Y Combinator's list of bets shows venture capital migrating from pure software to physical AI

Y Combinator published, for the Winter 2027 batch, its latest list of "Requests for Startups": eleven investment theses written by the fund's own partners, including Jared Friedman. The list isn't a blog curiosity. It's the document YC uses to tell the thousands of founders who apply each batch where it wants to put money, and by extension, where AI seed checks are moving.

The list's opening text is the central thesis, and it's worth quoting: Y Combinator states that AI's biggest bottlenecks are migrating beyond the models.

AI's biggest bottlenecks are moving beyond the models.

Y Combinator, opening text of the Requests for Startups, Winter 2027

Translated into investor terms: the race for a better model is no longer, by itself, where value gets captured. Value is migrating to whoever solves what comes after the model, whether that's manufacturing the hardware, financing the build, or selling to government.

The eleven categories, in three groups

Of the eleven theses listed, four bet on the physical economy that AI is pulling along behind it, two bet on government as a buyer market, and five bet on layers left over precisely because the big AI labs won't prioritize them.

The number behind the bet on physical infrastructure

Partner Diana Hu anchors the robotics thesis in concrete data: the United States installed 38,400 industrial robots in 2025, up 12%, but most are still imported from Japan and Europe. China manufactured 94% of the world's sintered rare-earth magnets in 2024, and when it restricted exports in April 2025, automakers had to cut production.

In July, the FCC began requiring that at least 65% of the component cost of connected mobile robots be domestic for US authorization, a percentage that rises to 75% by 2029. The source doesn't specify the exact year of this regulatory change. It's a regulatory barrier designed to force local manufacturing, not a side effect.

The same bottleneck logic appears in the data center thesis, signed by Jared Friedman himself: power transformers have a three-year lead time, switchgear and circuit breakers are sold out, and the grain-oriented electrical steel used in transformer cores is manufactured in a single plant in the United States. Friedman calls the data center construction cycle the largest capital investment project in American history, bigger than railroads, highways, and the Manhattan Project combined.

The pattern that repeats: government as customer

Two theses, both signed by Gustaf Alströmer, bet on selling to the public sector. "Plaid for Government Benefits" proposes an income and residency verification layer that eliminates the rework of proving the same data for each social program separately; the thesis cites a GAO (Government Accountability Office) estimate that fraud costs the US federal government up to half a trillion dollars a year, and notes that starting in 2027, states will need to re-verify Medicaid beneficiaries every six months.

"AI for City Hall" targets the municipal permitting process. YC's math is direct: if every building permit in the United States were approved in 10% of the current time, American cities would collect $16 billion more a year in property taxes that would start being paid sooner. The thesis argues that selling to government is no longer taboo for AI founders, because some of YC's portfolio companies in this sector are growing faster than almost anything the fund has ever seen.

Where model commoditization opens space

The remaining five theses start from economic-incentive reasoning, not regulation. Jared Friedman argues that language models got worse at writing as they got better at code and math, because the same reward signal that optimizes for programming flattens voice and style, and the big labs will always prioritize code and agentic workflows over prose. That leaves room for a model built to be good at conversation, not at solving tasks.

Diana Hu proposes vertical foundation models, specialized in manufacturing, energy, and biology, that match or beat general-purpose models while using orders of magnitude less inference compute. She cites Google's TimesFM-3, a 330-million-parameter model that forecasts related time series without task-specific fine-tuning, as a reference for what's possible.

Jon Xu closes with "Own Your Intelligence": companies that today rent intelligence via API will increasingly want to own models trained on their own data and workflows, especially in regulated sectors where data governance makes renting too expensive for the performance delivered. And Brad Flora bets that the biggest budget and the strongest defensibility lie in bringing cutting-edge AI into networks with no internet access, such as defense, banks, hospitals, and nuclear plants, even though it's a slower and harder sale.

Still on labor, Brad Flora also signs "The $15-an-Hour Expert": instead of AI replacing workers, the thesis bets on AI that turns an ordinary worker, with no formal credential, into someone capable of doing a specialist's job, in areas like social benefits, taxes, insurance claims, medical billing, and immigration paperwork. For YC, wherever the credential is the bottleneck, this model works.

The counterpoint: a VC signal isn't a guarantee

It's worth being skeptical before reading the list as a treasure map. YC has a direct interest in steering the flow of applicants toward categories where its own recent portfolio already operates, and much of the thesis set is tied to specifically American public policy: the FCC rule on robots, the Medicaid re-verification calendar, the US's 19,000 municipalities. A Brazilian founder won't replicate these theses literally, because the regulatory trigger that justifies them doesn't exist here.

It's also worth remembering that RFS lists come out every batch and not every category turns into a relevant company; it's the fund itself admitting bets, not predicting the future with certainty. The physical-financing thesis ("Wall Street for the Physical Buildout") is, in fact, an admission that equity alone doesn't cover this bill: YC is asking founders to create debt, leasing, and securitization because the traditional venture capital model doesn't cover long-cycle hardware.

What this means for those building in Brazil

The five model-layer theses (writing, vertical models, owning intelligence, air-gapped, and the $15-an-hour expert) are geographically neutral: a Brazilian founder can pitch them to any global investor without needing to translate regulatory context. They're also the ones that require the least upfront capital, which makes them more accessible to those who aren't raising in dollars from day one.

The "government as customer" pattern is the most transferable in spirit, even without a literal translation. INSS (Brazil's social security institute), Receita Federal (Brazil's federal tax authority), and Brazilian city governments face the same kind of friction that "Plaid for Government Benefits" and "AI for City Hall" target, but anyone pitching this idea here needs to treat LGPD (Brazil's data protection law) and interoperability between Brazilian public systems as a product problem, not a legal footnote.

As for the capital-heavy physical theses, robots, data centers, space access, Brazil is unlikely to be the first market, but they open a supply gap for whoever manages to enter these global chains, especially in inputs like rare earths, whose Chinese concentration YC cites as a bottleneck. The practical takeaway is simple: before building a pitch on top of any item on the list, it's worth asking whether the barrier underpinning the thesis exists outside the United States, because that's where the investor's bet stops holding.