AI Insiders reported yesterday that Poolside’s Laguna S 2.1, a 118 billion parameter mixture-of-experts design that activates 8 billion parameters per token and carries a 1 million token context window, is beating open-weight systems many times its size. The more interesting number is not the parameter count. It is the roster: co-CEO Eiso Kant told Latent Space that Laguna S came out of a group of fewer than 70 researchers and roughly 35 engineers, a team he says totals under 115 people, including himself.
That team, per Kant, runs somewhere between 10,000 and 20,000 experiments a month, a figure he offered as an estimate rather than a precise count. Laguna S itself trained in eight weeks. A larger successor, an internally named Medium model, started its own training run the same week Laguna S shipped and is expected to take about 39 days, according to Kant’s account on the podcast.
The infrastructure story matters more than the headcount. Kant says Poolside built its original training code from the ground up rather than forking an existing framework. The company then assembled what it calls its Model Factory. That system streams data straight into a training run instead of first bundling it and moving it onto a cluster, the batch method Kant describes as still common practice at other labs. Data is versioned and kept immutable, so every run functions as a controlled comparison rather than a one-off experiment. Kant also says software agents now write code, launch training jobs, and evaluate results on researchers’ own machines, with those agents starting to touch the pipelines that will train the next generation of models.
On hardware, Kant says Poolside currently trains on a cluster of roughly 10,000 Nvidia H200 GPUs, still the Hopper generation rather than Blackwell, and expects to scale that cluster substantially. That is a concrete number, but it comes from the company itself. Poolside has not published an independent audit of its GPU count, its monthly experiment volume, or its headcount, and the Latent Space interview offers no third-party verification of any of these figures.
Kant’s own framing is worth stating plainly, because it doubles as his thesis. He argues that roughly 95 percent of model building reduces to two variables: data quality and compute efficiency. Kant told the hosts that people have put foundation model companies “on this pedestal,” when in his words the work is “doing the basics right.”
Taken at face value, that claim reframes the competitive question for the rest of 2026. A cluster of 10,000 H200s is a real capital outlay, but it sits an order of magnitude below what OpenAI, Anthropic, and Google run. Poolside’s headcount is a rounding error next to those labs’ research staffs. If a group under 115 people can ship a competitive 118B model in eight weeks on homemade tooling, that says something about where the real constraint now sits. Frontier capability looks less gated by who can raise the most capital, and more gated by who can build the better training pipeline.
That is an unverified claim from an interested party. It is also a testable one.
Watch whether Poolside’s Medium model and its successors hold their benchmark position once the company’s cluster scales well past today’s 10,000 H200s. That is the point where the small-team, homemade-pipeline story either compounds or runs into the same limits that made large compute budgets necessary in the first place.
From an interview with Poolside co-CEO Eiso Kant on Latent Space, July 2026.
