When most people talk about AI coding tools, they are talking about products: Cursor, Devin, Claude Code, GitHub Copilot. The editor, the agent, the assistant. What they are less often talking about is the layer underneath all of those products: the models that power them.

Poolside is betting the model layer is where the real value accumulates.

The company raised $500 million in a Series B led by Bain Capital Ventures in October 2024, at a $3 billion valuation. Nvidia, DST Global, StepStone Group, Citi Ventures, and HSBC Ventures participated. The round brought Poolside’s total funding to $626 million since its founding in May 2023, making it one of the best-capitalised AI coding companies in the world despite building something most developers never interact with directly.

By late 2025, Poolside was reportedly in discussions to raise further capital at a $12 billion valuation, a 4x increase in roughly twelve months.

What Poolside Actually Builds

Poolside builds code generation foundation models. Not an IDE like Cursor. Not an autonomous agent like Devin. Not a coding assistant like Claude Code. The models themselves.

The positioning is that specialised code generation models, trained specifically and extensively on software development tasks, will outperform general-purpose language models applied to code. The same argument that has been made across domain-specific AI: a model trained to be excellent at one thing beats a model trained to be generally capable at many things, within that one domain.

Whether that argument holds in the long run is genuinely uncertain. OpenAI, Anthropic, and Google have all invested heavily in making their general-purpose models excellent at code, and the evidence from benchmarks suggests the gap between specialist and generalist models on coding tasks is narrower than it was two years ago.

Poolside’s response to that is that the specialisation advantage is not just about benchmark performance. It is about understanding the full context of software engineering: how code evolves over time, how large codebases are structured, how tests relate to implementations, how debugging requires reasoning about system behaviour rather than just syntax. Training on code alone is not sufficient. Training on the full artefacts of how software is actually built is the harder and more valuable thing.

The Infrastructure Play

Poolside’s funding is partly a bet on the model and partly a bet on the infrastructure position. A company that owns a leading code generation model is not just selling a product. It is a potential partner for every code editor, every AI coding tool, and every enterprise AI platform that needs a reliable, high-quality foundation for software-related tasks.

That is a different business model from selling seats in an IDE. It is closer to the position that OpenAI occupies with its API: a foundation that other products are built on, with revenue that scales with the industry rather than with individual product adoption.

Nvidia’s participation in the Series B is notable in this context. Nvidia does not invest in AI companies for financial returns alone. Its investments tend to signal which companies are buying and will continue to buy significant compute. Poolside’s model training requirements are a meaningful compute customer, and Nvidia’s participation validates both the company’s infrastructure scale and the seriousness of its model ambitions.

A Different Bet in the Same Category

Cursor, Cognition, and Poolside are all betting on AI and code. But they are betting at different layers of the stack.

Cursor is betting that the IDE layer is where users and enterprises are willing to pay, and that owning the best developer experience creates the most durable business.

Cognition is betting that the autonomous agent layer, where the human is out of the loop for whole tasks, is where the value shifts as models improve.

Poolside is betting that the model layer, the foundation underneath both of those, is where the most durable value sits. If you own the best code generation model, you do not need to win the IDE battle or the agent battle. You become infrastructure for whoever does.

All three bets could turn out to be right simultaneously. The AI coding stack has room for a platform, an agent layer, and a foundation model layer to all be large businesses. But the bets reflect genuinely different views about where the market will concentrate, and watching which of them compounds fastest over the next two to three years will be one of the more informative data points about where enterprise software value accumulates in the AI era.

---

Enterprise DNA tracks the AI coding landscape as part of its research into how AI is changing software development economics. If you want to understand what this means for your engineering investment strategy, talk to us.