The chokehold has a counterexample, and it is MIT-licensed

The US chip export control regime rests on a supply-side chokehold: deny advanced silicon, deny frontier AI. On Monday, Meituan open-sourced a 1.6-trillion-parameter model that scores 59.5 on SWE-bench Pro, trained end-to-end on 50,000 domestic Chinese ASICs, and released it under an MIT licence. The chokehold thesis is not dead. It is, however, load-bearing on assumptions that just moved.

I want to separate the two events people are conflating. The training-hardware story is one thing; the licence is another. They point in different directions, and treating them as a single narrative produces sloppy conclusions in both directions — the triumphalist "Chinese chips have caught up" reading, and the dismissive "it's still not GPT-5" reading. Both miss what actually shifted.

The hardware event is narrower than the headlines. LongCat-2.0 was trained on a 50,000-card domestic ASIC cluster with zero Nvidia hardware in the stack.¹ That is a real data point against the strong version of the chokehold thesis, which held that frontier-class training was infeasible without advanced Western silicon in the near term. It is not a data point that generalises to the sector. Meituan is a food-delivery and local-commerce conglomerate that happens to have built out enough compute to run a 50,000-card cluster; that is not a resource most Chinese companies have, and it is emphatically not a resource most second-tier global actors have.

The right update is one level up the stack. The chokehold has moved from inference-and-training to training-infrastructure-at-scale. If you can assemble 50,000 domestic ASICs and the engineering talent to make them cohere, you can train a near-frontier model. If you cannot, you still cannot. The bar has dropped, but it has not dissolved.

The licence event is broader than the headlines. This is the part I think most analysts are underweighting. A 1.6T-parameter Mixture-of-Experts model with roughly 48B active parameters per token and a 1-million-token context window is now available under MIT.¹ No usage restrictions. No reporting requirements. No API dependency. Any enterprise with the inference capacity to serve it, which, at 48B active parameters, is a more accessible bar than 1.6B dense, can self-host it tomorrow.

MIT is not Llama's community licence with the awkward carve-outs. It is the licence you use when you have decided that the weights are a gift to the commons and you want no friction on adoption. For enterprise buyers who have spent the last eighteen months negotiating data residency, egress terms, and audit clauses with Anthropic and OpenAI, this is a structurally different offering. The negotiation goes away because there is nothing to negotiate.

The negotiation goes away because there is nothing to negotiate.

The pricing floor keeps dropping, and the market is slower to price it than the evidence warrants. DeepSeek-R1 opened this direction. Lindy's public migration made the enterprise case visible. LongCat-2.0 extends it to a larger model with a specific agentic-coding specialisation and a permissive licence, and it does so with two months of real production usage already on the board under the "Owl Alpha" alias on OpenRouter.¹

The question for Anthropic and OpenAI is not whether some coding workloads migrate to self-hosted open weights. Some already have. The question is what pace of migration their enterprise contracts can absorb before per-seat and per-token pricing concessions become structural rather than tactical. I do not think anyone at the frontier labs believes their current enterprise pricing survives 2027 in its present form. I think the disagreement is only about the slope.

The Owl Alpha story is the most interesting thing in the release, and almost nobody is writing about it. For approximately two months before Meituan attached its name to the model, LongCat-2.0 sat on OpenRouter as "Owl Alpha" and topped the usage rankings.¹ Developers were picking it on performance, blind to brand. This matters for two reasons.

First, it produces a cleaner signal than any benchmark. SWE-bench Pro and Terminal-Bench 2.1 are useful, and 59.5 and 70.8 respectively are strong scores, but they are contestable in the way benchmarks always are.¹ Aggregate developer preference under production conditions, over two months, with no branding to bias the choice, is a different kind of evidence. It is closer to a revealed-preference market test than to a leaderboard.

Second, it means Meituan had two months of real deployment telemetry before the public release. That is an asymmetric information advantage most model launches do not have. The team knew what the model was good at, where it failed, and what the price-performance envelope looked like in actual developer workflows, before anyone else could form an opinion. If more labs adopt stealth-alias deployment as a pre-release strategy, the gap between "benchmark good" and "usage good" becomes a competitive moat rather than a footnote.

I want to flag OpenRouter's usage rankings are not an independently audited dataset, and claims of leadership rest on the platform's own reporting. Directionally informative, not verified market share. But directionally informative is enough for the point I am making.

The commercial logic of the release is worth sitting with, because Meituan has not explained it. Why does a food-delivery conglomerate open-source a 1.6T-parameter model under MIT? There is no stated rationale in the release materials.¹ The candidate motives are unsatisfying individually and more plausible in combination: talent attraction (the AI research market in China is competitive, and MIT-licensed frontier weights are a recruiting flag), ecosystem positioning (Meituan wants developers building on top of its stack for reasons that may become clearer later), domestic regulatory goodwill (open-source contributions play well with Beijing's stated preferences on AI diffusion), and distribution for a future managed service (open the weights, sell the hosting).

None of these individually justifies the capital outlay of training a 1.6T model on 50,000 ASICs. All of them together might. The honest reading is that we do not know Meituan's commercial theory, and analysts confidently asserting one are guessing. I will note that Chinese frontier labs have shown a consistent willingness to release capable weights under permissive licences at a scale that Western labs have not matched, and the strategic pattern here is real even if any single release's rationale is opaque.

The counter-case to my own reading. The strongest version of the "this changes less than it appears" argument runs like this. SWE-bench Pro 59.5 is near-frontier, not frontier. The best proprietary coding models still outperform it on the hardest tasks. The 50,000-card ASIC cluster is not reproducible outside Meituan's specific infrastructure. Open weights do not transfer training capability. The MIT licence is generous but self-hosting a 1.6T MoE at production quality is a serious engineering undertaking that most enterprises will outsource to a managed provider anyway, which puts them back in a vendor relationship.

I take this seriously. I think it is right about the specifics and wrong about the trend. Every one of these frictions is falling. Serving infrastructure for open-weight MoE models is maturing quickly; the gap between best proprietary and best open on coding benchmarks has been closing quarter over quarter; and the number of Chinese and non-Chinese actors with meaningful domestic compute clusters is growing. If you had told me in early 2025 that by mid-2026 a food-delivery company would open-source a 1.6T MoE trained on domestic silicon, I would have said "eventually, but not this year." The compression of the timeline is the story.

What I would watch next. Three things. First, whether other Chinese conglomerates with adjacent compute footprints, the cloud arms of ByteDance, Alibaba, Tencent, release comparable open-weight models in the second half of 2026. Second, whether US enterprise buyers begin citing self-hosted LongCat-2.0 in Anthropic and OpenAI renewal negotiations as a genuine BATNA (best alternative to a negotiated agreement) rather than a theoretical one. Third, whether the stealth-alias deployment pattern becomes standard practice, and what that does to the credibility of published benchmarks as a signal of model quality.

The chokehold thesis needed a counterexample to be tested. It has one. The right response is not to declare the policy dead, and not to wave the counterexample away as a one-off. The right response is to say that the ground has moved, name where it has moved to, and stop repeating the 2024 version of the argument in mid-2026.

Footnotes