Putin Promises Digital Sovereignty. The Digital World Thinks in English.

Fact-checkers went straight for the hardware and found what they expected to find. Russia's largest chipmaker, Mikron, advertises a 90-nanometre process, which was state of the art around 2004. The "domestic" Elbrus and Baikal processors were, until 2022, manufactured in Taiwan by TSMC, the same company that has since moved to 2 nanometres. All of this is true and all of it matters. But the argument about nanometres distracts from a question far more awkward for the Kremlin. Suppose, for a moment, that Russia learns tomorrow to print chips as well as the Taiwanese do. Would that make it a digital sovereign? No. Because sovereignty in the digital world does not come down to silicon. It comes down to language. And the language of that world is one, and it is not Russian.

Start with the simplest thing, the one every student sees in the first hour of a programming course. Every language in which modern software is written, from operating systems to banking apps to neural networks, speaks English. Not "mostly", not "predominantly", but entirely. If, else, while, return, import, class, print: these are not technical terms, they are ordinary English words, and without them you cannot write a single line of C, Python, Java or Rust. A coder in Shenzhen and a coder in Novosibirsk type exactly the same thing: for i in range. They may hate America, believe in a special national path, wear a badge with the leader's portrait, but the moment their fingers touch the keyboard they are obliged to think in the language of Silicon Valley. They have no other tool.

The objection writes itself: that is only the top layer, words for humans, while the machine understands ones and zeros. It is a comfortable illusion. Go all the way down to the metal. Every processor executes instructions that have mnemonics, the short names compilers and programmers use: MOV, ADD, JMP, CALL, RET. They are abbreviations of the English words move, add, jump, call, return. That is how Intel works, how ARM works, and how the "open" RISC-V architecture works, the one Moscow and Beijing now pray to as their salvation from Western dependence. RISC-V was designed at the University of California, Berkeley, and its entire specification is written in English. China's Loongson, which Beijing presents as a wholly homegrown architecture, is documented in English and its instructions carry the same English names. Russia's Elbrus has documentation in Russian, but its compiler emits the same Latin mnemonics, and that compiler is assembled from Western components. Below that level there are only transistors, and transistors have no language. So there is no crutch, no intermediate layer anywhere in nature where a Cyrillic letter or a Chinese character means anything to the machine. The Russian keywords in 1C, the country's ubiquitous accounting platform, survive exactly until compilation, after which they become the same bytecode run by a virtual machine written in C++.

Now climb back up from the hardware to what Russia calls its "domestic operating systems". Astra Linux, RED OS, Alt: all of them are builds of Linux, the kernel a Finn named Linus Torvalds began writing in C in 1991 and which thousands of programmers around the world have developed ever since, debating every line in English. China's Kylin and Deepin are built the same way. Even North Korea's Red Star OS is Linux. When Huawei, under sanctions, announced its own operating system, HarmonyOS, and its own programming language, ArkTS, it turned out that ArkTS is built on TypeScript, a Microsoft creation. The internet on which all of this runs rests on protocols whose standards have been published since 1969 in documents called RFCs, Requests for Comments, and they are still published only in English. A website address, even a Cyrillic one, is converted by the machine into Latin characters before it is sent, or the network will not understand it.

Then there is the science underneath it all. In modern computer science, a paper not published in English does not exist. It will not be cited, not replicated, not included in anyone's training data. This applies not only to mathematicians in California but to the Chinese, who today publish more in this field than anyone. The technical reports of DeepSeek and Alibaba, of which Beijing is rightly proud, are written in English and posted on arXiv, an American preprint server. Their code lives on GitHub, which belongs to Microsoft. They were trained on Nvidia GPUs using PyTorch, a framework created by Meta. Russia's GigaChat and YandexGPT are built on the same Nvidia hardware and the same PyTorch, from a company a Russian court has designated an "extremist organisation". Picture it: sovereign Russian artificial intelligence, built on a tool made by extremists.

This is where it gets interesting. One could argue: fine, the tools are Western, but the result is ours, the machine answers in Russian and Chinese, so it must think in Russian and Chinese. It does not. That is no longer a columnist's conjecture but an experimental finding. In 2024 a team at the Swiss Federal Institute of Technology in Lausanne published a paper whose title says it all: "Do Llamas Work in English?" The subject was Meta's Llama model. The researchers asked it questions in French, Chinese and Russian and used a special technique to peer into the network's intermediate layers, where the answer takes shape. They found that halfway through, the model arrives at an English word and only in the final steps translates it into the language of the question. The model's internal "concept space" turned out to lean towards English. The paper won an Outstanding Paper Award at the field's main conference. In 2025 researchers from Oxford and Google DeepMind went further and gave their paper an even blunter title: "Do Multilingual LLMs Think in English?" The answer was yes. The key semantic decisions, the choice of nouns and verbs, are made in a representation closest to English, regardless of the language of the question or the answer. For the model, every other language is an input and output panel, an interface, not a mind.

The strongest admission came from the Chinese themselves. In the technical report on DeepSeek-R1, the model that crashed Nvidia's share price in January 2025 and was hailed as the triumph of Chinese engineering, the authors state plainly that while the model was being trained to reason, its internal chain of thought kept mixing languages, and to force it to think in the user's language they had to introduce a special training penalty for "language inconsistency". They then add that this penalty slightly reduced the quality of the reasoning. Consider that. China's flagship artificial intelligence, left to its own devices, drifts into English. It has to be compelled to think in Chinese, and it pays for that in intelligence. If this is digital sovereignty, it is an expensive one.

Why this happened needs no conspiracy theory. In the dataset used to train GPT-3, English made up about 93 percent of all words and Russian less than a third of one percent. In Common Crawl, the corpus on which nearly every large model is trained in one way or another, Chinese and Russian models included, English accounts for almost half and Russian for about 5 percent. But volume is not really the point. The point is that everything from which a machine assembles its understanding of the world was written in English: all the science, all the code, all the documentation, all the patents, all the textbooks from which the people who wrote the code once learned. English sits inside the neural network not because there is a lot of it, but because everything the network contains was created in it.

Now imagine Russia seriously set out to become what Putin declared it to be. Not in the sense of blocking YouTube and herding officials into the state-sponsored Max messenger, but for real: building a digital universe that does not depend on decisions taken in Washington. What would that take? A programming language of its own, not a wrapper around C but a language in which a compiler, an operating system, a browser and a database have been written from scratch. A processor architecture of its own, with its own instruction set and its own compiler. Network protocols of its own, and an internet of its own for them to run on. A corpus of scientific texts of its own, large enough to train an artificial intelligence that thinks in Russian without being beaten into it. And millions of people to learn all of this, because everyone who can program today can do so in English. In other words, it would take a new Cyril and Methodius to invent a digital alphabet for the Russian world, and then centuries for a literature to appear in it. The brothers from Thessaloniki created their alphabet in the 9th century; the first secular book printed in it appeared more than 700 years later.

History, as it happens, has already run this experiment, and it ran it in Moscow. In the 1960s the Soviet Union had computers of its own; Sergei Lebedev's BESM-6 was considered one of the best machines in the world. In 1967 the Soviet government decided not to develop its own line but to copy IBM's System/360, and the ES EVM series was born. They copied the architecture, the instruction set and the English mnemonics along with it. The native school was wound down, and for 20 years the country lived by reproducing someone else's designs, several years behind the original. It ended, as everyone knows, with most Soviet citizens seeing their first personal computer as an import. That is the answer to what happens if Russia tries to "sovereignise" itself in earnest: it will not create an alternative, it will simply drop out of the environment where everything new is made, and chase it with clones. Not for 20 years, as last time, because then the gap was a few years, whereas now it is measured in decades and trillions of dollars. A hundred years, if one counts honestly.

But that is not even the main point. English did not become the language of the digital world because it was imposed. Nobody imposed it: Linux is free, the internet's standards are open, anyone can read arXiv, PyTorch can be downloaded in Shenzhen and in Skolkovo alike. English won because the environment it lives in runs on one rule for everyone: anyone may come, take, and make it better. That is how nations win when the law is the same for all, and how they lose when the law is one thing for insiders and another for everyone else. Digital sovereignty as Putin understands it means the exact opposite: shut the door and announce that everything necessary is already inside. Inside are Mikron with its 90 nanometres and 4 banned messengers in which the ministries still sit. Outside, the whole world, China included, goes on thinking in English, and the smartest machines humanity has ever built think in it too. So who, exactly, has sovereignised whom?