Amazon (AMZN.US) places another order for 2 million Nvidia GPUs, underscoring strong demand for AI computing power

Zhitong Finance APP has learned that at the very moment Nvidia (NVDA.US) delivered its record $96.2 billion quarterly report, one of its largest customers, Amazon (AMZN.US), put a powerful exclamation mark on that performance with the "largest GPU order in history." On August 26 (local time), Amazon Web Services (AWS) and Nvidia announced the expansion of their strategic partnership, planning to deploy an additional 2 million high-end Nvidia GPUs throughout AWS's global infrastructure between 2027 and 2028. This order is double Amazon's March announcement to purchase 1 million GPUs, integrating Nvidia Blackwell Ultra, Rubin, and Rubin Ultra chips into AWS's AI portfolio.

Order Overview: 2 Million GPUs, from Blackwell Ultra to Rubin Ultra

This deal is unprecedented in scale for the AI infrastructure sector. Amazon will complete the deployment in stages during 2027 and 2028, with the chipset lineup covering Nvidia's three most advanced generations:

Blackwell Ultra: The 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 within the Rubin series.

The unit price of these GPUs is expected to be at least several tens of thousands of dollars each. At a minimum of several tens of thousands per GPU, the contract value for 2 million units is conservatively in the tens of billions of dollars range. Including the 1 million GPU procurement plan announced in March, AWS's committed Nvidia GPU purchase amount for 2026 alone has reached 3 million units.

From "Buying Chips" to "Full-stack Collaboration": Vera CPU Debuts in AWS

This is more than just GPU procurement—it's a full-stack technology alignment. AWS will, for the first time, deploy Nvidia Vera CPU in its infrastructure—a high-performance processor designed specifically for agent-based AI workloads. Nvidia stated that some AWS Vera CPUs will be integrated with Rubin, while others will be used as standalone products.

The two companies will also deepen cooperation on NVLink Fusion, Nvidia’s high-speed chip interconnect technology, integrating it with Nvidia's new custom high-bandwidth memory (NVHBM). This cooperation will enable Amazon's self-developed Trainium chips to access faster and more energy-efficient memory. Meanwhile, AWS and Nvidia also plan to deliver an "AI factory" of 100,000 GPUs to the U.S. government on AWS infrastructure rated for Impact Level 6 security.

AWS: A Major Nvidia Buyer

This deal confirms a critical trend: even though Amazon is heavily investing in its Trainium AI chip, its demand for Nvidia GPUs is still rapidly accelerating.

Nvidia CEO Jensen Huang stated, "Demand is outpacing every forecast. For 16 years, we have worked together to expand Nvidia computing at cloud scale. Now, we are extending our partnership into full-stack technology — spanning GPU, CPU, networking, open models, and software."

Nvidia CFO Colette Kress revealed during the earnings call that capital expenditure by the "top five hyperscale customers" is expected to rise from $800 billion in 2026 to $1.3 trillion in 2027. AWS’s position as Nvidia's largest accelerator customer is further solidified by this transaction.

This expansion is built upon AWS's announcement at GTC 2026 of deploying over 1 million GPUs—making the new 2-million-GPU commitment a direct tripling of that figure.

Competition and Cooperation: Dual-track Strategy with Self-developed Chips and Nvidia

The most dramatic aspect of this order is that Amazon is actively pursuing a "replacement strategy" for Nvidia even as it expands collaboration. In recent years, Amazon has been doubling down on its self-developed AI chip business. Trainium AI accelerators have attracted clients like Anthropic to pay tens of billions of dollars for use. In June this year, Amazon further confirmed it is negotiating with potential customers to sell Trainium chips directly to enterprise data centers, moving beyond AWS exclusivity to external sales.

However, an order for 2 million GPUs shows that self-developed chips and Nvidia GPUs are not an "either-or" choice—but a parallel expansion strategy. AWS CEO Matt Garman stated that this expansion provides "more options for frontier laboratories, enterprises, and government to build and deploy AI on AWS." Institutional forecasts indicate Nvidia will retain at least 70% of AI training workloads—and the scale of demand that AWS secures in this deal is difficult for competitors' custom chips to match.

Conclusion

With its supplementary order for 2 million GPUs, Amazon AWS has made a decisive statement in the global AI infrastructure race: even with pressures from self-developed chips, regulatory scrutiny, and geopolitical factors, 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 cloud to federal government AI factory—this collaboration between AWS and Nvidia propels the competition for AI infrastructure from the "tens of billions" era to the "trillion-dollar age."