Garry Tan says this YC batch has more defense companies than all of last year, and the era of cost-plus procurement is finally ending

Defense surge

The current YC batch has more defense companies than all of last year combined, according to Garry Tan. He credits a genuine shift in government procurement: the Department of Defense is now willing to hear pitches from teams of 10 or 20 people and buy the best product, a model that would have been unthinkable for most of the past several decades.

Tan's argument against the old model is structural, not personal. Cost-plus procurement was designed to incentivize innovation but instead locked defense primes into a system that rewards margin management over performance. The analogy he reaches for is Siri — capable engineers trapped in an organizational structure that cannot ship what users actually want. In consumer tech, that's a nuisance. In defense, Tan argues, it costs lives.

The people changing this, in his telling, are figures now inside the Trump administration who have opened the DIU and what he calls the "Department of War" to competitive, outcome-based buying. He name-checks Emil Michael as one of the people driving that shift.

“There's more defense here in this batch than the entire last year... I think that there is actually an awareness that cost plus was originally designed to be something that created more innovation. And here we are fifty, eighty years later — they are actually, for the first time in decades, open to a team of 10 or 20 people, whether it's in Boston, in Austin, even down the street in San Francisco.”

Hard tech and AI

On the broader hard tech boom, Tan argues AI has removed most of the traditional barriers: supply chain research, vendor discovery, mechanical and electrical engineering guidance are all accessible through AI tools now. That's pulling more founders into hardware who would previously have self-selected out.

The more interesting tension he raises is around software companies that look easy to clone. His view is that the "anyone could vibe-code this" criticism misses the point — the team that actually shipped it, acquired customers, and built institutional understanding of the user has a real lead that's hard to close quickly. That edge may not last forever, but on an 18-to-36-month horizon, he thinks it holds.

G Brain

Tan's current personal project is G Brain, which he describes as aiming to be the Postgres for agents — a persistent memory layer that lets agents pull the right context from an arbitrarily large knowledge base. The framing: a human holds roughly seven items in working memory, an LLM holds something like three books' worth of context, but a real business contains something closer to the Library of Alexandria. G Brain is meant to bridge that gap by loading whichever subset of a company's knowledge is actually relevant to the task at hand. He's testing it with Codex and is positioning it as an internal YC infrastructure project, with tools eventually being built into Bookface.

The nontechnical label

Tan endorses Paul Graham's recent argument that nontechnical founders should stop calling themselves that. His point is that anyone who sits down with Claude Code can meaningfully contribute to a codebase today — something that wasn't true a year ago. YC's working thesis is that the binding constraint on startup creation is no longer capital or technical skill, it's agency: having a clear idea of what you want to build and the will to start.