Venture capital trades the AI model for physical infrastructure

The official Y Combinator Requests for Startups list for the Winter 2027 batch shifts its focus from language models to hardware, government, and financial engineering. For those using the YC seal as leverage with Brazilian funds, the signal has moved elsewhere.

Eleven bets, one pattern

Y Combinator published the list of Requests for Startups (RFS) for the Winter 2027 batch, the document in which the accelerator publicly states what kind of company it wants to see born in the coming months. There are eleven categories, signed by names like Jared Friedman, Gustaf Alströmer, and Diana Hu.

The opening text already delivers the thesis: AI's bottlenecks are ceasing to be the model and becoming hardware, infrastructure, the financial system, and government. It's a punchy line, but it carries weight: historically, YC made its name on pure-software bets, light capital, and 80% margins. This list is not that.

Of eleven items, at least six call for physical capital, a long sales cycle with government, or a financial instrument that doesn't exist yet. The other five remain in software's comfortable territory, just more specialized. The table below summarizes the split.

The bottleneck is no longer the model, it's steel

The most explicit case is Robots Made in America, signed by Diana Hu. The United States installed 38,400 industrial robots in 2025, up 12%, but most still come imported from Japan and Europe, and even those assembled domestically depend on Chinese rare-earth magnets. China accounted for 94% of the world's sintered rare-earth magnets in 2024 and restricted exports in April 2025, cutting assembler production.

In July, the FCC stopped authorizing new models of connected mobile robots that don't have at least 65% of component costs sourced domestically, a percentage that rises to 75% in 2029. Assembling imported parts no longer works. It's regulation turning into an investment ruler, something the software VC never needed to read carefully.

The same pattern appears in Data Center Supply Chain, by Jared Friedman: power transformers have a three-year lead time, disconnect switches and liquid cooling systems are sold out, and the grain-oriented electrical steel used in transformer cores comes from a single factory in the United States, unable to meet demand. Friedman sums up the scale of the bet in a line worth quoting.

The AI data center buildout is the largest capital investment project in American history - bigger than the railroads, the highways, and the Manhattan Project - combined.

Jared Friedman

When the VC becomes a bank

The most revealing piece on the list isn't even about technology: it's Wall Street for the Physical Buildout, by Brad Flora. He admits, bluntly, that the financing model for this physical economy doesn't exist yet. Banks today evaluate a robot as if it were a forklift, and founders are inventing, out of necessity, instruments like credit backed by robots that generate 500,000 dollars a year, or options on GPU capacity.

This is an important confession: venture capital equity, optimized for power-law returns on high-margin software, wasn't designed to finance steel and silicon. Every major infrastructure buildout, from railroads to electricity to telecom, generated its own financial companies alongside it. YC is saying it wants to finance whoever builds those companies, not just whoever builds the robots.

Silicon Valley bets on the State

Another not-so-obvious front is selling to government. In Plaid for Government Benefits, Gustaf Alströmer draws a direct comparison to the banking infrastructure Plaid built for fintechs.

Before Plaid, every fintech app built its own bank connections. After Plaid, all it took was one tap. The government needs the same layer.

Gustaf Alströmer

The number backing this thesis comes from the GAO (the US Government Accountability Office): benefits fraud costs the federal government up to half a trillion dollars a year, and starting in 2027 states will need to re-verify the eligibility of millions of Medicaid beneficiaries every six months.

In AI for City Hall, the same Alströmer adds another number: if every construction permit in the United States were approved in 10% of the current time, cities would collect 16 billion dollars more per year in taxes. YC has already funded companies selling AI to government that, according to the text, grow faster than almost anything the accelerator has seen before.

The counterargument: not everything has turned into concrete and steel

Reading this as YC abandoning software would be an overstatement. Three of the eleven categories remain anchored in model and data: LLMs That Write Well seeks a model that writes better because, according to Jared Friedman, the training that improves code worsens current models' writing style. Vertical Foundation Models, by Diana Hu, bets on small, specialized models (she cites Google's TimesFM-3, with 330 million parameters, as a benchmark for forecasting without task-specific fine-tuning). Own Your Intelligence, by Jon Xu, argues that regulated companies will want to own and train their own models instead of renting tokens from a frontier lab.

These three items are, at bottom, the same VC as always: light capital, pure-equity returns, no FCC, no transformer, no city hall. YC hasn't changed its thesis, it split the bet. And it's worth asking whether the physical part of the list is really a structural shift or whether it's the accelerator surfing the moment's headline, the same one that puts a humanoid robot and a data center on every business magazine cover. Ten-year funds, designed to return capital via IPO or software acquisition, still haven't proven they know how to price an asset that takes five years to generate cash.

What this means for those raising money in Brazil

Citing YC's RFS as leverage with a Brazilian fund only works if the founder knows which of the eleven categories they're citing, because they call for completely different types of capital. Govtech has a direct analogy here, from Cadastro Único (Brazil's unified social registry) to Gov.br (Brazil's unified government services platform) and Serpro (Brazil's state-owned IT company), and it's the slice of the list easiest to replicate with local sales cycles and regulation. Hardware and data centers, on the other hand, call for the financial instrument that YC itself says doesn't exist yet in the United States, and Brazil's venture debt and physical-asset leasing market is even smaller than America's.

In short: the seal