's the number that makes a 2-million-GPU order rational rather than speculative capacity-hoarding.

The constraint that doesn't move at the same pace is packaging, not silicon. TSMC's CoWoS advanced-packaging process -- the step that assembles compute dies with high-bandwidth memory into a finished accelerator package -- is fully allocated through at least mid-2027. TSMC's own CoWoS lines are projected to reach only 120,000-140,000 wafers/month by year-end, with outsourced assembly partners ASE and Amkor adding perhaps another 50,000-60,000 on top. Ordering more GPUs doesn't create more packaging capacity; it just moves further back in a queue that's already full.

Nvidia and Google separately launched an AI Energy Management program with Emerald AI, which points at the other constraint quietly running alongside the chip one: at gigawatt-scale deployment, the grid itself becomes as much a design constraint on how fast an "AI factory" can actually come online as the chip supply is. Two million more GPUs on order is a demand signal, not a guaranteed delivery schedule -- packaging and power are the two numbers to watch for whether that demand actually turns into deployed compute on the announced timeline.