Two Million More Chips, One Familiar Pattern
On August 26, 2026, AWS and NVIDIA announced a major expansion of their partnership. AWS will deploy 2 million additional NVIDIA GPUs - Blackwell Ultra, Rubin, and Rubin Ultra - across its global infrastructure through 2027-2028. The move extends a commitment first outlined at NVIDIA's GTC 2026 conference, where AWS announced plans to add more than 1 million GPUs starting in 2026. Demand has since exceeded those original expectations, prompting the larger figure. NVIDIA Vera CPU-based infrastructure is also coming to AWS, purpose-built for agentic AI workloads that need high-performance CPU compute running alongside the accelerators themselves.
AWS CEO Matt Garman said customers want the freedom to choose the best tools for their AI workloads, whatever combination that turns out to be. NVIDIA CEO Jensen Huang said the two companies have scaled NVIDIA computing in the cloud together for 16 years, describing the expansion as a continuation of that history rather than a new direction. The partnership itself spans more than raw chip counts: it covers AI factories, CPUs, networking hardware, open models, data processing, and robotics, positioning the relationship as infrastructure-wide rather than limited to a single product line.
The One Region Named By Number
Inside the same announcement, AWS and NVIDIA named one specific, numbered carve-out: 100,000 GPUs dedicated to secure data centers built for the U.S. government. Those facilities are designed to handle workloads classified at Impact Level 6 (IL6) and above, the tier the U.S. Department of Defense uses for its most sensitive unclassified and classified information. The figure appears explicitly in the announcement, attached to a named customer segment and a named security standard, not folded into a broader regional total.
That level of specificity matters as a signal. When a company names a customer segment with an exact GPU count inside a capacity announcement, it is making a public commitment with a number attached to it, one a customer can hold the company to later. The remaining 1.9 million GPUs are described only as going to "AWS Global Infrastructure," with no other region or market receiving a comparable named, numbered line. Europe is not mentioned once anywhere in the announcement, not as a region, a market, or a planned allocation.
What Brandenburg Was Supposed To Prove
AWS operates a separate initiative, the AWS European Sovereign Cloud, built specifically to meet EU data-residency and operational-independence requirements that a standard global region cannot satisfy on its own. Its first region went live in Brandenburg, Germany in January 2026, giving EU customers a cloud environment intended to keep data, operations, and access decisions inside the EU. AWS has said further regions are planned in the Netherlands, Belgium, and Portugal, extending the same sovereignty model across more of the bloc over the coming years.
The sovereign-cloud pitch was always about legal and operational independence: where data physically sits, who can access it, and under what jurisdiction, not about guaranteed compute abundance. This announcement is a reminder that those are two separate promises. A sovereign region guarantees jurisdiction, not GPU capacity, and capacity itself still gets allocated globally by demand and margin, not by which regions made sovereignty commitments first. A company can be fully compliant with EU sovereignty requirements in Brandenburg and still receive none of a specific hardware expansion announced the same week.
Reading A Capacity Announcement Like A Buyer
The reading method is straightforward once it is named directly. A named, numbered carve-out, like the 100,000-GPU U.S. government commitment, signals where a company is prioritizing capacity right now, because naming a number in public creates an expectation the company then has to meet. Everything folded into a broader category like "global infrastructure" is deployed at the vendor's discretion, landing wherever demand and margin justify it first. That location may or may not be the region a specific customer is counting on when they sign a multi-year commitment.
For EU enterprise and public-sector buyers holding AWS GPU commitments or planning expansions, the practical instruction is not to read a "global infrastructure expansion" headline as a promise of EU-region capacity. Ask AWS account teams for region-specific capacity commitments and delivery timelines in writing, rather than inferring availability from a global total that names other regions by number but leaves Europe unnamed. A number attached to your own region and workload, written into a contract, is worth more to a buyer than a favorable-sounding global total, and it is exactly the kind of detail a national data-protection authority would expect an organization to have on file regardless.
The Supply Squeeze Behind The Announcement
The same week as this deal, NVIDIA's own finance leadership said memory costs are climbing faster than the company had expected, a trend it now expects to extend through fiscal 2028. The underlying components behind Blackwell Ultra and Rubin-class GPUs are getting more expensive to produce even as commitments to build more of them keep multiplying across the industry. NVIDIA disclosed the memory-cost pressure separately from the AWS announcement, but the two land in the same week and describe the same supply chain.
When a constrained input like memory meets an expanding list of buyers requesting public capacity commitments, allocation will not land evenly across every region that wants it. Buyers willing to accept whatever region AWS assigns are positioned to receive capacity faster than buyers who require a specific, contractually guaranteed EU region for legal or operational reasons. That is exactly the choice this announcement leaves EU buyers facing, a trade-off between speed and jurisdiction, without ever saying so directly in the text of the announcement itself.
Read next: Grid, Water, Labor Now Strain Under One Amazon Cluster | AWS's Anti-Nvidia Chip Adopts Nvidia's Memory
