Digital Sovereignty

China Trained a Frontier AI With No US Chips

Meituan trained a 1.6 trillion parameter model end to end on 50,000 domestic Chinese chips. The export-control moat just gained a public counterexample, and Europe's excuse got weaker.

Digital SovereigntyBy Servola Tech Desk2026-07-053 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

China Trained a Frontier AI With No US Chips

Key takeaways

Meituan trained LongCat-2.0, a 1.6 trillion parameter model, entirely on more than 50,000 domestic Chinese chips - the first time a frontier-scale model has done both training and inference without US silicon, which weakens the core premise of export controls.
Earlier Chinese efforts used home-grown chips only for inference; doing the far harder pre-training on domestic hardware is the line that was just crossed, and it was open-sourced.
For Europe the uncomfortable read is that the barrier to a sovereign AI stack was never only the physics of chips, it was the capital and the will to fund the whole stack.
For owners, models built outside the US and Nvidia supply chain are now viable and cheaper, which reshapes procurement and shortens the shelf life of any advantage built purely on export controls.

A food-delivery company published the counterexample

The company that broke the assumption was not a chip champion or a state lab. It was Meituan, the Chinese food-delivery giant, which on 30 June 2026 released LongCat-2.0, a 1.6 trillion parameter mixture-of-experts model with a million-token context window. The claim that matters is not the size. It is where the model was built. Meituan says the entire run - pre-training and inference - happened on a cluster of more than 50,000 domestic Chinese AI chips, on roughly 35 trillion tokens.

To see why that lands, separate two things export controls have quietly conflated. Running a finished model on local chips is inference, and China had already shown it. Building the model from scratch on local chips is pre-training, and that is the compute-hungry, failure-prone part that everyone assumed still needed Nvidia. LongCat-2.0 did both, and Meituan open-sourced the result.

Why the export-control logic just cracked

The United States built its AI containment strategy on a single premise: deny the most capable chips and you slow the most capable models. That premise held while domestic Chinese silicon could only serve models trained elsewhere. A trillion-parameter model trained end to end on home-grown hardware is a public demonstration that the premise has an expiry date.

The point is not that the export controls achieved nothing. They raised costs and bought time. The point is that a moat built on a hardware bottleneck erodes the moment the other side proves the bottleneck can be routed around, and that proof is now downloadable. Advantage that depends on the other side being unable to do a thing does not survive the day they do it.

Europe's uncomfortable mirror

Europe has spent two years explaining its distance from the AI frontier in terms of access - to chips, to hyperscale capital, to US platforms. LongCat-2.0 reframes that explanation. If a delivery company in a sanctioned supply chain can assemble the compute and the talent to train a frontier model on non-US silicon, then the constraint on a European sovereign stack was never purely the export-control regime it also lives under. It was the decision to fund semiconductors, energy and model training as one connected programme rather than a set of grants.

This is where initiatives such as EuroStack and the EU Cloud and AI Development Act meet reality. The blueprint exists. What LongCat-2.0 removes is the alibi that the physics makes it impossible. It does not; it makes it expensive and slow, which is a question of political will and capital, not of what is technically achievable.

What owners should take from this

Two practical shifts follow. First, capable models are increasingly being built outside the US and Nvidia supply chain, and LongCat-2.0 is priced to undercut the Western frontier labs. For a buyer, that widens the menu and pressures the price of the incumbents you already licence. Ignoring non-US models on reflex is now a procurement blind spot, not a safeguard.

Second, treat any advantage a supplier claims from export controls as time-limited. If your vendor's edge rests on rivals being unable to get chips, price in the day that stops being true. The durable questions are the ordinary ones: does the model do the job, who controls the weights, and where does your data go.

Frequently asked questions

What is actually new about LongCat-2.0?

It is described as the first frontier-scale model (1.6 trillion parameters) to complete both pre-training and inference entirely on domestic Chinese chips - more than 50,000 of them. Earlier Chinese models used local chips only for inference; the pre-training was the harder line that had not been crossed publicly before.

Does this mean US export controls failed?

Not entirely. They raised China's costs and delayed progress. But the strategy assumed frontier training required US chips, and a trillion-parameter model trained end to end on domestic hardware is a public counterexample to that assumption. A moat built on a hardware bottleneck weakens once the bottleneck is shown to be avoidable.

Why should a European business care about a Chinese model?

Because it changes procurement. Capable models are now being built outside the US and Nvidia supply chain and priced below the Western frontier labs, which widens the buyer's menu and pressures incumbent pricing. It also shows that a sovereign AI stack is a question of capital and will, not of what is physically possible.

The lesson for Europe is not that China is ahead; it is that the excuse of impossibility is gone. A sovereign stack was always a matter of capital and resolve, and someone in a harder position just proved it can be built. The blueprint is on the table, and the alibi is off it.

LongCat-2.0MeituanChinaexport controlsEuroStacksovereign AI

Europe's Chip Sovereignty Win Is the Boring Silicon

Infineon opened a 5 billion euro 300mm fab in Dresden on 2 July 2026, three months early, part-funded by the EU Chips Act. The real sovereignty win is in power and analog chips - the workhorse silicon in every EV, factory and data centre, not the leading-edge logic race.

4 min read

Digital Sovereignty

Europe Bets 20 Billion on Its Own AI

The EU is putting a 20 billion euro InvestAI fund behind up to five AI gigafactories, with the call opening in July. Its first serious attempt to buy sovereign compute, not just regulate.

3 min read2 views

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.

2 min read14 views

← All articles

China Trained a Frontier AI With No US Chips

A food-delivery company published the counterexample

Why the export-control logic just cracked

Europe's uncomfortable mirror

What owners should take from this

Frequently asked questions

Europe's Chip Sovereignty Win Is the Boring Silicon

Europe Bets 20 Billion on Its Own AI

The EU Council Revived the Rule That Scans Private Messages

The Operator Brief