The Delivery
On June 30, a Chinese food delivery company announced it had trained a 1.6-trillion-parameter AI model entirely on domestic chips. It had been running anonymously on the world's largest model marketplace for two months, beating GPT-5.5 on software...
On June 30, a Chinese food delivery company announced it had trained a 1.6-trillion-parameter AI model entirely on domestic chips. It had been running anonymously on the world's largest model marketplace for two months, beating GPT-5.5 on software engineering benchmarks, before anyone knew where it came from or what it ran on. The export controls didn't prevent the model. They prevented the revenue.
On June 30, 2026, Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter language model trained entirely on domestic Chinese chips. Meituan is not an AI lab. It is China's largest food delivery platform, processing tens of millions of meal orders per day, with quarterly revenue of 91 billion yuan and a market capitalization of roughly $56 billion. Its primary competitors are Ele.me and Didi. The model it just released outscored GPT-5.5 on SWE-bench Pro, the industry standard benchmark for real-world software engineering.
Before anyone knew what LongCat-2.0 was, they were already using it. For two months before the public reveal, Meituan deployed the model anonymously on OpenRouter under the alias Owl Alpha. It accumulated 10.1 trillion monthly tokens of throughput. It ranked first on Hermes Agent, second on Claude Code, and third on OpenClaw by call volume. Developers chose it for production coding work based solely on performance. When Meituan revealed the identity on June 30, the model was already embedded in thousands of agent workflows worldwide.
The approach was a controlled experiment run in reverse. Rather than announce a branded model and wait for adoption, Meituan let the market discover it blind. The alias stripped the origin. No one selecting Owl Alpha for their coding agent knew they were running a Chinese model on Chinese-made chips. They knew only that it worked, and that it was cheap.
What it runs on is the point. LongCat-2.0 uses a mixture-of-experts architecture with approximately 48 billion parameters active per token, giving it frontier-scale knowledge at the inference cost of a much smaller dense model. It supports a native one-million-token context window. The architectural details matter less than where it was trained: a cluster of more than 50,000 domestic ASICs organized into Atlas-950 SuperPods, running Huawei's HCCL communication library, the Chinese equivalent of Nvidia's NCCL. The chips are based on Huawei's Ascend 910C, manufactured on SMIC's 7nm process. Each delivers roughly 60 percent of an Nvidia H100's performance. Meituan used 50,000 of them.
On SWE-bench Pro, LongCat-2.0 scored 59.5. GPT-5.5 scored 58.6. Claude Opus 4.7 and 4.8 still lead, and on FORTE, which measures agent performance across 15 professions under a 45-minute time limit, LongCat-2.0's 73.2 ties Claude Opus 4.6 but trails GPT-5.5's 77.8. It is not the best model in the world. It is near-frontier. And it was trained on hardware the United States has spent four years trying to keep out of Chinese hands.
The pricing makes the competitive position sharper. LongCat-2.0 costs $0.75 per million input tokens and $2.95 per million output tokens, with cached context reads free. GPT-5.5 costs $5 and $30. Claude Sonnet 5 costs $2 and $10 at its introductory rate. LongCat-2.0 is six to seven times cheaper than GPT-5.5 on input and ten times cheaper on output. Meituan announced the release under the MIT license, the most permissive open-source license available, meaning any company can take it, modify it, and sell products built on it without paying Meituan anything. The weights have not yet shipped. Both the Hugging Face and GitHub repositories still read "model weights coming soon."
The export controls this undermines are not abstract. On January 15, 2026, the Bureau of Industry and Security published its most comprehensive final rule on semiconductor exports to China, introducing total processing power thresholds and shifting key chips including Nvidia's H200 and AMD's MI325X from presumption of denial to case-by-case review. The Council on Foreign Relations called the framework "strategically incoherent and unenforceable." Chatham House was blunter: chip smuggling is widespread, third countries serve as gray markets, and demand growth makes the controls difficult to enforce at any scale.
LongCat-2.0 does not represent a smuggling success. It represents something the controls were not designed to prevent: a frontier-scale model trained without any controlled hardware at all. The Ascend 910C is not a smuggled H100. It is a different chip, manufactured domestically on a domestic process node, running a different software stack, achieving 60 percent of the performance at whatever scale the Chinese market can supply. Fifty thousand chips doing 60 percent of the work is the equivalent of 30,000 H100s. That was enough to reach the frontier.
What makes this different from DeepSeek is the scope of the domestic hardware. DeepSeek V3 trained on 2,048 Nvidia H800 GPUs, the China-compliant variant that shipped legally before further restrictions. DeepSeek V4-Pro was pre-trained on Nvidia hardware, with a Huawei-led team completing post-training on 1,000 Ascend 910C chips afterward. LongCat-2.0 is the first publicly confirmed frontier-scale model to complete its entire pre-training pipeline on domestic Chinese chips, with no Nvidia hardware at any stage. And instead of keeping it proprietary, Meituan is giving it away.
The timing of the release carries its own signal. On June 26, four days before LongCat-2.0 went public, OpenAI released GPT-5.6 in three variants called Sol, Terra, and Luna, then immediately restricted access to roughly 20 companies after the US government requested a limited rollout under the June 2 Trump executive order on frontier AI assessment. America's most powerful new model launched behind a gate. Four days later, China's near-frontier model launched under MIT.
On July 2, two days after LongCat-2.0 was announced, The Information reported that Anthropic is in preliminary talks with Samsung Electronics to manufacture a custom AI accelerator on Samsung's 2nm process. Anthropic hired Clive Chan in early June, the second engineer ever on OpenAI's custom chip team, where he spent two and a half years helping build the inference accelerator OpenAI unveiled as Jalapeño on June 24. The American AI labs are now designing their own chips to reduce dependence on Nvidia. The Chinese food delivery company already trained on chips that are not Nvidia's at all.
The original policy argument for export controls rested on a premise that felt intuitive: deny China the best chips, and you deny China the ability to train the best models. LongCat-2.0 did not disprove the premise. Claude Opus 4.8 still leads most benchmarks. What LongCat-2.0 proved is that "the best" was never the threshold that mattered. Good enough, cheap enough, and open is a different competitive position than best and closed. The controls were designed to prevent a chip from crossing a border. They have no mechanism to prevent a model from crossing one. The food delivery company understood this before the regulators did.