The Zhitong Finance App learned that at the same time that Nvidia (NVDA.US) handed over a record quarterly report of 96.2 billion US dollars, Amazon (AMZN.US), one of its largest customers, wrote the strongest footnote to this report card with a “largest GPU order in history.” On August 26, local time, Amazon Cloud Services (AWS) and Nvidia announced the expansion of strategic cooperation. The two sides plan to deploy an additional 2 million Nvidia high-end GPUs in the AWS global infrastructure from 2027 to 2028. This order is double the purchase plan for 1 million GPUs announced by Amazon in March of this year, incorporating Nvidia's Blackwell Ultra, Rubin, and Rubin Ultra chips into the AWS AI landscape.
Order panorama: 2 million GPUs, from Blackwell Ultra to Rubin Ultra
The volume of this deal is unprecedented in the field of AI infrastructure. Amazon will complete the phased deployment in 2027 and 2028, and the chip lineup covers Nvidia's three strongest current products:
Blackwell Ultra: an enhanced version of Nvidia's current flagship AI chip;
Rubin: the core GPU of Nvidia's next-generation AI platform;
Rubin Ultra: The flagship product in the Rubin series.
The list price for each of these GPUs is expected to be at least tens of thousands of dollars. At a minimum estimate of tens of thousands of dollars each, the contract value for 2 million GPUs is conservatively at the level of tens of billions of dollars. Coupled with the purchase plan of 1 million GPUs announced in March of this year, AWS has already committed to purchasing 3 million Nvidia GPUs in 2026 alone.
From “buying a chip” to “full-stack collaboration”: Vera CPUs are included in AWS for the first time
This is not only a GPU purchase, but also a full-stack technology binding. For the first time, AWS will deploy the Nvidia Vera CPU, a high-performance processor designed for agent-based AI workloads, in its infrastructure. Nvidia said that some of AWS's Vera CPUs will be integrated with Rubin, and some will be used as standalone products.
The two sides will also deepen cooperation on NVLink Fusion high-speed chip interconnection technology, so that it can operate in conjunction with Nvidia's new custom high-bandwidth memory (NVHBM), so that Amazon's self-developed Trainium chips can access faster and more energy-efficient memory. Meanwhile, AWS and Nvidia plan to deliver an AI factory with 100,000 GPUs to the US government on AWS infrastructure that meets the Impact Level 6 security rating.
AWS: Nvidia's Big Buyer
The deal confirms a key trend: even though Amazon invests heavily in its self-developed AI chip Trainium, its demand for Nvidia GPUs is still growing at an accelerated pace.
Nvidia CEO Hwang In-hoon said in a statement: “Demand is ahead of every forecast. For 16 years, we've worked together to scale Nvidia's computing in the cloud. Now we're expanding our partnership to the full range of technologies—including GPUs, CPUs, networking, open models, and software.”
Nvidia CFO Colette Kress revealed during an earnings call that the capital expenditure of the “top five hyperscale customers” is expected to increase from $800 billion in 2026 to $1.3 trillion in 2027. AWS's position as the largest buyer of Nvidia's accelerators is being further consolidated in this deal.
This expansion builds on AWS's plans to deploy more than 1 million GPUs announced at GTC 2026 — and this new commitment of 2 million directly triples that number.
Competitive game: self-developed chips and Nvidia's “dual track parallel”
The most dramatic backdrop to this order is that Amazon is simultaneously advancing an “replacement strategy” for Nvidia. In recent years, Amazon has continued to increase its self-developed AI chip business. Its Trainium AI accelerator has attracted customers such as Anthropic to pay hundreds of billions of dollars to use it. In June of this year, Amazon confirmed that it is in negotiations with potential customers and plans to sell Trainium chips directly to enterprise data centers, moving from being exclusive to AWS to overseas sales.
However, the order for 2 million GPUs shows that self-developed chips are not a “choose one of two” relationship with Nvidia GPUs, but rather a parallel expansion strategy. AWS CEO Matt Garman said the expansion “provides cutting-edge laboratories, enterprises, and governments with more ways to build and deploy AI on AWS.” The agency predicts that Nvidia will still hold at least 70% of the AI training workload — and the scale of demand locked in the AWS deal is hard to match until now for competitors' custom chips.
Conclusions
With an additional order for 2 million GPUs, Amazon AWS wrote a clear footnote to the global AI infrastructure race: even in the face of self-developed chips, regulatory scrutiny, and multiple geopolitical pressures, the “arms race” for AI computing power is far from slowing down. From Blackwell Ultra to Rubin Ultra, from GPUs to Vera CPUs, from commercial clouds to federal government AI factories — this collaboration between AWS and Nvidia has pushed the AI infrastructure competition from the “10 billion level” to the “trillion dollar era.”