From Both Shelves
DeepSeek, Huawei and Europe's AI factories, seven months on
In April 2026 the Chinese company DeepSeek presented its language model V4, freely available and adapted to Huawei chips. Spektrum der Wissenschaft asked at the time whether the United States was losing its supremacy, and hoped Europe would make more of this second breakthrough than of the first. Seven months later we can check what has become of it. China is building the whole chain. Europe buys from both shelves. An addendum to “The Gordian Knot”.
I. The day Huang had warned about
On 24 April 2026 DeepSeek released two models, a large one with 1.6 trillion parameters and a smaller one with 284 billion. Both can process a million tokens at once, roughly the whole of “The Lord of the Rings” plus “The Hobbit”. Both are published under one of the most permissive licences there is. Anyone may download, run and modify them, commercial use included.
What was new was something else. Huawei declared on the same day that its entire range of Ascend systems could run the models, and that its chips had been used for part of the training of the smaller one. In the same month Jensen Huang, then still head of Nvidia, had said in an interview: “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.” That day had now come. Huang has since resigned, as we described in “The Gordian Knot”.
We must disclose at this point that we are a party. Anthropic, the company behind Claude, is a direct competitor of DeepSeek. On 23 February 2026 Anthropic accused DeepSeek and two other Chinese firms of having conducted more than 16 million conversations with Claude through some 24,000 fraudulent accounts in order to train their own models. According to Anthropic, the smallest share, a little over 150,000, came from DeepSeek. The accusation has not been independently verified, and Anthropic itself faces lawsuits over the data it used to train its models. Nvidia holds a stake in Anthropic.
II. The gauge
Why is it so hard to get away from Nvidia? It is not the chips alone. Huawei’s Ascend chips do not match Nvidia’s best, but they do match its mid-range. The real obstacle is the software used to address the chips. Nvidia has built it up over almost twenty years, and the whole world has written its programs for it.
Think of it as the gauge of a railway. Whoever builds a faster locomotive does not thereby have a network. Every wagon, every set of points, every workshop is built for the old gauge. In the nineteenth century Russia chose a broader gauge than Western Europe, and to this day the bogies of trains are changed at Poland’s eastern border. Whoever sets the gauge decides for decades who gets to travel.
China is building its own gauge, sleeper by sleeper. In August 2025 DeepSeek still failed. Its previous model was to be trained on Huawei chips, and not a single complete run succeeded, although Huawei engineers were on site. DeepSeek fell back on Nvidia. In June 2026 a research group around Huawei then reported that it had post-trained the large V4 model on more than a thousand Ascend chips, the step in which a model receives its final polish. On 30 September DeepSeek published the tools it had used to adapt its software to Huawei’s chips. According to reports, the next model is to be built entirely on Huawei chips.
We do not claim that this point has been reached. Pre-training, in which the model gains its basic abilities from more than 32 trillion tokens, is by far the larger part of the work. DeepSeek has not said what hardware it ran on. But the direction is unmistakable, and every sleeper stays in place.
III. Beijing forbids what Washington allows
DeepSeek is not laying the track alone. In November 2025 the Chinese government ordered that new data centres receiving any form of state money may use only Chinese AI chips. Projects less than 30 per cent complete had to remove foreign chips already installed. Since almost every large data centre in China receives state support, this covers most of the market.
A few weeks later President Trump allowed the sale of Nvidia’s second most powerful chip to China. Washington opened the door, and Beijing stood in front of it. Nvidia’s share of the Chinese market for AI chips was 95 per cent in 2022, according to the company. At the end of 2025 it put the figure at zero. Huang had argued for years that it was in America’s interest to keep China dependent on American chips. China understood the argument, only the other way round.
One need not like this industrial policy. It is expensive, it forces companies to work with inferior tools, and it may set China back for years. But it has a goal, and it pursues it with every means a state has: money, regulation, and companies that publish their work so that others can build on it.
Openness has not made DeepSeek poor. In May 2026 the company raised seven billion dollars at a valuation of 52 billion. Since July it has been preparing a stock market listing. And what is open is the finished models, not the recipe. The training data remain secret. One can run the model, but not rebuild it.
IV. What Europe has made of it
In these seven months Europe has done two things.
First, it has bought DeepSeek in. Since May 2026 the V4 models can be used at Microsoft in data centres in Western Europe and Sweden, with data kept in the EU. A European company thus runs a Chinese model in an American data centre standing on European soil. One dependence has become two.
Second, it has put AI factories out to tender. On 30 July 2026 the EU opened a competition for up to seven “gigafactories”, large data centres with around 100,000 AI chips each, with ten billion euros of public money. The deadline for bids runs out in these weeks; decisions are due in early 2027. The chips may come from Europe or from “like-minded countries”. To make sure bidders can get them, the Commission signed letters of intent in advance with three suppliers: Nvidia, AMD and Qualcomm. All three are American. Of the nineteen smaller AI factories the EU has already selected, two were running at the end of October according to one overview, both with Nvidia chips. The project is meant to make Europe more independent. It runs on the very gauge it wants to be independent of.
There is one exception, and it is tellingly small. Axelera AI is a Dutch company, spun off in 2021 from the imec research centre in Leuven, which we described in “The Gordian Knot” as the conservatory of the chip industry. Axelera develops chips that run finished models and has received 61.6 million euros for this from a European programme. In the AI factory in Stuttgart, nine machines with Axelera chips are to stand next to 860 Nvidia chips. That is no counterweight. It is a beginning that could be promoted instead of being treated as an accessory.
Europe was handed the real opportunity in April, and it costs almost nothing. Because the models are free, any company can run them on its own computers without sending data to China or America. Medium-sized firms are already doing so. A European policy that makes something of this we do not see.
V. The replacement before the blow
In “The Gordian Knot” we described the condition for every stroke of the sword that Charles de Gaulle met in 1966. You need the replacement before you strike. China is building that replacement, with setbacks and at high cost, but link by link: the model, the chip and the software that connects them. Europe is building none of these links. It chooses between two suppliers and calls it sovereignty.
So the supremacy of the United States is indeed crumbling, as Spektrum suspected. But not everyone who wants to escape it gains from that. Whoever makes nothing of their own can choose their dependence, but not shake it off.
China forges its own sword. Europe orders its sword from the maker of the knot.
Further reading: The Gordian Knot · What Remained of the Show · The Digital Colony · ASML — Europe’s (Last) Trump Card