Amazon wants to sell its AI chips to challenge Nvidia

Amazon is taking its AI chips beyond AWS and aiming more directly at Nvidia

Amazon could take a major strategic step in the battle over hardware for artificial intelligence. According to TechCrunch, citing The Information, AWS is reportedly in talks to market its AI chips beyond its own cloud. In other words, Amazon would no longer be content to use its in-house accelerators as an internal lever for its cloud services: the group would also seek to sell them more broadly, with a market dominated by Nvidia in its sights.

The stakes go far beyond a simple catalog expansion. For several years, major cloud providers have been developing their own semiconductors to reduce costs, better control their supply chains, and optimize the performance of certain workloads. But the prospect of Amazon turning this internal capability into a standalone commercial business changes the scale of things. It would place the group in more direct confrontation with Nvidia, currently a central player in global AI infrastructure, while increasing competitive pressure on other chip suppliers for data centers.

The economic signal is explicit. According to the report relayed by TechCrunch, Andy Jassy, Amazon’s chief executive, presents this business as a potential $50 billion opportunity. Such a figure indicates that, for Amazon, AI chips are no longer just an internal optimization tool intended to support AWS, but potentially a business line in their own right. This fits into a broader movement: generative AI has turned compute accelerators into strategic assets, on the same level as data centers, energy, or networks.

The timing is particularly significant. After the initial rush around training large models, the market is increasingly focusing on inference, that is, running models in production at scale. It is on this ground that costs, hardware availability, and energy efficiency become decisive. If Amazon commercially opens up its chips, it will not simply be a matter of offering a technical alternative: the group could help redefine the value chain of AI data centers, moving from the status of integrated user to that of supplier of components and platforms.

For European and French companies, this development deserves particular attention. The local AI market depends heavily on imported infrastructure, particularly American infrastructure, and remains subject to the pricing, availability, and commercial priority decisions of hyperscalers. Any credible diversification in the face of Nvidia can alter cost structures, access to compute capacity, and the architecture strategies of companies deploying AI services in production.

Chips designed for AWS, in a long-term strategy

The possibility of broader commercialization is not coming out of nowhere. Amazon has been developing its own chips for the cloud for several years. This strategy is part of a deeper trend among major technology players: instead of depending exclusively on general-purpose suppliers, they design components tailored to their specific needs. In AWS’s case, the initial objective was clear: to gain greater control over the costs and performance of its infrastructure.

Amazon has gradually built a portfolio of in-house silicon for different uses. The group has notably highlighted its Graviton processors for general-purpose workloads, while its AI effort has taken shape around chips dedicated to training and inference. This vertical logic is consistent with AWS’s DNA: optimize the full stack, from hardware to software services, in order to offer more predictable performance and more competitive prices.

Until now, this integration primarily served Amazon’s cloud offering. The in-house chips were a means of differentiation for AWS, not necessarily a product sold independently to customers outside the Amazon cloud perimeter. That is precisely what makes the discussions reported by TechCrunch so important. If Amazon chooses to open up commercialization of its accelerators, the company will no longer position itself only as an infrastructure operator, but also as a more direct supplier of AI hardware.

This development recalls a shift often seen in tech: a capability developed for internal use becomes, once mature, a commercial product. Historically, AWS itself was born from this logic of industrializing tools initially built for Amazon’s own needs. The same pattern applied to AI chips would have considerable significance, because it would affect a sector where barriers to entry are very high, development cycles are long, and credibility is measured in performance, availability, and software capability alike.

It is nevertheless necessary to remain cautious about the exact scope of the initiative. The source mentions discussions, not a detailed launch with a timeline, product list, commercial terms, or named partners. At this stage, this is therefore more of a strategic signal than an exhaustive roadmap. But that signal is strong enough to shed light on Amazon’s ambitions: to move out of a defensive posture, where chips serve above all to reduce dependence on external suppliers, and adopt an offensive posture in the market for data center accelerators.

This ambition also reflects the transformation of hardware’s role in the AI economy. Long regarded as a relatively interchangeable layer for the cloud, silicon has once again become a major differentiating factor. The success of large models has shown that accelerator availability was not just a matter of raw performance, but a determining element in the ability to launch products, attract customers, and support growth. In this context, owning its own chips may no longer be enough: it must also decide whether that control should remain an internal advantage or become a commercial business.

Why Amazon wants to reduce Nvidia’s grip on AI infrastructure

The implicit target of this strategy is obvious. Nvidia occupies a central position in the AI ecosystem, particularly in data centers. Its GPUs have become the benchmark for model training and retain major weight in inference, even if that segment is now attracting more competitors. This dominance does not rest on hardware alone. It also stems from the maturity of the software ecosystem, developers’ familiarity with its tools, and Nvidia’s ability to deliver a complete platform, from silicon to optimization libraries.

For hyperscalers, this situation presents advantages but also risks. The main advantage is rapid access to high-performance chips that are widely adopted. The main risk is dependence. When one player concentrates such a large share of the value, it acquires considerable power over prices, available volumes, delivery schedules, and, more broadly, over the market’s technological evolution. Developing internal alternatives has therefore become an almost natural response for the major cloud providers.

Amazon is not the only one following this path. It is well established that other technology giants are also investing in their own AI chips. This competitive dynamic does not mean Nvidia is immediately threatened in its core market, but it does reflect a structural change: the biggest buyers of AI hardware are seeking to become, at least partially, their own suppliers. If Amazon also decides to sell its chips externally, the relationship changes further. The company is no longer merely limiting its dependence; it is trying to capture part of the value currently absorbed by Nvidia.

The $50 billion figure mentioned by Andy Jassy gives an idea of the scale of the bet. Such an amount suggests that Amazon does not see this market as a simple marginal extension of AWS, but as a leading industrial opportunity. This matches the new hierarchy of value in AI: models attract media attention, but durable revenue and the most strategic margins are often found in the infrastructure layers, where compute capacity, interconnects, storage, and now specialized accelerators are sold.

The most favorable ground for an Amazon offensive appears to be inference. In training very large models, Nvidia benefits from a considerable lead in ecosystem and adoption. By contrast, inference opens up more room for differentiation. Customers there often seek a better trade-off between cost, energy consumption, latency, and throughput. That is precisely where an integrated provider like AWS can highlight its strengths: it controls the cloud environment, knows its customers’ consumption patterns, and can design its chips for targeted uses rather than for maximum versatility.

Another factor reinforces Amazon’s interest: demand for AI keeps increasing, but companies are paying closer and closer attention to the bill. The era in which raw power was almost exclusively prioritized is giving way to a phase of economic optimization. If Amazon can offer competitive accelerators for certain workloads at lower cost or with better availability, the commercial argument could be powerful, especially for companies deploying AI services at scale and seeking to stabilize their infrastructure spending.

A battle shifting from training toward inference and the value chain

One of the most important aspects of the matter is the gradual shift in competition. During the first phase of the rise of generative AI, attention focused on training the largest models. It was this race that put Nvidia at the center of the game, with GPUs being essential to absorb the colossal compute needs of labs and major technology groups. But as models move into production, inference is becoming an equally strategic battleground, perhaps even larger in volume over the long term.

Inference corresponds to the concrete use of models: conversational assistants, text generation, augmented search, classification, summarization, information extraction, document automation, translation, or recommendation. Once a model is trained, it must be run again and again to respond to user requests. At scale, this represents a continuous, potentially massive cost. The challenge is therefore no longer simply to have very powerful chips, but to run repetitive workloads profitably.

In this context, the value chain of AI data centers is being reshaped. Hardware can no longer be thought of in isolation. Its value depends on its integration with software frameworks, deployment tools, orchestration layers, managed services, and optimization mechanisms. This is an area where AWS has a structural advantage: Amazon does not just sell silicon, but a complete cloud services environment. Commercializing its chips could therefore allow it to offer a more vertical proposition, in which hardware would be tightly coupled with operations, provisioning, and the platform’s AI tools.

This integration can reshuffle the cards in several ways. First, it can shift competition away from the sole level of raw performance toward that of total cost of use. Next, it can favor more specialized architectures, optimized for families of models or specific scenarios. Finally, it can reduce the role of intermediaries in the value chain, to the benefit of players capable of simultaneously mastering chips, data centers, and software services.

Amazon’s move also illustrates a growing tension between two business models. The first is that of an independent chip supplier, which sells its hardware to the entire market. The second is that of an integrated hyperscaler, which develops its own components first for internal use, then possibly for its customers. If the major clouds gradually shift toward this second model at scale, the sector’s competitive structure could evolve durably. Chipmakers would no longer sell only to cloud operators: they would also have to face operators that have become competitors on the silicon front.

For Nvidia, this does not mean an immediate loss of leadership. The company retains solid advantages, notably its head start, its ecosystem, and its installed base. But the potential commercial opening of Amazon’s chips shows that the next phase of competition will not concern only who has the best chips, but who controls the complete architecture of AI in production. On this front, hyperscalers have particular power: they see real-world usage, manage infrastructure at scale, and can adjust hardware and software according to the needs observed among their customers.

What this could change for companies, clouds, and the French-speaking market

For user companies, the main interest of such a development lies in diversification. Today, many organizations deploying AI depend heavily on a small number of suppliers for access to accelerators. This concentration can create tensions over prices, availability, and architecture choices. If Amazon manages to offer a credible proposition beyond its cloud, that could broaden the options available to large enterprises, operators, and certain integrators.

In the French-speaking world, this issue is particularly sensitive. French and European companies consume technologies designed outside the continent on a massive scale, whether models, cloud platforms, or hardware components. They therefore bear the industrial trade-offs made in the United States and Asia, without always having comparable levers at their disposal. The arrival of new AI chip offerings, even if they come from another American giant, could nevertheless create more competition in access to compute capacity, which is a concrete issue for CIOs, data teams, and digital service providers.

For AWS customers, the broader commercialization of Amazon chips could also strengthen the attractiveness of the group’s ecosystem. If Amazon is able to offer continuity between hardware, cloud, and AI services, some customers could be encouraged to standardize their deployments more heavily on that technology stack. This raises a classic cloud question: diversification of component suppliers can paradoxically increase vertical integration around a hyperscaler. In other words, reducing dependence on Nvidia does not necessarily mean reducing dependence on a platform.

For other cloud players, Amazon’s move adds further pressure. Hyperscalers are engaged in a race for differentiation in which customized hardware is becoming a central commercial argument. Whoever better controls inference costs, capacity availability, and service optimization can gain market share in AI workloads. Customer companies, for their part, will have to arbitrate among several promises: performance, price, software compatibility, model portability, and the risk of technological lock-in.

The French market could view this development from two angles. On the one hand, stronger competition among accelerator suppliers is likely to improve conditions of access to AI infrastructure. On the other, it confirms that strategic value is increasingly concentrated in the hands of players capable of simultaneously financing chip design, data center operations, and the development of global cloud services. For European digital sovereignty initiatives, the message is twofold: dependence on Nvidia can be put into perspective, but dependence on American hyperscalers remains intact, or may even be reinforced if they also control the silicon.

The implications for integrators, software vendors, and AI start-ups must also be considered. If the market opens up to more hardware architectures, software players will have to optimize their applications for several targets, not just for today’s dominant environments. This may represent an adaptation cost, but also an opportunity to negotiate infrastructure choices more effectively. In sectors where inference is intensive, such as customer service, document search, content analysis, or certain industrial uses, the cost of compute remains a determining variable in the business model.

An offensive that could durably reshape the economics of AI hardware

In the short term, the information reported by TechCrunch should be read for what it is: a project under discussion, not a shift already materialized in all its operational details. But strategically, the direction of travel is clear. Amazon now sees its AI chips as an asset that could be monetized more broadly, in a market the group considers large enough to mention $50 billion. That wording alone says a great deal about the maturity the subject has reached.

In the medium term, the commercial opening of Amazon’s chips could accelerate several underlying trends. The first is the relative fragmentation of the accelerator market, with more options depending on use cases. The second is the rise of inference as the main field of economic competition. The third is the growing integration between hardware and cloud services, which favors players capable of managing the entire chain, from silicon design to application deployment.

This dynamic could also change the way AI companies are evaluated. For a long time, attention focused on the models themselves and on applications visible to the general public. Yet the most structuring battle may be taking place in the less visible layers: data center power supply, interconnection, cooling, advanced packaging, and of course specialized accelerators. By seeking to sell its chips, Amazon is sending the message that the next great source of AI rents could lie in industrialized infrastructure, not only in ownership of the models.

For Nvidia, the issue is less one of a brutal loss of momentum than of a gradual erosion of its indispensable character in certain segments. If inference becomes the main volume driver and hyperscalers impose their own chips on a growing share of deployments, the market’s center of gravity could shift. Nvidia would remain a major player, but in an environment where its biggest customers are also competitors. This is a classic configuration in industrial tech, but relatively new at this scale in AI.

For Amazon, the challenge will be multifaceted. Selling chips is not just a matter of posting good performance. It must convince on reliability, support, the software ecosystem, the roadmap, production capacity, and the sustainability of the investment. That is where the difference is measured between a chip designed for internal use and a platform capable of appealing to a broader market. The potential mentioned by Andy Jassy shows the ambition, but turning that ambition into a durable business will depend on execution and ecosystem buy-in.

From a longer-term perspective, the Amazon case illustrates a profound transformation of the digital industry. Hyperscalers are no longer just buyers of components; they are becoming full architects of the global computing stack. If this trend is confirmed, the boundary between cloud provider, chipmaker, and AI platform supplier will continue to blur. For the French-speaking market, this means infrastructure decisions will no longer be able to be considered separately from choices about industrial dependence. The real change may not only be that Amazon wants to challenge Nvidia, but that the value of AI is increasingly concentrating among those who simultaneously control chips, the cloud, and large-scale uses.