Poolside.

Somewhere in San Francisco, a model is reading its own homework.

It is late. The cluster hums - ten thousand NVIDIA GPUs, give or take, draw power somewhere on the AWS spine. A model called malibu writes a snippet of Python, runs it, watches it fail, rewrites it, and tries again. No one is babysitting. There is no human grader scoring it like a high-school essay. The compiler does that. So does the test suite. And every cycle, the model gets a little less wrong.

This is not OpenAI. This is not Anthropic. It is not even, despite the obvious resume comparison, GitHub. It is Poolside - a two-year-old company that has somehow already raised $626 million, signed a strategic deal with AWS, and convinced Nvidia to commit up to a billion dollars more. The pitch is unfashionably narrow: build the best foundation model in the world for one job, software engineering, and sell the entire stack - models, training infrastructure, IDE assistant - to enterprises that would rather not ship their source code to a chatbot in Seattle.

Autocomplete is not engineering.

The dirty secret of the first wave of AI coding tools is that they were, mostly, very good typists. They guessed the next token in a function the way your phone guesses the next word in a text. Useful, occasionally clever, often wrong in ways that only show up at 2 a.m. on a Sunday.

Software engineering, the actual job, is something else. It is reasoning across a sprawling codebase. It is tracking why a service crashes on Tuesdays. It is reading a half-written test and inferring the contract the author never bothered to write down. A model that finishes lines is not a model that finishes work.

So Poolside picked a fight with the data wall. Public code on the internet is finite. Stack Overflow runs out. GitHub runs out. The companies racing to train the next frontier model on the open web were going to hit a ceiling, then renegotiate licensing terms with publishers, then hit it again. Poolside's answer was to generate its own training data by having the model write code, execute it, and learn from the results. They call this Reinforcement Learning from Code Execution Feedback, or RLCEF, an acronym that is mercifully easier to read than it is to say.

Two engineers who had seen this movie before.

Jason Warner spent five years at GitHub, the last several as CTO, and was there when Copilot went from internal demo to default tab-key habit for millions of developers. Eiso Kant had been building developer tools for over a decade - Source{d}, Athenian - the kind of companies you only know about if you also build developer tools. They met in 2017, kept in touch, and in late 2022, around the time everyone else was making mood boards about ChatGPT, they decided to start a company.

The bet, on paper, was almost impolite in its confidence. Most startups in 2023 were arguing they did not need to train their own foundation models. Poolside argued the opposite for one vertical - code - and then went further. The founders have said publicly that most companies should not train their own foundation models. They are, charmingly, including the bulk of their would-be competitors in that statement.

Caption: Two birthdays, four rounds, one cluster the size of a small country's grid demand.

One stack, deployed where your code already lives.

Poolside ships three things and pretends, politely, that they are one. There is malibu, the flagship foundation model, sized for the hardest engineering work - refactoring across packages, generating tests that actually exercise edge cases, debugging code that nobody currently working at the company wrote. There is point, a smaller, faster model meant for the kind of inline completion you want returning in milliseconds, not seconds. And there is the Poolside Assistant, the layer of plugins and CLIs and chat that lets a developer actually use the thing inside VS Code, JetBrains, Visual Studio, Coder, or the terminal.

The trick is where it runs. Poolside is built to be deployed inside a customer's environment - their VPC, their on-prem cluster, their AWS tenancy - so the model that learns from a bank's code does not, at any point, leave the bank. This is less glamorous than a public chatbot. It is also why regulated industries return the calls.

Capital is not validation. Customers are.

Capital is the easy part to write about, so we will get it out of the way. Poolside has raised, by public reporting, around $626 million across seed, Series A and Series B, with an additional commitment of up to $1 billion from Nvidia announced in October 2025. The Series B alone was $500 million at a $3 billion valuation, led by Bain Capital Ventures with participation from Nvidia, eBay Ventures, DST Global, Citi Ventures, and a parade of others.

The harder part is product traction, because Poolside sells inside firewalls and most of its customers do not love being named on a slide. Public reporting and the company itself describe a customer base concentrated in Global 2000 enterprises and public-sector agencies - financial services, defense, and other industries where the phrase "send our source code to a third party" is grounds for a very uncomfortable meeting with legal.

AGI, but useful.

Poolside is unusually direct about wanting to build artificial general intelligence. It is also unusually direct about doing so through software, which is, on balance, the most measurable domain humans have invented. Code either runs or it does not. Tests pass or fail. Latency is a number. There is no rubric debate.

That choice of domain matters. A model that can plan, write, test, and ship a working software system is doing something a lot like reasoning. Poolside's argument is that if you can teach a model to do this rigorously, on real codebases, with the compiler as judge, you are working on AGI - you are just doing it where the grading is honest.

The interesting question is not whether AI writes code. It is who owns the model that does.

For most of the past decade, the answer to "where does our intelligence live" has been roughly: in someone else's cloud. The Poolside bet is that this answer is unstable. Banks, governments, defense primes, large industrials - they will not be content for their software production capacity to depend on an external vendor's chat API forever. The intelligence layer will get pulled inside the firewall, the same way databases and search did before it.

If Poolside is right, the AI infrastructure stack of 2030 looks less like a chatbot subscription and more like a deployed system - a model the customer fine-tunes on their own code, runs on their own GPUs, and improves over time without sending anything to anyone. If Poolside is wrong, well, the cluster will at least have made some excellent music.

That is what Poolside is - not a chatbot, not an autocomplete, not even, exactly, a research lab. A company building the layer of software production that, if it works, no one will think to talk about. The highest compliment in infrastructure is to become invisible. Poolside is angling for it.

Links, sources, and the videos worth your time.