Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

Liam Fedus and Ekin Doğuş Çubuk left OpenAI to build Periodic Labs, which runs physical chemistry labs operated by AI. The pitch is simple: you cannot think your way to a new material. You have to make it and measure it. Frontier labs like OpenAI and Google DeepMind do not own furnaces, so they cannot do this work. That gap is the business.

The useful idea here generalizes well beyond materials science. Çubuk's point is that scientific discovery is, almost by definition, what a model has not been trained on. Fedus makes the matching point: if better reasoning at inference time were enough, labs would have stopped training new models after GPT-4 and just gotten clever at test time. They didn't. You still have to compress new knowledge into the weights, which means a reasoning model is only as good as what it has already absorbed. On anything genuinely new to it, the model needs fresh measured evidence, not a longer think.

The other real insight is about where the bottleneck sits. Fedus and Çubuk argue the hard step is characterization: figuring out what you actually made. Reading reality is harder than acting on it.

On competitive advantage, Fedus says their edge over frontier models is proprietary physical-experiment data, not a better base model. That lets them beat GPT-class models on materials tasks "even under the highest reasoning efforts," at lower compute. Which means the base model is a commodity to them, and the value is data nobody else has collected. That is the lesson for any operator: when open-weight models like Llama or Qwen are 80% as good for a fraction of the cost, your advantage is never the model. It is the measured, hard-to-collect data your competitor cannot scrape.

The slow-flywheel problem is real, though. A software rival iterates a thousand times a day. Periodic iterates when the oven cools. Proprietary physical data is either the strongest kind of moat (impossible to copy) or the slowest (gathered at the speed of furnaces while competitors compound daily). Right now almost nobody is collecting the same data, so slow-but-exclusive wins. The day a well-funded rival builds the same lab, slow becomes the problem.

There is one concrete pattern worth lifting out for Tuesday. Periodic embeds its model directly into the instrument with full context of intent and measurement history, rather than running it as a fixed script. Their example: the AI caught a loading error in a machine by noticing inconsistencies across measurements no hard-coded rule would have flagged. Swap "electron microscope" for your data pipeline or your QA step and the pattern holds. Give the model the full intent and history behind a task and it catches things your rules miss.

Çubuk's physics point is worth keeping too. Today's chips sit within three to four orders of magnitude of the Landauer limit, the thermodynamic floor for the energy cost of erasing one bit. Squeezing more performance out of existing materials is getting expensive fast. That is why new materials are a high-value target, and why the customers who care most are chip makers and data-center operators paying the electricity bill on a GPU fleet. If materials discovery accelerates, NVIDIA, TSMC, and anyone running large GPU clusters are the downstream beneficiaries.

The call: Periodic's primary disclosed revenue through 2026 will be forward-deployed engineering and services (teams embedded at customer sites, integrating tools into secure environments, fine-tuning on private data), not outcome-based or product licensing. Confidence: medium.

Fedus and Çubuk said this plainly themselves. Outcome-based pricing was named only as a long-run aspiration, compared to the GitHub Copilot-to-Codex arc. Attributing a discovered material or a yield improvement to your model requires measured results over furnace-speed cycles. A company does not flip from services to outcome pricing in a year when its iteration loop runs at the speed of physical experiments. We are right if Periodic's next public funding round or company statement still describes revenue mainly as embedded engineering. We are wrong if it reports material revenue from outcome-based pricing or a licensed product. Revisit by 2027-03-31.

Two ex-OpenAI people, Liam Fedus and Ekin Doğuş Çubuk, started a company called Periodic Labs that builds physical chemistry labs run by AI. Their pitch: you cannot think your way to a new material. You have to actually make it and measure it, and the frontier labs (OpenAI, Google DeepMind) do not own furnaces, so they cannot do this. The claim that matters for the reader: there are valuable AI problems where more model intelligence alone does not win, because the bottleneck is getting new data from the real world.

This is a "what does it mean" briefing, not a decision. Nothing here is hard to undo. There's no deadline. The reader is not buying from Periodic. What they're buying is a mental model: where does a smarter model help you, and where do you need your own data instead.

The Skeptic

Strip the romance. "Synthesis superintelligence" is a fundraising phrase. The real business today is forward-deployed engineers sitting inside Taiwanese chip plants, fine-tuning models on the customer's private process data. That is consulting with a model attached. Good business, old shape. The grand claim, that nobody will "zero-shot the room-temperature superconductor," is also the safest claim in the world. Nobody has done it with anything, so it can't be falsified for years. And the moat argument cuts both ways: if your edge is a furnace that runs on furnace time, your data flywheel turns slowly. A software rival iterates a thousand times a day. Periodic iterates when the oven cools.

The Researcher

The genuinely useful idea here is about where models get stuck. Çubuk says scientific discovery is "almost by definition what you haven't been trained on." That is correct and it generalizes. Fedus makes the matching point: if better reasoning at answer-time were enough, labs would have stopped training after GPT-4 and just gotten clever at test time. They didn't, because you still have to compress new knowledge into the weights. Translation for the reader: a reasoning model is only as good as what it already absorbed. On anything truly new to it, it needs fresh measured evidence, not a longer think. Their other real insight is that the hard step is characterization, figuring out what you actually made, not making it. The bottleneck is reading reality, not acting on it.

The Open-Source Advocate

Notice what the moat actually is. It is not the model. Fedus says their edge is proprietary physical-experiment data that lets them beat frontier models "even under the highest reasoning efforts," at lower compute. That means the base model is a commodity to them. The value is the data nobody else has. This is the lesson for any company: when the open-weight models (Llama, Qwen, Mistral) are 80% as good for a fraction of the cost, your advantage is never the model anymore. It is the measured, timestamped, hard-to-collect data you own and your competitor can't scrape. Periodic is an extreme version of a bet every operator should already be making.

The Compute Pragmatist

Çubuk's number is the one worth carrying out of this episode. Today's compute sits within three to four orders of magnitude of the Landauer limit, the hard thermodynamic floor for the energy it takes to erase one bit. Plain version: chips are now close enough to the physics ceiling that squeezing more performance out of the same materials is getting expensive, which is exactly why new materials are a high-value target. That is the real reason this company points at semiconductors and not, say, batteries first. The customers who feel the power and cooling wall hardest are the chip makers and the data-center builders. If materials discovery actually accelerates, the beneficiaries are NVIDIA, TSMC, and anyone paying the electricity bill on a GPU fleet.

The Builder

What could you use from this on Tuesday? The portable idea is embedding a model directly into your instrument or your tool with full context of intent and history, rather than running it as a dumb script. Their example: the AI caught a loading error in a machine by noticing inconsistencies across measurements no fixed rule would have flagged. Swap "electron microscope" for "your data pipeline" or "your QA step" and the pattern holds. Give the model the full intent and history behind a task, and it catches things your rules miss. The rest of this, the autonomous lab, the furnaces, is capital expense someone else carries. You integrate the pattern, not the furnace.

Where they disagree

The Skeptic and the Open-Source Advocate are looking at the same fact and scoring it opposite. Proprietary physical data as a moat is either the strongest kind of advantage (impossible to scrape, impossible to copy) or the slowest, because you gather it at the speed of ovens and electron microscopes while a software competitor compounds daily. Both are true. It depends entirely on whether anyone else is even trying to collect the same data. Right now, almost nobody is, so slow-but-exclusive wins. The day a well-funded rival builds the same lab, slow becomes the problem.

The second split is the Researcher versus the Skeptic on the "you can't zero-shot it" claim. The Researcher thinks it's a real, useful boundary on what models can do. The Skeptic thinks it's unfalsifiable marketing. They're both right too: the principle is sound, and it's also conveniently unprovable for years.

What this hinges on

One belief: in domains where the internet has no answer, does owning fresh real-world data beat having a smarter model? The council leans hard yes. That's the takeaway worth keeping. For the reader, forget Periodic's furnaces. The move is to audit where your own product depends on data that exists only inside your four walls, and to stop treating the model as your advantage. The model is rented. The data is yours.

Prediction: Periodic Labs' primary disclosed revenue through 2026 will be forward-deployed engineering and services (teams embedded at customer sites fine-tuning models on private data), not outcome-based or product licensing, as described in its next public funding or company announcement reported by late 2026 or early 2027.

Confidence: Medium. The founders described this as today's model themselves; outcome pricing was named only as a long-run aspiration.

Why: Fedus and Çubuk said plainly that revenue today comes from embedding engineers inside semiconductor customers, integrating tools into secure environments, and fine-tuning on customer data, with outcome-based pricing floated only as a future goal compared to the GitHub Copilot-to-Codex path. Outcome pricing in hard science requires attributing a discovered material or a yield improvement to your model, which takes measured results over furnace-speed cycles, and those cycles are slow by their own account. A company does not flip from services to outcome-based pricing in a single year when its iteration loop runs at the speed of physical experiments. The opposite, a shift to product or outcome pricing this fast, would require discovery wins already landed and attributable, which nothing in this conversation claims exist yet.

Revisit by 2027-03-31: We're right if Periodic's next public funding round or company statement still describes its revenue mainly as embedded engineering and services. We're wrong if it reports material revenue from outcome-based pricing or a licensed product.