A Food-Delivery Company Just Built a Frontier AI Without a Single Nvidia Chip. That’s the Story.
In a nutshell
US export controls have restricted China's access to advanced AI chips since late 2022, on a single strategic premise: that denying China the best hardware would keep it a step behind at the AI frontier. On June 30, a Chinese food-delivery company published a result that puts that premise in serious doubt. gafam.ai reads what it means — and why it lands hardest in Europe.
What Meituan Released
Meituan — the Chinese giant best known for delivering dinner — open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model, on June 30, under a permissive MIT license.
It has a one-million-token context window and is built for agentic coding. Because of its Mixture-of-Experts design, it activates only a slice of its parameters on any given token — averaging around 48 billion of the 1.6 trillion — which keeps a model this large comparatively cheap to run.
It was not an unknown quantity at launch. LongCat-2.0 had quietly topped the OpenRouter usage charts for around two months under the anonymous codename "Owl Alpha" before Meituan revealed its identity. Developers were already using it heavily before they knew what it was.
The Claim That Actually Matters — Handled Carefully
The headline is not the size. It is the hardware. And here gafam.ai applies discipline, because the central claims come from Meituan itself.
Meituan says LongCat-2.0 is the first trillion-parameter model to complete both full training and inference entirely on domestic Chinese chips — a cluster of more than 50,000 home-made ASICs, with no Nvidia GPUs involved at any stage. If accurate, that is a genuine milestone: proof that near-frontier AI can be built at scale without the American hardware that US export controls were designed to withhold.
Two honest caveats belong right next to that claim. First, the performance numbers are self-reported: Meituan says LongCat-2.0 scores 59.5 on the SWE-bench Pro coding benchmark versus GPT-5.5's 58.6 — a margin under a single point, not independently verified, and best read as "competitive," not "superior." Second, as of the reporting, the full model weights were not yet public — the pages said "coming soon" — so independent replication of the hardware claim is not yet possible. This is a serious, well-documented claim from a major company, not an established fact. The distinction matters.
Why the Hardware Milestone
Outweighs the Benchmark
Even setting the benchmark aside entirely, the hardware claim is the consequential one — because it speaks to the effectiveness of the single most important lever in US-China tech policy.
The entire logic of chip export controls is that frontier AI requires cutting-edge Nvidia silicon, and that denying it keeps China behind. LongCat-2.0 is Meituan's argument that the logic has a hole in it. The company frames the model explicitly as proof that Chinese firms can reach frontier scale without Nvidia's CUDA-based chips. If Chinese conglomerates can iterate trillion-parameter models on homegrown ASICs, the export-control strategy does not just fail to contain China — it actively accelerates the domestic-chip self-sufficiency it was meant to prevent. Restriction becomes the mother of substitution.
The Timing — Two Opposite Approaches, Same Week
The context makes the contrast unmistakable. This open, MIT-licensed, download-it-freely Chinese model arrived in the same window that Washington was tightening access to American models — the GPT-5.6 restriction to government-approved partners, the Fable 5 and Mythos saga, the Annex A named-entity lists we have documented all month.
Two systems, two opposite instincts, one week. America is gating its best models behind government approval and export licences. China is giving a frontier-scale model away under the most permissive licence that exists. Whatever one thinks of the motives on either side, the strategic asymmetry is striking: one approach makes frontier AI a controlled privilege, the other makes it a free download.
The European Perspective
LongCat-2.0 crystallises the impossible choice that a month of gafam.ai reporting has been circling. Europe now faces three doors, and this release throws all three into sharp relief. Behind the first door are the American frontier models — the most capable, but increasingly gated behind US government approval, export directives and named-entity lists, as the Fable 5 saga showed.
Behind the second door are Chinese open-weight models like LongCat-2.0 — free, permissively licensed, frontier-adjacent, and now demonstrably independent of the American hardware stack — but carrying unavoidable questions about data security, trust and strategic dependence on Beijing.
Behind the third door is a sovereign European capability that does not yet meaningfully exist. The cost reckoning we covered pushes European businesses toward the cheap Chinese door. The security establishment pushes back toward the expensive American one. And the absence of a European option means Europe is perpetually choosing between two forms of dependence rather than exercising autonomy. What LongCat-2.0 adds to this picture is decisive: it proves that the open-weight path is not a toy tier but a frontier-capable one, and that it is being defined, right now, by China rather than by Europe.
The uncomfortable truth is that the open-weight frontier — the one route that could give Europe sovereign AI without matching American capex — is currently a Chinese achievement, not a European one. The EUROPA consortium's open model is the right idea; LongCat-2.0 is the measure of how far ahead someone else already is. Europe does not lack the strategy. It lacks the execution, and the clock is being set by others. gafam.ai will be watching.
We are not first. We are right.
SOURCES
— SiliconANGLE: China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips
— South China Morning Post: China claims biggest AI model trained on local chips, as Meituan releases LongCat-2.0
— VentureBeat: Meituan open sources LongCat-2.0, the 1.6T near-frontier agentic coding model — trained entirely on Chinese chips
— Cryptopolitan: Meituan open-sources 1.6-trillion-parameter LongCat-2.0 AI model trained on Chinese chips
— Fello AI: LongCat-2.0: China's 1.6T Open-Source Coding Model (benchmark caveats)
🔒 This analysis is for GAFAM Intelligence members only.
Already a member? Log in here