Andy Jassy is buying more Nvidia chips for AWS’s cloud business than most companies will ever touch, and he still says it won’t be enough. AWS announced plans to deploy another 2 million Nvidia GPUs across its cloud infrastructure in 2027 and 2028, on top of the million-plus GPUs it committed to back in March. Combined, the two agreements now cover more than 3 million planned chips headed into AWS’s global cloud footprint.
The new deployment brings in Nvidia’s Blackwell Ultra, Rubin, and Rubin Ultra GPUs, along with matching CPUs, networking gear, and interconnect technology built for cloud-scale AI workloads. A separate slice of the deal sets aside 100,000 GPUs for secure AWS infrastructure supporting US federal and national-security workloads, systems the companies say will underpin government AI operations directly.
What makes the timing notable is how far ahead customers have already locked in capacity with AWS. Jassy said customers have already reserved most of the compute AWS plans to add for 2027, with bookings stretching into 2028. Even so, he doesn’t think it will be enough. “We will still not have enough capacity to meet all of the demand we have in 2026,” he said on the company’s earnings call.
The spending numbers back that up. AWS closed the second quarter with a $496 billion contracted backlog, up sharply from $364 billion three months earlier, while cloud revenue hit $42.2 billion, up 37% year-over-year. Amazon’s cash capital expenditure for the quarter reached $53.1 billion, nearly double what it spent a year earlier, and the company raised its full-year 2026 capex forecast to roughly $220 billion in July, citing rising memory costs as a major driver.
AWS isn’t relying on Nvidia alone, either. The company continues expanding Trainium, its own AI chip line built through Annapurna Labs specifically for cloud workloads. Trainium3 started shipping early this year, and Amazon says customers have already nearly fully subscribed to it, with more than $225 billion in revenue commitments behind the chip. Trainium4, arriving in 2027, will support Nvidia’s NVLink Fusion interconnect, letting AWS’s own silicon and Nvidia’s hardware run inside the same rack-scale systems.
Nvidia, for its part, says it remains supply-constrained even as demand keeps climbing across the cloud industry, and AWS is betting the capacity it locks in now will justify the wait.
