The Balance of AI Power: Anton Leicht on Politics, Pacing Deals, and Muddling Through Well

Full analysis

This is a two-hour geopolitics conversation between Anton Leicht of the Carnegie Endowment and Nathan Labenz. Strip away the training-pause diplomacy and the space-compute speculation, and one claim actually touches the tools you buy: the cheap open-weight model you were counting on as a hedge against frontier API prices may stop keeping up in the domains you care about. That's the thread worth pulling.

How hard is this to undo? Nothing here demands a decision this week. But if you're building a product roadmap around "open models will catch up," that's a bet that gets expensive to unwind if it's wrong. What's actually being decided, for a reader, is whether to keep treating open-weight models as a real cost hedge or start pricing in a permanent gap for specialized work. Nothing sets a deadline. This is a watch-the-trend situation, not a this-quarter one.

The Skeptic

Most of this episode is untestable. A US-China training pause "on the merits," orbital data centers by 2029, Europe wielding ASML as a cudgel. None of it changes what you ship. Leicht himself rates real extinction risk low and pads his 10% "doom" number with "political doom," which means totalitarian lock-in, not a rogue model. That's a category built to sound scary while conceding the technical case is thin. The auditor-access commitments from OpenAI and Anthropic are real and already happened, but "employee-like access for Metr and Redwood" is a voluntary press-release commitment with no enforcement behind it. Treat it as marketing until a government mandates it or an audit actually catches something.

The Open-Source Advocate

Here's the one claim that matters, and it cuts against my usual optimism. Leicht argues the open catch-up game depended on distillation: you take a frontier model's outputs through its API and train a cheaper open model to copy them. Cheap, fast, and it kept Llama, Qwen, and Mistral within striking distance. His point is that the frontier capability is moving into places you can't distill from, specifically the expensive reinforcement learning and domain-specific training environments (think Anthropic building a whole life-sciences setup to teach a model biology). If the good stuff lives there, copying the public API gets you a model that's polished but can't do the specialized work. For general chat and coding, open weights stay fine. For bio, materials, deep domain reasoning, the gap could widen instead of close.

The Researcher

The distillation argument is the real content, and it's plausible but unproven. We've watched the open-to-frontier gap shrink for two years on general benchmarks. Qwen and DeepMind's own releases keep landing near the top. Leicht's claim is narrower: that specialized reinforcement-learning pipelines don't leak through the API the way raw knowledge does. That's a genuine mechanism, not hand-waving. But it's untested at the level that matters to you. Nobody has published a clean measurement showing open models falling further behind specifically in bio or materials while holding pace on coding. Until someone does, this is a well-reasoned hypothesis, not a fact you can plan a budget around. Leicht also thinks current AI valuations require pharma and materials-science R&D contracts, not "Bob in accounting" using a chatbot. If those premium contracts don't show up, he expects valuations to crack.

The Compute Pragmatist

Follow the money on that valuation point, because it decides your API bills. Leicht's read is that the buildout only pencils out if labs land multi-million-dollar drug-discovery and materials contracts. If those don't materialize, the frontier labs are sitting on capacity they overbuilt, and overbuilt capacity gets discounted. That's good news for you in the near term: excess compute means cheaper inference, more price wars, more free tiers. The space-compute stuff is a 2029-plus curiosity, gated by how many rockets SpaceX can launch, and irrelevant to any real-time work because the round trip to orbit kills latency. Ignore it for planning. What's live is the possibility that the whole frontier build is chasing a revenue story that hasn't arrived, and the correction, if it comes, lands on prices before it lands on capability.

The Builder

On Tuesday morning, none of this changes my stack. But two things go on my checklist. First, if I'm using an open-weight model for anything domain-specific, I should actually measure it against the frontier version on my task, not on a generic benchmark. Leicht's argument predicts the gap shows up exactly in the specialized work, so run your own eval on your own data before you assume the cheap model holds. Second, auditor access becoming normal means enterprise buyers will start asking for audit attestations. If you sell AI features into big companies or government, expect "which evaluators have looked at your model provider" to show up in procurement questionnaires within a year or two.

The tensions

The real disagreement is between the Open-Source Advocate and the Researcher, and it's the whole story. Leicht says the open catch-up breaks down in specialized domains. The Researcher says two years of benchmarks show the opposite trend and nobody has measured the specialized case cleanly. Both are right about what they can see. The Advocate sees a real mechanism; the Researcher sees no data yet confirming it bites.

The second tension: the Skeptic treats the auditor-access commitments as marketing, the Builder treats them as an early procurement signal. Those aren't in conflict so much as different time horizons. Voluntary today, contract requirement in eighteen months.

What it hinges on

One belief decides whether this episode matters to you: does frontier capability in specialized domains detach from what you can copy through a public API? If yes, your open-weight cost hedge shrinks to general tasks and you pay frontier prices for anything specialized. If no, the open ecosystem keeps you covered and this was two hours of geopolitics. You don't have to guess. Measure it. Take your actual domain task, run it on the best open model and the best frontier model, and watch that gap over the next few release cycles. If it widens, Leicht was right and you budget accordingly. If it holds, keep hedging.

Prediction: By the end of 2026, at least one open-weight model (Qwen, DeepMind's Gemma line, Llama, Mistral, or DeepSeek) will score within 10 points of the top frontier model on a public general-purpose coding or reasoning benchmark, while no open-weight model matches frontier performance on a published specialized bio or life-sciences evaluation.

Confidence: Medium. The general-purpose trend is well established; the specialized gap is inferred from Leicht's mechanism, not yet measured.

Why: Open models have closed to within single digits of the frontier on general coding and reasoning benchmarks for two straight years, and that trend has no reason to reverse by year-end, so the first half is close to safe. The second half rests on Leicht's argument that specialized capability now comes from proprietary reinforcement-learning setups that don't leak through the public API the way general knowledge does, which means the copying method that powered open catch-up can't reach it. The opposite outcome, an open model matching the frontier on a published bio evaluation, is unlikely mainly because almost nobody publishes open bio evals at all, and the labs building those domain pipelines guard them precisely because they're the moat. That thinness is the weak point in the call, which is why it's Medium and not High.

Revisit by 2026-12-31: We're right if an open-weight model sits within 10 points of the frontier on a general coding or reasoning benchmark while none matches the frontier on a published specialized bio or life-sciences eval. We're wrong if an open-weight model matches frontier performance on such a specialized eval, or if open models fall more than 10 points behind on general coding and reasoning too.

The useful move here isn't reacting to the geopolitics. It's running your own eval on your own domain and watching whether that gap moves. That's a Tuesday task, and it settles the only question in the episode that touches your bills.