Amazon Web Services and Nvidia expanded their partnership on August 26, and the headline number is two million additional GPUs landing across AWS data centers in 2027 and 2028. Five months ago the figure was one million. The revision says more about the demand curve than any earnings slide could.
But the GPU count is the least interesting part of this deal. What Nvidia actually sold AWS was the rest of the stack — CPUs, interconnect, memory technology, open models, robotics tooling. That is a different kind of relationship than “customer buys chips.”
What was actually signed
The two companies laid out a full-stack expansion, not a hardware order:
- Two million additional GPUs across AWS global infrastructure in 2027–2028, spanning Blackwell Ultra, Rubin, and Rubin Ultra silicon
- Vera CPU infrastructure coming to AWS — Nvidia’s general-purpose processor push, and a direct shot at the traditional server-CPU market
- NVLink Fusion extended with custom high-bandwidth memory (NVHBM), developed with Amazon’s Annapurna Labs so Trainium and Nvidia GPUs can sit inside a common rack-scale architecture
- 100,000 GPUs for U.S. government AI factories running on secure AWS infrastructure for federal and national-security workloads
- Nemotron open models on Bedrock and SageMaker, plus Nvidia’s physical-AI stack — Omniverse, Cosmos, Isaac, Jetson — powering Amazon’s warehouse robotics fleet
Neither company disclosed financial terms. At prevailing per-unit economics, the GPU portion alone runs into the tens of billions.
The number that frames it
Nvidia dropped this during its Q2 FY2027 call, and the earnings gave the announcement its context:
| Metric | Q2 FY2027 | Change |
| Total revenue | $96.2B | +106% YoY |
| Data center revenue | $89.0B | +117% YoY |
| Gross margin | 75.0% | — |
| Q3 guidance | $108.0B (±2%) | — |
Supply and manufacturing commitments now sit at $279 billion, up from $119 billion a quarter earlier. Jensen Huang’s framing on the call was blunt: AI has hit its inflection point, tokens are profitable, and compute is now revenue.
The company also guided fiscal 2028 to roughly 70% growth — while calling that number supply-constrained rather than demand-limited. Read that twice. The ceiling isn’t customers. It’s memory and manufacturing capacity.
The part everyone glosses over
Amazon is simultaneously Nvidia’s largest-scale partner and one of its more credible challengers.
Trainium is a direct alternative to Nvidia’s deep-learning silicon. Graviton is an Arm-based CPU aimed squarely at Intel and AMD. Amazon’s custom chip business crossed a $25 billion annualized run rate, backed by roughly $225 billion in commitments from AI labs including Anthropic and OpenAI.
And yet the same week, Amazon signed for two million more Nvidia GPUs — and agreed to let Nvidia’s memory technology and interconnect sit inside Trainium racks.
This is not capitulation. It’s a read on the market. Custom silicon wins on cost-per-token for known, stable workloads. Nvidia still wins on time-to-capability for everything else — frontier training, novel architectures, and the software ecosystem customers already build against. AWS is not choosing. It is renting optionality on both sides and letting customers pick.
If you sell into infrastructure, that dual-track posture is now the default enterprise stance, not an exception.
Why this matters if you’re building or marketing in AI
Compute scarcity is a real constraint, not a talking point. When the supplier says demand exceeds what it can ship through FY2028, expect capacity pricing to stay firm and reserved-instance economics to matter more than they did in 2024.
“Agentic and physical AI” is now the framing both parties chose. Not chatbots. Robotics, simulation, world models, and long-running agents are where the infrastructure spend is being justified. Positioning still anchored to text generation is a generation behind the buyer’s roadmap.
Government AI infrastructure is a live category. A dedicated 100,000-GPU secure-cloud build for federal workloads signals a procurement channel that barely existed at scale two years ago.
Coopetition is the operating model. Your partners are your competitors, your competitors are your suppliers, and enterprise buyers have stopped treating that as a red flag.
What to watch next
- Rubin’s ramp. Production shipments have started. Q3 is the first real read on whether next-gen demand holds.
- Memory costs. Nvidia guided gross margins down toward 71–72% by Q4 on component pricing. That pressure travels downstream to cloud list prices eventually.
- Trainium’s external sales. AWS has said it is in talks to sell Trainium to third parties. If that closes, the partnership dynamic shifts again.
- Whether compute converts to profit. Hundreds of billions are being deployed on the thesis that more compute means more profitable tokens. That thesis is now large enough to be tested in public.
Doubling an order inside five months isn’t a purchasing decision. It’s a forecast correction. AWS looked at its demand signal, found it running ahead of a one-million-GPU plan it had made in March, and doubled it — while hedging with its own silicon at the same time.
The companies building the infrastructure are behaving like people who believe the demand is real. Whether the returns arrive on the same timeline is the open question of the next eighteen months.
