1.6 Trillion Parameters, Not One Nvidia GPU

2026-11-15·6 min read·OKCodex Editorial

1.6 Trillion Parameters, Not One Nvidia GPU

Three signals, one thread

The week's AI news looks unrelated: Meituan released weights for a 1.6-trillion-parameter agent model trained entirely on Chinese hardware; reporters traced nearly $12 billion in Nvidia chip orders routed to Chinese firms; and Australia's central bank found AI is hitting 15- to 24-year-old entry-level workers first. Put together, they describe one thing: the capability ceiling of AI agents is being reshaped by the structure of compute supply — and the bill is arriving first at the youngest end of the workforce.

LongCat-2.5: what matters beyond the score

According to The Decoder, Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and trained entirely on Chinese hardware. It scores 88.4 on Terminal Bench, landing just behind Claude Opus 5. The sample deployment described in the source runs on 194 Huawei Ascend chips.

The interesting part is not the 88.4 — benchmark scores inflate every year. It is three structural facts.

First, this is the training side, not inference. Domestic silicon has mostly been discussed as a way to cut inference costs; here it carried a trillion-parameter training run. Second, it is built for agents, not chat. Terminal Bench measures multi-step task completion in a terminal environment, which is the core metric for agent products. Third, the weights are open, so downstream teams can fine-tune and deploy in their own environments.

The caveat matters just as much: the source summary does not give training cluster size, duration or cost, and the 194 Ascend chips describe a sample deployment, not the training configuration. Those three numbers decide whether this is reproducible. A stack that can train a trillion-parameter model once, at a cost nobody can repeat, is a different thing from a domestic training pipeline that scales.

Supply-side noise is the domestic stack's tailwind

The Nvidia-to-China reporting from the same period is not pretty. The Wall Street Journal reported that Megaspeed — a firm linked to Alice Huang, described as Jensen Huang's former partner — arranged nearly $12 billion in Nvidia chip orders for Chinese companies between 2024 and 2025. Short seller Culper Research estimates Nvidia's actual China revenue share exceeds 20 percent. TechCrunch separately reported that Nvidia's security chief quietly left in August, and that Jensen Huang has stepped down over his relationships, including one with an executive at a company suspected of funneling chips to China. A petition signed by Elon Musk and other tech executives calling on Nvidia to reinstate him had passed 115,000 signatures as of Saturday afternoon.

These are press reports and third-party estimates, not settled findings. But for anyone buying compute, the conclusion is already clear: cross-border supply and compliance risk is rising, not falling. That is why 'trained entirely on Chinese hardware' deserves to be recorded at all.

Demand side: the stronger the agent, the harder entry-level work is hit

Australia's central bank supplies the other end of the evidence. Per The Decoder, it sees early signs that AI is hitting young entry-level workers in highly exposed jobs: since ChatGPT launched, job postings and employment among 15- to 24-year-olds in software development and customer service have dropped more sharply than among older workers.

That tracks with where model capability is heading. Anthropic's Claude Fable 5.2 illustrates the pattern from another angle: rather than computing pixels, the language model writes Python code that places hundreds of thousands of simulated brushstrokes on a virtual canvas, complete with bristle texture and paint application. The model delegates execution to code — a drawing technique, but also the general shape of an agent. Once the action space is code and the terminal, the first tasks to go are the ones that amount to following steps.

What to do about it

Engineering leads: keep reproducible benchmarks such as Terminal Bench in your selection process, but do not buy on the total score. Build a 20-50 task evaluation set from your own workflows and look at multi-step failure recovery. Cost out domestic-silicon porting early — it will not show up in a benchmark, it will show up in your schedule.

Infrastructure and procurement: assume supply keeps fluctuating. Redundancy across stacks for critical training workloads belongs in architecture review, not in the post-incident memo.

Teams and individuals: if you manage people with 0-3 years of experience, restructure their work now. Hand scriptable steps to agents; keep judgement, acceptance and cross-system coordination with humans. Entry-level jobs will not vanish, but jobs that are only entry-level work will.

Conclusion

The real significance of LongCat-2.5 is not that it trails by a few points, but that it shows a trillion-parameter agent model can be trained off the Nvidia stack while compute supply is repeatedly disrupted by policy. That is a testable starting point, not a finish line. Watch next for published training costs, domestic accelerators beyond Ascend entering the training side, and the lagging labour-market feedback from stronger agents. For readers, the practical move is singular: make 'compute substitutability' and 'how scriptable is this task' explicit fields in this year's architecture reviews and headcount plans.

Sources

Share:Weibo

Subscribe to weekly briefing

Comments

Loading comments…

Leave a comment

0 / 500

Comments appear after review.

Find tools for your next task

Start with scanned documents, code review or research. Get a shortlist with matching evidence; no model is run.

Find tools for my task →
Analyze this topic with two models

DeepSeek and Qwen are preselected with the article topic. You choose when to run them.

Try it →

Keep reading

Mentioned

Agents mentioned in this article