The State of AI market

A structural analysis of announced AI financing activity, April 2025 – April 2026

Contents

1. Eight things this report found
2. What this dataset is — and what it is not
3. The market in one page
4. What kind of capital moved?
5. Where the money went
6. How mature is the market?
7. Where is the activity?
8. How concentrated is it?
9. Who keeps raising?
10. What we can see and what we cannot
11. The signals that matter most
12. Appendix A: Valuation and round economics
13. Appendix B: Definitions and methodology

1. Eight things this report found

One. The dataset contains 8,202 announced AI financing events across 7,118 companies over twelve months. The disclosed dollar total is $431.8 billion. The median round is $2 million. Those two numbers — the headline and the median — belong to essentially different markets.

Two. Eighty-one rounds of $500 million or more, barely 1% of all announcements, account for 81% of disclosed dollars. The top ten rounds alone account for 58%. The visible AI capital market is not skewed — it is almost entirely a mega-round phenomenon.

Three. Not all of that money is venture capital. Roughly 30% of disclosed dollars — about $131 billion — comes from debt facilities, strategic investments, public-market offerings, and secondary transactions. Treating every row as "AI funding" without separating these regimes produces a number that blends fundamentally different economic events.

Four. Applications make up two-thirds of AI companies. Infrastructure absorbs five-sixths of the capital. Foundation model developers — 158 companies, 2.2% of the ecosystem — captured 51% of all disclosed dollars. Compute and data center companies added another 23%. The AI funding boom and the AI company boom are happening in different parts of the stack.

Five. The United States accounts for half the companies and 83% of the capital. Within the U.S., the Bay Area accounts for 39% of companies and 77% of the dollars. The market has a global footprint and an American — specifically Californian — capital center.

Six. Twelve percent of organizations raised more than once during the window. That 12% absorbed 82% of all disclosed capital. The top single repeat raiser — OpenAI — absorbed 37% of all repeat-raiser capital by itself.

Seven. More than a third of disclosed dollars sit in rounds with no clean venture-stage label. The market's largest capital events — $122 billion for OpenAI, $14.3 billion for Scale, $11.6 billion for Crusoe — are also the ones that resist the seed-to-growth framework most completely.

Eight. The market's concentration story, repeat-fundraising story, infrastructure story, and geography story are not four findings. They are one finding. The same 15–20 companies appear at the top of every list this report constructed. Remove them, and the headline total drops by more than half — but the broad ecosystem of 7,000+ early-stage companies barely changes.

2. What this dataset is — and what it is not

This report draws on 8,202 announced financing events across 7,118 AI companies, tracked between April 2025 and April 2026. The dataset spans venture equity, debt facilities, government grants, public-market transactions, and secondary sales — 73% of rounds disclose a raised amount, 11% disclose a pre-money valuation, and headquarters country is known for 95%. Dollar totals are disclosed lower bounds; the actual figure is higher and unknowable from this data alone.

With those ground rules set, here is what the data shows.

3. The market in one page

The headline number is $431.8 billion in disclosed financing across 8,202 announced rounds. The number that actually matters is $2 million — the median round size. Those two figures, sitting side by side, tell you almost everything you need to know about this market before you read another page.

Most of these 8,202 rounds were small. Nearly half of all disclosed rounds came in under $5 million. Seed and pre-seed rounds alone made up 61% of all announcements. This is a market with thousands of companies raising modest amounts of money to build things.

And yet, 81 rounds of $500 million or more — exactly 1% of announcements — accounted for 81% of all disclosed dollars. Ten rounds accounted for 58%. The entire disclosed dollar total is, in practical terms, a story about a few dozen very large checks written to a very small number of companies.

Two markets in one dataset

The easiest way to see this is to look at what happened month by month.

Round counts were remarkably stable. Every full month landed between 607 and 813 announcements. No boom month, no dead month. The AI financing ecosystem was active and consistent throughout the year.

Disclosed dollars were the opposite — wildly volatile. February 2026 alone contributed $177 billion, roughly 41% of the entire year's disclosed total. Remove mega rounds, though, and the volatility disappears. The non-mega market ran at a steady $2.6–4.1 billion per month.

That gap between the two lines is the central tension of this entire report. There is a broad, steady, seed-heavy AI financing ecosystem — and then there is a small, volatile, mega-round-driven capital layer sitting on top of it. They show up in the same dataset, but they are essentially different markets.

Where the rounds are vs where the money is

The split shows up just as clearly when you look at stage.

Pre-seed and seed rounds make up 61% of all announcements but account for just 3.5% of disclosed dollars. Late venture and growth rounds make up just 2.2% of announcements but capture 24% of disclosed dollars. And the single largest dollar bucket — 36% of all disclosed capital — sits in rounds with no clean stage label at all.

This is not a market where capital is gradually distributed across a healthy stage ladder. It is a market where the broad base and the capital peak are almost entirely disconnected.

Mega rounds — 3.5% of announcements — hold 90.2% of disclosed dollars. That one number should anchor everything that follows. Every chapter in this report is, in some way, an attempt to explain what is inside that 90% and what is outside it.

4. What kind of capital moved?

That $431.8 billion is not one kind of money. The dataset contains venture equity, debt facilities, government grants, public-market offerings, corporate rounds, and secondary transactions — instruments that look identical in a spreadsheet row but represent fundamentally different economic events.

Private venture equity makes up 82% of all announcements but only 70% of disclosed dollars. Roughly $131 billion — about 30% of the total — comes from something other than classic venture rounds.

Debt is the standout non-venture regime: just 2.6% of activity but $69 billion in disclosed capital — money for data centers and infrastructure, not early-stage product development. Liquidity and secondary transactions added $21.2 billion across only 30 rounds, but that money typically went to existing shareholders, not the company. Grants and non-dilutive capital account for nearly 10% of all announcements but just 0.08% of disclosed dollars — breadth without capital weight.

5. Where the money went

If you looked only at how many companies raised money, you would conclude that AI is an applications market. If you looked only at how much money was raised, you would conclude it is an infrastructure market. Both would be right. That is the single most important structural fact about AI financing today.

The stack: thousands of apps, billions for infrastructure

Every organization in the dataset was classified by its position in the AI stack — from low-level compute and foundation models up through developer tools and into horizontal and vertical applications. The chart below shows what each layer looks like by company count versus disclosed dollars.

Vertical applications — companies building AI for healthcare workflows, legal research, financial compliance, and hundreds of other specific domains — represent 30% of all organizations. Add horizontal applications and the application layer accounts for two-thirds of the ecosystem by company count. This is a genuinely broad market. Thousands of teams are building AI products for real end users.

But those thousands of application companies collectively received just 12% of disclosed dollars.

Meanwhile, foundation model developers — 158 companies, just 2.2% of the ecosystem — absorbed 51% of all disclosed capital. Compute and data center companies added another 23% on a similarly tiny organizational base. Two layers of the stack, representing fewer than 360 companies combined, captured three-quarters of all visible money.

A typical vertical application company raised $1.4 million. A typical compute company raised $26.5 million — nearly 19× more. These are not companies competing in the same economic league.

The aggregate view

Applications: 66.5% of companies, 64.8% of rounds, 11.9% of disclosed dollars.
Infrastructure: 21.6% of companies, 22.5% of rounds, 83.9% of disclosed dollars.

That is not a gradual tilt. It is a structural inversion. The market that most founders experience and the market that headlines describe are almost entirely different pools of activity.

Who is "AI-core" and does it matter?

AI-native products — companies using AI as a core capability but defined by the product they deliver — make up 67% of companies and receive 12% of disclosed dollars. AI-core companies — those building foundational AI technology itself — make up 14% of companies and receive 77% of disclosed dollars.

The broad ecosystem is building with AI. The capital is flowing to companies building AI itself.

Where are these companies selling?

Enterprise horizontal dominates both breadth and capital — 48% of companies and 88% of disclosed dollars. That figure is somewhat misleading: many infrastructure companies land here simply because enterprise is where their revenue comes from. Outside enterprise, end markets are broad but financially thin — healthcare has 12% of companies and 2.7% of dollars; financial services 8% and 2%; consumer 6% and 0.5%.

What this means

The "AI funding boom" is really two different markets wearing the same label. One is broad, early-stage, and modestly funded — thousands of application companies building AI into specific workflows and industries. The other is narrow, capital-intensive, and enormous — a small number of foundation model developers and compute providers absorbing the overwhelming majority of disclosed dollars. Infrastructure is expensive, and that cost is real. But it does mean the two markets are almost entirely disconnected.

6. How mature is the market?

If you count rounds, the AI market looks very young. If you follow the dollars, it does not look young at all. And if you look closely, a surprising amount of the capital sits in rounds that resist stage labeling entirely.

The stage ladder by the numbers

Each round in the dataset was mapped to a normalized lifecycle stage — from pre-seed/angel through seed, early venture, late venture/growth, and into private equity/public/liquidity territory. Two additional buckets matter: alternative/non-stage-specific (grants, assistance programs) and unknown (rounds where neither the instrument nor the label gave a reliable stage signal).

The first four rows tell a clean story. Seed and pre-seed make up 61% of all announcements — the vast majority of visible activity — but just 3.5% of disclosed dollars. Late venture/growth is a sliver of activity at 2.2% of rounds, yet captures 24% of disclosed dollars with a median round of $101 million. The stage ladder works the way you would expect: each step up is dramatically more capital-intensive. A seed company raises $3M. A late-stage company raises $101M. That is a 34× jump.

But then look at the bottom two rows. Unknown-stage rounds account for 36% of all disclosed dollars — more than any single named stage. Add alternative/non-stage-specific and over half of the market's disclosed capital sits outside the clean venture ladder. That is not a data quality footnote. It means some of the largest financing events in AI — the ones that move the headline numbers — are precisely the ones that do not fit neatly into the seed-to-growth progression that most people imagine when they think about startup funding.

Not every part of the market is at the same stage

The previous chapter showed that applications and infrastructure are essentially different markets by capital. The stage data shows they are at different points of maturity too.

Vertical and horizontal applications look unmistakably early. About two-thirds of their rounds are pre-seed or seed. They are broad, young, and raising small amounts. That tracks with what Section 5 already showed — the application layer is where the ecosystem is widest.

Foundation models tell a different story. They still have plenty of early-stage rounds by count, but 58% of their disclosed dollars sit in unknown-stage rounds alone. That is largely the gravitational pull of a few enormous financings — like OpenAI's $122 billion round — that defy clean stage labels. For compute, 42% of disclosed dollars sit in PE/public/liquidity and another 26% in alternative financing. These are not companies climbing the venture ladder. They are companies accessing capital markets that the traditional stage framework was never designed to describe.

The practical implication: if you use stage as a maturity signal, application categories look early and infrastructure categories look mature. But the infrastructure categories are not just "later stage" — they are partly stage-illegible, operating in financing structures that venture-stage labels cannot capture.

When companies raise again, they mostly stay put

Among the 857 companies that raised more than once during the window, we can track whether they moved up the stage ladder between their first and last observed round.

The diagonal tells the story. Of the 766 repeat raisers with comparable stage data, 70% stayed at the same stage. Only 30% advanced. The most common pattern was Seed → Seed (193 companies), followed by Pre-seed → Pre-seed (130) and Early venture → Early venture (129). The most common advancement was Pre-seed → Seed (111 companies), then Seed → Early venture (85).

This does not mean these companies are stuck. A twelve-month window is simply too short to observe full lifecycle progression for most startups. What it does tell us is that the repeat fundraising we see in this dataset is more recursive than developmental — companies raising again at the same stage, topping up, extending runways — rather than a neat march from seed to Series A to Series B.

What this means

The AI market's maturity profile mirrors its capital profile: lopsided. The broad base is genuinely early — thousands of seed and pre-seed companies in the first stages of building. The capital peak is genuinely later — but "later" often means outside the venture-stage framework entirely, in financing structures that look more like infrastructure project finance or public-market capital raises than traditional venture rounds.

And when companies raise again within the year, the dominant pattern is not graduation. It is repetition. The market is not climbing a ladder so much as circling within stages — especially at the early end — while a few capital-intensive players at the top operate in a stage category of their own.

7. Where is the activity?

AI companies are everywhere. AI capital is not.

The dataset spans 90+ countries. Companies from Europe and Asia-Pacific each make up roughly 18% of the organizational base — a genuinely global footprint. But when you shift from counting companies to counting dollars, the map collapses. The Americas account for 54% of organizations and 83% of disclosed dollars. Europe and Asia-Pacific, despite their combined 37% of companies, share just 15% of the capital between them.

The United States alone tells most of that story. It accounts for half the announcements and half the organizations — a large share, but not an overwhelming one. Then you look at dollars: 83% of all disclosed capital traces back to U.S.-headquartered companies. The gap between 50% participation and 83% capital is not subtle. It means the average disclosed round from a U.S. company is far larger than one from anywhere else, driven by the infrastructure mega-rounds that earlier chapters already identified.

Outside the U.S., the rankings scramble

The most interesting thing about the non-U.S. landscape is how different it looks depending on which metric you use.

The United Kingdom leads non-U.S. activity with 14% of non-U.S. rounds — a broad ecosystem of mostly early-stage companies. But it captures only 12% of non-U.S. disclosed dollars, putting it third by capital. China is second by activity at 10% of non-U.S. rounds, but its dollar figure comes with a major asterisk: only 39% of Chinese rounds disclose an amount, the lowest rate among major countries. Its visible capital position almost certainly understates reality.

The real surprise is Australia. It contributes just 2.8% of non-U.S. rounds but 16% of non-U.S. disclosed dollars — the largest non-U.S. capital share by a wide margin. Singapore shows a similar pattern at a smaller scale: 3.6% of non-U.S. rounds, 9.6% of non-U.S. dollars. Both countries are being pulled upward by a small number of large rounds rather than by broad ecosystem activity. Ireland is even more extreme — 0.9% of non-U.S. rounds, nearly 20% of non-U.S. disclosed dollars — almost certainly driven by one or two very large financings.

The pattern is familiar by now. A few large checks can make a country look capital-rich. Broad company formation tells a different, and usually more durable, story.

Within Europe, the EU accounts for 57% of European rounds and 72% of European disclosed dollars. The UK contributes 34% of rounds but 25% of dollars. The rest of Europe — Switzerland, Norway, and others — adds 10% of rounds and just 3% of capital.

Inside the U.S., one state owns the market

If the global picture concentrates into the U.S., the U.S. picture concentrates into California.

California represents 47% of U.S. rounds — dominant, but not monopolistic. New York is a clear second at 15%. Texas, Florida, Massachusetts, and Washington each contribute 3–5%. That is a recognizable multi-hub startup landscape.

Then look at dollars. California jumps to 79% of U.S. disclosed capital. New York drops to 3.6%. The only states that register meaningfully on the dollar chart are Colorado (6.3%) and New Jersey (6.0%) — both likely lifted by a small number of infrastructure-scale financings rather than broad ecosystem depth.

The metro level sharpens this further. San Francisco alone accounts for 28% of U.S. rounds and 61% of U.S. disclosed dollars. Add Palo Alto, San Jose, Mountain View, Menlo Park, Sunnyvale, and San Mateo — all fragments of the same Bay Area — and the combined figure reaches roughly 39% of U.S. rounds and 77% of U.S. disclosed dollars.

New York, the second-largest metro by company count, captures 3.5% of U.S. disclosed dollars. Denver, which barely registers by round count, shows 6.2% of dollars — another case of a few large rounds reshaping the capital map. The metro table also contains labels like Dover and Wilmington, which reflect legal incorporation addresses rather than actual AI hubs.

What this means

Geography follows the same structural logic as everything else in this report. The footprint is broad — 90+ countries, multiple U.S. hubs, a genuinely international early-stage ecosystem. The capital is narrow — overwhelmingly American, overwhelmingly Californian, overwhelmingly Bay Area. That is not because the rest of the world lacks AI companies. It is because the mega-round, infrastructure-heavy capital layer that dominates disclosed dollars is physically concentrated in a very small number of places.

For any country or city trying to build an AI ecosystem, the implication is straightforward: breadth of company formation is achievable and already happening globally. Capital depth — the kind that funds foundation models and data centers — remains extraordinarily concentrated.

8. How concentrated is it?

Earlier chapters kept circling around the same observation: a small number of very large rounds dominate the dollar totals. This chapter stops circling and measures it directly. The answer is more extreme than the earlier hints suggested — and it holds up even after you try to explain it away.

The headline numbers

Start with rounds. The top 10 disclosed rounds — ten individual financing events out of nearly 6,000 with known amounts — account for 58% of all disclosed dollars. The top 1% of rounds account for 79%. Mega rounds as a group, which make up 3.5% of all announcements, hold 90% of disclosed dollars.

That means the remaining 96.5% of rounds — the seed checks, the Series As, the grants, the debt facilities — collectively account for less than 10% of disclosed capital. The headline AI financing market is not skewed. It is almost entirely a mega-round phenomenon.

Remove the mega rounds and see what happens

The natural question is: what does the market look like without those mega rounds? Much flatter.

With mega rounds removed, the top 10 rounds account for just 2.2% of the remaining disclosed dollars. The top 1% account for 11%. That is still a power-law distribution — capital markets always are — but it is a recognizably normal one. The extreme headline concentration is not a feature of the broad AI market. It is a feature of a very thin mega-round layer sitting on top of it.

But this is not just a round-level illusion

You might think: maybe the headline concentration is just a handful of press-release-friendly mega rounds distorting an otherwise healthier picture. Move to the company level, where capital accumulates over multiple rounds, and perhaps things even out.

They do not. The top 10 companies by disclosed capital in-window absorb 68% of all disclosed dollars. The top 1% of companies — about 54 organizations — absorb 83%. And here is the test that matters: restrict the analysis to fresh capital only, excluding liquidity and secondary transactions. The top 10 companies still account for 68%. The top 1% still account for 82%.

The concentration is not an artifact of double-counting, liquidity transactions, or accounting quirks. It is structural. A very small number of companies genuinely absorbed most of the visible capital.

The curves make it visceral

The cumulative share chart shows all three views at once — rounds, organizations, and rounds excluding mega — and the shape of the curves says more than any table can.

The blue and red lines are nearly vertical at the left edge. By the time you reach 5% of entities, you have already passed 90% of disclosed dollars. The gold line — rounds excluding mega — rises much more gradually. That is the non-mega market: still unequal, but not grotesquely so.

The deepest pools are also the most concentrated

This is where the concentration story connects back to the company-type chapter. It is not just that infrastructure categories dominate disclosed dollars. It is that the capital within those categories is almost entirely held by a handful of names.

Foundation models account for 51% of the market's disclosed dollars. Within that category, the top 10 organizations hold 98% of the segment's capital. That is not concentration in the usual venture sense. That is essentially the entire segment.

Compute accounts for 23% of market dollars, with 83% held by its top 10. Data infrastructure accounts for 8%, with 81% held by its top 10. These are the market's three largest capital pools, and each of them is internally dominated by a very small set of winners.

The application categories look different. In vertical applications and horizontal applications, the top 10 organizations hold 36% and 35% of their respective segment capital. Still concentrated by most standards, but nothing like the infrastructure layers. The application market genuinely has a broader base of capital recipients.

Read across both columns. The categories with the most capital are also the ones where that capital is most concentrated. The categories with the broadest company base have the least capital to distribute. Concentration and dollar dominance reinforce each other.

Geography tells the same story

The previous chapter showed that capital is geographically concentrated. This chapter adds a second layer: even within each geography, a small number of companies carry the total.

Inside the U.S., the top 10 organizations account for 77% of U.S. disclosed capital. Inside Europe, the top 10 account for 75%. Inside Asia-Pacific, 72%. These are not markets where a broad ecosystem of companies shares the capital roughly evenly. They are markets where a few names define the visible total.

In smaller countries, this becomes even more extreme. In Ireland, a single organization accounts for 99% of disclosed capital. In Australia, one company holds 94%. In Singapore, one holds 84%. The previous chapter noted that some countries appeared capital-rich relative to their company count. This chapter explains why: they are not capital-rich markets. They are markets that happen to contain one or two very large recipients.

What this means

The AI financing market's concentration operates in layers, and each layer reinforces the others. A small number of mega rounds dominate the disclosed-dollar tape. A small number of companies absorb most of the capital even at the organization level. The categories where those companies sit — foundation models and compute — are themselves internally winner-take-most. And the geographies where they are headquartered inherit that concentration.

Remove any one of those layers and the market still looks concentrated. That is the difference between a market that happens to have a few outliers and a market whose entire capital structure is built around a few winners. This one is the latter.

9. Who keeps raising?

The previous chapter showed that a few dozen companies carry most of the capital. This chapter asks a different question: how many of those companies came back for more within the same year? The answer reframes the entire market.

The 12% that absorbed 82%

Of the 7,118 organizations in the dataset, 857 — exactly 12% — raised more than once during the window. That is a small minority. But those 857 companies account for 24% of all round rows and 82% of all disclosed capital.

Read the left panel as a staircase: 12% of companies, 24% of activity, 82% of dollars. Each step roughly doubles. The right panel shows that even inside the repeat-raiser pool, capital is not spread evenly. The single top repeat raiser accounts for 37% of all repeat-raiser disclosed capital. The top three account for 58%. The top ten account for 81%. So it is not that 857 companies are sharing the bulk of the market's capital. It is that a small handful of companies, raising repeatedly, dominate everything.

The typical repeat raiser is not hyperactive. 83% raised exactly twice in the window. Another 12% raised three times. Only 12 companies — xAI, CoreWeave, Crusoe, Firmus Technologies, and a few others — raised five or more times. The repeat-raiser effect is driven less by a large cohort raising constantly and more by a small number of names raising at enormous scale.

The names

The table makes it concrete.

These are not ten random companies. They are the same infrastructure and foundation-model names that dominated the concentration chapter, the company-type chapter, and the geography chapter. The market's concentration story, its repeat-fundraising story, and its infrastructure-versus-applications story are not three separate findings. They are the same finding, viewed from different angles.

CoreWeave raised eight times in twelve months. xAI raised six. Firmus Technologies raised five. These are not annual fundraising cycles. They are companies accessing capital markets on a near-quarterly basis, at scales that more closely resemble infrastructure project finance than traditional venture fundraising.

Where repeat raisers matter most — and least

The repeat-raiser effect is not uniform across the market. It is strongest in exactly the categories where capital concentration is already highest.

The left panel shows that repeat-raiser prevalence varies modestly — from about 9% of horizontal application companies to 22% of compute companies. The differences are real but not dramatic. Every category has repeat raisers.

The right panel is where the gap blows open. Repeat raisers account for 98% of disclosed capital in foundation models, 85% in data infrastructure, 75% in compute, and 66% in robotics. In those categories, almost all visible capital flows through companies that raised more than once.

The application layer is different. Repeat raisers account for 48% of capital in vertical applications, 35% in horizontal applications, and 24% in healthcare/biotech. These categories still have repeat raisers, but the capital base is much more distributed across one-time fundraisers. The broad application ecosystem that earlier chapters described is not just broad by company count — it is also broad by capital recipient. Infrastructure's capital pool, by contrast, is almost entirely a repeat-raiser phenomenon.

The geography split tells the same story from a different direction. Repeat-raiser prevalence is nearly identical across macro regions — roughly 12% everywhere. But repeat-raiser capital share differs enormously: 87% in the Americas, 82% in Asia-Pacific, but only 40% in Europe and 39% in the Middle East/Africa. The Americas and Asia-Pacific are not just capital-heavy because they have more companies. They are capital-heavy because the repeat-fundraising cohort within those regions absorbs nearly all the visible money.

They raise fast, but they do not always advance

The median repeat raiser spans 133 days from first to last announcement in the window, with a median gap of 104 days between rounds. That is roughly one fundraising event every three and a half months — fast enough to matter, but not unusual for high-growth companies in a capital-intensive market.

What is more revealing than speed is direction. Among repeat raisers with comparable stage data, 70% stayed at the same normalized stage between their first and last round. Only 30% advanced. The most common patterns were Seed → Seed (193 companies), Pre-seed → Pre-seed (130), and Early venture → Early venture (129). The most common actual advancement was Pre-seed → Seed (111 companies), then Seed → Early venture (85).

Section 6 already noted this pattern. Here it becomes clearer why it happens. Most repeat fundraising is not companies graduating from one stage to the next. It is companies raising again at the same stage — extending runways, topping up rounds, or closing follow-on tranches. The venture ladder, in the data at least, is more of a landing than a staircase.

What this means

Repeat fundraising is one of the main mechanisms through which the market's capital becomes concentrated. It is not the only mechanism — mega rounds, financing-regime mix, and the inherent capital intensity of infrastructure all contribute. But it is the most visible one at the company level.

The market has 7,118 organizations. Most of them raised once, at a modest scale, in the early stages. That is genuine ecosystem breadth. But 82% of the disclosed capital went to 857 companies that came back to the market at least twice — and within that group, a handful of foundation-model and compute names absorbed the vast majority. The broad market and the capital-heavy market overlap only slightly. Repeat fundraising is where you can see the seam between them most clearly.

10. What we can see and what we cannot

Every chapter so far has made claims about the market's structure. This chapter asks how much of that structure is actually visible — and where the data thins out enough that conclusions should carry a warning label.

The short version

The dataset is strong enough for market-structure analysis. It is not strong enough for market-wide valuation analysis. Those are different claims, and the gap between them matters.

Country is known for 95% of rounds. Organization websites are available for 96% of companies, and descriptions are high or medium quality for 97%. Company taxonomy confidence is high or medium for 86%. These fields — the ones that support the geography, company-type, and concentration chapters — are solid.

Round amounts are disclosed for 73% of rounds. That is workable, but it means every dollar figure in this report is a lower bound, and some segments are much more transparent than others.

Valuation is known for 11% of rounds. That is not workable as a broad lens. It is why the valuation chapter sits in the appendix, constrained to a 775-round subset, rather than claiming to describe the whole market.

Not all parts of the market are equally visible

The 73% average disclosure rate hides real variation. Some corners of the market show you almost everything. Others show you less than half.

By financing regime. Public-market rounds disclose amounts 97% of the time. Private venture equity is at 76%, and debt at 78% — both usable. But liquidity and secondary transactions disclose only 37%, and strategic equity only 47%. The regimes that carry the most interesting non-venture capital are also the least transparent about it.

By stage. Disclosure improves as rounds get larger. Late venture and growth rounds disclose 94% of the time. Early venture is at 89%. Seed drops to 72%, and alternative/non-stage-specific financing — grants, assistance programs — sits at just 56%. That means the early, broad base of the market is somewhat hazier than the later, capital-heavy layer.

The valuation panel on the right tells its own story. Late venture and growth rounds have 50% valuation coverage — enough for a credible subset analysis. Early venture drops to 20%. Seed is at 8%. Pre-seed is at 14%. For most of the market, valuation simply is not available.

By country. This is where the variation gets sharpest.

The United States, United Kingdom, India, Canada, France, and Australia all sit between 74% and 84% money disclosure — a tight, usable band. Israel is the strongest at 84%. China is the outlier at 39%. South Korea is also materially lower at 54%. Japan sits at 66%.

That has a direct implication for this report's geography chapter. When we said the U.S. captures 83% of disclosed dollars and China captures less than 2%, part of that gap is real and part of it reflects the fact that China discloses amounts on barely a third of its rounds. The U.S. capital position is robust. The Chinese capital position is almost certainly understated.

The taxonomy is useful, not perfect

The company-type chapters relied on an LLM-assisted classification of every organization. That classification is strong in the core categories and weaker at the edges.

The right panel is the one to focus on. Vertical applications, horizontal applications, healthcare, security, and data infrastructure all have low-confidence shares under 13%. These are the categories that carry most of the report's company-type analysis, and they are taxonomically solid.

Three categories are hazier. Other/adjacent — the residual bucket — is 96% low-confidence. That is by design; it catches companies that did not fit cleanly elsewhere. Services/implementation is 41% low-confidence, meaning nearly half the companies in that bucket may be misclassified. And foundation models is 24% low-confidence — higher than you might expect for a category that carries 51% of disclosed dollars.

None of this invalidates the earlier chapters. The structural finding — that applications dominate breadth while infrastructure dominate capital — holds regardless of how you treat the fuzzy edges. But it does mean that fine-grained claims about specific company categories should be weighted by confidence. The core buckets are trustworthy. The residual and edge buckets are approximate.

What this means for the report

The report's structural conclusions — on concentration, repeat fundraising, the infrastructure-vs-applications split, and geographic patterns — rest on the strongest data layers: round counts, country, company descriptions, and 73% money disclosure. Those findings are robust.

Dollar comparisons across regimes, stages, and countries are real but uneven. They are strongest for private venture equity in the U.S. and Europe, weaker for non-venture regimes and for East Asian markets with lower disclosure rates.

Valuation analysis is a narrow subset, not a market-wide view. The appendix treats it accordingly.

Where the data is clear, this report says so. Where it is hazy, the report has tried to say that too. A reader who trusts the structural findings and holds the dollar comparisons more lightly will have the right mental model for what this dataset can actually show.

11. The signals that matter most

The previous chapters each examined one dimension of the market — capital type, company category, stage, geography, concentration, repeat fundraising, disclosure. This chapter asks a different question: when you lay all of those dimensions on top of each other, which patterns keep showing up no matter how you look?

Three do.

One mismatch explains most of the market

Every chapter in this report has, in some way, been a restatement of one structural fact: the companies that make up the AI ecosystem and the companies that absorb the AI capital are almost entirely different groups.

The chart below distills that into a single image. For each AI stack category, it shows the gap between that category's share of organizations and its share of disclosed dollars. Bars to the left mean "more companies than capital." Bars to the right mean "more capital than companies."

Vertical applications: 30% of companies, 5% of dollars. A gap of −25 percentage points. Horizontal applications: 26% of companies, 4% of dollars. Gap of −22 points. These are thousands of real companies, building real products, raising real money — and collectively, they account for less capital than a single foundation-model developer's annual fundraising.

Foundation models: 2.2% of companies, 51% of dollars. A gap of +49 percentage points. Compute: 2.8% of companies, 23% of dollars. Gap of +20 points. Two categories, roughly 360 companies between them, absorbing three-quarters of all visible capital.

Healthcare, security, robotics, data infrastructure — they all sit near the center line. Meaningful categories with meaningful activity, but neither defining the ecosystem's breadth nor capturing its capital depth. The extremes belong to applications on one end and infrastructure on the other.

This is not a new finding. Section 5 showed it. Section 8 showed it. Section 9 showed it. But this chart is the cleanest expression of the underlying structure: one market is wide, the other is deep, and they barely overlap.

The rounds that break every framework

One of the quiet themes of this report has been that the market's largest financing events often resist the labels we use to organize them. The stage chapter noted that 36% of disclosed dollars sit in "unknown" stage rounds. The financing-regime chapter noted that 30% of dollars sit outside private venture equity. This section puts names on those patterns.

Start with stage outliers. Even within cleanly labeled stages, some rounds are so far from the norm that they function as one-company categories.

Fab 34 JV raised $14.2 billion in a round classified as private equity / public / liquidity, where the stage median is $12 million. That is 1,183× the median. Thinking Machines Lab raised $2 billion in what the data calls a seed round — 665× the seed median of $3 million. CoreWeave's $7.5 billion PE/public/liquidity round was 625× its stage median. Anthropic's $30 billion late-venture round was 297× the late-venture median.

These are not statistical noise. They are the rounds that define the market's capital structure, and they happen to sit inside stage categories designed for companies raising three or four orders of magnitude less. The stage ladder is a useful organizational tool for 99% of the dataset. For the 1% that matters most in dollar terms, it is not just approximate — it is misleading.

Then there are the rounds that don't sit in a clean stage at all. OpenAI's $122 billion round is classified as "unknown" stage. Crusoe AI Data Center's $11.6 billion round is "unknown." Scale's $14.3 billion round is "alternative / non-stage-specific." Firmus Technologies' $10 billion round is "alternative." Databricks' $5 billion round is "unknown." These five rounds alone total $163 billion — 38% of the entire dataset's disclosed dollars — and none of them has a clean venture-stage label.

The market's most important capital events are also its least classifiable. That is not a data quality problem. It is a signal about what kind of market this has become. When the largest rounds involve debt facilities for data centers, multi-tranche strategic investments, and equity raises so large they defy series labeling, the traditional venture framework is not wrong — it is just describing a different market than the one where the money actually sits.

The same names, everywhere

There is one more pattern worth making explicit, even though it has been visible since the concentration chapter. The market's different structural stories — concentration, repeat fundraising, mega rounds, infrastructure dominance, geographic gravity — are not independent findings. They are the same finding.

OpenAI appears in the top-10 by disclosed capital, in the repeat-raiser list, in the mega-round list, in the stage-outlier list, and in the foundation-model concentration table. Anthropic appears in all five. xAI appears in all five. CoreWeave appears in all five. Scale, Databricks, Crusoe, Firmus Technologies — the same names rotate through every analytical lens this report has applied.

That convergence is the market's deepest structural feature. It is not that AI financing is concentrated and repeat-heavy and infrastructure-dominated and geographically narrow. It is that all four of those descriptions point at the same small cluster of companies. Remove roughly 15–20 names from the dataset, and the headline dollar total drops by more than half. The monthly volatility disappears. The geographic concentration softens. The infrastructure-versus-applications split moderates. The repeat-raiser capital share drops dramatically.

The broad market — 7,000+ companies raising seed and pre-seed rounds across 90+ countries — would barely notice. It exists largely independent of those names. But the capital market that produces the headlines would be unrecognizable.

Where the picture goes dark

A final signal, and an honest one. The places where this report's evidence is weakest are not random. They cluster in ways that matter.

China discloses round amounts on only 39% of its financing events — half the rate of most major markets. It has 4.6% of the dataset's organizations and 388 rounds of activity, making it the fourth-largest non-U.S. market by company count. But its visible dollar position — 1.7% of disclosed capital — almost certainly understates reality by a wide margin. South Korea, at 54%, and Japan, at 66%, are also materially below the global average. Any cross-country dollar comparison involving East Asian markets should carry that caveat.

On the taxonomy side, the residual "other/adjacent" bucket is 96% low-confidence — essentially an admission that those companies could not be confidently classified. That is by design. More surprising is that foundation models, the category carrying 51% of disclosed dollars, has a 24% low-confidence share. Services/implementation sits at 41%. The core application categories — vertical applications, horizontal applications, healthcare, security — are all under 13% low-confidence. The report's thematic conclusions lean on those cleaner categories, and they hold. But the categories closest to the capital center are also somewhat fuzzier than the ones at the edges.

None of this invalidates what earlier chapters found. It sharpens the reader's sense of where to trust the numbers fully and where to hold them more lightly.

The takeaway

The visible AI financing market between April 2025 and April 2026 had genuine breadth — thousands of companies, dozens of countries, a steady rhythm of seed and early-stage activity that never paused. It also had extraordinary depth — hundreds of billions of dollars flowing into a small number of infrastructure and core-model companies that raised repeatedly, at enormous scale, in financing structures that often defy traditional venture labeling.

Those two markets coexist in the same dataset but barely interact. The broad market would exist without the capital-heavy one. The capital-heavy one would exist without the broad one. The fact that we call both of them "AI funding" is a linguistic convenience, not a structural description.

If this report leaves a reader with one mental model, it should be this: the AI financing landscape is not a pyramid with a wide base gradually tapering to a narrow peak. It is two separate layers — a wide, shallow pool and a narrow, very deep well — that happen to share a label.

Appendix A: Valuation and round economics

This appendix exists because valuation data is too thin for a main chapter but too interesting to ignore entirely.

Only 11% of rounds in the dataset disclose a pre-money valuation. To say anything defensible, we constrain the analysis to a narrow subset: private-market, fresh-capital, equity-like rounds with both a known round amount and a disclosed pre-money valuation. No debt, no grants, no liquidity transactions, no public-market rounds. That leaves 775 rounds — 9.5% of the full dataset.

That is a small window. But it is an honest one, and the patterns inside it are sharp enough to be worth reporting.

How small the window really is

The right panel matters most. Late venture / growth rounds have valuation coverage of 53% — enough to work with. Early venture has 22%. Seed drops to 9%. Pre-seed is 14%. So later-stage valuation findings rest on a reasonable base. Earlier-stage findings should be read as directional, not definitive.

What valuations look like by stage

Within the subset, the stage ladder works the way you'd expect — only steeper than most people assume.

Pre-seed to seed is a 3.4× jump in median valuation. Seed to early venture is 13.7×. Early venture to late venture is 14.7×. From pre-seed to late venture/growth, the total span is roughly 680×. A company that enters the window at a $4 million pre-money and reaches late-stage territory is being priced at nearly 700 times its entry valuation. That is the full range of the visible AI venture ladder in twelve months of data.

How much do companies give up?

The round-economics view asks a different question: not what are companies worth, but how much of themselves are they selling at each stage?

Early venture stands out. At a median round-to-pre-money ratio of 22.3%, these companies are raising a much larger fraction of their valuation in a single round than companies at any other stage. Late venture raises more in absolute dollars — $198 million versus $32 million — but against a much larger valuation base, so the proportional bite is smaller.

The 75th percentile for early venture is 69% round-to-pre-money, which means a meaningful share of early-venture rounds in the subset involve very aggressive raises relative to disclosed valuation. Those tails are precisely why this appendix uses medians, not averages.

Infrastructure prices differently

The company-type split shows up here too, even inside this narrow subset.

Foundation model companies in the valuation subset carry a median pre-money of $1.32 billion. Vertical application companies carry $6.5 million. That is a 200× gap. Even within a constrained, disclosure-selected subset, the infrastructure-vs-applications divide that defined the main report shows up just as starkly in how these companies are priced.

The coverage column provides its own signal. Foundation models and compute have the highest valuation coverage within their defensible denominators — 31% and 26% respectively. Application categories sit around 13–15%. Infrastructure companies are not just valued higher; they are more likely to disclose that valuation. The valuation data we can see is disproportionately drawn from the capital-heavy end of the market.

What this appendix can and cannot say

This is a selected subset. The 775 rounds that survive the filter are not a representative sample of all AI financing. They skew toward private venture equity (747 of 775 rounds), toward later stages, and toward companies willing to disclose valuation. The findings here describe the visible portion of the valuation landscape — useful for directional structure, not for market-wide claims.

Within that constraint, two things are clear. First, the AI venture ladder is very steep: each stage jump represents roughly an order-of-magnitude increase in valuation, and early venture is the stage where companies give up the most equity relative to their price. Second, the infrastructure/core-model layer prices at a fundamentally different level than the application layer — a gap that is visible even inside a narrow, disclosure-filtered lens.

Appendix B: Definitions and methodology

This appendix documents how the dataset was cleaned, enriched, and analyzed. It is the reference layer behind every number in the report. Readers who want to verify a specific claim or understand a specific definition should find it here.

Dataset source and window

The source file is funding_2025-04-01_to_2026-04-10.csv, drawn from Crunchbase's announced AI financing activity. It contains 8,202 rows, each representing one announced financing event. The date range spans April 1, 2025 through April 10, 2026. April 2026 is a partial month and is marked as such wherever monthly analysis appears.

All analysis in this report treats the dataset as announced financing activity — not as a census of all AI capital deployed globally. Some rounds may be announced weeks or months after they close. Some may never be announced at all. The dataset captures what was publicly reported, which is a useful but incomplete view.

Organization identity

Organizations are identified by Organization Name URL, a stable Crunchbase URL slug. This is more reliable than organization name alone because 41 names in the dataset map to multiple URLs (distinct legal entities sharing a name).

Using the URL key, the dataset contains 7,118 unique organizations. The raw name count is 7,076. This report uses 7,118 throughout.

Financing regime

Every raw Funding Type value (26 distinct types in the dataset) was mapped to one of seven normalized financing regimes. The full mapping:

Each regime also carries economic-form flags:

• fresh_capital_to_company: true for all regimes except liquidity/secondary, where the capital goes to existing shareholders.
• equity_like / debt_like: indicates the instrument's economic form. Convertible notes are flagged as both. Some ambiguous types (Funding Round, Non-equity Assistance) have blank flags where the form could not be confidently assigned.
• public_market / private_market: whether the issuer is publicly traded.

The full mapping table with all flags is stored in data_model/financing_type_mapping.csv.

Normalized lifecycle stage

Raw Funding Stage is incomplete (missing for 24% of rounds) and too coarse on its own. The report uses a derived lifecycle stage built from both Funding Type and Funding Stage:

Stage is kept analytically separate from financing regime. A debt facility is "alternative / non-stage-specific" by stage even though it is "debt" by regime. The full mapping is in data_model/lifecycle_stage_mapping.csv.

Company taxonomy

Every organization was classified on three axes using an LLM-assisted process:

Axis A — AI stack / layer (10 categories): Compute / chips / data center / energy, Foundation models / core model developers, Data infrastructure / MLOps / developer tools, Horizontal applications, Vertical applications, Robotics / autonomy / embodied AI, Healthcare / biotech / scientific AI, Security / safety, Services / implementation, Other / adjacent.

Axis B — End market (11 categories): Enterprise horizontal, Consumer, Healthcare / life sciences, Financial services, Industrial / manufacturing, Defense / public sector, Education, Media / creator, Climate / energy, Mobility / logistics, Other.

Axis C — AI purity (4 categories): AI-core, AI-native product, AI-enabled but not AI-core, AI-adjacent / ambiguous.

Classification used organization descriptions as primary evidence, with a structured system prompt enforcing single-label assignment per axis and explicit tie-break rules. Each classification carries a confidence rating:

The taxonomy is a first-pass LLM classification. It has not been individually audited for every company. The report's strongest claims lean on categories with large sample sizes and low shares of low-confidence labels. Residual categories — particularly "other/adjacent" (96% low-confidence) and "services/implementation" (41% low-confidence) — should be treated as approximate.

The "infrastructure vs applications" aggregate used throughout the report groups categories as follows:

• Infrastructure: Compute / chips / data center / energy, Foundation models / core model developers, Data infrastructure / MLOps / developer tools, Security / safety.
• Applications: Horizontal applications, Vertical applications, Healthcare / biotech / scientific AI, Robotics / autonomy / embodied AI.
• Not classified into either: Services / implementation, Other / adjacent.

Geography

Geography is based on organization headquarters, not deployment or customer location. The hierarchy:

• Country: cleaned and standardized from the raw Country column, with fallback to parsed Organization Location. Known for 95.2% of rounds.
• Macro region: Americas, Europe, Asia-Pacific, Middle East / Africa.
• Region: finer subdivisions — North America, European Union, United Kingdom, Rest of Europe, East Asia, South Asia, Southeast Asia, Oceania, Middle East, Africa, Latin America, Central Asia.
• Europe subgroup: EU, UK, Rest of Europe.
• U.S. state and metro: derived from location strings for U.S.-based organizations. Metro labels come from the city component of the location field and should be interpreted cautiously — some (e.g., Dover, Wilmington) reflect legal incorporation addresses rather than operating hubs.

Size buckets and mega rounds

Disclosed round sizes are bucketed as: <$1M, $1–5M, $5–15M, $15–50M, $50–100M, $100–500M, $500M+, undisclosed.

Mega round is defined as any round with disclosed Money Raised of $100 million or more. There are 285 mega rounds in the dataset (3.5% of all announcements).

Repeat raisers

An organization is a repeat raiser if it appears in more than one financing event during the window, using the stable URL key. There are 857 repeat-raiser organizations (12.0% of all organizations), accounting for 1,941 round rows (23.7% of all rows).

Stage movement for repeat raisers compares the normalized lifecycle stage of each organization's first and last round in the window. Organizations where both endpoints are "unknown" or "alternative / non-stage-specific" are excluded from the comparable set for stage-progression analysis.

Disclosure and completeness

• Money disclosed: Money Raised is non-null and parseable. 5,977 of 8,202 rounds (72.9%).
• Valuation known: Pre-Money Valuation is non-null and parseable. 899 of 8,202 rounds (11.0%).
• Country known: 7,809 of 8,202 rounds (95.2%).
• Website known: 6,805 of 7,118 organizations (95.6%).
• Description quality: rated high, medium, or low based on description length and richness. High or medium for 97.2% of organizations.

All dollar figures in this report are disclosed lower bounds. They sum only the rounds where Money Raised is known. Undisclosed rounds are excluded from dollar totals, never treated as zero. Every dollar-based comparison in the report should be read with the relevant disclosure rate in mind.

What the dataset cannot answer

This report does not and cannot address:

• The true total of AI capital deployed globally. The dataset captures announced activity, not all activity. The 27% of rounds with undisclosed amounts, plus rounds never publicly reported, mean the real total is higher and unknowable from this data.
• Investor behavior, syndicate structure, or lead investors. The dataset does not include investor-level fields.
• Returns, markups, or exit outcomes. No performance data is present.
• Revenue, profitability, traction, or product-market fit. The dataset describes financing events, not business quality.
• Causal explanations. The report describes patterns — which categories attracted capital, how concentrated the market was, where disclosure was weak. It does not claim to explain why capital moved the way it did.

Dataset: Crunchbase announced AI financing activity, 2025-04-01 to 2026-04-10. All dollar figures are disclosed lower bounds unless otherwise noted.