Global venture funding in H1 2026 was $510 billion — more than all of 2025's $440B, in six months (Crunchbase). It is the number every deck in this sector leads with, and it is the least useful number available.
OpenAI and Anthropic alone took $217B of it — 43% of all global H1 2026 venture funding (same source). Strip those two out and the headline has already lost nearly half its mass to two counterparties who are not buying what most vendors sell.
There are two buyer pools with almost nothing in common: ~10–20 labs with effectively unbounded budgets and formal security review, and ~8,000–11,000 ordinary funded startups whose individual budgets have not grown much at all. Almost no vendor serves both well. Mercor at ~91% lab revenue and Vanta at 16,000 startup logos are the two poles, and they are different companies in every respect.
Buyer one: the labs
Ten to twenty entities, and the money is real to a degree that is hard to convey.
OpenAI is at $40B annualised as of July 2026, projecting roughly $27B of cash burn in 2026 and ~$63B in 2027, with a 2026 R&D compute budget of ~$19B, roughly double 2025 (Sacra; Epoch AI). Anthropic went from $47B annualised in May 2026 to $65B in July 2026, has 1,000+ customers spending $1M+ annually as of April 2026, and discussed spending over $1 billion a year on RL environments alone as of September 2025 (Sacra; The Information via Epoch AI).
Against that, the contract sizes are exactly what you would expect:
| What | Price |
|---|---|
| Normal data/environment contract band | Six to seven figures per quarter; one founder: "often seven figures per quarter or more" |
| A neolab researcher's typical deal | $300K–$500K |
| Per task | $200–$2,000 typical; $20K/task rare but possible for complex SWE work |
| A website replica | ~$20K |
| A high-fidelity Slack clone | $300K+ |
| Exclusivity | 4–5x the non-exclusive price — labs pay largely to deny the asset to rivals |
| Surge AI's described frontier-lab contracts | Eight and nine figures |
| Google's spend with Scale AI | ~$150M in 2024, ~$200M planned for 2025 — with a single vendor |
Sources: Epoch AI, State of RL Envs; Inc via Yahoo Finance; Sacra.
The procurement bar matches the cheque. OpenAI's trust centre lists SOC 2 Type 2, ISO/IEC 27001, 27017, 27018, 27701 and 42001, PCI DSS v4.0.1, CSA STAR, GDPR, TX-RAMP and FedRAMP 20x, plus a published Supplier Code of Conduct and a documented supply-chain risk assessment process (trust.openai.com). A buyer holding ISO 42001 and FedRAMP will run a formal third-party risk assessment on anything touching training data. And they multi-source deliberately: Mercor claims all top-five labs, Surge AI lists ~12 frontier labs. Being one of three approved vendors is the pattern; sole-source is not. See Expert data for frontier labs.
Buyer two: the funded startups
The addressable count is the number the thesis actually needs, and no published source states it.
The $10M+ and $50M+ round counts below are derived from quarterly Crunchbase figures, not published totals. They are [UNVERIFIED] and should be treated as order-of-magnitude only. Verifying them against Crunchbase Pro or PitchBook is cheap and high-value, and has not been done.
- $10M+ rounds. Early-stage Q1 2026 was $41.3B across ~1,800 deals (mean $23M). Annualised, roughly 5,000–7,000 companies globally per year → ~8,000–11,000 distinct companies over 18 months.
[UNVERIFIED] - $50M+ rounds. Late-stage Q1 2026 was 584 deals for $246.6B; adding large Series Bs gives an order of magnitude of 1,500–2,200 per year, ~2,500–3,500 over 18 months.
[UNVERIFIED] - $100M+ rounds. This one is published: 158 companies globally in Q1 2026 (Crunchbase). Annualising gives ~500–600 a year, ~700–900 over 18 months.
The price these companies pay is the more useful anchor. Vanta runs ~$19K of ARR per customer across 16,000 customers ($300M ARR, April 2026, +69% YoY) and is "the de facto solution for three-quarters of YC companies" (Sacra). That $19K is the empirical price point for a horizontal service sold to funded startups at scale: above credit-card self-serve, comfortably below anything that triggers procurement — a founder or head-of-engineering decision. It is priced on volume of work absorbed, not seats.
Everything else about this buyer is constraint. Median VC-backed runway is ~14 months, which is a hard ceiling on the contract length they will sign. And the substitute is already deployed: offshore headcount allocation rose from 24% to 30% year-on-year at a stated cost arbitrage of 40–50% (SaaStr on ICONIQ). A vendor is not competing against an SF salary. It is competing against an offshore FTE at half of one. That is the pricing reality behind Contingency recruiting marketplaces, Paid creators, clipping and UGC ad ops and every function-level vertical in the atlas.
Why the $510B is not a TAM
Because the distribution has no middle.
- Megadeals ($100M+) captured 87.5% of all capital in H1 2026 (PitchBook).
- $235B went to 158 companies raising $100M+ in Q1 2026 out of ~$300B and ~6,000 funded startups. 2.6% of funded companies took 78% of the money.
- Four rounds were 65% of global Q1 2026 funding: OpenAI $122B, Anthropic $30B, xAI $20B, Waymo $16B. Four of the five largest venture rounds in history closed in one quarter.
- The top 1% of companies captured a third of all capital; the bottom 50% received 7% (SVB via SaaStr).
- Carta's primary-round data, Q1 2023 → Q4 2025: dollars +130% ($12.8B → $29.5B) while deal count rose 3% (802 → 823). The money grew; the number of buyers did not.
That last line is the whole argument. The pool of companies that can write you a cheque has been roughly flat for three years; the cheques a handful of them can write have doubled. A TAM computed by multiplying "$510B of venture funding" by an assumed services share is measuring the wrong quantity — most of that money is in accounts that will either never buy from you or will buy at a scale where you are one of three vendors and they set the terms.
PitchBook says it directly: "the market is setting records at the very top while contracting in almost every other segment underneath it." Underneath the megadeals, US VC fundraising fell ~20% YoY to its lowest since 2019, first-time fund formation is on pace for its lowest year since 2016, and NA Q2 2026 seed was down 27% YoY with early-stage deal count at a five-quarter low.
The counter-case
Three facts that a model of this buyer has to survive.
Brex is the cleanest natural experiment. Its entire business was selling corporate cards to VC-backed startups. When startup spend fell, its annualised revenue bottomed at $312M in 2022 before recovering to $700M by August 2025 — and its strategic response was to stop serving SMBs and startups and move upmarket to enterprise (Sacra). Any model of this buyer needs to survive a 30–40% revenue drawdown without a change of strategy, because the company that knew this segment best did not.
The customer dies at a known rate. Carta's Class of 2018, tracked over seven years: 62% shut down, 15% reach Series B, 1.1–1.3% raise at $1B+ (SaaStr on Carta). On a five-to-seven-year view roughly two-thirds of your logos churn for reasons entirely outside your control. No amount of customer success fixes a customer that ceases to exist. The offset is real — NDR of ~110–120% and survivors growing 3–5x a year — but the segment only works if expansion on survivors outruns death.
And the lab buyer's concentration is not a risk, it is a cliff. Appen had 80% of revenue in five clients and a ~US$4.3B peak; Alphabet terminated a ~US$83M contract in January 2024 and the shares fell 40–41% in a day, on the way to a 97% drawdown. Sama issued redundancy notices to 1,108 Nairobi employees in April 2026 after Meta ended one contract (TechCabal). Scale AI cut 200 FTEs and ~500 contractor relationships in July 2025 after Meta's $14.3B investment triggered a customer exodus. See One customer is a binary event.
The lab pool has unlimited budget for the right thing and sits in two counterparties holding 43% of world venture funding. The startup pool is broad, cheap per unit and mortal — ~$19K ACV, 14-month runway, 62% eventual failure — and cannot be served as high-touch bespoke services. Sell into the segment that is expanding, and price for the segment that is contracting.
Where the record is thin
There is no published benchmark anywhere for professional-services or outsourced spend as a share of startup opex, and probably no public one exists. The 24%→30% offshore-headcount figure is the best available proxy. Also missing: any S&M/R&D/G&A split by ARR band (it sits in gated ICONIQ and Bessemer PDFs); non-payment and bad-debt rates when selling to startups; and any dataset attributing vendor churn to customer death. Who signs a data contract at a lab, by title, is also unestablished — Epoch's reporting implies research-led budget authority, but no source names a role.