Sequoia Capital investors Sonya Huang, Pat Grady, and Sonali Singh argued in an X thread posted August 17 that application companies should stop defaulting to frontier APIs and start building proprietary models for specific parts of their product. The post lands as AI infrastructure spending has become the defining line item on startup budgets. Several of the companies the investors cite as proof points, including Harvey and Fireworks AI, are Sequoia portfolio companies.
The thread opens by invoking two public remarks from AI executives. Palantir chief executive Alex Karp has urged enterprises to “own the means of production,” the investors write, and Microsoft chief executive Satya Nadella has framed frontier API spending as a double cost: money paid to the lab, plus the competitive value of the prompts and proprietary data handed over to make that model useful. Huang, Grady, and Singh use those two statements to set up their own claim, that open-weight models such as Kimi K3 and GLM 5.2 have closed enough of the gap with frontier systems that owning a model is now a live option rather than a performance sacrifice.
According to the investors, four conditions determine when renting stops making sense: cost, when inference spending scales directly with usage and erodes margin; speed, in latency-sensitive domains such as coding autocomplete or security; proprietary data, when a company’s own feedback loop is what improves the system; and control, as labs push upward into product and application companies push downward into training. The thread traces those conditions to a Sequoia-hosted event where Harvey, Mercor, LangChain, Trajectory Labs, and Fireworks AI walked investors through what they describe as a post-training playbook: evals, harness and context engineering, post-training, and an online learning loop that turns production trajectories into training data.
As evidence the playbook works, the investors point to Harvey, the legal AI startup, which built an internal benchmark spanning more than 1,200 legal tasks across two dozen practice areas, graded against upward of 75,000 rubric criteria written by practicing lawyers.
What the thread does not weigh is the cost of the playbook itself. Evals, a custom harness, post-training runs, and a working trajectory-capture pipeline are four separate engineering disciplines, each requiring machine learning staff most application-layer startups have not hired. Harvey’s own case study, a research team of seven, is presented as proof the approach is lean; it is closer to proof that doing this well still requires a dedicated team most seed and Series A companies do not have budget for. Meeting even two of the four stated conditions cleanly, let alone all four, is the exception among the companies actually shipping AI products today, not the rule the thread implies.
The source of the advice matters too. The thread was written by investors whose portfolio includes the harness, orchestration, and inference vendors, LangChain and Fireworks AI among them, that a company would need to hire or buy from to execute this roadmap. Advice to “own your intelligence” tends to be advice that sells more infrastructure, not less, regardless of whether the underlying economics favor the switch for a given team.
AI Insiders reported today that Anthropic captures 65.1 percent of spending on Vercel’s AI Gateway while serving 30 percent of the platform’s tokens, a pricing gap that illustrates exactly the cost pressure this thread responds to. It also shows why frontier pricing power, not a shortage of open-weight capability, is the more immediate problem for high-usage products.
Founders evaluating this advice should run the actual unit economics of their highest-volume AI feature before greenlighting a post-training team: if inference cost per active user is flat or shrinking, renting remains the cheaper bet.
Sonya Huang, Pat Grady, and Sonali Singh of Sequoia Capital, posted on X, August 17, 2026.
