On stage, a demonstration is all it takes: an assistant answers, translates, writes code. Behind the scenes, it takes chips, servers, cables and an uninterrupted power supply. Competition in artificial intelligence also plays out in this industrial engine room. Specialist conferences such as RAISE Summit offer a lens through which to understand this shift in power: inventing a better model is not enough; developers must also be able to train it, distribute it and fund every use.
Looking ahead to September 2026, this perspective sheds light on the choices facing businesses and governments alike. The following analysis draws on publicly documented trends and events established before that date; its projections are forward-looking. It is not a report on announcements or presentations from a 2026 edition of the summit.
Chips: AI’s first tollgate
In a data center dedicated to AI, the graphics processing unit, or GPU, has become a strategic resource. Originally designed to process images, it performs the mathematical operations required by neural networks in parallel. Training a large model requires thousands of accelerators to work together. Their number matters, but so do their memory, interconnections and the reliability of the system as a whole.
Nvidia’s position therefore rests on more than its components. It also depends on CUDA, its software environment, optimized libraries and the expertise developers have accumulated. Switching suppliers can require rewriting, testing and optimizing some of the software. A cheaper, readily available chip is not automatically an equivalent alternative.
AMD, specialized accelerators and chips designed by major cloud providers are broadening the options. Google has been using its TPUs for several generations; Amazon is developing Trainium and Inferentia, among other chips. These solutions can improve the economics of certain tasks. They also sharpen a dilemma: reducing dependence on a manufacturer can increase dependence on a platform.
A global supply chain, national decisions
Production involves several bottlenecks: advanced manufacturing, high-bandwidth memory and sophisticated packaging. TSMC occupies a central position in this ecosystem, while a handful of manufacturers dominate the memory needed by accelerators. Placing more orders does not instantly create capacity: expanding a factory and qualifying its processes takes years.
US restrictions on exports of advanced chips to China, introduced in 2022 and subsequently tightened, have shown that access also depends on geopolitical decisions. For a laboratory, regulations can change which equipment is available and which partnerships are possible. Computing power is now as much a matter of industrial policy as a line item in an IT budget.
Electricity: the factor the demo leaves out
Once servers have been purchased, they still need power and a way to dissipate their heat. An available building does not guarantee a sufficient grid connection. Grid operators must accommodate these new loads, sometimes alongside competing industrial needs. The timeline for a transformer or power line can then become more restrictive than the software schedule.
In its Electricity 2024 report, the International Energy Agency was already pointing to an expected rapid increase in demand from data centers, AI and cryptocurrencies. These categories must nevertheless be distinguished: not all data center consumption can be attributed to AI. Dramatic comparisons often obscure this distinction and the uncertainties surrounding future uses.
Cooling adds a local constraint. Depending on the facility, it requires water, electricity or both. A site’s appeal therefore depends on the climate, the grid, water resources and public acceptance of the project. A commitment to buy renewable electricity over the course of a year does not necessarily mean that machines run on carbon-free power every hour.
For France, a largely low-carbon electricity mix is a potential asset. But that advantage is no substitute for grid connections, investment or operational expertise. Looking toward 2026 and beyond, competition could favor locations able to offer genuinely available capacity rather than merely announce future megawatts.
The cloud sells computing power—and dependence
For a startup, renting GPUs avoids a substantial upfront investment. The cloud also provides storage, networking, security and deployment tools. This simplicity explains its central role. However, it gives major operators control over pricing, quotas, available regions and the terms of access to new machines.
Ties between cloud providers and laboratories illustrate this interdependence: Microsoft with OpenAI, and Amazon and Google with Anthropic. Without making assumptions about specific contractual provisions, the industrial logic is clear. Laboratories gain resources and funding; platforms attract models, developers and customers to their infrastructure.
Competition then plays out on several levels at once. A company can offer an excellent model while depending on a competitor to run it. Cloud credits make it easier to get started, but do not guarantee a sustainable bill as usage grows. Migrating means moving data and rebuilding integrations.
The real economic test begins after training
Training attracts attention because it demands resources on a staggering scale. Yet inference—each response generated by a model—can become the decisive cost at scale. An assistant with little usage is relatively inexpensive to run. The same service integrated into millions of daily workflows changes the equation entirely, especially if each request triggers several operations.
This is why smaller models, quantization, caching and selecting a model to suit the task are becoming strategically important. A document classification tool does not always need the most powerful system. Competitive advantage can come from an efficient architecture capable of meeting a quality threshold with less computation.
Open-weight models broaden the options for hosting and customization. They do not make infrastructure free, and their licenses vary. Nevertheless, they can reduce certain dependencies. For a buyer, comparing advertised performance alone would be insufficient: operating costs, latency, privacy and the ability to switch providers must also be examined.
Sovereignty is measured by the ability to choose
In discussions about AI, owning domestic data centers can easily become a symbol. Operational sovereignty is more demanding: who controls the equipment, software, access and maintenance? Where does the data reside? What happens if a supplier changes its prices or discontinues a service?
What next? The next phase could reward control over models’ total cost more than their sheer size. For businesses, that means testing multiple providers and measuring each use case. For public authorities, it means coordinating energy, industry and competition policy. Behind the conference stages, the decisive question will remain simple: who has the physical resources to turn an AI promise into a sustainable service?
