Kleiner Perkins' Leigh Marie Braswell on the AI coding wars, Windsurf, and what comes next

Summary

Leigh Marie Braswell, partner at Kleiner Perkins and Windsurf board member, makes a case that the AI coding market is bigger and more durable than most investors assumed — and that the M&A and competitive dynamics now playing out are only going to intensify.

The Windsurf story

Braswell describes being genuinely surprised by how fast tools like Windsurf scaled. Her prior instinct when GitHub Copilot launched was that the market was effectively closed — one dominant player, done. The opposite happened. Cursor and Windsurf carved out real product differentiation, the addressable market expanded beyond professional developers into prosumers with no coding background, and enterprise adoption followed bottoms-up. Her father-in-law, unprompted, told her she should check out Windsurf. That anecdote carries the point: the category has broken out of Silicon Valley.

Windsurf's enterprise motion runs both ways — a dedicated enterprise sales team pulling deals top-down, and prosumer adoption pushing bottom-up into company accounts.

Pricing and billing

On pricing, Braswell sees per-seat as the floor, with usage or credits layered on top to capture inference costs. Outcome-based billing — charging per resolved ticket or approved line of code — is happening in customer support but hasn't taken hold in coding. She doesn't rule it out, but the current market is competitive enough that per-seat-plus-usage is where most players are landing.

Model provider tension

Braswell declines to speak directly to Windsurf's reported difficulty accessing Anthropic's newest models, but she doesn't deflect the underlying tension. Every major player in coding — the IDEs, the foundation model labs, the hyperscalers — is converging on the same prize. She expects all of them to end up competing against each other within six to twelve months, and argues the companies that can combine synchronous assistant interaction with asynchronous task delegation from a single interface will have a structural edge. No product does that cleanly yet.

The model providers' incentive to own the consumer relationship in coding is real. Braswell frames it as something the industry will have to resolve, not something that goes away.

TAM and the labor replacement question

When founders pitch labor-replacement TAMs — adding up global payroll for a category and calling it the opportunity — Braswell's instinct is to pressure-test rather than dismiss. The coding market taught her that once software gets dramatically better, buyers spend more, not less, and the frame shifts from incumbent software spend to work replacement. Both markets stack. She thinks that dynamic probably applies across the categories now seeing heavy competition: legal AI, medical documentation, sales tooling, customer support.

Kleiner's portfolio reflects those bets. The firm backed Harvey's Series B, Ambiance (medical scribe, seed and Series B), and Nooks in sales AI. Partner Amun incubated Glean before Braswell joined. She flags customer support as the one major battleground where Kleiner doesn't yet have a generalized position.

Competing with hyperscalers

Braswell pushes back gently on the idea that Google IO or Microsoft Build announcements are category-killers for startups. Her read, consistent with Ben Thompson's post-IO argument she references, is that the hyperscalers are most dangerous where a market directly threatens their core — search, social, prosumer. In vertical enterprise applications, startups that move fast and find a defensible wedge can still win, because building internally is slower than acquiring or partnering with a team that already has the expertise and distribution.

She points to Neon's partnership with Databricks as the cleaner model: a database infrastructure team with deep expertise pairing with a platform that needs persistent memory and storage for AI agents. The logic is straightforward enough that she says she's now actively mapping which categories make similar sense for each hyperscaler.

Video generation

On VO3, Braswell is bullish but her more durable point is about platform risk. Because so many labs — closed and open source — are now producing competitive video models, a startup that stays model-agnostic and offers best-in-class selection across providers may have more durability than one tied to any single model. She's less worried about platform concentration in video than the conversation might suggest.

The M&A read

The Windsurf acquisition sits at the center of a broader argument: the "don't build a ChatGPT wrapper" meme was wrong, or at least premature. Application-layer companies with real distribution and scaled revenue are exactly what hyperscalers need, and building those capabilities internally is slower than ever as the pace of AI development accelerates. Braswell expects more consolidation, and says the companies with the sharpest team, product differentiation, and bottoms-up distribution are the ones that get picked up — or win outright.