Reading note. This article is based on a research announcement attributed to IBM regarding a sub-nanometer process, sometimes referred to as NanoStack, said to pack on the order of 100 billion transistors onto an area comparable to a fingernail. These figures are reported and, at the time of writing, have not been confirmed by independent measurements on mass-produced silicon. We treat them as a directional signal, not an established fact. A lab demonstration is not mass production.
In one sentence
While the public talks about prompts and chatbots, the decisive battle of AI is being fought in matter: a few layers of atoms, etched and stacked. IBM's announcement, if confirmed, does not merely say "we can etch finer"; it is a reminder that the next step of artificial intelligence will depend less on the models than on the machines that run them — and that whoever masters silicon masters a share of the AI of the next ten years.
1. What is being announced, and why now
For decades, the computing industry has advanced on a simple idea: pack ever more transistors into an ever smaller space. This is the spirit of Moore's law — Gordon Moore's 1965 observation that the number of transistors per chip doubles at regular intervals. That logic has gradually run into a physical limit: miniaturize far enough and you reach the scale of the atom, where you can no longer simply "shrink" as before.
The approach attributed to IBM is interesting precisely because it changes method rather than degree. Instead of thinking only in terms of surface, the process would introduce a vertical logic: transistors are no longer merely aligned on a plane, they are stacked. The industry moves from a flat-ground logic to a high-rise logic — when there is no more room at ground level, you build upward.
The shift looks technical; it is in fact structural. Miniaturization no longer depends only on etching fineness, but on architecture: how components are organized, connected and optimized relative to one another.
2. Two words of vocabulary
Before going further, two terms are worth setting down, because the rest follows from them.
Transistor — The smallest switch on a chip: it lets current through or blocks it, and by combining billions of these "yes/no" decisions a processor computes. The more you fit, the more you can compute.
3D stacking — Rather than spreading components across a single layer, you superimpose several floors of circuits connected vertically. You gain density without etching finer, and you bring elements closer together, which shortens the distance information has to travel. This idea — building upward — is at the heart of the announcement.
3. Why it matters for artificial intelligence
Modern AI is hungry. It consumes data, energy, GPUs, memory, cooling, data centers and billions in investment. Behind every conversational assistant, every generated image, every multimodal model lies a heavy, costly, energy-intensive material chain.
If the promise holds, this kind of chip could shift the equation. More transistors in a smaller space means potentially more compute, better energy efficiency, and a lower cost per operation. In other words: faster models, cheaper to run, less power-hungry.
The subject then goes beyond a mere product announcement. The question is not only whether IBM can beat Intel, TSMC or Samsung on an etching process. The real question is: who will control the material infrastructure of AI for the next ten years? Because AI is not a magic cloud. It rests on silicon, factories, lithography machines, supply chains, patents, engineers and geopolitical choices.
"The AI of the future may be smarter. But above all, it will have to be more frugal, faster, cheaper, and closer to the ground."
4. Leaving rootless AI behind
For two years, part of the market has sold AI as a matter of software, prompts, instant productivity and a "revolution" reachable in three webinars. It is comfortable, but incomplete. AI is not only an interface; it is a heavy industry.
This announcement is a reminder of something often forgotten: digital ruptures always have a material base. The internet did not explode because websites became prettier. It took cables, servers, routers, processors, smartphones, fiber and mobile networks. In the same way, AI will only become massive, durable and accessible if its cost of execution collapses — and that collapse will come largely from hardware. It is a movement we have been tracking for a while.
More powerful, more frugal chips open the way to a more distributed, more local, more embedded AI. An AI that does not always depend on a giant data center on the other side of the world, and that can run in devices, vehicles, industrial equipment, schools, local authorities or critical infrastructure.
This is where the subject becomes political.
5. Chips, sovereignty and strategic dependence
Semiconductors have become a strategic raw material, much as oil was in the 20th century. Whoever masters chips holds a share of the digital economy, of defense, health, mobility, education, finance and AI.
The announcement thus belongs to a worldwide industrial contest, where each player has its hand to play:
- United States — Want to keep their lead, anchoring their strategy on champions such as IBM, Intel or Nvidia.
- Taiwan — Remains central with TSMC, the world's leading foundry, and the nerve center — therefore the most exposed point — of the whole chain.
- South Korea — Plays its card with Samsung in advanced etching and memory.
- Europe — Tries to reduce its dependence, with means still far below those of the leaders.
- Japan and China — The former is returning to the race; the latter accelerates despite export restrictions.
In this context, a chip able to go below the nanometer is not only a scientific innovation: it is a strategic asset. Digital sovereignty is not limited to hosting your data "at home" or using open-source software. It begins much lower down — at the capacity to produce, buy, understand and integrate the components that will run tomorrow's critical systems.
6. Signals to watch
Keeping a cool head remains essential. Between a research demonstration and a chip available at scale lies a chasm: manufacturing yield, cost, reliability, integration into existing chains, machine availability, industrial partners, packaging, memory, cooling. A few concrete indicators will separate the announcement from reality:
- Independent confirmation — The claimed figures (density, transistor count) must be measured by third parties on real silicon. Until then, caution.
- Manufacturing yield — A lab feat only matters industrially if a sufficient share of produced chips are functional. This is often where promises stumble.
- The time horizon — IBM reportedly speaks of several years. This technology will not reach our computers tomorrow; the announced pace will tell whether it is a trajectory or a flash.
- Production partners — IBM designs, but who will manufacture at volume? The names of associated foundries will say much about the seriousness of the schedule.
- The real energy gain — Density is not enough. The true test, for AI, is the cost per operation and consumption per watt.
The future of chips will probably not be a mere continuation of the past. It will be made of 3D stacking, new materials, chiplets, memory closer to compute, specialized architectures and ever finer energy trade-offs.
7. A situated word
We write from Réunion Island, 9,000 km from the clean rooms where this race is decided. From here, IBM's announcement is not first of all a matter of stacked transistors.
What interests us is the possible relocation of compute. For an island territory dependent on submarine links to reach the major data centers, the idea that a useful AI might one day run on site, offline, with no meter ticking, is not a comfort detail: it is a matter of sovereignty and frugality. More frugal chips mean the prospect of a frugal lab that ceases to be at the mercy of the latency and inference prices of a distant provider.
While some sell AI as a marketing layer laid over slides, the announcement is a reminder that the revolution is still manufactured in laboratories, clean rooms and production lines. Tomorrow's battle will not only be "who has the best chatbot?", but "who owns the chips, the energy, the factories, the talent and the infrastructure to run AI at scale?".
The real fight is not fought only in the cloud. It is fought in the atom.
Sources and further reading
- IBM Research — Public communications on the semiconductor roadmap (nanosheets, 3D stacking, advanced nodes). Primary source to weigh against reported announcements.
- Gordon Moore (1965) — Founding article setting out "Moore's law", the historical framing of the density/cost trade-off.
- RTX Spark: Nvidia and Microsoft's "Apple Silicon" moment — Our analysis on relocating compute to the edge of the network.
- The collapse of the cost of the token — Our article on the economic dynamic that makes AI massive.
This document is updated if new elements appear. Last revision: 原 June 25, 2026.