Amazon’s decision to add 2 million more Nvidia GPUs to AWS is not merely a bet on rising demand. It is a test of whether cloud providers can reduce their dependence on Nvidia while customers continue to demand the company’s newest and most capable systems.
Amazon and Nvidia are expanding an artificial intelligence infrastructure partnership that was already among the largest in the industry. AWS has agreed to add another 2 million Nvidia GPUs to its data centers, building on an earlier commitment to deploy more than 1 million chips.
The new systems are expected to include Nvidia’s Blackwell Ultra, Rubin and Rubin Ultra GPUs. Deliveries and deployment are planned across AWS facilities in 2027 and 2028, according to TechCrunch, which first reported the expanded commitment. Neither company disclosed the financial terms. At current prices for high end AI accelerators and the networking, storage and cooling systems required to operate them, an order of this scale could be worth tens of billions of dollars.
The more important question is not how large the order is. It is why Amazon is increasing its exposure to Nvidia so sharply while AWS is simultaneously investing billions of dollars in custom chips intended to reduce that dependence.
The answer points to a difficult reality for the cloud industry. Hyperscalers can design their own accelerators, optimize software around them and offer customers lower cost options. But when customers want access to the latest generation of systems for training and serving advanced models, Nvidia remains the default supplier. The commercial value of independence is therefore limited if customers continue to measure cloud platforms by the availability of Nvidia hardware.
Demand is moving faster than earlier forecasts
Amazon’s original commitment to deploy more than 1 million Nvidia GPUs was already a major infrastructure decision. The new order effectively adds another 2 million units only months later.
That timing is significant. Nvidia said demand had exceeded expectations roughly five months after Amazon’s earlier commitment, according to the TechCrunch report. The increase suggests that AWS and its customers are revising their computing requirements upward faster than planned, rather than simply executing a fixed multiyear procurement schedule.
AI infrastructure demand is difficult to forecast because it depends on several variables at once. Model developers may need large clusters during training, then require a different mix of computing capacity for inference when products reach users. New models can also increase demand even when the number of users remains stable. A more capable model often requires more computing for each request, particularly when companies add long context windows, multimodal processing, reasoning systems or agentic workflows.
This creates an unusual market dynamic. Demand does not necessarily rise only because more people are using AI. It can rise because each interaction becomes more computationally expensive.
Nvidia’s reported quarterly data center revenue of $89 billion, up 117 percent year over year, provides the clearest commercial evidence of the current spending cycle. Such growth would be difficult to sustain without strong orders from the largest cloud providers and model companies. Amazon’s expanded purchase indicates that the major buyers still expect AI workloads to grow quickly enough to justify additional capacity several years into the future.
Yet an order is not the same as utilization. Cloud companies can reserve capacity to ensure access to scarce hardware, support future customer commitments or prevent competitors from gaining a supply advantage. The investment may be strategically rational even if some of the machines are not fully utilized immediately. It also means the industry must distinguish between genuine end customer demand and infrastructure spending based on expectations.
That distinction will become more important as deliveries extend into 2027 and 2028. By then, the economics of model inference, the competitive position of AI applications and the availability of custom chips could look very different.
Nvidia is becoming an ecosystem, not just a chip supplier
The AWS deal reaches beyond GPUs. The companies are also working to integrate Nvidia networking hardware, open models, CPUs, data processing software and robotics platforms.
That broader relationship matters because Nvidia’s advantage is not based solely on the performance of an individual accelerator. Its strategic position comes from the system built around the chip. Networking determines how efficiently thousands of processors work together. Software determines how easily developers can use the hardware. Model tools, deployment frameworks and optimized libraries influence switching costs.
By integrating more of Nvidia’s stack into AWS, Amazon can offer customers a packaged environment for advanced AI workloads. Customers may be able to access Nvidia computing, networking and software through familiar AWS services rather than building and managing the infrastructure themselves. For AWS, that can make the platform more attractive to companies that want the capabilities of Nvidia systems without owning the underlying data centers.
The partnership also extends into physical AI. Amazon plans to use Nvidia’s Omniverse, Cosmos, Isaac and Jetson platforms for its robot fleet. AWS will make Nvidia’s Nemotron open models available through Bedrock and SageMaker.
The robotics connection could become strategically important for Amazon because its logistics operations provide a large commercial environment in which AI controlled machines can be tested and deployed. Warehouse robots, delivery systems and industrial automation require more than language models. They need simulation, perception, planning, real time processing and specialized hardware. Nvidia’s physical AI tools are designed to connect those functions.
For Nvidia, Amazon offers more than a large customer. It provides a major distribution channel for software and platforms that could expand Nvidia’s addressable market beyond data center training and inference. For AWS, the partnership supplies tools that could accelerate robotics development without requiring Amazon to build every component internally.
The arrangement therefore benefits both companies, but not equally in every area. Nvidia gains additional hardware revenue and broader control over the software environment. AWS gains access to a mature technology stack and a way to satisfy customers that specifically want Nvidia systems. The risk for Amazon is that the partnership strengthens the very supplier it is trying to challenge.
Trainium shows the limits of vertical integration
AWS has spent years developing Trainium, its own AI accelerator, as an alternative to Nvidia hardware. Custom chips can give a cloud provider more control over cost, supply and product design. They can also be optimized for the provider’s software and for specific types of workloads.
AWS has discussed selling Trainium capacity to outside companies, and its custom chip business has reportedly reached a $25 billion annualized revenue run rate, driven by AI related commitments. That is a substantial achievement. It suggests that Amazon’s chip strategy is not merely experimental and that some customers are willing to use non Nvidia systems when the economics or availability are attractive.
Trainium can help AWS differentiate its cloud service. If Amazon can offer comparable performance at a lower total cost, it could improve margins and reduce the amount it pays to external suppliers. Custom accelerators may also allow AWS to tailor systems for inference, where the cost of serving millions or billions of model requests can matter more than peak training performance.
But the expanded Nvidia order shows that custom chips do not eliminate the need for merchant hardware. They address only part of the market.
Customers building advanced models often want the latest Nvidia architecture because their software is already optimized for it, their engineers are familiar with the development environment and their models may have been trained or tested on Nvidia systems. Porting workloads to another accelerator can involve more than replacing one chip with another. It may require changes to kernels, compilers, libraries, distributed training systems and operational processes.
The decision is also influenced by risk. A customer may accept a higher price for Nvidia capacity if it reduces the chance of delays, performance problems or engineering work during a critical product launch. In a market where companies are competing to release models and AI services quickly, time can be worth more than the theoretical savings from a custom accelerator.
This creates a two track strategy for AWS. It can use Trainium where workloads are predictable, cost sensitive and sufficiently compatible with its software environment. It can rely on Nvidia where performance, availability of tools and customer preference are more important. That approach is commercially sensible, but it does not produce full independence.
The economics of avoiding Nvidia are complicated
Cloud providers have several reasons to pursue their own chips. Nvidia captures a significant share of the value created by AI infrastructure, and its pricing power is reinforced by strong demand. Hyperscalers would prefer to retain more of that value inside their own platforms.
Custom silicon can also improve capacity planning. A cloud provider that controls chip design can coordinate the accelerator with its servers, networking systems, cooling architecture and software. It can decide which workloads to prioritize and potentially reduce reliance on a constrained external supply chain.
However, the savings are not automatic. Designing an accelerator requires large research and development investments. The provider must fund software tools, developer support and system integration. It must also absorb the risk that a chip will arrive late, perform below expectations or become obsolete as model architectures change.
Nvidia spreads those costs across a broad customer base and a large ecosystem. Its scale allows it to invest aggressively in successive architectures, networking products and software platforms. A hyperscaler may design an efficient chip for its own workloads, but it cannot easily replicate the entire ecosystem that makes Nvidia hardware widely usable.
The new AWS commitment reflects this tradeoff. Amazon is willing to invest in its own silicon, but it is not willing to let that effort prevent AWS from offering Nvidia’s newest systems. Customer demand comes first. If customers believe that the best workloads require Nvidia, AWS must provide Nvidia capacity or risk losing those workloads to Microsoft Azure, Google Cloud or specialized providers.
That competitive pressure may be more important than the cost of any individual accelerator. Cloud companies do not compete only on price. They compete on the ability to attract model developers, startups and large enterprises that need reliable access to advanced computing. A platform with cheaper chips but insufficient access to the preferred hardware may lose valuable customers.
The capacity race could create new risks
The scale of the investment raises a second question: whether AI revenue will grow quickly enough to support the infrastructure being built.
The industry has strong reasons to expect demand to continue expanding. Enterprises are embedding AI into software, customer service, search, coding, analytics and operational systems. Model providers are adding new capabilities that require more computation. Robotics and industrial applications could eventually create another large source of demand.
Still, many AI businesses remain in an investment phase. Companies are spending heavily on training, inference and data center capacity before they have demonstrated durable margins. Some services may eventually generate substantial revenue, but others could remain difficult to monetize if customers are unwilling to pay enough to cover their computing costs.
High utilization is not guaranteed simply because total demand is rising. Workloads may be concentrated during training periods, leaving capacity underused at other times. Inference demand can be unpredictable. Customers may reserve capacity for future needs but delay deployment because of product changes, regulatory concerns or weak adoption.
There is also a risk that model efficiency improves faster than expected. Better algorithms, smaller models, quantization and specialized inference systems could reduce the amount of computing required for some tasks. That would not necessarily end the AI infrastructure boom, since lower costs could encourage wider use. But it could change which hardware is most valuable and leave some planned capacity less attractive.
The reverse risk is equally serious. If reasoning systems and autonomous agents require far more computation per task, demand could exceed even the current investment plans. In that scenario, AWS’s Nvidia order would look less like overbuilding and more like a necessary effort to secure supply.
The difficulty is that both outcomes can be plausible at the same time. Overall demand can grow rapidly while individual facilities, models or customers produce disappointing returns. Infrastructure providers must therefore manage not only the size of their investments but also their timing and flexibility.
Who gains the stronger position?
Nvidia appears to gain the clearest near term advantage. It receives a major commitment from one of the world’s largest cloud companies and expands its role from accelerator supplier to strategic platform partner. The inclusion of networking, software, models and robotics tools increases the number of points at which Nvidia can influence customer decisions.
Amazon also gains. AWS can offer a wider range of systems, meet customer demand for Nvidia hardware and use the partnership to accelerate physical AI initiatives. Its ability to combine Nvidia infrastructure with Trainium gives it more options than a cloud provider dependent on only one architecture.
But the deal also exposes AWS to a strategic contradiction. The more customers associate advanced AI on AWS with Nvidia systems, the harder it may be for Amazon to make Trainium the default choice. AWS can still use custom chips to improve economics behind the scenes, but it may have less power to change customer expectations.
For Nvidia, the danger is that its biggest customers remain determined to build alternatives. Amazon, Google and Microsoft are all investing in custom silicon or specialized systems. If those efforts eventually reach sufficient scale and software compatibility, Nvidia’s share of cloud infrastructure could decline even while the overall market grows.
That transition is not immediate. The expanded order suggests the current phase of AI development still favors Nvidia’s integrated platform. Custom chips are gaining ground, but they have not yet displaced the ecosystem customers already trust.
The most important result of Amazon’s decision is therefore not the number of GPUs. It is the evidence that the AI infrastructure market is becoming more dependent on scale, compatibility and speed. AWS can design its own chips, but it still needs Nvidia to satisfy the most demanding workloads and protect its position against rival clouds.
The partnership shows that vertical integration has a practical limit. Owning more of the stack can improve margins and control, but only if customers are willing to follow. For now, customer demand is pulling AWS back toward Nvidia’s platform, even as Amazon invests in the tools that could eventually make that dependence less necessary.
That tension will define the next stage of the AI infrastructure race. The winners will not simply be the companies that build the most chips. They will be the companies that match capacity to real demand, turn expensive hardware into profitable services and preserve strategic flexibility when the next generation of computing arrives.
