Amazon orders 2 million Nvidia GPUs amid AI demand

Amazon faces the accelerated computing bottleneck

Amazon is reportedly preparing a substantial expansion of its purchases of Nvidia graphics processors. According to TechCrunch, whose headline reads “Amazon just tripled its order of Nvidia chips over ‘surging demand’”, the group could add around 2 million Nvidia GPUs to its infrastructure over the next two years. The move would represent a tripling of the order initially envisaged, according to information reported by the US media outlet.

The figure should be treated cautiously, as this is information reported by TechCrunch rather than a detailed announcement directly from Amazon or Nvidia. But its scale, as well as the reason cited — rising demand — sheds light on a reality that has become central to the technology industry: accelerated computing capacity is one of the main limiting factors in the deployment of artificial intelligence.

Since the explosion of generative AI, major cloud providers have no longer simply sold storage, virtual machines, and database services. They must be able to reserve, install, power, cool, and operate very large volumes of accelerators. These components are needed to train language models, multimodal models, computer vision systems and, increasingly, to run these models at scale once they are made available to users.

For Amazon, this pressure is distinctive. The group is both the operator of AWS, one of the world's leading cloud infrastructure providers, a major Nvidia customer, and a designer of its own accelerators, notably Trainium for training and Inferentia for inference. The reported increase in Nvidia GPU orders does not necessarily call this internal strategy into question. Rather, it indicates that chips developed by AWS are not, on their own, enough to meet the diversity and volume of its customers' needs.

Demand does not come solely from startups specializing in foundation models. It also comes from major software publishers, companies that train or adapt their own models, research players, content creation platforms, and organizations seeking to integrate generative assistants into their business tools. Some customers want Nvidia GPUs because their software, development frameworks, and technical teams have been built around that environment. Others are primarily seeking available capacity, regardless of the chip supplier.

This distinction is essential. In traditional cloud computing, competition was largely based on price, regional availability, security, managed services, and administration tools. In AI, these criteria remain important, but they are preceded by a more fundamental question: are accelerators actually available, in the desired region, at the desired time, with the required interconnects and memory? When a customer cannot access this infrastructure, detailed comparisons of prices or features become secondary.

The term “GPU” also encompasses varied technical realities. Originally intended for graphics rendering, GPUs became established in scientific computing and then in deep learning thanks to their ability to perform a very large number of operations in parallel. Modern AI models rely on repeated mathematical operations, notably matrix multiplications. The parallel architecture of these processors makes them particularly suited to these workloads, although specialized chips such as ASICs developed by some cloud providers can be designed to optimize certain scenarios.

The volume mentioned by TechCrunch therefore serves as a reminder that the AI battle is not being fought solely in the laboratories that design models. It is being fought in data centers, semiconductor supply chains, electrical capacity contracts, and in the ability to turn highly sought-after components into cloud services usable by customers. For Amazon, as for its competitors, the challenge is to prevent the scarcity of computing from becoming a lasting commercial obstacle.

What TechCrunch reports: an order tripled by demand

TechCrunch's central information is clear: Amazon reportedly decided to triple its order of Nvidia chips in response to demand described as growing. The outlet states that this expansion would involve adding 2 million Nvidia GPUs to the group's infrastructure over the next two years. Without additional public details on the precise delivery schedule, the processor families involved, financial amounts, or allocation among different data centers, it would be risky to go beyond what has been reported.

What nevertheless emerges is the strategic nature of the decision. An order of this scale is not a simple one-off capacity adjustment. Installing accelerators in data centers involves major infrastructure choices: rack design, very-high-speed networking, power supply, cooling, orchestration software, customer isolation mechanisms, and monitoring tools. GPUs are only the visible part of a much larger whole, often referred to as an “AI cluster.”

Requirements also differ depending on use cases. Training large models typically requires very many accelerators to work together within the same system, with fast communications between them. Performance therefore does not depend solely on the power of an individual chip. It also rests on interconnects, memory bandwidth, network topology, and the software's ability to distribute computing efficiently.

Inference, meaning the execution of an already trained model to respond to users or process data, poses a different equation. It may require less computing per request than full training, but it must be carried out with low latency, potentially for a very large number of users. With the spread of conversational assistants, image generation features, document summarization, and augmented search, inference has become a source of infrastructure consumption in its own right.

TechCrunch's wording, which refers to booming demand, can thus encompass several simultaneous trends. There are Amazon's own needs, as it integrates AI into different activities. There are those of AWS customers, which consume accelerated computing instances or managed AI services. Finally, there is the precautionary effect: a cloud provider must order and deploy capacity long before it knows precisely which customers will use it and at what pace.

In this market, the risk of underinvestment is high. Insufficient capacity can lead a strategic customer to turn to a competitor or distribute workloads across several clouds. The risk of overinvestment also exists, as accelerators are costly to acquire and operate. But for hyperscalers, AI computing capacity has become a resource as structuring as general-purpose servers were during the major expansion of public cloud computing.

This dynamic explains why the reported announcement should not be interpreted as a simple commercial victory for Nvidia or as Amazon abandoning its own chips. Above all, it illustrates that major platforms are seeking to multiply their sources of computing, architectures, and options offered to customers. In an environment where availability is tight and needs are evolving very quickly, diversification is a form of industrial insurance.

It is also necessary to distinguish the announced number of chips from the final capacity accessible to users. Several stages occur between ordering, manufacturing, system assembly, deployment in data centers, and the commercial launch of cloud instances. The figure of 2 million reported by TechCrunch indicates the scale of Amazon's procurement ambitions, but does not by itself provide information about the services that will be available, their prices, locations, or launch schedule.

For Nvidia, this potential demand comes at a time when its accelerators have become a central part of the AI economy. The group does not only supply chips: its influence also stems from CUDA, its parallel computing software platform, and a set of libraries and tools used by many developers, researchers, and software publishers. This hardware-software combination makes migrations to alternative architectures possible, but rarely immediate.

Why Nvidia remains indispensable despite Trainium and Inferentia

Hyperscalers' dependence on Nvidia is one of the most visible paradoxes of the current AI race. Amazon, Google, Microsoft, and other major operators all have an interest in reducing their exposure to a single supplier, gaining greater control over costs, and adapting chips to their services. Yet they continue to deploy Nvidia GPUs at scale. The reason lies as much in hardware performance as in the software ecosystem and the market's working habits.

Nvidia began developing CUDA in the mid-2000s, well before generative models reached the general public. This head start allowed its environment to become a benchmark in accelerated computing. With the development of deep learning in the 2010s, many major tools were optimized for Nvidia GPUs. Engineering teams, academic researchers, and companies accumulated expertise in this technical stack.

For an organization seeking to train a model, choosing infrastructure is not limited to a chip's stated specifications. It must assess library compatibility, compiler maturity, the availability of debugging tools, methods for distributing computing, real-world performance on the models used, and internal skills. Nvidia benefits here from an inertia advantage: a significant share of the ecosystem has been built around its technologies.

Amazon Web Services is not absent from this field. AWS has presented Inferentia as a chip dedicated to inference workloads and Trainium as a chip dedicated to training. These products serve several goals. They can offer AWS customers an alternative to GPUs for certain workloads. They also give Amazon more control over its infrastructure and roadmap. Finally, they create useful competitive pressure on Nvidia in a commercial relationship in which AWS remains one of its major customers.

But designing an accelerator and building an ecosystem comparable to CUDA's are two very different challenges. Theoretical performance alone does not guarantee adoption. Customers must be able to port their models, adapt their pipelines, validate result accuracy, train teams, and ensure that the environment meets their production requirements. For companies already committed to the Nvidia ecosystem, the operational cost of a change can be significant.

AWS offerings are therefore likely to coexist rather than completely replace one another. Trainium may be relevant when a customer chooses to optimize its training for the AWS environment. Inferentia may suit certain inference cases where cost-performance is the priority. Nvidia GPUs nevertheless retain considerable appeal for workloads that must remain close to the industry's most widely used tools or that require broad compatibility with existing software.

The same phenomenon is visible among Amazon's competitors. Google has developed its TPUs, used notably for certain internal workloads and offered in its cloud. Microsoft is a major buyer of Nvidia GPUs for its Azure infrastructure and AI services. Meta has also invested heavily in accelerated computing infrastructure to train and serve its models. Each company is trying to build proprietary capabilities, but none of these strategies has eliminated Nvidia's role in the broadest deployments.

This coexistence of general-purpose chips, specialized GPUs, and in-house accelerators reflects a real fragmentation of needs. Very large language models, recommendation systems, industrial vision, scientific data analysis, and conversational applications do not all impose the same constraints. Some uses are sensitive to cost per request, others to training speed, and still others to memory capacity or ease of deployment.

The additional order reported by TechCrunch therefore suggests less a failure of Amazon's internal chips than a pragmatic trade-off. AWS needs a diversified offering capable of meeting highly varied demand without forcing all customers into a technology migration. In the cloud, the breadth of the catalog and immediate availability are competitive advantages. A platform that can offer several architectures reduces the risk that an AI project will be delayed by the absence of a specific type of hardware.

This reality nevertheless raises a major economic question. If cloud providers use both their own accelerators and Nvidia GPUs, pressure on capital expenditure remains high. Costs do not only concern purchasing chips. They include electrical infrastructure, cooling, networking, maintenance, and depreciation of hardware whose generations evolve quickly. The AI market has thus become one in which financial and industrial capacity matters almost as much as software quality.

The GPU race, a bottleneck for the entire AI chain

Massive reliance on GPUs is transforming how digital offerings are built. For a long time, a company could launch a web service with a limited number of general-purpose servers, then gradually increase capacity based on its audience. Generative AI applications change this model. An organization seeking to train a high-performing model may need considerable infrastructure before generating a single unit of revenue. An organization that wants to serve millions of requests must likewise anticipate demand and reserve computing in advance.

The relative scarcity of accelerators affects pricing, access delays, and market concentration. Organizations able to sign large procurement contracts or build their own data centers benefit from a structural advantage. Smaller companies often have to rent resources from a cloud provider, use intermediaries, or heavily optimize their models to reduce computing consumption.

This constraint can encourage software innovation. Quantization, which consists of reducing the numerical precision used by a model under certain conditions, is one avenue used to decrease hardware requirements. Distillation makes it possible to create more compact models from larger models. Fine-tuning existing models can be less costly than complete training from scratch. Inference optimization, caching, and routing requests to models of different sizes also play a role.

But these techniques do not eliminate the need for computing. They alter its distribution. A more efficient model can be deployed more broadly, increasing the total number of requests. New uses then become profitable: writing assistance, automation of document processes, code analysis, internal search, customer service, content generation, or specialized tools for regulated professions. Improved efficiency can therefore fuel a new increase in demand for infrastructure.

The challenge is also physical. AI accelerators concentrated in data centers consume electricity and generate heat. Their deployment requires appropriate energy capacity and cooling systems. Chip order announcements are often presented from the perspective of computing performance, but they also point to real estate, energy, and territorial issues. A cloud provider cannot install additional clusters indefinitely without the necessary buildings, connections, and equipment.

For hyperscalers, planning is therefore becoming more complex. They must anticipate component delivery cycles, changes in demand, site availability, energy constraints, and regulators' expectations. They must also arbitrate between shared infrastructure, used by multiple customers, and capacity reserved for very large users. In some cases, a customer may favor reserving dedicated capacity to secure strategic projects.

Amazon is not alone in this competition. Microsoft Azure, Google Cloud, and Oracle Cloud Infrastructure are among the platforms offering accelerated computing resources for AI. Data center builders and operators are also being drawn into the movement, along with manufacturers of memory, networking equipment, and cooling systems. Nvidia occupies a visible position in this chain, but final availability depends on many industrial links.

The rise of open models adds further pressure. When a model is made available to the community, thousands of organizations may seek to adapt, evaluate, or run it. This spreads innovation, but also distributes demand for computing far beyond a small number of laboratories. Cloud providers then find themselves serving both very large contracts and a multitude of more fragmented workloads.

In this context, the addition of 2 million Nvidia GPUs reported by TechCrunch appears as an indicator of capacity rather than merely an indicator of component sales. It reflects the need for Amazon to maintain a sufficient reserve of power so as not to let AI demand shift to other platforms. Commercial value does not depend only on the installed chip: it depends on the ability to make it available in a reliable, documented, secure service integrated with the rest of the cloud.

The risk of concentration should finally be emphasized. When a significant part of the industry relies on the technologies of a limited number of component suppliers, any supply tension, production delay, or price change can have very broad effects. This is precisely why hyperscalers invest in internal alternatives, support different partners, and seek to develop software capable of operating on several architectures. The order mentioned by TechCrunch shows, however, that this diversification is not yet enough to sharply reduce Nvidia's role.

Consequences for France and Europe: access to computing, sovereignty, and project costs

For French and European businesses, Amazon's investments in Nvidia GPUs have a double meaning. On one hand, they may help expand the computing supply available through AWS, whose services are used by many organizations on the continent. On the other hand, they serve as a reminder that access to the resources needed to develop and operate AI models depends largely on infrastructure and technology chains controlled by major international groups.

The issue is not abstract for businesses. A laboratory, startup, or large group seeking to train a model must be able to finance and obtain access to accelerators. If it uses a cloud provider, it must examine costs, reservation conditions, data location, security mechanisms, and availability in relevant regions. If it chooses to acquire its own hardware, it must manage hosting, energy, operations teams, and technology renewal.

In many cases, the cloud remains the most accessible solution, as it avoids a very high upfront investment and enables resources to be mobilized temporarily. However, when demand explodes, dependence on the availability of major providers becomes more visible. The announcement reported by TechCrunch can thus be interpreted as a positive signal for customers seeking AWS capacity, but it does not guarantee that this capacity will be offered in every region or under identical conditions for all uses.

The location of processing is particularly important in Europe. Requirements related to data protection, contracts, security, and regulated sectors may lead organizations to favor deployments in certain geographic areas. The availability of accelerators in European cloud regions is therefore a concrete factor in competitiveness. A company cannot always freely move its data or processing to any data center to benefit from more abundant capacity.

This situation fuels debates on digital sovereignty and computing sovereignty. Europe has recognized expertise in AI research, semiconductors in certain segments, software, and many industrial sectors. But training very large models and serving generative applications at scale require infrastructure whose cost and complexity are considerable. The gap is not measured only by the number of models developed: it is also measured by the ability to run these models sustainably.

European cloud and infrastructure providers face a difficult equation. They can differentiate themselves through proximity, compliance, local support, or specialized offerings. But they must also access the same components and energy capacity as hyperscalers. When major global players secure very high volumes of GPUs, smaller operators must find their own paths: partnerships, market niches, resource pooling, software optimization, or targeted deployments.

For French startups, the consequence is mixed. Access to existing models and APIs makes it possible to launch products quickly without training a complete foundation model. This is a major advantage: not all young companies need tens or hundreds of thousands of accelerators to innovate. But those that want to develop proprietary model technology, process sensitive data at scale, or guarantee strong independence from platforms must incorporate computing costs into their strategy from the outset.

The debate also concerns the environmental footprint. Increasing AI capacity generates additional electricity needs and makes energy efficiency more important. For customers, choosing a smaller or better-optimized model can have an economic and operational impact. For providers, the challenge is to improve infrastructure utilization, limit idle resources, and deploy suitable cooling systems. GPU announcements therefore cannot be separated from European discussions about energy and data centers.

Finally, French organizations should avoid an overly binary reading that pits Nvidia GPUs against alternative chips. The right choice depends on the model, the volume of data, expected latency, available skills, desired portability, and compliance obligations. The most robust strategy may be hybrid: use managed services for certain needs, reserve accelerated computing for intensive phases, adopt more compact models for inference, and retain the ability to move workloads when possible.

Toward AI defined by the ability to industrialize computing

The order reported by TechCrunch comes at a time when the AI market is gradually moving from technological demonstration to industrialization. Generative AI's first public successes demonstrated the value of these tools, but their widespread deployment now requires addressing issues of cost, reliability, security, data governance, and hardware capacity. GPUs are at the heart of this transition because they directly determine the volume of computing available.

Amazon will have to reconcile several objectives. The first will be to meet immediate demand for Nvidia GPUs, requested by customers already using this environment. The second will be to continue developing Trainium and Inferentia in order to have technical and economic alternatives. The third will be to make very substantial infrastructure investments profitable in a sector where chip performance and model needs are evolving quickly.

The ability to offer several architectures could become an even more important advantage. The most advanced customers will likely seek to distribute their workloads according to their constraints: GPUs for certain training tasks or to maintain broad software compatibility, proprietary accelerators for optimized processing, and general-purpose processors for tasks that do not require massive computing. From this perspective, the best-positioned hyperscaler will not necessarily be the one with a single dominant chip, but the one that makes these choices easier to manage.

Nvidia, for its part, retains a considerable advantage thanks to the maturity of its ecosystem. But increased investment in in-house chips, specialized accelerators, and more portable software means that its customers are also seeking to reduce the risks associated with excessive dependence. Competition is not limited to silicon design. It extends to development tools, compilers, libraries, networks, cloud services, and support capacity.

For the French-speaking market, the question in the coming years will be less whether AI will use accelerated computing — it already does — than determining who will be able to access it, at what price, and with what guarantees. Companies that treat this resource as a simple IT expense risk underestimating its strategic importance. Those that precisely assess their needs, optimize their models, and preserve technical options will be better able to navigate a market where computing power is becoming a digital raw material.

If TechCrunch's information is confirmed in full, Amazon's addition of 2 million Nvidia GPUs will be one of the clearest signals of this industrialization. It will not merely show that AI demand is growing. It will show that even AWS, which develops its own accelerators, still considers massive access to Nvidia GPUs indispensable to meet market demand. Over the longer term, the real dividing line will probably not be between companies that use AI and those that do not, but between those that master the economics of computing — availability, efficiency, location, and cost — and those that are subject to it.