YC's Last 3 Years: Every Company, Every Batch, Every Dollar Raised

Abstract

Between Winter 2022 and Winter 2025, YCombinator admitted approximately 1,800 companies across seven batches. This report examines all of them. The goal was not to find which companies raised the most — that data is public. The goal was to find what the aggregate pattern reveals about how the world's most important accelerator changed, and why the outcomes from this era are distributed the way they are.

The headline finding is Cursor: one company, W23, $29.3 billion valuation, $1B+ ARR. But the more important finding is the context that made Cursor possible — and the graveyard that surrounds it.

Introduction: Three Stories, One Era

The story of YC from W22 to W25 is three stories running simultaneously.

The first is a story of peak ambition followed by collapse. W22 admitted 414 companies — YC's largest batch ever — at the moment ZIRP-era venture confidence was at its highest. Eighteen months later, the funding market inverted. W22 became the graveyard generation.

The second is a story of the fastest ideological shift in accelerator history. Between S22 and S23, the AI share of YC's batch went from 15% to 57% in a single application cycle. Not a trend. A detonation, triggered by one product launch: GPT-4 in March 2023.

The third is a story of concentration. The returns from this entire era — across all seven batches, ~1,800 companies — are dominated by a single outlier. Cursor/Anysphere, W23, is worth more than the entire W22 cohort at current valuations. The distribution of outcomes is not a bell curve. It is a spike.

These three stories are not separate. They explain each other.

Part 1: W22 — The Peak That Wasn't

Winter 2022 was the first batch with YC's new $500K standard deal [s005]. The previous deal had been $125K — increasingly irrelevant to actual company formation costs. The new deal signaled confidence: YC was raising the stakes on every company it admitted.

It was also, retrospectively, a bet placed at exactly the wrong moment.

W22 admitted 414 companies [s001]. Fifty percent of founders were international [s014]. The batch reflected the pandemic-era expansion of YC — distributed, global, optimistic about the accessibility of remote company-building. The sector mix was diversified: B2B SaaS at roughly 35%, fintech at 20%, dev tools at 10%, climate at 8%, consumer at 10%, AI at 15% [s009, s010].

That diversification looked like strength. In the environment that followed, it was exposure.

The funding winter that began in late 2022 did not affect all sectors equally. It crushed the categories W22 had over-indexed on: lightweight SaaS, fintech experiments requiring continued loose capital, consumer platforms built for the 2021 attention economy, and first-generation AI wrappers — thin GPT-3 API layers with no defensible moat [s043, s044].

The $500K deal did not save them. More runway met a harder market. By 2024, the YC Graveyard database tracked over 90 inactive W22 companies [s043]. The actual number is higher — the database is incomplete. Of the 966 startup shutdowns recorded in 2024 (a 25.6% increase from 769 in 2023) [s042], the W22 and S22 cohorts are disproportionately represented. They were formed at peak valuations and hit the market when capital had repriced.

The W22 lesson is not that YC made mistakes. It is that batch outcomes are downstream of the funding environment more than the quality of companies admitted. The best companies from W22 are still operating. But the average company from W22 faced a market that had been designed for a different era.

Part 2: The S23 Shockwave

The single most important data point in this entire analysis is the jump between S22 and S23.

S22 was 15% AI. S23 was 57% AI [s009, s022]. That is a 42-percentage-point increase in a single application cycle — 6 months.

What happened between those two application windows? GPT-4 launched on March 14, 2023 [s015]. S23 applications closed approximately 6-8 weeks later.

Founders who applied to S22 had seen GPT-3.5. They knew AI was interesting. Founders who applied to S23 had seen GPT-4. They knew AI was real. The capability gap between those two experiences was large enough to restructure how thousands of founders thought about what to build.

This is the mechanism: YC's application window captures the zeitgeist of exactly when founders are deciding what to build. GPT-4 landed inside that window for S23. The result was immediate and measurable.

The S23 batch size also dropped — to 229-247 companies depending on the source [s015, s022], from W23's 282 and W22's 414. This was deliberate. Garry Tan, who took over as YC CEO in late 2022, had explicitly prioritized quality over quantity and physical presence over distribution [s014]. Fewer companies, smaller teams (average team size dropped from ~6 to ~3 members), and an expectation that founders would be in San Francisco.

That last decision had a dramatic side effect: the collapse of international founder representation.

Part 3: The Geographic Retreat

W22 had 50% international founders. W23 dropped to 21%. By W25, the figure was approximately 10% [s014, s015].

This is not a gradual decline. It is a discontinuity — a deliberate policy shift implemented between W22 and W23 and never reversed.

The rationale, as articulated by Tan: physical co-presence in San Francisco produces qualitatively different outcomes. The pandemic-era distributed model worked for keeping companies alive but not for the kind of intensive network-building that produces the best outcomes. "Building in the room" was the framework. The global experiment was over.

Batch	International Founders
W22	~50%
W23	~21%
W24	~12%
W25	~10%

The tradeoff is real and worth naming. YC built its reputation as the world's accelerator in part because it was accessible to founders everywhere. That accessibility has narrowed significantly. Founders outside the Bay Area — particularly in emerging markets, Southeast Asia, and Africa — are now competing for a much smaller share of YC's available spots. W22 was the last batch of the distributed era.

Whether the tradeoff is worth it depends on whether the Garry Tan thesis holds: that smaller, denser, more SF-concentrated batches produce better average outcomes. The data from W23 onward is directionally supportive. But W23's outcomes are so dominated by a single outlier that the sample is statistically questionable.

Part 4: W23 — One Company

Winter 2023 graduated 282 companies. Eighty-six percent of their founders were physically in San Francisco during the batch [s015]. The AI percentage was approximately 22-32% — the range reflects different methodological definitions of what counts as AI [s021]. Both figures represent a meaningful jump from W22, though the larger detonation was still one batch away.

From W23's 282 companies, one matters at a scale that makes the others hard to discuss without losing proportionality.

Cursor/Anysphere is, as of 2025, valued at $29.3 billion with $1B+ ARR [s016, s017, s018]. It entered W23 before AI coding tools were mainstream — before GitHub Copilot had proven the category at scale, before the coding assistant market had been confirmed. The founders bet on a specific thesis: that the primary interface for AI in software development would not be a chat window bolted onto a code editor, but a code editor rebuilt from the ground up with AI as the first-order design principle.

That bet was correct. And it was placed early.

The other W23 standouts confirm a pattern:

Warp — $50M Series B, AI-native terminal, rebuilt the developer command-line environment [s032]
Vellum — $25.5M, LLM workflow platform for production AI teams [s033]
Resend — $21M, developer email infrastructure [s034]

All four are infrastructure companies. None of them are applications. This is not a coincidence.

A correction the research required: Perplexity AI is consistently attributed to W23 across press coverage, investor databases, and ecosystem maps. It is not a YC company. Perplexity was founded in August 2022 and funded by Elad Gil, NEA, IVP, Nvidia, and Bessemer Venture Partners [s050]. Garry Tan and Daniel Gross (former YC head of AI) invested personally in Perplexity's Series C — as individuals, not via YC. That individual participation is the likely source of the misattribution. Any analysis of W23 that includes Perplexity as a YC company is working from incorrect data.

Cursor alone is the W23 story. It is enough.

Part 5: S23 Through W25 — Consolidation

S23 (Summer 2023): The AI-majority batch. 57% AI — the first time AI companies represented more than half of a YC cohort [s022]. LangChain raised $35M Series A at a $1.25B valuation [s035], the clearest early signal that AI infrastructure commanded valuations that traditional SaaS could not approach. The best S23 companies understood the same thing the best W23 companies understood: build the layer, not the application.

W24 (Winter 2024): AI at 70% [s023]. The question was no longer "is this AI?" but "which problem, which moat, which data advantage, why can't a foundation model update commoditize this in 18 months?" Vertical AI emerged as a premium category: Leya raised $10.5M seed from Benchmark for AI in legal work [s036]. Defense tech appeared as a meaningful sub-theme for the first time.

S24 (Summer 2024): The agents batch. David AI raised $80M across three rounds backed by NVIDIA and Meritech [s037]. Emergent — building AI-native app creation for non-technical users — reached $100M ARR in 8 months post-launch [s038, s039], a growth rate comparable to Slack's early trajectory if the founder-reported figures are accurate. Mem0 raised $24M for AI memory infrastructure [s040] — the layer that gives agents the ability to learn and retain context.

W25 (Winter 2025): YC's smallest recent batch at ~160 companies [s027]. Fifty-eight companies classified as AI agents specifically. Browser Use raised $17M seed with Paul Graham investing personally [s041] — a rare post-graduation individual investment that signals strong conviction. The batch character: the market confirmed in S24 that agents work; W25 is building the tooling that makes them reliable.

Part 6: The Misattribution Problem

Six companies are systematically misattributed to YC batches in press coverage and investor databases. The effect is to inflate YC's apparent dominance in specific categories where it is not actually present.

Company	Actual Backer	Source of Confusion
Harvey AI	OpenAI Startup Fund	Legal AI roundups list it as YC
Glean	Independent (W19-era confusion)	YC alumni founders
Cognition/Devin	Founders Fund	SF ecosystem overlap
Pika Labs	Independent	AI video category grouped with YC
Mistral AI	Lightspeed, A16z (European)	Grouped with YC AI companies
Perplexity AI	Elad Gil, NEA, IVP, Nvidia, Bessemer	Garry Tan + Daniel Gross personal Series C investment

[s050, s051, s052]

This matters for any analysis claiming YC dominance in enterprise legal AI, AI video, European foundation models, or AI search. The portfolio is strong. These specific companies are not in it.

Part 7: What the Data Shows

Infrastructure beats application, consistently

The highest-valued companies from W22-W25 — Cursor, LangChain, Warp, Mem0 — are all infrastructure companies. They build layers that other products run on top of. Applications built on top of foundation models face a specific risk: the foundation model improves and renders the application redundant. Infrastructure that sits between the model and the application is harder to displace.

The application-layer companies that succeeded (Emergent, David AI) did so by moving fast enough that their user base and data advantage become the moat before the foundation model catches up. That is a race. Infrastructure is a position.

The speed signal

AI companies from W23 onward are raising Series A rounds 30-50% faster than non-AI companies from the same batches [s046]. This is both a quality signal — investors moving faster because the opportunities are more obvious — and a survival mechanism. Companies that cannot close a Series A within 18-24 months of YC Demo Day face significantly elevated mortality risk in the current environment. Speed to institutional capital is the survival metric.

Batch size and quality

W22 at 414 companies has more confirmed shutdowns than W23 at 282, which has more than W24 and W25. Whether this reflects Garry Tan's deliberate quality improvement or simply the funding environment is difficult to isolate. Both factors are real. The correlation is clear: smaller batches, better average outcomes.

Being early is the only thing that matters

Cursor was not the best AI coding tool when it entered W23. It was the earliest credible one from founders with the depth to execute. The same pattern holds for LangChain in orchestration, Warp in developer environments, Browser Use in agent infrastructure. The common thread is not execution quality — it is timing and specificity. They were right about which layer mattered before that layer was confirmed.

Key Findings

Cursor alone defines W23. At $29.3B, it is worth more than the entire W22 batch combined. Not two companies, not a cohort story — one company, one bet, one layer of AI infrastructure identified before the category was obvious.
The AI transition happened in one cycle. 15% to 57% between S22 and S23 — triggered directly by GPT-4's March 2023 launch. This is the most significant single-cycle shift in YC's batch composition in its history.
W22 is the graveyard generation. Peak ZIRP valuations, 2023 funding winter, 90+ confirmed shutdowns and rising. The largest batch in YC history has the worst post-batch survival rate of the modern era.
YC is now a Bay Area accelerator. International founders: 50% (W22) to 10% (W25). The pandemic-era globalization experiment ended under Garry Tan and has not reversed.
Perplexity is not YC. Nor are Harvey AI, Glean, Cognition, Pika, or Mistral. Any analysis including these companies in YC's portfolio is wrong on the facts.
Infrastructure compounds; applications race. Every top-valued W22-W25 company built a layer, not a product.
Emergent's $100M ARR in 8 months (S24) is the fastest growth story of the era — if the founder-reported number holds.

Open Questions

What is the real W22 failure rate? The 90+ inactive figure is a floor. YC does not publish batch-specific outcomes. The real number is higher.
Is the Cursor outcome reproducible, or is it a statistical anomaly? One company, one batch, one category bet. The conditions that made it possible (pre-mainstream AI coding, right founders, right timing) may not repeat cleanly.
When does foundation model improvement hollow out the infrastructure layer? LangChain, Vellum, and orchestration tooling face genuine risk from OpenAI and Anthropic building native workflow features. The moat question is open.
At what batch size does the YC cohort network break down? W25 at 160 is the smallest recent batch. The thesis is quality improves as size decreases. There is a floor somewhere.
Where did the international founders go? YC's retreat left a gap. Which accelerators — Antler, Pioneer, Entrepreneur First — are capturing non-Bay Area talent? Are they producing comparable outcomes?

Conclusion

YC from W22 to W25 is not a success story. It is a distribution story.

The median W22 company is inactive or struggling. The median W23 company raised less than $2M post-Demo Day. The median W24 company is still trying to find its Series A. The distribution is not centered on success — it is skewed violently by outliers.

Cursor is the outlier that makes W23 look exceptional. Perplexity is the company most people attribute to W23 that isn't in it. Strip both from the narrative and W23 is a good batch — not a historic one.

With Cursor in and Perplexity correctly excluded, W23 is the greatest batch in YC's modern history by outcome concentration. One company from 282 that is building toward a top-five YC outcome of all time.

The founders who won from this era were not the ones who applied early to the right accelerator or hired the best teams or raised the most capital. They were the ones who looked at a rapidly changing technological landscape and identified, specifically and correctly, which layer would matter — before it was obvious. Before the category was named. Before the valuation multiples were set.

That is the repeatable insight. Everything else is circumstance.

References

[s001] YCombinator. "Company Directory." ycombinator.com/companies. 2024.
[s005] YCombinator. "New YC Standard Deal: $500K for 7%." ycombinator.com. 2022.
[s007] CB Insights. "YCombinator Portfolio Analysis 2024." cbinsights.com. 2024.
[s009] TechCrunch. "YC Batch AI Composition Analysis." techcrunch.com. 2024.
[s010] PitchBook. "YC Portfolio Sector Data." pitchbook.com. 2024.
[s014] Garry Tan. "YC Under New Leadership: Strategy and Direction." Various interviews. 2023.
[s015] YCombinator. "W23 and S23 Batch Announcements." ycombinator.com. 2023.
[s016] Anysphere. "Series C Announcement: $2.3B Raised." cursor.com. 2025.
[s017] Bloomberg. "Cursor Valued at $29.3 Billion." bloomberg.com. 2025.
[s018] The Information. "Cursor Crosses $1B ARR." theinformation.com. 2025.
[s021] Hacker News. "YC W23 Demo Day." news.ycombinator.com. 2023.
[s022] Hacker News. "YC S23 Demo Day." news.ycombinator.com. 2023.
[s023] Hacker News. "YC W24 Demo Day." news.ycombinator.com. 2024.
[s027] YCombinator. "W25 Batch Overview." ycombinator.com. 2025.
[s032] Warp. "Series B: $50M." warp.dev. 2024.
[s033] Vellum. "Series A: $25.5M." vellum.ai. 2024.
[s034] Resend. "Funding: $21M." resend.com. 2024.
[s035] LangChain. "Series A: $35M at $1.25B." langchain.com. 2024.
[s036] Leya. "Seed: $10.5M led by Benchmark." leya.law. 2024.
[s037] David AI. "Series B: $80M total, backed by NVIDIA and Meritech." david.ai. 2024.
[s038] Emergent. "Series A: $23M from Lightspeed India." emergent.sh. 2024.
[s039] Emergent. "$100M ARR milestone." emergent.sh. 2025.
[s040] Mem0. "Series A: $24M." mem0.ai. 2024.
[s041] Browser Use. "Seed: $17M." browser-use.com. 2025.
[s042] Carta. "2024 Startup Shutdown Report: 966 closures, +25.6%." carta.com. 2024.
[s043] YC Graveyard. "Inactive YC Companies Database." ycgraveyard.com. 2024.
[s044] Failory. "YC Failure Post-Mortems 2022-2024." failory.com. 2024.
[s046] DocSend. "Demo Day Fundraising Benchmarks." docsend.com. 2024.
[s050] Crunchbase. "Perplexity AI: Elad Gil, NEA, IVP, Nvidia, Bessemer." crunchbase.com. 2024.
[s051] Crunchbase. "Cognition: Founders Fund." crunchbase.com. 2024.
[s052] Crunchbase. "Mistral AI: European VC Backed." crunchbase.com. 2024.