> ## Content Index > Fetch the complete content index at: https://www.edgewisely.com/llms.txt > Use this file to discover other available public pages before exploring further. # Amazon Triples Its Bet on Nvidia - URL: https://www.edgewisely.com/amazon-triples-its-bet-on-nvidia/ - Published: 2026-09-01T11:45:36.000Z - Updated: 2026-09-01T11:45:36.000Z - Description: How a five-month-old GPU order just tripled, and what that says about who really controls AI's supply chain - Author: John Karpentar - Tags: Cloud, Chips *How a five-month-old GPU order just tripled, and what that says about who really controls AI's supply chain* **AWS and Nvidia will add 2 million more GPUs to Amazon's cloud in 2027 and 2028, pushing a commitment made in March past 3 million chips before a single one of the new units has shipped.** ![AWS and Nvidia logos representing their expanded AI infrastructure partnership](https://storage.ghost.io/c/54/5a/545a66b3-60ef-480c-80ae-765bac52f6ec/content/images/2026/09/aws-nvidia-2m-gpu-cover.jpg) Image: [NVIDIA Newsroom](https://nvidianews.nvidia.com/news/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai?ref=edgewisely.com) On the morning of August 27, 2026, an Nvidia vice president named Ian Buck wheeled a server cart past the Spheres at Amazon's Seattle headquarters and handed it to two AWS executives, Willem Visser and Supreeth Sheshadri. Inside the cart sat the first Vera CPU server and Vera Rubin GPU that Amazon Web Services had ever received — hardware that, a day earlier, the two companies had said would anchor a dramatically larger partnership than either had described in March ([NVIDIA blog](https://blogs.nvidia.com/blog/vera-cpu-delivery/?ref=edgewisely.com)). The delivery was theater, but the number behind it was not. On August 26, AWS and Nvidia announced they would deploy 2 million additional Nvidia Blackwell Ultra, Rubin, and Rubin Ultra GPUs across AWS's global infrastructure in 2027 and 2028 — on top of the more than 1 million GPUs AWS had already committed to installing starting in 2026, a plan unveiled just five months earlier at Nvidia's GTC conference ([AWS press release](https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai?ref=edgewisely.com)). Add it up and AWS's committed Nvidia GPU order has gone from roughly 1 million chips to more than 3 million in less than half a year. ## What was announced The companies framed the expansion as touching nearly every layer of the AI stack, not just chip count. Alongside the 2 million additional GPUs, AWS and Nvidia said they would work to bring Nvidia's new Vera CPU-based infrastructure to AWS, extend Nvidia's NVLink Fusion interconnect with a custom high-bandwidth memory technology called NVHBM, and build AI factories for the U.S. government that include 100,000 GPUs for federal and national-security workloads cleared to run at Impact Level 6 and above ([AWS press release](https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai?ref=edgewisely.com)). "NVIDIA and AWS have built one of the great growth engines of the AI era, and demand is running ahead of every forecast," Nvidia founder and CEO Jensen Huang said in the joint announcement. AWS CEO Matt Garman framed the deal as a customer-choice argument: "Customers want the freedom to choose the best tools for their AI workloads, and they want confidence that everything works seamlessly together," he said, describing the goal as making AWS "the best place to run NVIDIA AI technologies." The scale escalation matters on its own, but the composition matters more. At GTC in March, AWS's pledge covered Blackwell-generation chips arriving through 2026\. The new order shifts forward to Blackwell Ultra and to Nvidia's next architecture, Rubin, plus a beefed-up Rubin Ultra variant — chips that in most cases have not yet shipped in volume to any customer. AWS is effectively pre-committing to two full silicon generations at once, a bet that Nvidia's roadmap holds and that demand for whatever ships in 2027 and 2028 will look like demand today. Two technical threads run underneath the headline number. The first is Vera, Nvidia's first custom CPU, built for the CPU-heavy work — orchestration, tool-calling, long-context retrieval — that agentic AI systems generate even when a GPU isn't doing the reasoning. Nvidia has been hand-delivering Vera systems all year, first to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure in May, and now to AWS. "Agentic AI is creating a new CPU moment in the AI factory," Buck said of the AWS delivery, "as models move from answering to acting" ([NVIDIA blog](https://blogs.nvidia.com/blog/vera-cpu-delivery/?ref=edgewisely.com)). The second thread is memory. Amazon's chip design unit, Annapurna Labs, is becoming the first partner to work with Nvidia on NVHBM, a new high-bandwidth memory architecture that moves the memory controller off the compute die and into the memory stack itself. Nvidia says the design yields up to 30% more memory bandwidth and 15% lower power draw than standard HBM4E, while freeing up chip area for more compute. Annapurna will apply the technology starting with its next-generation Trainium4 chip, letting Amazon's own silicon and Nvidia's GPUs share a common rack-scale architecture through NVLink Fusion. "We look forward to this technology collaboration to benefit future AWS infrastructure designs," said Nafea Bshara, Annapurna Labs' vice president, in Nvidia's announcement of the memory partnership ([NVIDIA blog](https://blogs.nvidia.com/blog/nvlink-fusion-nvhbm-custom-high-bandwidth-memory/?ref=edgewisely.com)). That detail is worth sitting with. Amazon spent years building Trainium specifically so it would not need to depend entirely on Nvidia. The company is now wiring Nvidia's proprietary memory and interconnect technology directly into that chip's next generation — buying more Nvidia GPUs and making its homegrown alternative more Nvidia-compatible at the same time. ## The market's split verdict Wall Street did not read the announcement as unambiguously good news for either company. Nvidia, which reported blockbuster fiscal second-quarter earnings the same week, jumped almost 9% the following trading day — a move driven primarily by its own results rather than the AWS order alone. Amazon's stock moved the other way, slipping roughly 2.5% to trade near $260, even as the company disclosed one of the largest hardware commitments in its history ([The Motley Fool](https://www.fool.com/investing/2026/08/28/amazon-just-committed-to-2-million-more-nvidia-gpus-and-the-stock-traded-lower/?ref=edgewisely.com)). The divergence makes a certain sense once you follow the cash. Every one of those 2 million GPUs is future revenue for Nvidia and future spending for Amazon. Amazon's 2026 capital expenditure estimate, which stood at about $200 billion in February and April, was raised to roughly $220 billion on the company's July earnings call, with CEO Andy Jassy citing the rising cost of memory. Amazon's trailing-12-month net capital spending had already climbed to $169 billion by the second quarter, up $66.1 billion year over year, and free cash flow had swung to a $7.6 billion outflow from an $18.2 billion inflow a year earlier ([The Motley Fool](https://www.fool.com/investing/2026/08/28/amazon-just-committed-to-2-million-more-nvidia-gpus-and-the-stock-traded-lower/?ref=edgewisely.com)). Against that spending, though, sits a demand picture that is arguably outrunning it. AWS revenue grew 37% year over year in the second quarter to $42.2 billion, the segment's fastest growth since 2021, and its backlog of signed-but-undelivered work reached $496 billion, expanding at a triple-digit rate. AWS's annualized AI revenue run rate has climbed past $25 billion, also growing at a triple-digit clip. On the July call, Jassy put it plainly: "Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027, too. In fact, the demand we already have for 2028 is striking," he said, as quoted by The Motley Fool ([The Motley Fool](https://www.fool.com/investing/2026/08/28/amazon-just-committed-to-2-million-more-nvidia-gpus-and-the-stock-traded-lower/?ref=edgewisely.com)). That is the case for reading Wednesday's announcement as confirmation rather than concern: a company does not triple a multiyear hardware order against demand it merely hopes will show up. It commits that kind of capital against demand it is already turning away. ## For AWS, Nvidia, and everyone renting compute For AWS, the calculus is about staying the default answer to "where do I run this workload." AWS already offers the widest range of GPU-based instances of any cloud provider, and the new agreement adds RTX PRO 4500 Blackwell Server Edition GPUs to Amazon EC2 G7 instances, which the companies say deliver 4.6 times the AI inference performance of the prior G6 generation. It also folds in GPU-accelerated data processing on Amazon EMR, which Nvidia and AWS say runs up to 3.7 times faster than CPU-based configurations, and GPU-accelerated vector indexing on Amazon OpenSearch that the companies describe as up to 9 times faster at a quarter of the cost ([AWS press release](https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai?ref=edgewisely.com)). None of that is about winning a single customer. It is about making switching costs high enough that customers never seriously look elsewhere. For Nvidia, the deal is a hedge against the one risk that actually threatens its dominance: hyperscalers building good enough silicon of their own. Trainium, Google's TPUs, and Microsoft's Maia chips all exist because cloud providers want leverage over their largest supplier. By putting NVLink Fusion and NVHBM inside Trainium4, Nvidia turns Amazon's alternative-chip program into a customer for Nvidia's interconnect and memory technology rather than a threat to Nvidia's GPU business. Nvidia does not need every chip in AWS's data centers to say Nvidia on it. It needs every chip, including AWS's own, to run on Nvidia's plumbing. For the enterprises, government agencies, and AI labs actually renting this capacity, the announcement is mostly about certainty of supply, not price. Getting Blackwell Ultra, Rubin, or Rubin Ultra capacity from a smaller cloud provider carries real allocation risk when demand exceeds supply industry-wide; a reservation inside a 3-million-GPU AWS order is a hedge against being last in line. The federal government slice — 100,000 GPUs cleared for Impact Level 6 workloads — extends that same logic to national-security customers who cannot use commodity cloud capacity at all and need a dedicated, security-cleared supply chain instead. ## Takeaways The pattern to watch is not the chip count itself but how fast it moved. A commitment made in March, built around one Nvidia architecture, was superseded in August by a commitment nearly three times larger, spanning two newer architectures that had not yet shipped when the first pledge was made. Infrastructure plans that look aggressive on the day they are announced can look conservative within a single fiscal year if the underlying demand curve keeps bending upward. The stock market's split reaction is also a signal worth taking seriously rather than dismissing as noise. When the buyer of $2 million-plus GPUs sees its own shares dip on the news, that is the market pricing near-term margin pressure against long-term demand certainty — and betting, at least for a few trading sessions, that the margin pressure is real even if the demand is too. Finally, watch what Amazon does with Trainium, not what it says about it. A company does not typically weave a rival's proprietary interconnect and memory technology into its flagship custom chip unless independence has quietly become more expensive than integration. ## Zoom out Every cloud provider now says the same thing publicly: we offer choice, we are not locked into any single vendor. AWS's own custom silicon program is the proof point it points to. But this week's announcement shows that choice and dependence can coexist inside the same balance sheet. Amazon can keep building Trainium and keep buying millions of Nvidia GPUs, and both of those facts can be true at once, because the real constraint in this market has stopped being which chip a company prefers. It is which chip is available, on what timeline, at what scale — and right now, for the buyer with $220 billion to spend, the answer keeps coming back to Nvidia. ## Frequently Asked Questions **How many Nvidia GPUs has AWS now committed to buying?** AWS's committed order has grown from more than 1 million GPUs, announced at Nvidia's GTC conference in March 2026, to more than 3 million after the August 26, 2026 announcement of 2 million additional Blackwell Ultra, Rubin, and Rubin Ultra GPUs for delivery in 2027-2028. **What is Nvidia's Vera CPU, and why does it matter for this deal?** Vera is Nvidia's first custom CPU, designed for the orchestration and tool-calling work that agentic AI systems generate outside of raw GPU computation. AWS received its first Vera CPU server and Vera Rubin GPU on August 27, 2026, following earlier deliveries to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure. **Why did Amazon's stock fall on news of a bigger Nvidia order?** Amazon had already raised its 2026 capital spending estimate to about $220 billion, and free cash flow had turned negative. Investors weighed the added spending commitment against AWS's accelerating revenue and backlog, and shares slipped roughly 2.5% even as AWS executives described demand as outrunning supply. **What does the 100,000-GPU federal component involve?** AWS and Nvidia plan to build AI factories for the U.S. government that include 100,000 GPUs running on AWS's secure infrastructure, supporting federal and national-security workloads classified at Impact Level 6 (IL6) and above. --- **Related on Edgewisely:** - [Nvidia's Earnings Reset the AI Growth Story](https://edgewisely.com/nvidia-q2-fy2027-earnings-data-center-record/?ref=edgewisely.com) - [Top 7 GPU Cloud Providers for AI Training and Inference in 2026](https://edgewisely.com/top-7-gpu-cloud-providers-for-ai-training-and-inference-2026/?ref=edgewisely.com) - [Britain's Bet Against Nvidia](https://edgewisely.com/britains-bet-against-nvidia/?ref=edgewisely.com)