Keith Rabois on founder mode, Airbnb's pivot, AI consolidation, and why micropayments will never work

Summary

Keith Rabois, managing partner at Khosla Ventures, covers a wide range of territory — from Airbnb's strategic pivot to AI consolidation to why micropayments will never work — and the throughline is a investor who has sharply updated his views on AI faster than almost anyone in his peer group.

AI is now the majority of what Rabois bets on

Of his last 11 investments, six are AI-based. A year ago, that number was zero. The shift isn't just in his portfolio — at the annual Khosla Ventures CEO Summit, which drew 150-plus CEOs and speakers including Sam Altman, John Collison, and the Databricks CEO, AI infused every conversation across every vertical, not just the dedicated AI track. Last year it was a slice of the room; this year it was the room.

The practical use cases he sees as no-brainers for any company are engineering productivity and customer support. Ramp's internal data puts engineer productivity up 46% from AI tooling, a figure Rabois cites as a clean, strictly measured signal.

Foundation model consolidation

Brad Gerstner of Altimeter, speaking at the summit, argued there will be only one successful foundation model research lab — not a category that VCs should be funding. Rabois largely agrees, but adds his own frame: the next generation of transformative companies probably won't look like an LLM or a research lab at all, just as OpenAI doesn't look like a search engine or a social network. OpenAI, in his view, is already dominating Google — ChatGPT will become the default interface for the majority of people on the planet, displacing Google Search.

Airbnb's pivot

Rabois thinks Ben Thompson's critique of Airbnb's services expansion — that the platform economics push users to disintermediate once they find a reliable local provider — misses a key nuance. The disintermediation risk is real for home-city use cases like a regular hair stylist, but Airbnb's core user is a traveler moving across cities. Someone going from New York to Cincinnati to Dallas to Utah has no incentive to go off-platform, and Rabois argues that curating local experiences for visitors rather than residents is where Brian Chesky's bet actually sits. Thompson's critique may be valid at the edges, but it's the wrong lens for the traveler use case.

On whether Airbnb needs an advertising product — Uber and Instacart have built meaningful ad revenue — Rabois is skeptical. Airbnb's take rate runs around 13%, high enough that the economic pressure to layer in ads doesn't exist the same way it does for low-margin platforms. He acknowledges advertising can be a genuine value-add when done well, citing a Google internal study showing users exposed to both paid and organic content report higher satisfaction than those seeing only organic results.

Micropayments will never work

Rabois disagrees with Ben Thompson's argument that micropayments could replace lost advertising revenue as AI agents erode the open web's economics. His case goes back to a lesson from PayPal circa 2000–2003, when he explored micropayments as a new market and Peter Thiel shut the idea down. The argument is simple: digital content has gross margins in the high 90s, so transaction fees — even at 10–30% — aren't the real problem. The problem is consumer friction. If a user has to do anything beyond a single click to pay for a short piece of content, they won't do it. Rabois is explicit that this is a demand problem, not an economics problem.

Stripe's original pitch to Rabois in 2010 was partly built on the idea that simpler payments would unlock micropayment-driven content monetization. It didn't happen. Substack works, but Rabois argues that's a differentiated-voice subscription business, not a micropayment system — and genuinely differentiated voices are extremely rare. He names Ben Thompson and Bill Simmons in his prime as the kind of writers who can command a premium, and notes the bar for everyone else is nearly impossible to clear.

AI safety is a dead debate in the US

Rabois is blunt: the AI safety debate as it existed in 2023–2024 is effectively over in the United States. The policy priority has shifted to accelerating AI development and ensuring Western, specifically American, AI dominates over China's. He reads Meta's inability to use safety concerns as cover for Llama 4's training issues as evidence the argument has lost its cultural legitimacy even as a deflection.

The carve-out he makes is model observability and bias — understanding why a model produces a given output and correcting it. He draws a parallel to TikTok's algorithm potentially steering user beliefs, and says foreign-controlled AI models pose a similar manipulation risk. He's watching startups in model interpretability closely, describing the technical challenge as something like brain surgery. That concern, though, is distinct from the existential-risk framing of traditional AI safety, which he dismisses as effectively anti-acceleration rhetoric.

American AI goes global for energy and capital, not markets

On Jensen Huang, Sam Altman, and Scale AI touring the Gulf and other sovereign wealth regions, Rabois frames the trips as resource acquisition, not market expansion. The primary need is power and manufacturing capacity at scale. Wealthy sovereign nations with surplus capital are being asked to fund energy infrastructure and advanced manufacturing that the US AI ecosystem needs. Exporting American AI as the default global standard is a secondary strategic argument, but the real driver is unlocking the physical infrastructure required to keep training frontier models.