Chips: the sinews of war

Let us pick up where we left off in the previous chapter: we know how to build data centres. What remains is knowing what we put inside them.

So what are these chips? The ones that run AI are in fact among the most advanced and most complex objects humanity has ever managed to produce. You are reading this article thanks to a chip; your phone works thanks to a chip. And for the anti-technologists among us, a good share of the medical equipment that would be used if anything happened to you runs on a chip. It is the foundation of our modern technology.

What is an electronic chip?

Electronic chips (or semiconductors) are the essential components of modern electronic devices. They are made from silicon, a semiconducting material, and contain millions, sometimes billions, of miniaturised transistors. Those transistors are assembled to perform precise functions: computation, data storage or peripheral management.

Put simply, without these chips, say goodbye to every electronic machine you own: your games console, your computer, a drone, your car, even your washing machine. They are everywhere, in many kinds, with wildly different levels of complexity depending on the use. Here we will focus on the two most talked-about right now: CPUs and GPUs.

The CPU (Central Processing Unit)

The CPU is often described as the "brain" of a computer. Its main functions:

• ▸executing the instructions of software and programs,
• ▸handling the calculations needed to run a computer system,
• ▸coordinating the different parts of a computer.

Structurally, the CPU is divided into several cores, each able to process instructions simultaneously. It contains computing units called ALUs (Arithmetic Logic Units) for mathematical and logical operations, and a control unit to orchestrate the data flows.

The GPU (Graphics Processing Unit)

The GPU specialises in massively parallel computation, in particular the kind needed for graphics rendering. Its main functions:

• ▸handling the calculations behind the display of images, videos and animations,
• ▸accelerating the complex calculations needed for video games, graphics software and virtual reality,
• ▸processing non-graphical tasks that require parallel computation: artificial intelligence, scientific simulation, or cryptocurrency mining.

Its structure is the opposite of the CPU: a very large number of cores, individually less powerful than those of a classic processor. It therefore excels at running very many small tasks at once.

The other types of chip

• ▸ASIC (Application-Specific Integrated Circuit): designed for a single task, for instance in routers or bank cards. This family also includes NPUs (Neural Processing Units), specialised in AI computation, and LPUs (Language Processing Units), designed for running LLMs.
• ▸FPGA (Field-Programmable Gate Array): a chip that can be reprogrammed after manufacturing, for specific uses.

With the arrival of LLMs, public attention has focused heavily on GPUs, now seen as "the cards you do AI with" (and behind them the giant NVIDIA). But it is worth remembering that GPUs were historically designed for graphics processing. Recently, new players have been promoting architectures better suited to artificial intelligence: LPUs, IPUs, NPUs. One of them deserves a detour: Groq (not Elon Musk's AI Grok, but Groq with a q ;) ). I will come back to it below.

You will see, then, why it is essential for a country to be able to source its own chips. Without them, there is no running our factories, our medical equipment, our weapons. This hardware is almost as important as energy or agriculture. And that dependency can indirectly affect our industries, and even our political decisions, as the ITAR regulation has shown.

What are these chips made of?

Just as you need materials to build a wind turbine, or uranium for a power plant, these chips need very specific inputs. Here is a necessarily incomplete list, but one that shows the essentials:

• ▸Silicon, the base material of transistors. Purified to 99.9999999% ("electronic grade"), it is turned into wafers, the discs on which circuits are etched.
• ▸Conductive metals: copper for interconnects, gold for critical connections, tungsten, aluminium, nickel, cobalt and molybdenum for certain alloys and protective layers.
• ▸Dopants, which modify the properties of silicon: phosphorus and arsenic (N-type, adding free electrons), boron (P-type, creating "holes").
• ▸Insulators: silicon dioxide, and hafnium oxide in modern processors to improve transistors while reducing power draw.
• ▸Heat dissipation materials: silicon carbide, graphite and graphene, ceramics.
• ▸Substrates and packaging: glass-fibre reinforced epoxy (FR-4), copper and aluminium alloys, plastics and polymers.

What stands out when you look at the exporting countries for these materials is the preponderance of China. With a special mention for Indonesia, the DRC, Turkey and Peru, which lead on certain materials. Even though these figures cover global exports, and therefore serve far more than chips alone, you quickly grasp how much certain shipping lanes matter.

So there are forms of "monopoly" on materials needed to build chips, which complicates logistics and weighs on relations between countries: the tariff discussions between China and the United States illustrate this constantly. Note in passing that some countries do not exploit their own reserves, France for example, the environmental impact of mining being judged too great or too costly to reduce.

Why advanced nodes matter

There are several design stages between these raw materials and the finished chip. But one of the variables presented as decisive is the notion of advanced nodes. Think of it as depositing a very thin layer of metal in a circuit, which will carry the electricity.

Using advanced nodes (the 7 nm, 5 nm, 3 nm and soon 2 nm processes) has a major impact on performance, power consumption, cost and design complexity. Other, thicker nodes exist: intermediate nodes and legacy nodes. Some are perfectly sufficient to run your washing machine. It all depends on the need and the expected power draw.

Let us take a less cheerful example. In the Ukrainian conflict, drones are everywhere. Piloted 100% by humans at first, we now see drones able to lock on and hold their course despite jamming. To make a drone fly, there is no strategic need for advanced nodes. To put image processing or air defence systems on board, you are into that territory. And optimising the energy needs of that task buys flight time without touching any other component.

Remember that a smaller node lets you fit more transistors into the same space, therefore more power for equal conditions. I will not go into the detail of the problems this creates for heat management or leakage currents, but bear in mind that a finer node offers more architectural options and better efficiency.

But why insist so much on nodes? Because there is a worldwide monopoly on the machines capable of etching the finest nodes, the ones that are indispensable to artificial intelligence. And that company is European: ASML.

The ASML case

In chip creation, you have a giant, TSMC, in Taiwan. But behind that giant sits a company without which none of this would be possible: ASML. It builds the EUV lithography machines, a method that uses extremely short ultraviolet wavelengths to etch patterns onto silicon wafers, making circuit miniaturisation possible.

You have to understand that ASML sells each machine for more than €150 million. It is by far the jewel of Dutch industry, and its chief executive is, with a touch of national pride, French: Christophe Fouquet. Its importance in the Netherlands is such that it shapes the country's political debate.

This company is clearly key to European soft power, and it would be essential to factor it into any strategy aiming to produce latest-generation chips at home.

The foundries

Foundries are the companies that specialise in manufacturing the chips designed by the architects (Apple, NVIDIA and others). They master manufacturing processes at nanometre scale. Here we are interested in leading-edge foundries, the ones able to go down to the finest nodes, which today are indispensable to new technologies.

It is hard to produce perfectly reliable statistics on this specific segment, most data being global. But the market is undeniably dominated by three players, TSMC, Samsung and Intel, with TSMC clearly above the rest.

A genuine geopolitical subject for eight years now, with the rise of crypto, then Covid, and now AI, the chip market is under real global strain. And there is a genuine chokepoint around advanced-node chips, located mainly in Taiwan. Taiwan actually uses this industrial asset as a defence against China.

• ▸the United States is seeking to reduce its dependence on TSMC by building production capacity on its own soil,
• ▸Germany has launched a foundry with a €10 billion investment (5 from TSMC, 5 from public funding), due to open in 2027,
• ▸Japan has created Rapidus, backed by the government and eight large Japanese groups (Toyota, NTT, Sony, NEC, SoftBank, Denso, Kioxia and MUFG Bank), to put the country back in the race for advanced semiconductors. And which buys its equipment from… ASML!

The chip architects

A chip architect (or integrated circuit architect) is the one who designs and plans the overall structure of the semiconductor. They do not manufacture it: they draw it, then hand production over to a foundry.

CPU architects

I have focused here on the server market, but bear in mind that there are different markets for mobile and for personal computers. On servers, leadership is shared between Intel (around 58%) and AMD (around 38%), far ahead of Apple and Qualcomm.

GPU architects

In this market, there is no possible doubt about NVIDIA's dominance.

But why NVIDIA, actually?

80% of the market is crushing. And the answer that comes to mind spontaneously is "because they make the best chips". That is true. It is not enough. AMD knows how to make very good chips, Intel has colossal resources, Google has been designing its own TPUs for a decade. If hardware alone were enough, this market would be far more contested.

NVIDIA's real advantage is not in the silicon. It is in the software. And it has a name: CUDA.

The fifteen-year bet

Back to the early 2000s. At that time, a GPU knows how to do one thing: display pixels. A researcher who wanted to divert that computing power to something else, simulating a fluid for instance, had to disguise their computation as an image. Literally: turn their equations into textures and triangles so the card would agree to process them, then read the results back from the output pixels. It was possible. It was horrible. Almost nobody did it.

In 2006, NVIDIA introduced CUDA (Compute Unified Device Architecture). The principle: write almost normal code, in C, that runs directly on the GPU. No more disguise; the graphics card becomes a general-purpose compute accelerator.

But the real stroke of genius was not technical, it was commercial: NVIDIA made CUDA available on all its cards. Including the €200 GeForce of the casual gamer, not just on wildly expensive professional hardware. Overnight, any student, in any lab in the world, already owned a machine at home capable of massively parallel computation. It cost a great deal, for a very long time: for years, analysts asked Jensen Huang why his gaming graphics card company was burning so much money on software research, for a market that did not exist.

2012: the market exists

That year, three researchers, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, blew apart the ImageNet image recognition competition with a neural network called AlexNet. Trained on two GeForce GTX 580 cards. Gamer cards. With CUDA.

That is the big bang of modern deep learning, and it happened on NVIDIA hardware, with NVIDIA tools, because NVIDIA was the only one to have put those tools in researchers' hands six years earlier.

The moat

Since then, the company has never stopped stacking. On top of CUDA: cuDNN (the computational building blocks of neural networks), NCCL (to make thousands of GPUs talk to each other), TensorRT (to speed up model execution). And above all that, PyTorch, the framework used by almost every AI researcher and engineer on the planet, and which, when you ask it for a matrix multiplication, ends up calling NVIDIA code.

This is what is called a moat: nearly twenty years of software, millions of lines of code, hundreds of thousands of trained developers, documentation, tutorials, and answers on Stack Overflow at three in the morning. A competitor can release a faster, cheaper chip tomorrow: if your code takes six months to port and then runs 30% slower for lack of optimisation, you will not buy it. That is exactly the wall AMD has been hitting for years, its equivalent, ROCm, being regularly described as functional but painful.

And NVIDIA has widened the moat far beyond the chip. By acquiring Mellanox around 2020, it bought the network that links servers together. Today it no longer really sells a card: it sells an entire rack, GPUs, network, software and tools included. The lock-in is no longer at component level, it is at system level.

The breach

Some good news all the same, and it is essential for what follows: the moat is being filled from the top. Today a researcher almost never writes CUDA by hand, they write PyTorch. And PyTorch is an abstraction layer: if someone makes the effort to wire it correctly onto other hardware, the user's code does not change. It is an enormous, thankless and invisible job. But it is a finite job, whereas catching up on twenty years of ecosystem by hand would be infinite.

That is the breach through which new entrants pass. And it is also the breach through which China passed.

The rise of new chips

The big tech firms are now launching their own chips, above all for AI: Meta with its MTIA, or AWS with its Trainium chips. It is worth remembering that GPUs were not originally designed to do AI. One player that stood out very early is Groq, designer of LPUs to optimise LLM usage. To gauge what is at stake with this type of chip: Groq raised $640 million from BlackRock, Cisco and Samsung, among others.

Broadly, the giants everyone is trying to compete with are: TSMC for foundries, NVIDIA for GPUs and Intel for CPUs.

The rise of Chinese chips

If you want to see what a semiconductor sovereignty strategy actually produces, there is an open-air laboratory: China. Except that it did not really choose to enter it.

From October 2022, the United States banned exports to China of the highest-performing AI chips, NVIDIA's A100 and H100. And every time NVIDIA designed a throttled version to stay within the rules (A800, H800, then H20), Washington tightened the dial. The stated objective: to keep several years of lead on compute.

The result is more ambiguous. Deprived of a supplier, Chinese companies did not stop doing AI: they were forced to buy Chinese. The sanctions therefore handed the domestic industry exactly what Europe most cruelly lacks: a captive, enormous and solvent domestic market.

Etching without ASML

Remember what we said about ASML: without an EUV lithography machine, no advanced node. China does not have access to them, the Netherlands having blocked those exports under American pressure.

It produced 7 nm anyway. The foundry SMIC manages it by multiplying passes with the older DUV machines: instead of etching the fine pattern in one go, it is broken down into several successive exposures. It works. It is slower, more expensive, and above all far less reliable. Yields on Huawei's Ascend chips were long estimated at between 20 and 40%, where a leading-edge foundry runs at 80-90%. Translation: out of ten chips etched, six to eight go in the bin.

So yes, you can do without ASML. But you pay for it in money, in capacity and in time, which incidentally confirms just how much this European company is a strategic asset we underestimate.

The brute force answer

Since a Chinese chip is worth less than an NVIDIA chip, Huawei made a deliberate choice: put in far more of them. Its CloudMatrix system, presented in 2025, packs 384 Ascend accelerators into a rack, where the NVIDIA equivalent lines up 72. The result: the Chinese system delivers more raw power than its American competitor. It also consumes roughly four times more electricity. In Europe that would be prohibitive and would make no sense, since ASML is European. In China, it is an industrial and political will that gives them this kind of latitude; for them, it is a perfectly rational trade-off.

The real bottleneck: memory

A detail that matters and is almost always forgotten: what limits Chinese AI chip production today is no longer etching, it is HBM (High Bandwidth Memory), the ultra-fast memory stacked right next to the processor, without which an AI accelerator is an engine with no fuel. SMIC would reportedly be able to output enough for more than a million chips a year, while Chinese HBM capacity would cap production at a few hundred thousand. It too has been under embargo since late 2024, and is dominated by a Korean-American trio: SK hynix, Samsung and Micron.

And the CUDA moat?

Huawei identified precisely the problem I described above. Facing CUDA, it built CANN, its own software layer, and MindSpore, its equivalent of PyTorch. And above all, it announced in 2025 that it was opening the CANN source code: when you are behind on an ecosystem, open source is the only known lever for attracting developers you cannot pay.

The result showed in April 2026, with DeepSeek V4, post-trained on Huawei Ascend chips, via CANN. The nuance matters: pre-training, by far the heaviest phase, is not involved. But the signal remains considerable: an entire stage in the life cycle of a leading-edge model took place outside the NVIDIA ecosystem. The software moat, reputed to be uncrossable, has just been breached. So there is no inevitability, only a cost of entry.

The turnaround

Meanwhile, NVIDIA's market share in China has collapsed. The H20 saga sums up the absurdity of the situation on its own: a chip throttled to comply with American rules, banned in April 2025 (nearly $4.5 billion written off at NVIDIA), re-authorised in July, then shunned by Beijing, which quietly advised its companies against buying it on security grounds. It was never really sold.

Since January 2026, Washington has opened the door again with a case-by-case review of the H200, while Congress is trying to lock it down instead. But the notable fact lies elsewhere: it is no longer only America refusing to sell, it is now China starting to refuse to buy. And Huawei is not alone: Cambricon, Biren, Moore Threads, Hygon, not counting the in-house chips of Alibaba and Baidu. The ecosystem now exists from top to bottom of the chain.

What this tells us

• ▸The technological barrier can be crossed. Not for free, not quickly, but crossed, including the software barrier, long considered the real lock.
• ▸What made this possible comes down to three ingredients, and none of them is genius: a captive domestic market, state-industry coordination sustained over a decade, and an external constraint that made the status quo impossible. Europe has the third, at least potentially. It lacks the first two.
• ▸The American lead is real, but it is no longer a rent. It rests on a software ecosystem. And a software ecosystem can be worked around.

So if China manages it under embargo, the real question is no longer "is Europe capable of it?", but "what is stopping us?".

What about French and European industry?

Broadly, France, but above all Europe, has the foundations for every stage I have just described, but there is work to do… a lot of work. Setting ASML aside, we have designers of specialised chips such as Kalray, manufacturers such as STMicroelectronics (even though its head office is in Switzerland), materials suppliers such as Soitec, and even research centres such as CEA-Leti in Grenoble (see the Yvan Casta video shared above).

• ▸STMicroelectronics: a Franco-Italian manufacturer specialising in chips for automotive, industrial electronics and IoT. It produces on mature technologies (28 nm, 40 nm) at Crolles, near Grenoble, on 300 mm wafers. That said, we remain a long way from what AI demands.
• ▸Soitec: based in Bernin, a supplier of silicon-on-insulator (SOI) substrates, used in smartphones, 5G and automotive. It works with TSMC, Samsung and GlobalFoundries.
• ▸CEA-Leti: an applied research centre in Grenoble, known for its work on 3D transistors, EUV lithography and heterogeneous integration.
• ▸Kalray: a designer of processors for critical applications, with its MPPA (Massively Parallel Processor Array). Like NVIDIA, Kalray does not manufacture its chips.

Broadly, our expertise does not go as far as that of our American and Chinese counterparts, who had to learn at breakneck speed. But that is precisely why we still have cards to play.

Some companies deserve our attention, for example:

• ▸VSora: they design hardware accelerators dedicated to AI inference in data centres. That makes them far more directly comparable to NVIDIA or AMD than SiPearl. Very recent news: in July 2026, Ardian Semiconductor took a minority stake in VSORA to accelerate its commercial strategy.
• ▸SiPearl: a chip designer, of course, with its RHEA1 — a high-performance CPU, not a GPU (and manufactured by TSMC).
• ▸Axelera AI and Innatera: two interesting Dutch companies. You can feel the proximity to ASML.

We should not forget another major European player: ARM. This British company specialises in designing processor architectures, but it was bought in 2016 by SoftBank (Japanese) for $32 billion. Unlike Intel or AMD, ARM does not produce hardware: it designs and licenses. Its architecture is today mainly used in mobile formats, which explains the presence of Samsung, Qualcomm and Apple among its customers. If you have a Mac with an M1 or M2 chip, congratulations: you are on an ARM architecture!

To gauge its strategic importance: the acquisition of ARM by NVIDIA, for $66 billion, was blocked by regulators. And the energy stakes around chips mean NVIDIA nevertheless plans to move to this architecture for its PCs.

What should we conclude?

I am going to give you my point of view with the information I have to hand. The purpose of this article is above all, for me, to set down what has been running around my head for many years. But I invite you to react and to bring information, so that we can dig into the subject together :)

We have seen that it takes a multitude of players to produce a chip: the architect, the foundry, sometimes licences. Chips are everywhere, but one of the strategic issues likely to grow most in importance over the coming years is that of chips dedicated to artificial intelligence. The question to ask is therefore: what long-term strategy should we put in place for our future strategic needs?

At first glance, three strategies look "obvious":

• 01The German strategy, bringing foreign companies onto your soil, with TSMC. It lets you benefit from the knowledge of existing companies, but it is the model most fragile to third-party decisions, and it leaves us dependent on licensing questions. A caveat though: today it seems we are talking about 12 nanometre nodes at the finest, which remains far from the 2 nm we discussed earlier.
• 02The Japanese strategy, creating a structure from scratch on your soil, with Rapidus. The advantage: a national champion. The disadvantage: the price, the time and the risk, for a start-up only in 2027, and in a country that nonetheless has a strong industrial history.
• 03The strategy of scaling up an existing player, via national or European support. It has merit and remains possible, but European mechanisms require precisely justifying the strategic interest of the investment, and above all they make it difficult to concentrate resources on a single national player. And one candidate I saw as possible is no longer based in an EU country, which narrows the options further.

But before choosing a strategy, we may need to define what we mean by sovereignty.

This is an important distinction. A sovereign Europe does not necessarily need to manufacture every wafer on its territory, any more than it needs to build every machine used in a foundry on its own. What it does need is to know how to identify the dependencies whose disruption could endanger its industry, its defence or its critical infrastructure, and to have a credible alternative when the risk becomes too great.

It is from this definition that I would like to take a step sideways.

Taking a step sideways

As those who know me will suspect, if I am telling you about these three "obvious" strategies, it is to tell you about the one that seems to me the most intelligent: taking a step sideways from the Japanese strategy. Let us start again from a few observations.

• ▸Europe has the intellectual capacity. Without any chauvinism, there are many European engineers in cutting-edge industries, and we suffer particularly from brain drain. Let us see that not as a fate but as an opportunity, as the founders of Mistral or Kyutai perceived. In the chip field, let us recall that ARM's Cortex-A9, which powered the very first iPhones, was designed in Sophia-Antipolis.
• ▸We already know how to produce certain types of chip thanks to the companies present on our territory. The main challenge concerns the strategic technologies of tomorrow, whose flagship is AI.
• ▸Etching fineness is only one variable among those that determine a chip's performance.

« Think of AI as a screw you want to drive in to assemble your IKEA furniture. Your GPU is a hammer whose performance we keep trying to improve, when what you want is a simple screwdriver. »

That is the bet of the companies of the moment, heavily marketed by Groq at its creation: "a specialised architecture can offset part of the disadvantage of a less advanced etching node", much like the interest in the ARM architecture.

So you can see where I am going. Honestly, I would have liked Groq to have gone into producing its own chips at some point, and for us to have sought to produce them in Europe. The fact that this did not happen is in itself a reason to explore what I think is a good idea: making a new Mistral for AI semiconductors.

So do not try to make NVIDIA-style GPUs, not right away. Genuinely aim for architectures specialised for AI, whether LPUs, NPUs or other dedicated architectures. And that for two very pragmatic reasons: the need (we do not want to run video games, but to build the AI of the future) and the cost (aiming at a precise target will reduce a bill that will already be heavy enough).

One clarification that follows directly from everything we have just seen: such a company would not be a silicon project, it would be a software project. Three levers seem decisive to me.

Lever 1: the product is not the chip

« The product must not be the chip. The product must be the platform. »

The chip is only the hardware on which the following runs:

• ▸the compiler;
• ▸the kernels;
• ▸the runtimes;
• ▸the libraries;
• ▸the drivers;
• ▸PyTorch;
• ▸vLLM;
• ▸Triton;
• ▸the profiling tools;
• ▸the agent frameworks;
• ▸the cloud tooling, and so on.

One of the major reasons for AMD's lag behind NVIDIA is not only the silicon: it is the considerable lead NVIDIA took on the software ecosystem, notably with CUDA. Huawei's breakthrough is not explained only by its chips, but by the fact that it built, then opened, its own software layer. A European player devoting 90% of its budget to hardware and 10% to compilers, libraries and PyTorch integration would reproduce exactly the mistake others have already paid for. The good news is that Europe has a real tradition in that area. And that it costs infinitely less than a foundry.

That is exactly the problem AMD is trying to solve today with ROCm, which bundles runtimes, compilers, libraries, tools and integration with PyTorch, JAX, vLLM or SGLang. And Huawei is doing something similar around Ascend, CANN and the open source ecosystem.

Lever 2: you need a first customer

Less technical, but just as decisive. None of these projects started on market demand alone. Huawei benefits from an immense Chinese market and an industrial policy reinforcing preference for domestic solutions, while Rapidus benefits from direct commitment by the state and major Japanese industrial groups. American foundries, for their part, have public procurement and defence.

A European AI chip does not need to convince the whole world on day one: it needs the continent's administrations, research bodies and sovereign clouds to commit to buying it. It sounds complex, and yet we already see states increasingly seeking to migrate from Microsoft to open source solutions based on Linux. It is probably the most effective, and least costly, political lever we have.

« A share of the computing capacity funded by European public money must be available on European accelerators. »

Lever 3: what we are still missing

On our territory we have cloud providers positioning themselves on AI (Scaleway, OVH), and an LLM champion in France (Mistral) which, understanding the stakes, has also moved into the data centre side, hence the well-known contract with Microsoft. What we are missing, then, are players to complete the picture: a company able to architect LPUs and NPUs and to master their software ecosystem, along with European capacity able to manufacture, assemble and integrate these architectures when it is strategically necessary.

And we should not reduce this ambition to the wafer alone. For modern AI accelerators, packaging, interconnects and memory have become decisive elements of performance. Chiplet-based architectures precisely allow certain functions to be decoupled and several components assembled within a single package. European sovereignty can therefore also play out here: we may not need to master every etching generation immediately if we master the architecture, the packaging, the interconnect and the software that give the final system its value.

A common thread: energy coherence

One last aspect runs through these three levers, glimpsed with ARM: energy. In Europe we need a coherent strategy between our industry and our regulatory objectives. The European Union is launching several regulations on the environmental and societal impact of companies and their suppliers (CSRD and CS3D). The paradox is that most companies are equipping themselves with artificial intelligence tools to improve their design, or even to comply with these regulations, while those very tools are themselves highly energy-hungry. Creating a player whose baseline standard is the pursuit of a good performance / consumption ratio and compliance with European social standards would be a political message of coherence.

Given the complexity of this industry, political will and work with private players are mandatory, as the investments around Rapidus in Japan prove. This industry could help reindustrialise the European continent while securing our sovereignty on the subject.

Appendices

What follows is not necessary to understand the article: it is the level of detail below, with the sources, for those who want to dig. The three appendices revisit and deepen passages already skimmed over above.

Appendix 1 — The composition of a chip

Silicon. It is the base material of the transistors that make up the integrated circuits of CPUs and GPUs. It is a semiconductor: depending on conditions, it conducts or insulates current, which is precisely what you need to build a transistor. It is purified to "electronic grade" (99.9999999% purity), then turned into wafers, the discs on which circuits are etched at nanometre scale.

Metals, for conductivity. They carry the current through the circuits and interconnects.

Copper (Cu) is used for interconnects between the layers of the integrated circuit: high conductivity and good heat dissipation.

Aluminium (Al) was once used for interconnects; it has been replaced by copper in modern processors, but is still used in certain components.

Gold (Au) is reserved for critical connections, such as the very fine wires linking the chip to its package (bond wires). It is highly corrosion-resistant.

Tungsten (W) goes into certain interconnect layers, for its mechanical strength and resistance to high temperatures. Material factsheet, MineralInfo

Nickel (Ni), cobalt (Co) and molybdenum (Mo) go into specific alloys or protective layers of transistors. Halt to nickel ore exports (French Treasury) · Molybdenum (L'Élémentarium)

Dopants. They modify the properties of silicon to turn it into a transistor. Phosphorus (P) and arsenic (As) are N-type dopants: they add free electrons that conduct electricity. Arsenic production (Statista)

Boron (B) is a P-type dopant: it creates "holes" that ease conduction through the absence of electrons. Main boron producing countries (Statista)

Gallium arsenide (GaAs) is used in specialised components, for better speed and efficiency. Material factsheet (Metoree)

Insulators. They separate the layers and parts of the circuit. Silicon dioxide (SiO₂) is the main insulator of integrated circuits: it protects the transistors and prevents short circuits. Hafnium oxide, used in modern processors, improves transistor performance while reducing power draw. Material factsheet (Metoree)

Heat dissipation. Without it, the chip overheats. Silicon carbide (SiC) is used for its superior thermal and electronic properties in certain components. Material factsheet (L'Élémentarium)

Graphite and graphene are used in certain chips to evacuate heat efficiently; ceramics, in substrates and packages, to insulate and evacuate heat. Main graphite producing countries (Statista) · special mention: French companies soon to be European graphene champions (Les Échos)

Substrates and packaging. The chip is mounted on a substrate, then protected: glass-fibre reinforced epoxy (FR-4) to carry the chip, copper and aluminium alloys for packages and heat sinks, plastics and polymers to insulate and encapsulate sensitive components.

Appendix 2 — Advanced nodes in detail

Transistor density. Advanced nodes allow more transistors to be integrated into a given area. More transistors means more functional units (compute cores, parallel processing units, embedded memory) and therefore better performance — or, for an equal transistor count, a smaller chip, which benefits compact devices. In exchange, the smaller the transistors, the more sensitive they become to physical phenomena such as leakage current and variability in electrical characteristics. And the increased density complicates heat dissipation.

Performance and consumption. Smaller transistors shorten the electrical paths, which improves clock speed; design advances (FinFET, GAA) allow better current management and reduced latency. On the energy side, the supply voltage drops, and so does consumption — essential for mobile devices, data centres and embedded systems. The downside: this voltage drop reduces the operating margin of the transistors, making chips more sensitive to environmental variation (heat, electrical noise).

New architectures. Advanced nodes open the way to designs that were previously out of reach:

• ▸3D stacking: vertical integration of circuits, such as chiplet-based architectures or 3D V-Cache.
• ▸Memory integration: more memory directly on the chip, as in modern GPUs.
• ▸AI and heterogeneous computing: hybrid architectures (CPU + GPU + AI accelerators) take advantage of advanced nodes to maximise performance.

Appendix 3 — French and European players

• ▸STMicroelectronics — semiconductor manufacturer, Franco-Italian, headquartered in Geneva but with a strong presence in France. One of the European leaders in chip design and manufacturing, specialising in automotive, industrial electronics, IoT and integrated solutions. It produces on mature technologies (28 nm, 40 nm) suited to those sectors, notably at Crolles, near Grenoble, on 300 mm wafers.
• ▸Soitec — materials supplier, based in Bernin, near Grenoble. Specialising in silicon-on-insulator (SOI) substrates and other advanced materials, used for smartphones, 5G, automotive and embedded systems. Its substrates improve chip performance and reduce power draw. It works with TSMC, Samsung and GlobalFoundries.
• ▸CEA-Leti — applied research centre, in Grenoble. Specialising in semiconductor innovation, known for its work on 3D transistors, EUV lithography and heterogeneous integration. It works in partnership with STMicroelectronics, Soitec and international players, and accelerates the innovation that keeps European industry competitive.
• ▸Kalray — designer of specialised processors, in Montbonnot-Saint-Martin, near Grenoble. Designs processors for critical and high-performance applications: the MPPA (Massively Parallel Processor Array), a DPU (Data Processing Unit) used in autonomous systems, data centres and artificial intelligence. Like NVIDIA, Kalray does not manufacture its chips.