Y Combinator, founded in 2005, is the accelerator every other copies, running two three-month batches a year and a standardised US$500,000 deal. Rather than concentrating capital in a handful of frontier labs, it underwrites AI as a volume business, funding hundreds of application-layer teams per year. Since 2023 most of each batch has been AI-native.
YC does not try to pick the one model company that wins. It writes the same cheque into hundreds of teams per year, accepts a high failure rate, and relies on batch scale plus follow-on rights to capture outliers — the opposite of the concentrated frontier bets made by growth funds.
Frontier training runs are far outside a $500K cheque. The portfolio clusters instead around products built on top of someone else's model: voice agents, coding tools, vertical copilots, and enterprise search where a small team can reach revenue inside a batch.
Speech, evaluation, memory and fine-tuning infrastructure recur batch after batch — Deepgram, AssemblyAI, Vellum, OpenPipe and Mem0 all sell picks and shovels to other AI builders, a category YC has funded for a decade.
The uniform deal and Demo Day remove price negotiation from the earliest stage. For founders the value is access to 1,000+ investors on a fixed date; for YC it is the ability to originate at scale without competing on valuation.
Y Combinator was founded in 2005 by Paul Graham, Jessica Livingston, Robert Morris and Trevor Blackwell. Since 2023 the large majority of each batch has described itself as AI-native, making YC the single widest funnel into early-stage AI. The standard deal is US$125K for 7% equity plus a US$375K uncapped SAFE.
A selection of YC-backed AI companies, from autonomy and data labelling to the voice, memory and evaluation tooling layers.
Data labelling and evaluation infrastructure for frontier labs and defence programmes; YC S16.
Self-driving vehicle developer acquired by General Motors; the batch's earliest large autonomy outcome. YC W14.
Speech-to-text and voice agent APIs sold to contact centres and voice product teams. YC W16.
Speech recognition and audio intelligence models delivered as a developer API. YC S17.
Contract lifecycle platform that added model-driven review and redlining to legal workflows. YC W15.
AI language tutor using speech models for open-ended conversation practice. YC S16.
Developer platform for building and deploying real-time voice agents.
Voice agent infrastructure for inbound and outbound phone operations.
Prompt management, evaluation and deployment tooling for LLM applications.
Fine-tuning and distillation platform that converts prompt traffic into cheaper task-specific models.
Open-source memory layer giving agents persistent, retrievable user context.
Open-source enterprise search and chat over internal company knowledge; formerly Danswer.
The pre-AI outcomes that made the batch model credible — not exhaustive, but the alumni most often cited by founders applying today.
Home-sharing marketplace; YC W09 and the accelerator's defining outcome.
Payments infrastructure for internet businesses. YC S09.
Listed cryptocurrency exchange. YC S12.
Local delivery marketplace. YC S13.
Grocery delivery and retail media platform. YC S12.
File sync and storage. YC S07.
Social discussion network funded in YC's first batch in 2005.
Live-streaming platform, launched out of Justin.tv. YC S07.
No rounds led by Y Combinator are in the tracker yet. Deals will appear here as they are recorded.
There is nothing to negotiate: US$125K for 7% equity plus a US$375K uncapped SAFE, identical for every company. Removing pricing from the earliest stage is the product — founders get a fixed-date market of 1,000+ investors at Demo Day, and YC gets to originate at scale without competing on valuation. A normal seed round is a negotiation; this is a listing.
Cheque physics. Frontier training runs cost billions — Moonshot's July 2026 Series F alone was a reported US$3.5B, 7,000x the US$500K deal — so YC underwrites the layer it can move: voice agents, coding tools, vertical copilots and enterprise search, where a small team can reach revenue inside a three-month batch. The model layer is someone else's balance sheet; YC funds what gets built on it.
Because the model prices selection as impossible at that stage. YC writes the same cheque into hundreds of teams per batch-year, accepts a high failure rate, and relies on batch scale plus follow-on rights to capture the outliers — the opposite of Conviction's concentrated 19 or NFDG's taste-driven book. Since 2023 the large majority of each batch has been AI-native, making YC the sector's widest funnel.
Picks-and-shovels is YC's oldest AI pattern, predating the LLM wave: Deepgram (W16) and AssemblyAI (S17) sold speech APIs years before ChatGPT; Vellum, OpenPipe and Mem0 now sell evaluation, fine-tuning and memory to other AI builders. A batch system notices what builders need, because the builders are the batch — the recurrence is demand observed twice a year, not a thesis written once.
Both, deliberately — and the programme is the scarce asset. The standardized US$500K cheque is the entry ticket; the three-month batch and Demo Day are what every other accelerator copies. Unlike a fund, YC does not manage to a portfolio thesis; it manages a repeatable selection process whose output — Airbnb, Stripe, Coinbase, DoorDash — became the credibility that attracts the next batch.
