Global AI Venture Capital
Oct 2024 – Apr 2026

Ten companies took 80% of the money.

Top 10 = $528B, accounting for 81% of the total. Top 30 = 90%. The remaining 373 companies split less than 20%.

Eighty percent of the capital went to ten companies. That's money in the bank, not pledges.

Concentration doesn't mean the door is closed. Outside the Top 30, another 120 early-stage companies raised Seed or Series A: Rillet (Vertical AI · $95M), Emergent (Agents & Applications · $93M), Aspora (Other Tech · $88M), Modus (Agents & Applications · $85M), Gimlet Labs (AI Infra · $80M). They're scattered across the long tail.

The question: within that 80%, how much is genuine consensus from multiple VCs, and how much is a single big bet creating a mirage?

Heat alone isn't a signal—the shape of heat is.

30 sub-sectors absorbed 92% of the capital. Only a handful pass all three tests: high deal count, high valuations, and multiple lead investors. The rest owe their heat to one VC betting big on one company.

Three metrics must all rank in the top quartile for a sector to qualify as genuinely hot: deal count, median valuation, and number of distinct lead investors. Only 3 sectors pass: General LLM, Coding Agents, Cloud Compute.

Databases is a textbook fake-hot: ranks #10 in deal count, #2 in valuation, but only 3 lead investors. The entire sector's valuation is propped up by Databricks ($134B) alone. Self-driving is the same pattern—Wayve single-handedly inflates the average. Healthcare is the opposite: plenty of deals, multiple leads, but valuations haven't taken off yet.

How big is a typical round in each sector at each stage?

Not "can you raise"—but how big at each stage.

52 sectors × 5 stages, median single-round raise. This isn't cohort survival—companies in the window are at different stages, so you can't read "how many make it." What you can read: at each stage, how big is a typical check.

Three sectors, three rhythms.Foundation Models runs mega-rounds throughout: Seed $220M, Series A $1.3B, Series B $627M, Series C $2.0B, Series D+ $13B. Small teams can't even get a ticket.

Coding Agents follows the SaaS growth playbook: Series A $55M, Series B jumps 4× to $218M, Series C doubles again to $400M, then growth slows post-D. Early rounds price in fast; later rounds price in customers and retention.

Vertical Healthcare takes the smallest steps, but each is backed by revenue: Series A $32M, Series B $141M, Series C $126M, Series D+ $363M. Other Agent Apps has the lowest barrier: Seed $30M, Series A $29M. Less capital, but competition is also scattered.

Now look at whose hands the money comes from.

Led by 5 VCs vs. led by 1—fundamentally different sectors.

29 firms × 46 sectors, 279 lead rounds. The column header shows how many independent firms have led in that sector. Higher number = more potential lead investors for your next round.

Only 31 sectors have been led by ≥ 3 independent VCs. The most concentrated is General LLM: 22 firms, 30 total leads. The top three are Shanghai Guotou Pioneer Fund×3, SoftBank Vision Fund×3, Lightspeed Venture Partners×2.

5 sectors have only one lead investor: Vertical · Education, Creative Tools, MLOps, Other. If that fund changes direction or closes, you're starting from scratch for your next round.

The most active firm is Andreessen Horowitz, with 45 leads across 22 sectors—16% of the total sample.

All those numbers are nominal. How much is real cash vs. compute credits on paper?

$250B in press releases—maybe $75B lands in OpenAI's bank.

639/666 deals are pure cash. The other 22 are compute-for-equity or hybrids, comprising 72% of nominal value. The gap between headline numbers and actual cash can be 2–5×.

Two ledgers: press-release figures and actual cash are different numbers. Crunchbase records press releases; this report tracks both—headline amounts and real cash equivalents, side by side.

Microsoft × OpenAI (Oct 2025): The press release said $250B. Zero immediate cash—all five-year Azure lock-in. Cash equivalent $75B, a 30% haircut.

Amazon × Anthropic (Apr 2026): Nominal $25B ($5.0B cash + extended commitment), but Anthropic must commit $100B back to AWS. Cash equivalent $5.0B, a 20% haircut. Compare that to the Nov 2024 deal, which only discounted 12%. In two years, the cash-to-compute ratio in cloud giant contracts shifted from 9:1 to 1:4.

NVIDIA × OpenAI (Sep 2025): Letter of intent for $100B cash + $100B tied to GB200 purchases. Cash equivalent ranges from $10B to $100B—the spread itself signals structural uncertainty.

For contrast: Microsoft × Anthropic (Nov 2025) was pure cash $5.0B, press release = cash equivalent. That's a normal round.

Every fund's portfolio is its answer:"What counts as AI."

AI-native: remove AI, the product doesn't exist. AI-augmented: AI is a feature module. Line up 35 investors by this ratio, and each bar is a fund's answer to "what is real AI."

Deal-weighted average: 76% AI-native. A majority bet on native.

But zoom in to individual funds and the split is sharper than labels suggest. Greylock Partners, often perceived as old SaaS money, has 88% AI-native (24/27)—more aggressive than any other Sand Hill firm. Andreessen Horowitz talks "building the future" but runs only 69% (58/83). Managing 83 bets naturally requires diversification.

Most focused: Menlo Ventures at 96%—only native platforms. Most indifferent to the AI-native label: Founders Fund (58%), SoftBank Vision Fund (44%), and Balderton Capital (39%). Founders Fund's logic: Anduril, Hadrian—AI is a component, not the product. SoftBank and Balderton buy revenue growth, not technical definitions.

Beyond the mainstream bets, are there overlooked mega-sectors?

Where mainstream AI narrative doesn't look—three sectors pulling in serious money.

Foundation models and agents dominate the headlines. But rank by single-round size, three sectors lead: Defense AI, Robotics & Embodied Intelligence, Vertical Legal. Combined: 45 companies, 65 deals, $23.3B. Each has at least one company valued above $10 billion.

What these three have in common: model quality doesn't matter much. Defense AI sells "don't break in combat." Embodied AI's bottleneck is motors and gearboxes, not algorithms. Legal AI wraps existing APIs with compliance guardrails, billing for work that used to cost $600/hour. Pricing in these businesses tracks contract value, not model capability.

Helsing is in Germany, Wayve in the UK, Agibot in China, Legora in Sweden. Not one of them is in the Bay Area. The AI capital map is more complex than "Silicon Valley plus China."

Money departs from the US, passes through Europe, flows back to the US.China mostly recycles internally.

666 investments split into 14 capital-flow routes. Foreign capital into the US: 88 deals (13%); China's internal circulation: 74 deals, with only 2 flowing out. Europe's Series A post-money median is one-third of Silicon Valley's—the most visible entry-price gap in the dataset.

Note on scope: this chapter aggregates all 666 investments (including 14 historical-context deals from 2024-02..2024-09). A strict 18-month window (2024-10..2026-04) yields 652, with CN→CN ≈ 68 / US→US ≈ 383 / US→EU ≈ 55 / EU→EU ≈ 49—same shape, same arbitrage thesis.

US→US: 389 deals, the largest corridor. SG→US 29, ME→US 21, JP→US 15, EU→US 21. Global capital flows one-way into Silicon Valley.

Series A post-money median: US $1.5B (n=23), Europe $500M (only 33% of US, n=5), China $500M (n=3, very small sample). Same stage, same window—Europe and China price significantly below the US. This is the dataset's clearest entry-price gap. Capital sources: US→EU 56, plus EU local 50. China is a closed loop—no foreign capital inflow recorded in the dataset.

Stack the signals from the first nine chapters—eight sectors worth entering.

Five dimensions scored: deal density lead diversity recent acceleration entry round size valuation runway. Thresholds: ≥4 deals, ≥3 lead investors, ≤50 deals (excludes saturated mega-tracks). Take the top 8.

Calculated, not curated. #1 Vertical · Healthcare: score 0.86, 26 deals, 11 lead investors, recent acceleration 4.0×. The algorithm is public—change the weights and re-rank.

Common thread: no foundation models, no self-driving. Those are either saturated or dominated by one or two giants. Verticals dominate (2 vertical sectors + 3 agent apps)—all concrete workflows that general-purpose models can't directly replace. Entry barriers are moderate: median round size $29M, median lead count 7.

One last variable: who are the founders who got funded?

VC capital doesn't spread evenly across sectors—it clusters around a few pedigrees.

759 founders, 403 companies. Four panels: prior-company distribution, serial vs. first-time by region, PhD share by sector, top 10 schools.

Overall: 28% PhDs, 26% serial entrepreneurs. Foundation-model founders are majority PhDs; AI infrastructure about one-third; verticals see PhD shares drop notably. The two longest bars in Panel A: academia (publish then start) and acquired startups (cash out then re-enter).

PhD share by region: US 25%, Europe 33%, China 55%. Chinese founders have the highest academic credentials—Tsinghua/Peking PhDs plus GPU/robotics hard-tech combos are most common. US big-tech alumni and serial founders bring the average down.

Anthropic: 6/7 from OpenAI; Jared Kaplan from academia. Cursor: 4/4 studied or just graduated from MIT—no big-tech stints. Mistral's bench is DeepMind London + FAIR Paris. Moore Threads: all three founders from AMD China. Biren: one AMD, one SenseTime.

All data from public datasets: 666 investments, 403 companies, 36 investors, 759 founders. Source data, methods, and limitations are fully disclosed in Chapter XII.

How the data was built, and where the edges are.

Data window, coverage, limitations, source ledger, raw downloads.