The recap: what we had left hanging
In our article of 25 April on DeepSeek V4, we showed that China had just proved it could build a frontier AI model on Huawei chips, without Nvidia, and distribute it free of charge around the world. That release was not just a technical announcement. It was a geopolitical statement.
Three weeks later, the consequences are taking shape. The United States and China have each made major strategic moves that are drawing the new map of global AI power. And the situation is more complex (and more worrying) than one might think.
The dramatic reversal in American policy
On 13 January 2026, the US Department of Commerce announced a radical change: Nvidia H200 and AMD MI325X chips may now be sold to China, under case-by-case licensing. This is the exact opposite of what Joe Biden was doing, who over four years had built a "technology lock" strategy to prevent China from accessing the most powerful chips.
In February 2026, Beijing approved the purchase of 400,000 H200 chips by major Chinese AI companies. The United States takes 25% of the revenue from each sale, amounting to several billion dollars. Donald Trump announced it on Truth Social as a commercial victory. Many American experts, by contrast, see it as a major strategic failure.
To understand the scale of the change, here is the context: the H200 is roughly 6 times more powerful than the best Nvidia chip China could buy under Biden (the H20). And it is better than anything Huawei plans to produce for at least two years. The Council on Foreign Relations has calculated that exporting 3 million H200s to China in 2026 would give Beijing more AI compute power than it could produce domestically before 2028-2029. It is an invaluable strategic gift.
The embarrassing secret: Huawei is going backwards, not forwards
Here is the statistic that changes the whole conversation. According to the official roadmap published by Huawei in September 2025, Huawei's next chips (Ascend 950PR and 950DT, due in 2026) will be less powerful than its current best chip (Ascend 910C).
A regression. Not a stagnation. A regression. That never happens in the chip industry. It probably means Huawei cannot produce in sufficient quantity with its Chinese foundry SMIC, which remains stuck at the 7-nanometre node because of American sanctions on manufacturing equipment. Several experts even suspect that the majority of current Ascend 910C chips were produced illegally at TSMC in Taiwan, and not in China.
| Metric | Nvidia (USA) | Huawei (China) |
|---|---|---|
| Performance gap per chip (2026) | 5x more powerful | Baseline |
| Projected gap in 2027 | 17x more powerful | The gap is widening |
| Estimated annual production (2025) | ~10 million chips | 200,000 chips (2% of Nvidia) |
| First Huawei chip more powerful than H200 | Baseline | Not before late 2027 (Ascend 960) |
This technical reality changes the whole analysis. China is not "catching up" with the United States on chips. It is falling behind at an accelerating rate. That is why Beijing enthusiastically accepted the sale of the H200s: it needs them in order not to drop out of the race entirely.
Chinese ingenuity: compensating for weak hardware with software mastery
But China has not said its last word. Rather than trying to compete chip for chip with Nvidia, it has chosen a different approach: compensating for its hardware lag with software and architectural superiority.
The most striking example is DeepSeek V4. As we showed in our article of 25 April, V4 reaches 80.6% on SWE-Bench (the benchmark test for AI programming) using 27% of the computing power and 10% of the memory of its predecessor. That is pure engineering. Where the Americans compensate with raw power (more chips, more data centres), the Chinese compensate with algorithmic elegance.
The result is paradoxical: the best Chinese AI models are trained on inferior chips, but they are more efficient in final use. For an end user who just wants AI that works, the quality gap is minimal. For Beijing, that is exactly what is needed.
The real map of the conflict: Africa, South-East Asia and Latin America
Here is the angle nobody talks about enough. While Washington and Beijing fight over chips, the real strategic ground is being contested elsewhere: among the 6 billion people who live outside both the American and the Chinese bloc.
As we documented in our article on global AI adoption, DeepSeek has already conquered several markets on a massive scale that are out of reach for OpenAI: 49% in Cuba, 43% in Russia, 39% in Iran. But above all, Africa is adopting AI 2 to 4 times faster than the West, mainly through free Chinese models such as DeepSeek.
Why? Three reasons. ChatGPT Plus at $20/month is out of reach for a student in Lagos. DeepSeek is free, open source and runs on any phone. And the Chinese authorities are far more flexible on regulatory compliance than their American counterparts. For developing countries, the choice is rational.
According to a projection by the consultancy McKinsey published in March 2026, if China maintains its pace of adoption in Africa, more than 40% of global AI users could be using a Chinese model by 2028. That is more than Apple + Google + Microsoft combined today. It would no longer be a trade war, it would be a civilisational shift.
Where does Europe fit into all this?
Europe is in the most uncomfortable position. It depends technologically on the United States (for chips and most models), is ideologically more aligned with the Americans than with the Chinese, yet commercially very closely tied to China. At every turn of the conflict, Europe is called upon to pick a side, without having the means to build its own.
France has made its choice with Mistral AI, but as we showed in our article on France's lag, uptake in businesses remains a minority affair. Germany is pushing for "European sovereign AI" but without any industrial follow-through. The United Kingdom, post-Brexit, is mainly playing its card as a research hub.
The risk for Europe is simply becoming the commercial battleground between OpenAI/Anthropic/Google on one side, and DeepSeek/Qwen/Kimi on the other. Neither player nor referee. Just a market to be conquered.
The three scenarios for the next 24 months
If we look coldly at the current dynamic, three scenarios are taking shape for 2026-2028.
Scenario 1: the AI cold war intensifies. If the Trump administration returns to a strict policy (for example under pressure from Congress via the Chip Security Act), exports of H200s to China stop. Beijing accelerates its technological independence, Europe is called upon to choose, and technological decoupling takes shape over 5-10 years. This is the "silicon curtain" scenario.
Scenario 2: murky status quo continues. Trump maintains an opportunistic trade policy (sell the H200s but not the Blackwells, tax sales at 25%, open and close the tap depending on negotiations). China continues to build up its software expertise while the hardware remains American. Emerging markets gradually shift towards Chinese models. No clean break, but a slow erosion of American leadership.
Scenario 3: a geopolitical crisis rewrites the rules. A major crisis over Taiwan (where TSMC produces 90% of the world's AI chips), a mutual cyberattack, or an unforeseen geopolitical event forces a brutal acceleration of decoupling. This is the "shock" scenario, unlikely but with immense consequences.
What this changes in practice
For most readers, these issues may seem remote. They are not. Three concrete consequences will affect us within 18 months.
First, prices. Technological decoupling has a cost. Chips, data centres and models are becoming more expensive everywhere in the world because of logistical inefficiencies and taxes. This will feed through to the price of AI subscriptions, cloud services, and every product that depends on AI.
Second, the choice of tools. European companies will have to pick a side, whether they like it or not. Using DeepSeek raises questions of data sovereignty. Using ChatGPT creates strategic dependencies. No option is neutral.
Third, the pace of innovation. US-China competition is accelerating AI development on both sides. As long as this race continues, models will improve at a rate that renders today's tools obsolete within months. This is exactly what we described in our 20 April article on runaway acceleration.
The silicon curtain has fallen. It remains to be seen whether we can make of it anything other than a battlefield.