Sonya Huang on AI's 'not your weights, not your product' moment and why application companies are building their own labs

Sonya Huang on AI's 'not your weights, not your product' moment

Sonya Huang, general partner at Sequoia Capital, argues the clean separation between foundation model companies and application companies is dissolving. Application companies are building their own research capabilities, and the production of intelligence itself is democratizing — not just the revenue from it.

The growth numbers reinforce the point. Anthropic, OpenAI, and xAI are growing at a pace Huang says she's never seen before, but even companies below that tier — Open Evidence, Glean, Factory — are building business fundamentals that previously took a decade to develop. They're doing it in two to three years.

"Not your weights, not your product"

The strategic shift Huang is watching is companies moving from treating open-weight models as a cost rationalization exercise to treating them as an existential imperative. Her framing borrows from crypto: not your keys, not your crypto. The AI version is that if you're making API calls to someone else's model, you don't own the weights, you don't have steerability, and the data flywheel doesn't accrue to you.

She points to converging pressure from multiple directions. Alex Karp talks about sovereign intelligence and owning the means of production. Saatchia pushes proprietary data. Jensen Huang champions open weights. And application-layer startups are actually getting good at building their own research. Sequoia is invested in Fireworks AI, whose tagline is "own your intelligence." Cursor went on this same journey two years ago.

“The thing that's different now is like it is an existential and strategic imperative for them. There's this phrase — not your keys, not your crypto. I think the AI version of this meme is not your weights, not your product. Because fundamentally, if you don't own the weights, if you're just making an API call, you don't have ownership, you don't have steerability, the data flywheel doesn't accrue to you. The Harvey team has put out pretty extraordinary research — their entire research team is seven people.”

What an internal AI lab actually looks like

For a scaled SaaS company — Huang uses $500M ARR as the hypothetical — the answer is almost never pre-training your own models. The realistic path is taking open-weight models off the shelf and adapting them for a specific domain. That draws from two talent pools: people with post-training experience from the labs (a population that has grown significantly as lab post-training teams have scaled up), and generalist engineers smart enough to navigate the now-mature post-training tooling stack.

The tooling ecosystem is what makes this viable. Fireworks provides post-training infrastructure, LangTrain handles evals, Trajectory handles continual learning, and Mercor handles data factory work. A few years ago, only OpenAI and Anthropic had this infrastructure in-house. Now it's composable.

Harvey is Huang's sharpest example of how far a small team can go. Their legal AI research team is seven people, they just released an entire RL environment, and their benchmark is state of the art for legal. You're not competing with Meta on talent budget.

Fine-tuning as a service

For companies that don't want to build even a small internal team, a market of neo-labs is emerging — Huang mentions Thinking Machines and Tinker alongside Fireworks and Mercor as options for post-training as a service. The use case determines the vendor: reinforcement learning on online data points toward Fireworks; synthetic data generation for a domain where you can't train on customer data points toward Mercor.

Huang's caveat is that outsourcing can't replace internal ownership of the strategy. Someone smart inside the company needs to own the technical roadmap and make judgment calls about which parts of the stack to build versus buy. Harvey does this well — it uses five or six infrastructure partners intentionally, while keeping the strategic decisions in-house.

The price war

On the model price war, Huang invokes Jevons Paradox. Sequoia's portfolio companies are seeing gross margins rise even as AI usage accelerates, because cheaper inference makes their unit economics better. Inference providers like Fireworks are seeing improving cohort metrics despite price compression, because volume more than offsets price. Huang reads this as an everyone-wins dynamic, not a race to the bottom — at least for now.

The opacity of Anthropic and OpenAI as private companies is obscuring how fast the top of the market is still growing even with open-source alternatives available and prices falling. Huang expects clearer information once more of these players are public.