Amazon and Nvidia jointly published the official announcement of what AWS calls "a major expansion" of its partnership with Nvidia, also announced by AWS CEO Matt Garman on X: 2 million additional Nvidia GPUs deploying across AWS's global infrastructure in 2027–2028, Nvidia's new Vera CPUs coming to AWS as an added compute option, and 100,000 GPUs dedicated to new AI factories for the U.S. government, cleared for workloads classified at Impact Level 6 (IL6) and above — one of the highest federal security classifications. Neither company disclosed financial terms.
The number that matters is what it's added to, not the headline figure
The 2 million-GPU figure is genuinely large, but it's worth placing next to what AWS already committed to seven months earlier: at Nvidia's GTC conference in March 2026, AWS announced plans to deploy more than 1 million Nvidia GPUs across its regions starting that year. The new release states plainly that "since then, demand has exceeded those expectations" — this is explicitly additive, not a replacement for the original plan. Combined, AWS's disclosed Nvidia GPU commitment now sits at roughly 3 million units across the two announcements, which is the basis for reporting elsewhere describing this as AWS having tripled its order in under a year. Jensen Huang's own quote in the release backs the framing directly: "demand is running ahead of every forecast."
Amazon's own framing on Trainium is complementary — and this time it comes with an actual technical bridge, not just words
Amazon's release states the expansion "complements Amazon's own custom silicon, giving customers the freedom to choose the best compute for their specific workloads—whether that's NVIDIA GPUs, AWS Trainium chips, or both working together." That's Amazon's own choice of framing, not a neutral description, and it's worth treating as exactly that: a company describing its two chip strategies as compatible rather than in tension. What makes this specific announcement more concrete than a talking point is a real technical detail buried further down: AWS's chip-design arm, Annapurna Labs, will build next-generation Trainium chips around Nvidia's new custom high-bandwidth memory (NVHBM) technology and NVLink Fusion interconnect — the release states this lets Annapurna Labs "tap NVIDIA's custom memory technology and scale-up architecture" so that Trainium and Nvidia GPUs can sit together "within a common rack-scale architecture." That's a specific claim about physical integration, not just parallel product lines sold to different customers — worth distinguishing from the framing point above, since one is Amazon's chosen language and the other is a checkable engineering claim about how the hardware actually connects.
What Vera CPUs are specifically for
The release is precise about what problem Nvidia's Vera CPUs solve on AWS rather than leaving it as a generic "more compute" claim: Vera is described as built for "the CPU work behind agentic AI and reinforcement learning, including code execution, tool use, sandboxing, analytics, data pipelines, and orchestration" — the surrounding infrastructure work that keeps GPUs fed and agent loops running, distinct from the GPU-side model computation itself. That's a specific, checkable claim about where in an agentic AI pipeline this chip is meant to sit, not a vague performance claim.
The concrete customer numbers already in production are worth more than the forward-looking ones
The release lists specific, checkable performance figures from integrations already shipping today, not just the new commitment: GPU-accelerated data processing on Amazon EMR with Nvidia's cuDF library delivers "up to 3.7 times faster processing speeds and 30% better price-performance" than CPU-based Spark configurations, and GPU-accelerated vector indexing on Amazon OpenSearch Service delivers "up to 9 times faster indexing at a quarter of the cost." Both are specific multipliers tied to named AWS products rather than the more common "significantly faster" framing — worth noting as the more verifiable part of this announcement, since they describe results from infrastructure already deployed rather than the 2027–2028 buildout still to come.
The federal number's relationship to AWS's larger $50B pledge still isn't spelled out
The 100,000-GPU federal commitment lands inside a much larger, previously announced plan: in November 2025, Amazon said it would invest up to 50 billion, or a more concrete quantification of Nvidia's specific share within a plan already announced — worth treating that as an open detail rather than assuming either reading.
What to expect next
- Watch whether either company discloses actual dollar figures. The official release, like every other account of this deal, states no purchase price or contract value.
- Watch how the 100,000-GPU federal number reconciles with the $50 billion pledge. Whether this is new money or a more specific breakdown of an existing commitment is the detail worth AWS or a federal agency clarifying directly.
- Watch for real-world benchmarks of the NVHBM/NVLink Fusion-connected Trainium chips, since this is the first specific claim that Trainium and Nvidia GPUs will share memory and interconnect technology rather than just coexisting as separate product lines — whether that actually ships as described is a checkable, near-term claim.
References: Amazon — AWS and NVIDIA expand partnership for next-gen AI infrastructure, read directly in full · Matt Garman's announcement on X · related coverage: Frontier Arcade: trends & predictions