Amazon to Add 2 Million More NVIDIA GPUs: AWS's Massive Investment in "AI Infrastructure Supremacy"

The competition in generative AI is shifting from a stage where model performance is the focus to one where the ability to secure massive computational resources is the key.

Amazon has announced a large-scale plan that symbolizes this shift.

Amazon Web Services (AWS), the company's cloud division, has revealed plans to significantly expand its strategic collaboration with NVIDIA by deploying an additional 2 million NVIDIA GPUs across its global infrastructure between 2027 and 2028.

This includes not only the Blackwell Ultra but also future generations such as Rubin and Rubin Ultra.

Moreover, AWS has already indicated plans to add over 1 million NVIDIA GPUs starting in 2026. This announcement adds another 2 million to that number.

This is not merely news about Amazon purchasing a large number of GPUs.

It signifies Amazon's intention to embed AI as a common infrastructure across multiple businesses, including its cloud services, e-commerce, logistics, robotics, and government systems, with a massive computational base at its core.

The main focus of AI competition is shifting from "models" to "infrastructure"

When the generative AI boom began, market interest was concentrated on the performance of large language models developed by companies like OpenAI, Anthropic, and Google.

However, as AI services transition from the experimental stage to actual business and consumer services, the situation changes.

As AI agents begin to automate corporate tasks, write programs, analyze data, and handle customer interactions, the computational demand consumed by AI will surge.

Furthermore, as AI expands into fields known as "Physical AI," such as robotics, autonomous driving, and scientific research, the scale of computational resources required could become incomparable to the current level.

The expanded partnership between AWS and NVIDIA is precisely an anticipation of such a future.

The two companies are strengthening their collaboration not only in GPUs but also in CPUs, networking, AI models, data processing, and robotics, encompassing a "full stack" approach.

For cloud companies, owning the latest GPUs is not the only important factor.

They need to build an environment where hundreds of thousands or millions of GPUs are connected via high-speed networks, supplied with massive amounts of power, cooled, operated without failures, and made instantly available to businesses and research institutions when needed.

In other words, in the cloud competition of the AI era, the ability to operate a "massive AI factory" is starting to become more competitive than the semiconductors themselves.

The figure of 2 million GPUs announced by Amazon symbolizes that the scale of this competition has risen to a new level.

Why buy NVIDIA when you have your own chip "Trainium"?

What is particularly intriguing about this news is that Amazon is a company actively developing its own AI chips.

Through its subsidiary Annapurna Labs, Amazon is developing the AI accelerator "Trainium."

Generally speaking, expanding the use of its own chips should reduce dependency on NVIDIA.

Nevertheless, AWS is adding another 2 million NVIDIA GPUs.

At first glance, this might seem contradictory.

However, considering AWS's strategy as a choice between "NVIDIA or Trainium" might miss the essence.

Amazon's aim is likely not to depend on a specific semiconductor but to become a massive AI platform where customers can choose from a variety of computational resources according to their needs.

Among AI developers, the software environment centered around NVIDIA's CUDA is widely used. Therefore, if AWS wants to attract AI companies worldwide, it remains important to have a large stock of NVIDIA's latest GPUs.

On the other hand, Trainium can be leveraged for its price-performance ratio in specific applications.

AWS and NVIDIA are also working on combining NVIDIA's high-speed connection technology NVLink Fusion with Amazon's custom silicon.

In other words, Amazon's vision might not be "AWS competing with NVIDIA," but rather a "massive AI computing market incorporating both NVIDIA and its own semiconductors."

The "AI Factory" with 100,000 GPUs, and the U.S. government as a key customer

Another aspect not to be overlooked in this partnership is the AI infrastructure for the U.S. government.

AWS and NVIDIA have revealed plans to build a highly secure AI factory for U.S. government agencies, including infrastructure on the scale of 100,000 GPUs.

As AI is used for national security, cybersecurity, information analysis, and research and development, the computational resources required by government agencies could increase significantly.

This is where AWS's strengths become crucial.

AWS has built a long-standing track record not only in enterprise cloud services but also in cloud infrastructure for government and public sectors.

Combining this with NVIDIA's AI infrastructure could make the government AI market a huge growth area for Amazon.

The AI infrastructure competition is not just about enterprise cloud competition with Microsoft Azure and Google Cloud.

It is expanding into a larger market of who will provide national-scale AI infrastructure.

For NVIDIA, it's also a counter to the "demand slowdown theory"

Of course, this contract is extremely important for NVIDIA as well.

In the AI semiconductor market, there have long been concerns that the capital investment by giant tech companies might eventually peak.

Google has its TPU, Amazon has Trainium, and Microsoft is also developing its own AI semiconductors.

Therefore, there is a view that hyperscalers might reduce their dependency on NVIDIA GPUs in the long term.

However, the fact that Amazon, which is developing its own semiconductors, has decided to add 2 million NVIDIA GPUs, including next-generation ones, is noteworthy.

At least for now, the spread of proprietary chips and the demand for NVIDIA are not necessarily mutually exclusive.

Because the AI market itself is rapidly expanding, even if proprietary semiconductors increase, the demand for GPUs might be growing at an even faster rate.

NVIDIA's CEO, Jensen Huang, also recognizes that the AI demand is expanding at a pace that exceeds previous expectations in collaboration with AWS.

If this trend continues, the biggest growth factor for NVIDIA will not be "eliminating competitors" but "the continuous expansion of the AI computing market as a whole."

The next battleground for AI, "agents," is transforming Amazon's e-commerce

What is even more important for Amazon is that AI infrastructure investment does not end with AWS's revenue.

In Amazon's core e-commerce business, AI agents might be starting to change consumer behavior.

In a U.S. online retail survey conducted by Evercore ISI, 57% of users utilizing Alexa's AI features reported purchasing products they were previously unaware of.

Furthermore, 36% reported that their purchase volume increased due to the use of AI features.

These are very important figures for Amazon.

In traditional e-commerce, the structure was centered around users searching for products they had already decided to buy, which Amazon would then sell.

However, with AI agents, they can understand users' vague requests, compare options, suggest new products, and guide them to purchase.

This means Amazon could transform from a "place to find desired products" to a "place to consult AI on what to buy."

If this becomes a reality, AI will become not just a cost-saving technology for Amazon but a sales device that generates consumption itself.

Evercore ISI analyst Mark Mahaney, based on such survey results, raised Amazon's target stock price from $315 to $355. On the day of the report, Amazon's stock rose, with AI agents being recognized as a new growth driver for the retail business.

Amazon is capturing both "cloud AI" and "shopping AI"

A broad view of this series of news reveals that Amazon's AI strategy encompasses two massive revenue opportunities.

The first is AWS.

It earns by providing GPUs, AI models, storage, and networks to companies, AI startups, governments, and research institutions worldwide as the "infrastructure to run AI."

The second is Amazon.com.

Through Alexa and shopping agents, it AI-izes the discovery, comparison, and purchase of consumer products, expanding e-commerce transactions.

In addition, there are businesses like advertising, logistics, and robotics in between.

For example, if AI increases product purchases, the value of advertising might rise.

If order volumes increase, the processing capacity of logistics centers becomes important, promoting the introduction of AI robots.

If AWS's computational resources are used to operate robots and AI, it leads back to cloud investment.

If Amazon's massive businesses of "cloud, e-commerce, advertising, and logistics" are interconnected through AI, an economic zone that other AI companies cannot easily replicate will be formed.

The decision by Amazon Robotics to utilize NVIDIA's Physical AI platform symbolizes this vision.

On social media, "extremely bullish" and "caution over huge investments" intersect

The reaction to this announcement on social media is not one-sided.

NVIDIA's official X account emphasized not only the 2 million GPUs but also the expansion of collaboration to include Vera CPUs, networking, Nemotron, and Amazon's robotics, positioning it as "not just a GPU sale, but an integration of the entire AI stack with AWS."

Posts by investors and technology stakeholders on X also reflect the sentiment that "the scale of 2 million GPUs is no longer just a product news but AI infrastructure investment itself has become macroeconomic in scale."

There are also opinions appreciating the strength of NVIDIA's competitiveness, noting the point that "Amazon is procuring a large amount of NVIDIA despite having Trainium."

In the Amazon-related community on Reddit, while there are reactions like "long-term strong material," there are also posts expressing dissatisfaction that "even with such good news, the stock price does not react."

There are also comments expressing caution about the massive investment, questioning whether it is a "cyclical flow of funds," and posts estimating how much AWS can recover in future revenue relative to the GPU purchase amount.

In another stock community, there are bullish opinions that independently estimate the long-term AWS revenue from the 2 million GPUs, suggesting it should be viewed as a pipeline generating future cloud revenue rather than a one-time equipment purchase.

However, such individual revenue estimates are not official forecasts and depend on multiple assumptions such as GPU prices and utilization rates, so they are not figures that can be directly used for corporate value calculations.

On the other hand, individual investor sentiment towards Amazon's stock is not entirely bullish.

On Stocktwits, even during the phase where Amazon's stock rose following Evercore's target price increase, individual investor sentiment was reported as "Bearish."

This reaction suggests that, apart from expectations for AI investment, the market is cautiously watching Amazon's massive capital investments, profit margins, and investment recovery.

The simple market rule of "the larger the AI investment, the higher the stock price" is no longer applicable, and "how much cash flow the investment will generate in how many years" has become more important than ever.

The biggest risk is not "whether you can buy GPUs" but "whether you can fully utilize them"

The figure of 2 million GPUs is enormous.

However, the real challenge for Amazon is not securing GPUs.

It is whether they can be continuously used at a high utilization rate and generate revenue exceeding the investment amount.

AI data centers require not only GPUs but also massive additional investments in land, buildings, power grids, substation facilities, cooling systems, and high-speed networks.

If AI data centers increase globally, there might be a shortage of power and equipment itself.

And if the efficiency of AI models advances rapidly, the possibility of "processing the same work with fewer GPUs" cannot be ignored.

On the other hand, if AI agents become widely adopted in corporate operations and consumer services, the usage might grow at a pace that surpasses efficiency improvements.

Ultimately, whether Amazon's investment in 2 million GPUs was the right decision will not be determined by the performance of the GPUs.

It will depend on how much AI computation the world comes to require.

The relationship between Amazon and NVIDIA is beginning to transcend "seller and buyer"

Judging from this announcement, the relationship between Amazon and NVIDIA is moving beyond the traditional framework of "GPU manufacturer and cloud provider."

GPU, CPU, networking, memory, AI models, search, data processing, government AI, robotics.

The scope of collaboration is expanding to almost all layers of the AI industry.

Amazon is developing its own chips while adopting a large number of NVIDIA products, and NVIDIA is selling GPUs while integrating its technology into AWS's cloud, data services, and robots.

Both companies are beginning to become partners in expanding the AI market itself while maintaining a competitive relationship.

And for Amazon, 2 million GPUs are not the ultimate goal.

The question is whether they can operate corporate AI agents, government AI, warehouse robots, and even AI-enable consumer shopping on that computational power.

Will AWS end as a "cloud that runs AI," or will it evolve into a foundation supporting the global AI economic sphere?

This massive partnership with NVIDIA seems to indicate that Amazon is beginning to seriously invest in that challenge