Sequoia Capital
Confidence 0.75 · 6 sources · last confirmed 2026-10-25
US venture-capital firm. In this wiki Sequoia appears as a publisher of frontier-AI interviews — its Training Data podcast and AI Ascent event series — rather than as a technology subject. Its conversations are recurring primary sources for how AI leaders frame the field.
What the wiki holds via Sequoia
- “Why the Model Eats the Harness” — Sonya Huang interviews Logan Kilpatrick (Google DeepMind); source of the model-eats-the-harness and jagged-superintelligence framings.
- “From Vibe Coding to Agentic Engineering” — Sequoia AI Ascent interview with Andrej Karpathy.
- “How Companies Are Building Their Own Intelligence” (August 2026) — Sonya Huang’s opening keynote at Sequoia’s Own Your Intelligence event, to roughly 80 portfolio founders and AI leaders.
The August 2026 shift: from publisher to advocate
The first two sources are Sequoia hosting other people’s arguments. The third is different in kind, and worth flagging on this page: it is Sequoia making a directional bet in its own voice and convening its portfolio to act on it. Huang’s thesis — “not your weights, not your product” — is that the defensible layer is migrating from the application surface to the intelligence underneath it (“the race for the application layer” becoming “the race for the intelligence layer”), and that with Kimi K3 and GLM-5.2 near the frontier, owning a model stack has flipped from a performance sacrifice to a potential performance edge. The event was structured as “half a rallying call and half technical how-to”, with workshops from Fireworks, LangChain, Mercor, Trajectory and Harvey.
Two things follow for how the wiki should read Sequoia sources. First, the firm is no longer only an attribution node here — on the sovereign-AI question it is a party with a position, and the companies named as evidence for that position are largely its own portfolio. Second, its house framing (neolabs, own-vs-rent, the application-companies-are-the-newest-labs claim) is now something the corpus should treat as an interested framing to be tested, not as neutral market observation. See the source page for the full scope warning.
Sonya Huang appears in two of the three sources — as interviewer in the Kilpatrick episode and as speaker in the keynote — but is not yet promoted to an entity page; she is not an author: on either (both are attributed to the channel per the wiki’s video convention). Worth promoting if a third Sequoia source turns on her specifically.
Otherwise Sequoia remains an attribution/publication node: for the interview sources, the substantive claims live on the interview subjects’ pages and the concept pages they touch.
A former partner backs a portfolio company on his own (added 2026-10-04)
Bloomberg Tech, September 2026: Roelof Botha led Flow Engineering’s Series A at Sequoia “just more than a year ago” and now, no longer at the firm, invests his own capital and sits on Flow’s board. He names reindustrialisation as part of the original thesis for the company, and hardware’s weight among the most valuable technology companies as the reason it lasts.
A partner on the diffusion gap and owning intelligence (added 2026-10-17)
Alfred Lin at the Stanford AI Club (October 2026): a Sequoia partner of fifteen-plus years states the firm’s application-layer thesis (“there is value in the application layer… this whole notion that software is dead is kind of BS”) and the own-your-intelligence argument from Sonya Huang’s keynote, with portfolio examples: Harvey, OpenEvidence, Profound, Rogo, Chai Discovery, LangChain, Reflection. He also explains why check sizes keep rising without hurting returns: the largest company’s value has grown “15x in 15 years,” so outcomes are bigger. Published on the Stanford AI Club’s channel, not Sequoia’s.
Training Data on robot data (added 2026-10-25)
Gao (Mecka) on Sequoia’s Training Data (Oct 2026): Sonya Huang interviews Josh Gao of Mecka on manufacturing training data for robotics: sensorized people, tactile gloves, a converged data mix, and deployment through a new kind of systems integrator. The wiki’s first source on robot training data.