Not subscribed? Sign up to get it in your inbox every week.

This week's issue looks a little different. I've been dealing with a family emergency and couldn't put together a new newsletter, so I'm resharing my deep dive on the YC Summer 2026 batch instead. I'll be back with a fresh issue soon. Thanks for your patience.

Original article here: https://x.com/chris__lu/status/2097515515551809566

i pulled every company in YC's Summer 2026 batch off the YC site: 236 companies and 470 founders, then went through each one and placed it on a layer of the AI stack.

last batch i called "the agent batch" because 95% touched AI and 70% were building agents.

this batch is still 91% AI, and almost nothing else about it is the same.

every company was placed into one layer, energy at the bottom and applications at the top, and then i ran Spring 26 batch through the same rubric so the comparison is like for like.

companies working below the model (energy, chips, data centers, model training, inference) were 8% of Spring. they're 20% of Summer.

applications went the other way, from 55% of the batch to 39%, and horizontal apps alone dropped from 58 companies to 32 while vertical apps held flat at 25%. everything that grew is either underneath the model or out in the physical world.

45% of Spring 26 companies shipped autonomous agents. 33% of Summer do.

"agent" was in 27% of Spring one-liners (19% in Summer).

agent-infrastructure companies halved as a share of the batch.

nobody stopped building agents. they stopped leading with the word. it's the assumption now, not the thesis.

read the pitches back to back and four clusters jump out that didn't exist at this size in Spring.

the compute buildout, 21 companies: Atomarine is putting nuclear-powered data centers on barges at sea, Ethos wants to manufacture silicon on the moon, Pacific is mass-producing micro data centers, and three separate companies (Computable, Stoa, Touchmark) are building exchanges for compute.

PRINCEPS is an insurance company for the compute economy. Proprio builds the robots that run the data center.

two companies, Frontier Computing and Parasma, are growing human neurons as a training substrate, and one of them taught a dish of neurons to play Frogger.

inference cost: 11 companies. five of them promise to cut your LLM bill by 80% or more, and three describe themselves as "OpenRouter for X."

training data and environments for frontier labs, 9 companies, plus 7 building RL environments. that category had 1 company last batch; YC's own "Reinforcement Learning" tag is now on 14.

8 companies name it as the thing they replace, and no other incumbent gets more than 4.

YC files every company under one industry, and "Industrials" went from 12% of Spring to 24% of Summer.

45 companies ship physical hardware, double last batch. 24 are robotics or physical AI (two humanoids, construction robots, factory robots, drone defense, robot data and eval companies), and 8 build chips or compute hardware where Spring had one.

founder backgrounds moved with it. Tesla, Palantir, Nvidia and SpaceX all rose as feeder companies, and the four hardware leads from Humane, including a Nest co-inventor and the iPhone antenna lead, are in the batch together building a camera that sees through walls.

two AI-native accounting firms, one of which says it's "killing Deloitte." an insurance carrier with zero underwriters, a radiology practice that's already acquiring a $4.3M practice, and a law firm. there's a freight brokerage, two debt collectors, two quant trading firms, an addiction clinic, a defense contractor, and a BPO with $500M of receivables already running through it.

they companies do the work and charge for the outcome. that's 9% of the batch.

Spring had 18 companies selling AI to sales and marketing teams.

Summer has 6, and customer support went from 5 to 1.

the industries that took their place all have physical operations: manufacturing 11 → 18
logistics 7 → 12
construction 5 → 11
insurance 5 → 9
accounting 1 → 6

i pulled the background of all 470 founders, and this is the youngest batch i've measured.

37% are students or graduated in the last two years, and 59 companies, a quarter of the batch, are all-student teams. one company's three founders are all 17.

dropouts tripled as a share, from 3% to 9%, while repeat founders fell from 32% to 23%; 84% have a technical background.

the young teams aren't in the app layer either. a third of the model-and-training companies and a third of the chip companies are all-student teams, which is the frontier, not the safe part of the map.

Berkeley overtook Stanford: UC Berkeley 39, MIT 32, Stanford 25, Harvard 20, where Spring's top three were Stanford, Berkeley, MIT. Amazon is the #1 feeder for the second batch running (Amazon/AWS 30, Google/DeepMind 15, Meta 13, Apple 12).

only 19 founders came from an AI lab. that's 4%, and zero of them came from Anthropic

eight whole founding teams share a prior employer. Humane → Applied Electrodynamics (4 founders), Windsurf → Illume Labs, Scythe Robotics' founders → Agency Tool Company, and three prior YC companies (Drapr S20, Lilac Labs S24, Sable S19) are back as new teams.

what didn't change: 60% pairs, 19% solo, 74% San Francisco, 71% one-word names, 25% .ai. the one naming shift is that "Labs" went from 8 companies to 14.

two core bets:

founders are treating the app layer as crowded, and they're probably right.

the edge moved to places a wrapper can't follow: training data, environments, tokens, silicon, robots.

the compute buildout is now a startup category and no longer only a hyperscaler story. 21 companies are selling into it, from nuclear barges to compute insurance, and 11 more are selling ways to cut inference cost

the thing to watch is the "AI-native firm." the accounting firm, the carrier, the practice, the brokerage. their software looks like everyone else's; what they're actually building is a services business with a different cost structure, and that's harder to compete with than a feature.

last batch everyone was building an agent. this batch assumes the agent, and asks what you're running it on, what you trained it with, and whether you own the business it works inside.

analyzed from public YC data plus enriched founder backgrounds across 236 companies and 470 founders

Till next time,
Chris