The Fine-Tuning Gap Is Now a Chasm
When a single base model wins gold at both IMO and IOI via a published, reusable recipe, the base-model moat stops being a moat and starts being a commodity.
NVIDIA just proved that gold medals are a fine-tuning problem.
Nemotron 3 — a single base model family — won gold at both IMO 2026 and IOI 2026. Different competitions, different skills, different specialization stacks, same foundation. The IOI run scored 535.4 out of 600, clearing the gold threshold of 361.12 and beating the top human score of 498.27. The IMO system submitted proofs graded by official IMO graders and hit 30 out of 42, above the gold line of 29. NVIDIA published the models, the data, and the recipes on Hugging Face. No trade secrets. No proprietary moat claimed.
Read that last sentence again. Because it’s the part everyone is skipping over.
The dominant narrative about frontier AI is that the lab with the biggest model wins. Scale is destiny. Compute is the moat. This is the story that justified Anthropic’s multi-billion-dollar raises, that explains why Microsoft is spending $80 billion on data centers this year, and that makes the hyperscalers look like inevitabilities. Everyone says the base model is the prize.
The opposite is closer to true — or at least, it’s becoming true faster than the incumbents are pricing in.
What NVIDIA demonstrated isn’t that Nemotron 3 Ultra is smarter than GPT-5 or Claude 4. The IOI result was “unofficial, unsupervised” and the system was not included in the official ranking. NVIDIA is careful to say so. What they demonstrated is a reusable specialization recipe: supervised fine-tuning plus reinforcement learning plus feedback-driven inference, applied twice, to two genuinely distinct tasks, both reaching world-class performance. The same base, two gold medals, open weights. That’s not a benchmark result. That’s a manufacturing process.
Here’s what that changes. If specialization is a repeatable, publishable process rather than a mysterious art, then the value of a good base model becomes more like the value of a good steel mill than the value of a diamond mine. Steel mills are valuable. But the margin in steel is thin, because the process is known, the inputs are commodity-adjacent, and anyone who can run the process at scale competes with you directly. Diamonds are scarce by nature. If the frontier labs thought they were selling diamonds, they may discover they’ve been building steel mills.
The IOI result is particularly sharp. The system operated “under the same time, internet-access, and submission constraints as human contestants.” Real conditions. Not a cherry-picked benchmark designed to flatter the model. And the recipe — GenCorrect layered on top of Nemotron-3-Ultra-CC with SFT — is now documented and public. Any team with the right hardware and enough time can run it. The IMO result required a generate-verify-refine pipeline using multiple checkpoints. Also documented. Also public.
The labs that are betting their valuations on proprietary model advantage should be paying attention to what “open models, data, and recipes” actually means at this level of performance. It means the specialization layer is no longer a secret sauce. It’s a known process available to every well-funded team on the planet — and to every government with a chip stockpile and a reason to care about AI capability.
The geopolitical dimension is not subtle. Olympiad-level mathematical reasoning and competitive programming are proxies for the skills that matter in cryptography, logistics optimization, and autonomous systems. NVIDIA publishing this recipe isn’t just a Hugging Face flex. It’s a signal about where the capability frontier actually lives, and who can reach it.
The trillion-dollar question — literally, given what’s been committed to data center buildout — is whether the value in AI accrues to the model or to the application layer. NVIDIA’s answer, delivered via two gold medals and an open recipe, is: the application layer. The base model is infrastructure. Infrastructure gets commoditized. The margin lives in knowing what to build on top of it, and how fast you can build it.
If that’s right, the companies that should be terrified aren’t the startups. They’re the labs that raised at valuations premised on base-model moats — and the investors who bought that story.
The race isn’t over. But the laps are getting shorter, and the lead car just published its pit strategy.