10,000 GB300s for Four Weeks: Reflection’s Open-Model Ledger

Misha Laskin, co-founder and CEO of Reflection AI, comes on *No Priors* with Sarah Guo and Elad Gil just after shipping Beam, the company’s first open-weight model. For the previous twelve months the mission had been frontier open intelligence, widely accessible. He reaches for a Reid Hoffman line: they were assembling a plane while flying it. A year earlier Reflection had been around 30 people. Building a real rocket-ship model, he says, takes on the order of a hundred-plus researchers and engineers. The company is now around 300, with teams for pre-training, mid-training, and reinforcement learning, and it has trained its first models end to end.

Laskin was a researcher at Google DeepMind and holds a physics PhD. He and his co-founder Ioannis — Yanis in the conversation, a founding engineer at DeepMind, a key contributor to AlphaGo, and the person who led the reinforcement-learning work on Gemini 1 and 1.5 — started the company about two and a half years ago. Asked what was hardest, compute, talent, or data, he refuses the ranking. He gets the same question at all-hands. Everything is extremely hard. You have to get about 30 things right: hiring, retention, a mission that keeps people rowing the same direction, data, compute, and infrastructure that makes the compute actually usable. At a big lab those tools already exist. As a researcher at DeepMind he took them for granted because they just worked. As a founder he had to build them.

Beam is a 500-billion-parameter model with 23 billion active parameters, trained for coding and agent tasks. Pre-training used 6,000 GB300s for a few weeks; with infrastructure efficiencies they can now do that run in about 12 days, maybe less. Reinforcement learning used a little over 10,000 GB300s for four weeks. More FLOPs went into RL than into pre-training. Laskin is happy to say so out loud. The nice thing about an open-model lab, he notes, is that you can talk about the work openly. RL jobs are more complex: inference at scale, sandboxes for agents, and training at once. He has not seen a 10,000-GB300, four-week RL run documented in open source before.

The efficiency story sits on two clocks. When it was only human researchers, pre-training bought about 7x compute-efficiency gains per year, measured against a concrete loss: you change something and reach the same loss with less compute. Across the full system, especially RL, he puts the figure closer to 30x per year, depending on how good the model is at improving itself, roughly four or more times faster than researchers alone, and likely to accelerate. Intelligence extracted per training FLOP is rising. Revenue per FLOP is rising too, because more intelligence is packed into each one. That is why he does not expect the CapEx asymptote tomorrow, even though frontier labs are already putting hundreds of billions of dollars in and trillions in the next year or two seem unlikely.

The plot that pulled him into AI was AlphaGo’s: it never stopped improving, and they cut it off because it had already beaten a world champion. Apply that recipe to something economically valuable, and training becomes an economic question — how much money do you want to put in to keep going. He says the field is there. Beam’s RL plots, which will be in the tech report, keep going up. It is a matter of compute. Two years ago that was not true. The work now feels more like an engineering discipline than a wild-west science. The axes are set: pre-training, synthetic data, reinforcement learning. Hardware has co-optimized around what works, a hardware lottery that makes unrelated ideas harder to fund. He is not seeing a slowdown in pre-training efficiency, and he sees a lot of headroom in RL, but as engineering discovery rather than a new science.

The original bet was not to pre-train. Gemini 1 and 1.5 were mostly chat: no real coding, nothing agentic, about 95 percent of compute in pre-training. Laskin and Ioannis were RL researchers. Alignment for chat worked, with a ceiling. The idea was to take that recipe into mathematics and coding and make the systems agentic, as an independent lab, more capital-efficiently, on top of someone else’s open weights. Mistral had a small model. Llama 2 had just landed. Llama 3 was coming. They assumed a high-quality Western open base would keep showing up.

RL worked faster than the ImageNet-to-scalable-RL analogy suggested. That journey had taken about four years after 2012. Every year, Laskin says, moves faster than his previous prediction. About a year in, the good open models were coming from China. There were not really good Western ones. There were enterprise reasons and geopolitical reasons to change course. There was also a research fact: to make RL work well at scale, you actually need to pre-train the model. The stages are tightly coupled. So they combined the RL bet with building models end to end.

Catching up to the frontier, he argues, is cheaper than sitting on it, because talent lets you cut exploration and execute known ideas. The price of catching up still rises as the frontier moves. On the order of a year or 18 months ago it was hundreds of millions of dollars. Six months ago, single-digit billions. Heading into next year, on the order of ten — tens of billions. His heuristic is a 4x compute multiplier each generation: 100,000 H100s, then a big run such as Astra on 100,000 Blackwells, then around 100,000 Vera Rubin GPUs. CapEx cannot rise forever relative to revenue. He agrees an asymptote is coming. He does not think it is here, because efficiency and dollars per FLOP are both still climbing.

Code and agentic work are the foundation, Laskin says, and the intelligence is jagged. He was surprised how quickly Beam reached a capability other labs had only recently found, which looks like generalization. At the same time, different versions of Terminal Bench do not transfer cleanly across harnesses. Once a general agent works and you have data for a new harness, it adapts fast. The economically valuable move is then to find pools of data, or more precisely evaluations. You are not always training on the customer’s data. If you can stand up a good eval for their tasks, you can generate synthetic data that approximates the distribution and get generalization. Finance KYC and compliance, cyber defense, and legal agent workflows all follow that pattern.

What stands out on Beam is reasoning efficiency: not only whether the agent can do the thing, but how quickly. Ten minutes in a chat window is a different product from one minute. Beam tends to be three to four times more efficient than models in the same capability class, and on the order of 10x against some larger models. The cause, as he tells it, is a strong pre-training base for reasoning, then RL at what they believe is the largest scale yet in open source. The more RL you run, the higher the capability and the faster the system solves the problem — the same arc as AlphaGo, which meandered at first and searched smartly by the time it reached Lee Sedol. The first AlphaGo systems also imitated expert amateur games before RL. You need both. The byproduct is a workhorse for enterprises and sovereign customers.

Commercializing open and closed models is more similar than it looks, Laskin says. You are trying to maximize inference. The difference is rental versus ownership. A token is a slice of a whole stack — agent harness, model, inference software, cluster management, GPUs — amortized into a price. People start by renting apartments. As they grow up they want a house, even if it is a headache. AI has matured to the point where enterprises spend enough on rented intelligence that they want control. The open checkpoint is permissive, but customers still need everything around it. Reflection’s job with large enterprises and sovereigns is to supply those tools. Services are how you walk in, unlock a valuable use case, and drive inference demand. Services, in that sense, are a demand driver for the inference business.

Six months ago, on gateways such as OpenRouter and Vercel, the mix was majority closed, minority open, about 70/30. It has flipped, almost exactly, to 70/30 open. He thinks that accelerates, toward something like operating systems: 95 percent-plus of servers run Linux, and Microsoft and Apple are still among the most valuable companies on earth. Most token demand will go open. A lot of economic value will go open too, because even a fully permissive model still needs expensive GPUs and a different kind of cloud than CPUs. Dedicated open-model inference providers tell him 90 percent-plus of their tokens are already customized, but today’s biggest customers are AI natives — Cursor, Cognition, Harvey — whose whole product is one customization. He expects enterprise to run the other way: customized systems first, not customized weights. Take an open model, do not fine-tune yet, wrap a harness around a KYC flow, and only later touch the weights.

AI natives went closed, then open, then deep customization quickly, because their products are data collectors. Enterprises architect around existing products and tools, so the friction is higher. In practice they do not sprint straight to ownership. They ramp significant closed-model workloads, then consider open. Spending $100 million-plus a year on closed models, which he says is not uncommon, is the point at which the cost conversation starts. Compute shortage adds an on-prem twist: hyperscalers cannot always supply enough, rates are hard, and an enterprise might buy bare metal from someone like Dell and still need a partner between the iron and a working deployment. Just selling inference into that setting, he has not seen work. You have to walk in around the highest-value thing they already want to scale.

Durable revenue, in Laskin’s formula, is intelligence density times compute times trust — how good you are at solving the customer’s problem. Compute is scarce, and he does not see that as temporary. Arguments that this layer or that layer is getting commoditized miss the point: everything across the stack is hyper-competitive, including model margins under open-closed tension. AI natives already negotiate closed deals aggressively because the open ecosystem exists. So long as great intelligence density is accessible, many companies can succeed. The closed labs’ version of winning is that revenue shoots up and they amass more compute than anyone else. Serving great intelligence, holding a lot of compute, and having customer trust is what it takes.

Two years ago the notable open weights were Western: Mistral in Europe, Meta as a Europe-U.S. mix. Then Chinese open weights rose, in some accounts by distilling frontier closed models at industrial scale. Laskin’s first move is to call that a massive benefit. Western companies built more durable businesses because of it. Imagine the world where it did not happen, and there are simply no open models, and a tall-poppy application gets subsumed into a closed provider that needs the revenue. You also do not want a monopolar world: one or two closed labs, or all ownable intelligence coming from one country. Ideally there would be 10 or 20 open sources. CapEx is too high for that. At least two would be great. Competitive tension is the point, not West versus China as a morality play.

Chinese labs, he says, will keep being good. They distill closed models at industrial scale. They train on copyrighted PDFs that would not be free under Western rules. They have had direct or indirect state support; the companies made no revenue for a long time, which Gil describes as a Chinese-government subsidy to Western enterprise. Laskin agrees, and notes they are now turning into businesses: open abroad, and inside China the equivalent of closed labs. Why keep releasing weights in the West? His hypothesis is that China is a captive market where the open-model builders already are the intelligence providers, and that continuing to publish is geopolitically useful. Other countries building on your stuff is leverage, the way 5G, fiber, and Belt and Road locked people into infrastructure. Open models are Trojan horses for the software and hardware that follow. An allied country might run a Chinese model on American chips today; Huawei-class vendors will offer a full stack and lock the supply chain. America used the same playbook with the internet, open protocols, the dollar, and a navy that makes free trade work. On open source, he says, the United States has been on the back foot. No longer, or at least not for long. The competitors in China are really good, and the U.S. ecosystem is only now forming.

Closed labs’ main criticism of open models is safety and controllability. Laskin calls the concern legitimate and the dominant worldview dogmatic. If you designed safety from first principles, before closed versus open, you would notice that software went through this. In the early 1990s, strong encryption was closed. After catastrophic failures — a handful of engineers, unintended consequences, easy hacks — the protocols opened, and cybersecurity as a field was born. The default state of the world, he says, is that openness is safety. Models are software with a longer tail of vulnerabilities because they are black boxes. Linus’s law says that with enough eyeballs, all bugs become shallow. He believes most security and safety holes go the same way. Today a few hundred safety researchers inside closed labs cannot cover the tail. The empirical picture he offers: a very powerful closed model hacked into another company, and the only remediation was open models. Hugging Face, in a separate incident, reverted to open source because closed-lab guardrails blocked the defensive use.

Gil splits “safety” into cyber, bioweapons, and existential risk, and objects to treating them as one thing. Laskin answers with a spectrum from empirical reality to theory. Cyber capabilities that were sci-fi last year are real; that does not license folding in atmosphere-on-fire nuclear fears or 1990s grey-goo nanotech. Culture humanizes AI, so public perception peaks at doomsday. Leaders quoting a 10 percent chance we all die are quoting a made-up number. As a scientist he wants a distribution peaked on reality: misalignment and unintended consequences in the next six months, dinner conversation for the rest. Alignment science so far has been boring. There is no magical equation. Early pre-trained models were toxic, people thought that could not be fixed, and enough post-training data patched it and generalized until jailbreaks got hard. The daily work is whack-a-mole: find a hole, build data, patch, repeat, often by asking a language model to detect the failure. That is an argument for a hundred thousand people looking, not a few hundred.

The harder claim is that even if the tools are manageable, ordinary people should not have them. Laskin is unconvinced. A model trained to be safe is not trivial to turn into a weapon; it is not impossible either. Cybersecurity already lives with offense and defense on the same internet, and defenses tend to outweigh attacks. You cannot strip offensive capability without stripping defensive capability, which is what the Hugging Face episode showed. Absolutist “no one should have this except us” has rarely worked outside a few cases with almost no upside. What excites him two to five years out is science. He has been feeding language models his physics PhD. Two years ago they could only chat. A year ago they answered at undergraduate homework level. Six months ago they solved it at PhD level. Now they add information he had not considered. If you have the question, you can do a PhD in a week. Theoretical work is a chalkboard and will move first. Life science, materials, and chemistry will be slower, but still dramatically faster, especially where you can close a proprietary experimental loop.

He toured Stargate some weeks back and noticed the parking lots. Data-center projects, he says, create tens of thousands of high-paying jobs, the way a factory made a twentieth-century town. They are noisy and should sit away from houses, but the tax revenue is real. Research direction on Beam is led by Ioannis. The final model trained on a SpaceX cluster they received in July: a few weeks of pre-training, then synthetic data and RL, then ship, which means known high-impact bets did not all make it in. You exhaust those, then widen the risk, with small-scale probes, knowing some effects only appear at large scale. Models in the loop are a fast, eager colleague: hyperparameter search, granular infrastructure, more room for a scientist’s intuition. That is why annual improvement may have gone from roughly 7x to several times that. You need fewer people for the same effort next year, and you expand ambition instead, so they still feel understaffed. Three thousand researchers would not help. AlphaGo was on the order of ten people; a large effort now is on the order of a hundred. The missing headcount is applied: pods of five to ten targeting a capability, then the same eval-and-harness skill deployed as a forward-deployed engineer on real enterprise problems. Reflection would take as many of those people as it can get. The bottleneck is training, not demand.