A Model Pre-Trained Entirely on Non-Nvidia Silicon Will Rank Top 10 on a Major Public Capability Leaderboard by End of 2027

Prediction Statement

By December 31, 2027, at least one large language model whose full pre-training was performed entirely on non-Nvidia silicon — a domestic Chinese accelerator base such as the ASIC superpods used for Meituan's LongCat-2.0, or an equivalent home-grown platform — will rank in the top 10 of a major, widely-cited public capability leaderboard (for example the Artificial Analysis Intelligence Index, LMArena/Chatbot Arena text, or a comparable independent aggregate), at a moment when the developer has publicly stated the training was done on such hardware and that claim is not credibly refuted.

This is a claim about capability parity following the hardware decoupling — not merely that a domestically-trained model exists (LongCat-2.0 already established that on June 30, 2026), but that one is good enough to sit among the best models in the world on a neutral scoreboard.

The Case For

The decoupling already happened. The hard part — completing a trillion-parameter pre-training run on domestic silicon at all — has been demonstrated. What remains is closing the quality gap, and quality on these leaderboards has been converging across labs for two years. The distance from "trained domestically" to "trained domestically and top-10" is smaller than the distance from zero to "trained domestically."

Efficiency gaps close fastest where they are largest. The inefficiency of a first-generation domestic run lives mostly in the software layer — compilers, communication libraries, operational tooling — which improves far faster than silicon. Each subsequent run benefits from a more mature stack, so the capability per run should climb steeply from a low base.

Open weights recruit the world. LongCat-2.0 shipped open, and open models attract thousands of fine-tunes, distillations, and evaluations. A capable open base that becomes a default starting point tends to produce a derivative that lands high on a leaderboard even if the base model itself does not.

The incentive is national, not commercial. A less-efficient path is acceptable when the objective is sovereignty rather than margin, so the effort will continue to be funded through the inefficiency.

The Case Against

Leaderboard tops are a moving target. The frontier keeps rising; a top-10 slot in late 2027 requires beating whatever the leading labs ship next, not today's models. Convergence could stall if the incumbents open a new capability gap.

Provenance disputes. "Trained entirely on non-Nvidia silicon" is a claim with large surface area, and a high-profile leaderboard entry will draw scrutiny. If the claim is credibly contested, the prediction does not resolve true even if the model scores well.

Efficiency may bind harder than expected. If the domestic base is substantially less efficient, the compute needed to reach top-10 quality may not be marshaled within the window, even with national backing.

What Would Falsify This

The prediction resolves false if, as of December 31, 2027, no model meeting the non-Nvidia-pre-training condition appears in the top 10 of any of the named (or clearly equivalent) public leaderboards under an uncontested provenance claim. It resolves true on the first such appearance that persists long enough to be recorded by the leaderboard operator.

Confidence Rationale

Confidence is set at 60. The structural direction is strongly favorable — the decoupling is real and quality is converging — but the specific bar (top 10 on a neutral board, uncontested provenance, within roughly 18 months) is demanding enough, and provenance disputes likely enough, to keep this short of a high-confidence call. This is the capability-parity follow-through to the training decoupling analysis; the hardware question is settled, and this prediction is about how fast the quality question follows.