Vintages: how AI's funding cycle decoupled from reality

Disclosed round count peaked at 84 in 2024 and has fallen every year since. Disclosed capital raised has tripled across the same window. The two facts together describe a market that is no longer one market.

In 2024, 84 disclosed AI rounds closed. In 2025, 50. In the first four months of 2026, 25. Round count is collapsing. By the same accounting, total capital raised went from $80.9 billion in 2024 to $124.6 billion in 2025 to $184.4 billion in the first four months of 2026. The 2026 four-month total is already larger than the prior two years combined.

Both numbers describe the same dataset. The first says the AI funding boom ended in 2024. The second says it has barely started. The way to hold both together is to stop talking about "the AI funding boom" as a single thing. There are at least three vintages in play right now, and they look so different that comparing them is almost an error of category.

I have spent the last two months building a structured database of every disclosed AI funding round since 2017. The numbers above are extracted from the rounds file in the Nextomoro Atlas, which now covers 209 labs and 287 rounds. The rest of this essay walks through what the data actually says about the 2024, 2025, and 2026 vintages, and what the divergence between them is doing to the labs that are not on the rich list.

The 2024 vintage

2024 was the year everyone got funded. The dataset records 84 disclosed rounds, more than the prior three years combined. The median round closed at $200 million. Stage distribution looked healthy: 20 Series A rounds, 15 Series B, 12 Series C, plus a fat tail of seeds and later stages. Seventeen rounds cleared $1 billion. Four labs raised at $10 billion or more in a single transaction: Databricks ($10B Series J), Anthropic ($8B from Amazon), CoreWeave ($7.5B in Blackstone-led debt), and OpenAI ($6.6B Series G at $157B post-money).

The frontier seven (OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, Cohere, xAI) absorbed roughly 45% of disclosed 2024 capital. That share is large, but it left the other 55% spread across roughly forty different companies.

Look at what a typical month of 2024 financing actually contained. In March 2024 alone, the dataset records Cognition AI's $175 million Series A at $2 billion led by Founders Fund (Devin still in its viral demo phase) and Hippocratic AI's $53 million Series A led by Premji Invest. In May, Suno took a $125 million Series B led by Lightspeed at a $500 million valuation, and DeepL closed a $300 million Series C led by Index Ventures at $2 billion. In June, Mistral raised $640 million in a Series B at $6 billion, Pika took $80 million from Spark, and Reka raised $103 million from DST. In August, Black Forest Labs closed its $31 million seed led by Andreessen Horowitz on the strength of the FLUX launch, Codeium took $150 million from General Catalyst at $1.25 billion, and Moonshot raised $300 million from Tencent. In October, Cohere took $200 million in a Series C at $5.5 billion, OpenAI closed its $6.6 billion Series G at $157 billion, and Poolside raised $500 million from Bain. Eight different categories of AI company, eight different lead investors, eight credible rounds in a single month.

That cadence was the texture of 2024. Series A's at $50-200 million were normal. Series B's at $80-640 million were normal. The cap tables contained Andreessen Horowitz on 16 rounds across the year, plus Sequoia, Accel, Founders Fund, Lightspeed, Kleiner, General Catalyst, Bain, Index, and DST as repeat lead investors. Microsoft, Google, Amazon, and Nvidia took the strategic seats but did not yet dominate the cap tables. The conventional venture machine was funding a wide field of companies.

If you only saw 2024, you would have written this market up as the largest venture cycle in software history and reasonably expected 2025 to be larger. It was not.

The 2025 separation

By round count, 2025 was a 40% decline. By capital, it was a 50% increase. The median round jumped from $200 million to $250 million. Fifteen rounds cleared $1 billion, but they accounted for $117.6 billion of total capital raised. The megaround cluster swallowed almost everything else.

A 2025 megaround did not look like a 2024 megaround. OpenAI raised $40 billion in March, then $6.6 billion more in an October tender at $500 billion post-money. Anthropic raised $13 billion in September at $183B. xAI took $10 billion at $200B. Scale AI received $14.3 billion from Meta as part of an acqui-hire that valued the company at $29 billion and effectively transferred Alexandr Wang and most of the Scale leadership team into Meta's MSL group. Meta itself was reported to be receiving $14.3 billion in strategic AI investment, of which the Scale transaction was the visible component.

Strip the megarounds out and 2025 looks like a respectable year: 35 sub-billion-dollar rounds, fairly distributed across stages, with credible labs raising at credible valuations. Cursor took $2.3 billion at $29.3 billion. Anduril raised $2.5 billion at $30.5 billion. Reflection raised $2 billion at $8 billion led by Nvidia. Mistral did $2 billion at $14 billion led by ASML. Thinking Machines closed its $2 billion seed at $12 billion led by Andreessen Horowitz, the largest seed round ever recorded. SSI added another $2 billion at $32 billion.

The 2025 vintage split into two recognisable tiers, and the rounds in each tier looked quite different. The growth tier funded labs with revenue. ElevenLabs took $180 million in a January Series C led by Andreessen Horowitz at $3.3 billion. Synthesia closed a $180 million Series D led by NEA at $2.1 billion. Together AI raised $305 million in February at $3.3 billion. Hippocratic AI took its $141 million Series B from Kleiner Perkins. Perplexity ran a Series E in May at $14 billion led by Accel ($500 million), then added $200 million more at $20 billion in September. Runway took $308 million from General Atlantic at $3 billion. Suno closed a $250 million Series C from Menlo at $2.45 billion. Black Forest Labs closed a $300 million Series B at $3.25 billion led by Salesforce Ventures. Each of these is a real growth round into a revenue-generating specialty company. None of them broke $500 million.

The star-founder tier looked nothing like that. Periodic Labs took $300 million in a seed led by Andreessen Horowitz at $1.3 billion. Lila Sciences raised $235 million in a September Series A at $1.2 billion led by Braidwell, then $115 million more at $1.3 billion in October led by NVentures. Thinking Machines and SSI took their multi-billion-dollar rounds. AMI was already raising through the second half. The same investors who wrote $80 million Series A checks in 2024 were now writing $200-300 million seeds at billion-dollar valuations.

These are not small companies. They are also not the story. The story is that the megaround cluster pulled away from everything underneath it. In 2024, the largest single round (Databricks) was 12% of total annual disclosed capital. In 2025, OpenAI's $40 billion round alone was 32% of the year. The distribution went from "fat-tailed venture" to "one round dominates everything."

The 50 round count is what tells you something else broke. Forty fewer rounds closed. Most of the labs that had raised in 2024 did not raise again in 2025. The seed-and-Series-A funnel kept narrowing. The cycle had already shifted from "everyone gets funded" to "almost nobody gets funded, but the ones that do get funded enormously."

The 2026 vertical

Then 2026 happened.

In the first four months, 25 disclosed rounds. Eight of them cleared $1 billion. Those eight totalled $178.8 billion in raised capital, which is more than the entire 2024 and 2025 vintages combined. The median 2026 round closed at $619 million, more than triple the 2024 median. The frontier seven captured 94.1% of all 2026 capital.

The single round that dominates the year did not look like a venture round at all. In March 2026, OpenAI announced a $122 billion private financing led by SoftBank at an $852 billion post-money valuation. The headline number is larger than the entire annual disclosed AI capital base of 2024. The post-money is larger than the market capitalisation of all but four U.S. public companies. The lead investor is a Japanese conglomerate whose AI exposure has become indistinguishable from its sovereign-wealth-fund posture. This is no longer a market that venture capital is the right institutional vehicle to participate in.

The Anthropic Series G in February 2026 went $30 billion at $380 billion post-money, led by GIC, the Singaporean sovereign wealth fund. xAI took $20 billion at $230 billion in January. Cerebras returned to market with a Series H at $23 billion post-money led by Tiger Global. Moonshot, the Chinese frontier lab, closed its Series F at $18 billion. World Labs raised a $1 billion Series A. AMI, Yann LeCun's post-Meta lab, closed a $1 billion seed at $4 billion post-money less than four months after announcing.

What didn't make the 2026 megaround list is just as instructive. ElevenLabs took $500 million in a February Series D led by Sequoia at $11 billion. Cursor closed its Series E in April at $2 billion led by Andreessen Horowitz, on top of its $2.3 billion Series D from November 2025. Mistral added $830 million in debt in March. Synthesia did $200 million at $4 billion led by GV in January. DeepSeek took $300 million at $10 billion in April. Goodfire closed a $150 million Series B at $1.25 billion led by B Capital. The Cohere-Aleph Alpha merger announced in April 2026 was anchored by a $600 million Series E and a $530 million parallel commitment from Schwarz Group, the German retail conglomerate behind Lidl and Kaufland, in what is functionally the first European sovereign-AI consolidation underwritten by a non-state strategic. StepFun took $717 million in a January Series B+ led by Shanghai State-owned Capital Investment, which is the recognisable new face of Chinese frontier financing: state and state-adjacent capital writing nine-figure checks in a market where private VC has largely retreated.

These rounds are substantial. Each one is also less than 2% of the year's total disclosed capital. In 2024, a $500 million round was a marquee event. In 2026, a $500 million round is a footnote.

The stage mix tells you what the 2026 vintage actually is. Of 25 rounds in the first four months, six are Series E, two are Series F, one is Series G, one is Series H. Two are debt. Two are IPOs, both on the Hong Kong Stock Exchange in early January (Z.AI and MiniMax). One is an acquisition. Exactly one is a Seed round (AMI), and exactly one is a Series A (World Labs). Both of those "early stage" rounds were billion-dollar checks at multibillion-dollar valuations into companies whose founders had OpenAI, DeepMind, or Meta on their CVs.

The 2026 vintage has effectively eliminated the early-stage AI round. Anything below $100 million is invisible in the disclosed dataset, either because it isn't being announced or because it isn't happening at the volume that prior years produced. What's left is a barbell: nation-state-scale capital allocations into the frontier seven, and oversized "seed" rounds into a small number of star founders.

Strategic capital takes the wheel

The shift from venture to strategic capital is the most concrete way to see what changed. In 2024, $21.9 billion of disclosed capital was led by a corporate or sovereign strategic. In 2025, $16.6 billion. In the first four months of 2026, $123.1 billion.

Almost all of the 2026 strategic figure is the SoftBank-led OpenAI round. Strip that and the strategic figure collapses back to a normal year. But the round happened. SoftBank wrote a check that is larger than the cumulative disclosed venture capital deployed across every AI lab not named OpenAI in the entire dataset. This is what "phase change" means.

The strategics in the OpenAI cap table are SoftBank, Microsoft, Nvidia, Khosla Ventures (which now writes SoftBank-scale checks alongside its venture funds), Thrive Capital, Altimeter, Fidelity, MGX (the Abu Dhabi sovereign vehicle), and Tiger Global. The Anthropic Series G was led by GIC, with Amazon and Google as the largest existing strategic holders. The xAI Series E was led by, among others, MGX again, with significant participation from the SpaceX-related capital network. Cohere's Series E in April 2026 was anchored by Schwarz Group, the German retail conglomerate (Lidl, Kaufland), as part of the Aleph Alpha merger.

These are not venture firms. They are nation-states, hyperscalers, and the largest non-tech industrial companies on earth. The frontier seven has graduated from a venture-funded category into a sovereign-allocated one. Comparing this to the 2024 vintage is comparing the United States Treasury market to a Series B.

Andreessen Horowitz remains the most active venture investor in the dataset by lead count (16 rounds), but a16z has not led any of the four largest 2025 or 2026 rounds. Sequoia, Accel, Founders Fund, and Kleiner Perkins all show up in the smaller late-stage rounds. They are no longer the marginal capital provider for the frontier. They are competing for what is left in the second tier.

The exodus labs as venture's last stand

The labs that came out of the 2024-2026 frontier exodus are the most interesting venture-style story in the dataset, because they are the last generation of AI labs where venture-scale capital was the natural fit.

Look at the rounds: SSI raised $1 billion at $5 billion in seed, then $2 billion at $32 billion in Series B. Thinking Machines closed $2 billion at $12 billion in a single seed. AMI raised $1 billion at $4 billion in seed. Reflection took $2 billion at $8 billion in Series B. World Labs raised $230 million across two rounds and then $1 billion at the start of 2026. Magic raised $400 million at $1.5 billion in early 2025. Periodic, FutureHouse, Lila Sciences, Goodfire all closed seven-to-nine-figure rounds at billion-dollar valuations.

These rounds look extraordinary by any historical standard. They look modest next to the OpenAI Series G. The exodus cohort has roughly $7 billion in cumulative funding across fifteen labs at a paper valuation around $80 billion. OpenAI raised eighteen times that figure in a single transaction. The exodus labs are the most that the conventional venture machine can put up. The frontier has moved to a different capital base.

Andreessen Horowitz, Sequoia, Founders Fund, and Greenoaks are the most credible institutional buyers of the exodus cohort. They underwrite founders, not products. The rounds price reputation, not revenue. The thesis is that one or two of these labs becomes a frontier player, and the math works on the basis of that hit rate. This is recognisable venture logic operating at a scale (multi-billion-dollar valuations into pre-product companies) that was previously reserved for unicorn-stage growth rounds with seven-figure ARR underneath them.

If the 2026 vintage tells me anything, the exodus pattern is the high-water mark for venture-funded frontier AI. The next cohort of star researchers leaving big labs in 2027 will encounter a capital market where any check below $5 billion looks small, and where the natural buyers of $5 billion checks are sovereign wealth funds. The exodus labs may be the last generation that gets to be a venture story.

The 78% who haven't raised

The other half of the K is the labs that aren't on the rich list. Of 209 labs in the Nextomoro database, 163 of them (78%) have not had a disclosed funding round in 2025 or 2026. That is two consecutive years without raising for almost four out of five companies the dataset is tracking.

A subset of those 163 are profitable and don't need to raise. ElevenLabs took its February 2026 Series D at $11 billion specifically because the company decided more capital was strategic, not because it needed runway. A subset are early enough that they are still operating on prior rounds. Many of the open-source and academic labs (EleutherAI, Allen Institute, MILA, BigCode, BigScience) operate on grant cycles that don't show up as venture rounds.

But many of the 163 are simply stuck. They raised in 2022 or 2023 at valuations the 2024-2026 market won't underwrite. The late-2023 cohort that priced rounds at "AI premium" multiples now has to either grow into the valuation or take a down round. The 2025 dataset shows almost nobody choosing the second option. The dataset does not yet show anyone choosing the first.

The concentration at the top compounds the problem. The top single lab (OpenAI) holds 43.1% of all cumulative disclosed AI funding in the dataset. The top three (OpenAI, Anthropic, xAI) hold 66.0%. The top seven hold 79.6%. The top twenty hold 92%. Fifty labs hold 98% of the capital. The remaining 159 labs share less than 2% of the disclosed funding base.

This is not a bell curve. It is barely even a power law. It is a dumbbell with an exceptionally heavy ball at the OpenAI end.

For the labs in the long tail, the capital question is now an existential one. The 2026 funding environment doesn't write small Series A's. The acquirers (Microsoft, Meta, Google, Amazon, Nvidia, Cohere, increasingly Salesforce and Schwarz) are choosy. The IPO window has reopened in narrow form (CoreWeave on Nasdaq in March 2025, Z.AI and MiniMax on the Hong Kong exchange in January 2026) but only for revenue-substantial businesses with the right listing geography. The 163 labs that haven't raised in two years are facing a choice between accepting a down round, finding an acquirer, or quietly winding down. The dataset already records four "acquired" labs and one "shutdown." The honest forecast is that several dozen more will hit one of those statuses by the end of 2027.

What "raising in 2026" actually means

Vintage is a useful word for venture rounds because it tells you what kind of company emerged from a particular year of capital. A 2010-vintage SaaS company looks different from a 2014-vintage one and looks completely different from a 2021-vintage one. The same is now true of AI labs.

A 2024-vintage AI lab raised at $50-200 million across a Series A or B. The lab probably had a product or a credible model in development. The cap table contained Sequoia, Andreessen, or Founders Fund as lead. The valuation cleared $1 billion if things were going well, $300 million if they weren't. The lab had 12-18 months of runway and a path to a Series C in 2026.

A 2025-vintage AI lab is a more bimodal animal. Either it is a star-founder seed at $1-2 billion in capital and $5-30 billion in valuation (the SSI, Thinking Machines, AMI cohort), or it is a Series C-D growth round at $100-500 million for a company that has revenue and a defensible market position (Cursor, ElevenLabs, Synthesia, Suno, Hippocratic, Decart, Lila). The middle has been hollowed out. The conventional Series A-B at $30-100 million is largely missing from the disclosed 2025 data.

A 2026-vintage AI lab does not exist as a category in the conventional sense. The 2026 disclosed dataset contains the OpenAI sovereign-scale round, the Anthropic and xAI sovereign-scale rounds, a small number of star-founder mega-seeds (AMI, World Labs), late-stage growth rounds for revenue-generating specialty labs (Cursor's Series E, Cerebras' Series H, Synthesia's Series E, Runway's Series E), and one Asia-Pacific IPO. There is no 2026-vintage early-stage AI lab in the disclosed data, because almost no early-stage AI lab is disclosing rounds in 2026.

This does not mean the rounds are not happening. The vast majority of seed-stage AI investing happens below the disclosure threshold and won't show up in datasets like this one until the company is already at Series A or beyond. What it means is that the headline 2026 capital flow is no longer underwriting the formation of new labs. It is underwriting the continued existence of an extremely small number of incumbents.

Geography of the vintages

The 2024 vintage was geographically distributed. U.S. labs took the largest share (107 of 209 labs in the dataset are headquartered in the U.S.) but the year produced credible rounds in France (Mistral, Kyutai, H Company), Germany (Aleph Alpha, DeepL), the U.K. (Wayve, Stability), Israel (AI21, Decart), Canada (Cohere), Korea (Upstage, Naver), and China (across the 22 China-tagged labs).

The 2025 vintage tightened. The U.S. share of disclosed capital climbed. European rounds shrank. Aleph Alpha effectively gave up the consumer market and pivoted to enterprise sovereign-AI in 2024, then merged with Cohere in April 2026 in a transaction underwritten by Schwarz Group. The European frontier consolidated to a small number of names. China continued to fund domestically (Moonshot, Z.AI, StepFun, Baichuan) but with smaller individual rounds and a heavier mix of state and state-adjacent capital.

The 2026 vintage is barely geographic at all. It is dominated by the OpenAI round (U.S.-headquartered, Japanese-led, Abu Dhabi participation), the Anthropic round (U.S., Singapore-led), the xAI round (U.S., diverse strategic), the Cohere-Aleph Alpha merger (Canada-Germany, German-led), and the Moonshot Series F (China). Geography matters less now than it did because the cap tables are all global. Every frontier lab is being funded by some combination of U.S., Middle Eastern, East Asian, and European capital. The "national champion AI lab" framework has not produced a single 2026 round below $1 billion outside of state subsidy programs.

Why now

Three things converged to produce the 2026 vertical.

First, the AI revenue numbers came in. OpenAI's reported annualised revenue cleared $5 billion in 2024 and was reported at $15-20 billion by early 2026. Anthropic's annualised revenue cleared $1 billion in late 2024 and was reported at $5+ billion by Q1 2026. The capital markets concluded that frontier AI is a real revenue category, not a subsidy. The 2025-2026 megarounds price that conclusion. They are not pricing potential. They are pricing observed revenue trajectories that justify, on certain assumptions about gross margin and growth duration, the valuations being written.

Second, the supply of capital at the relevant scale narrowed. There are perhaps fifteen institutions on earth that can write a $10 billion check into a single AI company: SoftBank, Microsoft, Google, Amazon, Nvidia, Meta, GIC, Temasek, MGX, PIF, KIA, the largest U.S. pension funds, Berkshire, BlackRock, and a handful of others. Once the rounds got big enough to require those checks, the round dynamics changed. There are no auctions for $50 billion of primary capital. There are bilateral negotiations between labs and the small set of institutions that can clear the price. The OpenAI round took roughly a year to assemble.

Third, the early-stage AI venture machine ran into the Cambrian-explosion ceiling. By the end of 2024, almost every plausible AI vertical had a venture-funded leader and three to five challengers. The marginal seed-stage AI investment now has to either compete with an incumbent that already has product-market fit, or find a niche the 2024 cohort did not cover. Both options are harder than they were a year ago. The result is fewer rounds at the bottom of the funnel, and the appearance (which the dataset confirms) that "AI funding has dried up" even as the headline numbers go vertical.

What to watch

Five things over the next twelve months that would change the picture.

1. Whether SoftBank actually funds the OpenAI round. The $122 billion is committed but not entirely funded. SoftBank's track record is to announce headline numbers and then deliver in tranches over multiple quarters as it raises the corresponding debt or asset sales. If SoftBank delivers in full, the 2026 vintage gets a second leg. If it delivers in tranches and the AI revenue trajectory disappoints, the Q4 2026 OpenAI valuation could be the first major frontier-lab markdown of the cycle.

2. The first sub-frontier megaround to break the pattern. Of the eight 2026 megarounds, seven went to the frontier seven. The one exception was AMI's $1 billion seed, which is technically a frontier-adjacent founder-led round. The next sub-frontier megaround, into a specialty lab or a sovereign-AI program, will tell you whether the megaround pool is broadening or whether 94% concentration is the new normal.

3. The acquisition wave. Of the 163 labs that haven't raised in two years, the most likely outcome for forty or fifty of them is acquisition by a hyperscaler, a sovereign-AI vehicle, or one of the larger specialty labs. The first wave of acquisitions in late 2026 will set the price floor for the dormant cohort and tell you whether the AI lab population shrinks by 10%, 25%, or more by 2027.

4. Whether sovereign capital becomes structural. SoftBank, GIC, MGX, PIF, the Korean K-Sovereign program, Schwarz Group as proxy for German strategic interest. The 2026 cap tables read as sovereign-influenced even when the headline lead is a private institution. If sovereign-led rounds remain the backbone of frontier financing through 2027, the regulatory and geopolitical shape of the AI industry changes substantively. Frontier labs will be operating with cap tables that look more like national strategic assets than venture-backed startups.

5. The Series A revival, if it happens. The most underdiscussed signal in the 2026 dataset is the near-disappearance of the AI Series A. If a wave of $20-50 million Series A rounds returns in the second half of 2026, it means the venture machine is finding new categories below the frontier. If it doesn't, the AI lab population stops growing at the bottom and starts shrinking through the middle.

The honest summary

The 2026 vintage looks nothing like the 2024 vintage and almost nothing like the 2025 one. The headline numbers describe a market where capital is more abundant than it has ever been. The round-level data describes a market where almost nobody can raise. Both are true. They are true at different points on the dumbbell.

I started this piece by noting that the 2024 round count was the peak. That fact is not a marker of a market top. It is a marker of when the AI funding market stopped being a market and started being something else: a sovereign-scale capital allocation toward a small number of incumbents, with a vestigial venture market handling the founders that the incumbents have not yet absorbed.

The shape of the next two years is downstream of that single change. Watch the cap tables, not the round counts. The names of the lead investors are the new editorial signal. When SoftBank, GIC, MGX, and Schwarz Group are leading the rounds that matter, "the AI funding cycle" is the wrong frame. What I see is the early shape of an AI capital base. The vintages of 2024, 2025, and 2026 are the first three years of a transition the dataset is now tracking in real time.

The 2024 vintage funded a generation of labs. The 2025 vintage separated the frontier from the rest. The 2026 vintage stopped pretending the rest existed.

Sources used in this piece:

• The Nextomoro Atlas dataset (rounds.json, labs.json, manifest.json) extracted 2026-04-30, covering 287 disclosed rounds across 209 labs since 2017.
• The Nextomoro 2026-Q1 to 2026-Q2 quarterly diff for revisions, new rounds, and leadership changes since the prior refresh.
• Public funding announcements via Crunchbase, PitchBook, TechCrunch, The Information, and primary lab blog posts cross-referenced for each megaround cited.
• Reported revenue figures for OpenAI and Anthropic from The Information and Reuters coverage of Q4 2025 and Q1 2026.
• The companion essay "The Frontier Lab Exodus" for the underlying read on the 2024-2026 founder departures cohort.

Last updated: April 30, 2026. All round amounts and valuations are at announced figures, converted to USD at the rate effective at announcement. Disclosed only; the dataset does not estimate undisclosed financing. Send corrections.