This year a Chinese laboratory delivered a frontier model rivaling OpenAI's o1 at an estimated $5.6 million in training cost — a fraction of its Western peers. OpenAI completed its structural conversion from a capped-profit entity into a public-benefit corporation, and in May closed approximately forty billion dollars in secondary at a valuation near three hundred billion. The agent layer, long treated as a feature, became its own category. Three years into what we now think of as the AI cycle, this letter is about what has actually changed and what has not.
On the $5.6M Training Run That Reset Everyone's Compute Math
DeepSeek's January release of R1 demonstrated that frontier-class reasoning could be produced at training costs roughly two orders of magnitude below what the consensus believed necessary. The number itself matters less than the implication — the moat that several Western foundation-model companies had been pricing into their late-stage rounds was structurally smaller than the price implied. We had been skeptical of those rounds since 2023. The skepticism is now broadly shared.
The longer-term implication, which we believe is still under-modeled, is that the cost curve of frontier capability is falling faster than the application layer can absorb. Capability is becoming a commodity. The defensibility of AI businesses, over the next five years, will rest on workflow integration, distribution, and proprietary data — not on model parameters. We have re-underwritten every AI-native commitment in our portfolio against this thesis over the past nine months. The exercise has produced two adjustments — one position we have increased on the basis of the workflow analysis, and one position we have reduced because the company's defensibility had been understood to rest on a model advantage that we no longer believe is durable. Both adjustments were specific. Neither implies a change in our overall AI weighting.
On OpenAI's Public-Benefit Conversion, Read Carefully
The structural transition from capped-profit to public-benefit corporation, completed mid-year, resolves an internal contradiction that had been governing the company since 2019. The contradiction was that no investor would commit at the scale required for frontier development if the upside was capped, and no employee would remain if the mission was abandoned. The new structure resolves both, on terms that — we will say plainly — favor the equity holders more than the original mission would have implied.
We do not say this critically. We say it as a forecast. The institutional structure of the most important AI companies will, over the next decade, look more like the institutional structure of late-stage internet companies than the founders of those companies originally intended. This is a cycle we have seen before. The 2002 letter described it in the context of the previous wave of mission-driven companies converting to commercial governance under capital pressure; the 2025 conversion is a more sophisticated version of the same dynamic, but the dynamic is the same. The companies whose mission survives will be the ones whose governance was structured, before the capital arrived, to make the conversion difficult enough that the conversion itself was a deliberate decision rather than a path of least resistance. OpenAI's conversion was, by any reasonable measure, deliberate. Whether the eventual mission alignment survives the conversion is a question for the 2030 letter.
On the $300 Billion Secondary, and What It Says About Liquidity Without IPOs
The May secondary placed approximately forty billion dollars of equity at a valuation near three hundred billion, with no IPO required. The transaction is the clearest evidence yet that the liquidity premium of public markets has compressed for the largest private companies. They no longer need to go public to monetize; they no longer need to go public to recruit; they no longer need to go public to access institutional capital.
The implication for our portfolio is structural. The companies we are now backing may, in their successful cases, never enter public markets. We are pricing this into how we think about hold periods. The 1997 founding memo committed the firm to minimum holding periods of seven years; the 2025 reality suggests that the meaningful holding periods for our most consequential 2025 commitments may be twenty years or longer, with liquidity events occurring within that window via secondary rather than via IPO. The structural implications are significant. We are revising our internal templates accordingly.
On the Agent Layer, in Plain Language
An agent is, in current usage, a software system that decomposes a multi-step task and executes the steps autonomously, calling on models and tools as needed. The category was, until this year, largely speculative. It is no longer. Several agent products shipped in 2025 are producing measurable economic outcomes for their users, in categories ranging from software development to enterprise procurement to consumer task automation.
The category will be larger than the consensus currently models. It is also, in our judgment, more difficult to build well than the consensus currently models. The companies producing reliable agent outcomes are companies that have invested heavily in the integration layer — the connections between agent systems and the underlying services they need to call — and have done so in ways that are not obvious from the company's external description. We are being selective. We have made three commitments in the agent category in 2025 and have declined approximately fifteen. The selection ratio reflects the volume of low-rigor companies entering the category; we expect the ratio to remain unfavorable through 2027 or 2028.
On the Companies We Are Now Underwriting Differently
Our underwriting framework, three years into this cycle, has updated in two specific ways. We now require, at first commitment, that the company have a defensibility that does not rest on its model layer. And we now require, at first commitment, that the company have a credible path to operating margins that do not depend on the price of inference falling further than it already has. The two requirements taken together constitute a meaningful tightening of our framework. The tightening reflects, in essence, that the AI cycle is mature enough that the speculative premium for being early has compressed, and the underwriting standards must compress correspondingly.
The 2025 vintage of our portfolio will, in consequence, be smaller in count than the 2023 vintage and larger in average commitment size. The change is deliberate. The 2025-2026 cohort of AI-native companies will, we predict, be the most concentrated in winners of any AI vintage. We intend to be heavily weighted toward those winners.
A Closing Note
The cycle is mid-flight. The next three years will, we believe, separate the companies of consequence from the companies of momentum. We intend to be present for the separation.
The Partners
Winzheng Family Investment Fund · December 2025