Poolside for $6B, Cognition at $48B: AI coding is segmenting, not consolidating

Two deals three weeks apart tell you how the AI coding market is actually structured. On August 21, Nvidia signed a non-exclusive licence with Poolside for $6B plus a $1B investment at a $12B pre-money valuation, according to Newcomer; about 109 people received offers to join Nvidia, and the company keeps operating. On September 8, Cognition, the company behind Devin, raised more than $2B in a Series E at a $48B valuation, less than four months after a round at $26B. Its run-rate revenue grew from $492M to roughly $900M; Goldman Sachs is scaling Devin to thousands of engineers, and Cognizant reports that AI now generates about 30% of its code. One deal is a licence plus acquihire. The other is a growth round on a buyer base that is paying. Neither is consolidation.

Two templates, not one market

The Poolside structure is new enough to name. Nvidia did not buy the company; it bought a non-exclusive licence to the models and made offers to the team. Poolside continues, presumably with a smaller team and a large cheque. That is a template for what a chip vendor wants from a coding lab: the model and the people, not the product or the customers.

Cognition's round is the opposite bet. Investors are paying for a product with enterprise revenue and a distribution story (Goldman, Cognizant). If AI coding were winner-take-all, the money would not be flowing into Cognition at the same time as into Cursor and into Anthropic's Claude Code business. It flows into all of them because they win different segments: autonomous agents for backlog work, IDE-native assistance, terminal-first agents for engineers who live in the shell. I mapped those segments in the 2026 tooling comparison, and the deals since have confirmed the map rather than redrawn it.

A market where the chip vendor licenses one player and the banks fund another is a market with several winners and no throne.

Why vendor-neutral beats picking a winner

The enterprise data from the same window points the same way. AT&T, per The Information, cut costs on part of its coding and AI workloads by up to 56% with roughly a 2% drop in quality by routing simpler tasks to open models and keeping frontier models for the hard work, at about 45 billion tokens a day, with 40% of queries already on open models and a target of 60-70%. Goldman Sachs now describes that "model optimisation layer" as the key bottleneck in enterprise AI. Nobody who routes 45 billion tokens a day wants a single vendor deciding their unit economics.

There is also a resilience argument. GitHub's outage on August 17 ran about eight hours, was a capacity failure rather than a bad deploy, and was made worse by retry storms around Copilot. Every agentic coding pipeline that assumed GitHub was always there stalled. Picking one vendor for everything creates that single point of failure by design.

What a neutral adoption looks like in practice

• Standardise the process, not the tool. Review rules, test gates and prompt conventions live in the repo so any agent can follow them.
• Route by task. Cheap open models for summaries and boilerplate, frontier agents for the hard change. The routing playbook is the mechanism; the AT&T numbers are the proof it scales.
• Measure quality per task, not per vendor. A 2% quality drop is acceptable in some workflows and a firing offence in others; you only know which if you measure.
• Keep two agents installed. The switching cost is your negotiating position and your outage plan.
• Renegotiate annually. At these valuations pricing will move; an intro price expiring on a fixed date is the normal case, not the exception.

The limits of this reading

Every number here is a company or investor figure. Run-rate revenue is not audited revenue, $48B is what the last buyer paid and not what the next one will, and the Poolside terms come from Newcomer rather than from Nvidia. "30% of code generated by AI" at Cognizant says nothing about how much of that code survived review. And "segmenting" is an inference from where capital is flowing today; a downturn in AI funding would consolidate this market fast, in which case the neutral posture above is still the right one, for a different reason.