Nobody writes this page. Every deck in the sector talks about gross margin; none of them mention that a labour middleman is, structurally, a bank that lends its customers money at 0% and funds the loan with its own equity.
The mechanic is simple and it never varies. You pay the crowd weekly or biweekly, because that is what it takes to keep people working. The buyer pays you net-30, net-45 or net-60 — sometimes net-90 — and starts the clock on invoice acceptance, not delivery. In between sits a permanent, growing hole. Even Upwork, which holds escrow and takes an 18.7% marketplace rate, flags it in the risk factors of its FY2025 10-K: losses from "clients failing to pay invoices, particularly in instances where we advance payments to talent for invoiced services on behalf of the client."
The arithmetic, properly
Two numbers get confused here, so separate them first.
The receivable is everything the customer owes you: 100% of billed revenue, outstanding for the payment term. On a $100M book at net-60 that is a $16.4M asset ($100M × 60/365).
The funding requirement is smaller, and it is the number that matters. You only have to find cash for the money that actually left the building — the crowd payments. The margin portion of the invoice never leaves your account. That gives:
Trapped cash ≈ revenue × (1 − gross margin) × (customer days − payout days) ÷ 365
At 25% gross margin, weekly payouts and net-60 terms: 0.75 × 53 ÷ 365 = 10.9% of annual revenue, permanently.
Work it through on a $100M gross book. Crowd cost is $75M a year, or $205,479 a day. Money goes out seven days after the work is done and comes back 60 days after the invoice. Fifty-three days of daily outflow are in the air at all times: $10.9M. That is not a peak. It is the floor, and it never comes back while the business is running.
Change the inputs — these are the only four that matter:
| Customer terms (payout at day 7) | 25% GM | 33% GM | 40% GM |
|---|---|---|---|
| Net-30 | 4.7% | 4.2% | 3.8% |
| Net-45 | 7.8% | 7.0% | 6.2% |
| Net-60 | 10.9% | 9.7% | 8.7% |
| Net-60 + 15-day acceptance lag | 14.0% | 12.5% | 11.2% |
| Net-90 | 17.1% | 15.2% | 13.6% |
All percentages are of annual gross revenue [derived arithmetic, illustrative]. Read the table twice. The first reading tells you that terms matter more than margin: moving a customer from net-60 to net-30 frees more cash than an eight-point margin improvement. The second tells you why the Expert data for frontier labs cohort's 27–33% gross margins are more dangerous than they look — a thin-margin business funds a larger share of every invoice.
Now add growth, which is where this stops being an accounting curiosity.
The gap scales linearly with revenue. Going from $100M to $300M in a year does not require funding $10.9M once; it requires holding $32.7M by year end, having found the extra $21.8M during the year, out of the growth you have not yet collected on. Triple, and you fund the gap three times over. This is why fast-growing, profitable, high-demand businesses in this sector run out of cash. Nothing is wrong with them. They are lending faster than they are earning.
For scale: Mercor was paying out more than $1.5M a day by October 2025 (TechCrunch) — roughly $550M a year of crowd payments. Apply net-60 terms to a payout cycle that fast and the funded position runs to the high tens of millions [derived, illustrative — Mercor's actual terms are not public].
The payout side has its own cost, separate from the gap. Paying weekly rather than monthly multiplies fixed per-payout fees by about 4.3x. PayPal Payouts charges 2% domestically (capped at $1.00) plus a 3.00% currency-conversion spread; Wise Business starts at 0.23% on the mid-market rate; Deel prices contractor management from $49 per contractor per month — $2.94M a year for a 5,000-person crowd in tooling alone. At crowd scale the binding cost is FX spread times payout frequency. Fast payment is a real recruiting weapon — Paraform advertises no 60-day hold as a reason recruiters stay — but it is bought twice: once in rails, once in the widened gap.
How to fund it, and what each option costs
Deposits, milestones and advance billing. The cheapest capital in the world and the first thing to try. It costs negotiating leverage, and it is easiest to win at the start of a relationship and nearly impossible to retrofit.
Buying shorter terms with a discount. Standard early-payment terms are the mechanism. 2/10 net 60 — 2% off if they pay in ten days — means paying 2% for fifty days of money, an annualised cost of roughly 14.9% (2 ÷ 98 × 365 ÷ 50) [derived arithmetic]. That is expensive money, but it is cheaper than equity and it is instant.
Receivables financing or factoring. Structurally the right instrument: you are financing a short-dated receivable from a very well-capitalised obligor. Two cautions. First, concentration cuts against you — a lender looking at a book that is 91% two customers sees one obligor, not a diversified pool, and prices or declines accordingly. Second, "well-capitalised" is not "investment grade"; private frontier labs have no public credit rating.
No factoring or receivables-financing rate for this sector appears anywhere in the research. Advance rates, discount margins and whether lenders will take concentrated private-lab receivables at all are unknown here. This is a cheap gap to close with two phone calls and it should be closed before it is planned around.
Equity. The default, and by a wide margin the most expensive. Capital in this model buys speed of headcount and speed of client acquisition — and it buys them at a valuation only a very specific future justifies. Using a round priced at tens of times net revenue to fund a 53-day receivable is the most expensive lending facility ever constructed. The counter-examples are instructive: Surge AI was bootstrapped and profitable from launch and passed $1B of revenue — basis unstated, almost certainly gross — before taking a dollar; Toptal raised $1.4M in 2012 and nothing since. Both grew slower. Both kept the company.
Slowing crowd payouts. Free. See below.
The trap
Closing the gap on the supply side is the one option that costs nothing on the day and everything afterwards, and the sector has now run the experiment twice in public.
Handshake. Contractors on OpenAI projects had accounts suspended between late December 2025 and January 2026 with pay withheld. Business Insider interviewed five; four were unpaid. Dozens more surfaced on Reddit, at least two lawsuits followed, and one court ruled a contractor was owed $6,475. The company's stated grounds were credential discrepancies, task times three to four times benchmark, and work performed outside the US. Its support line is the part worth quoting:
"This decision is final. There is no appeal process, and any work associated with this violation is not eligible for payment."
(Business Insider via AOL.) A community tracker separately documents a May 2026 payment episode on "Project HH" with workers receiving 20–50% of earned pay [WEAK — community data, not press]. See Handshake AI.
Mercor. Read the docket, not the headlines. Between 1 April and 26 May 2026, CourtListener's RECAP records a wave of federal filings against Mercor entities — Esson, Gill and Deboni (1 April), Lofton (2 April), Massman (7 April), Ramos (15 April), Ananthula (21 April), White (8 May), and Currey v. Mercor Global (26 May) — across N.D. Cal., N.D. Tex. and M.D. Fla. The nature-of-suit codes are 360 (personal injury, other), 380 (personal property) and 190 (contract), and the stated causes include diversity fraud and breach of contract. These are not wage-and-hour claims. They are non-payment and account-deactivation claims — money earned and not paid. [UNVERIFIED — the complaints themselves were not read; nature-of-suit codes are the only evidence of what they allege.]
That distinction is the whole point. The sector already knew it had a misclassification problem; see The law is about to arrive. What the 2026 dockets show is a second, separate legal surface that opens the moment payment timing becomes a cash-management lever. And the statutory regime is tightening around exactly that surface: New York's Freelance Isn't Free Act (GBL Article 44-A, effective 28 August 2024) and California's SB 988 (effective 1 January 2025) both impose written contracts, payment deadlines and — in New York — double damages plus fees [WEAK on the specific parameters]. From 2 December 2026, Article 11 of the EU Platform Work Directive requires a written statement of reasons for any refusal to pay for work performed, delivered no later than the date it takes effect, with a right to human review. "There is no appeal process" becomes unlawful for EU-resident contributors on that date.
Then the commercial cost, which arrives faster than the legal one. These supply pools are shared across platforms and they talk to each other in public. Pay cuts, deactivations and withheld pay are the dominant complaint themes across every community tracker in this sector, and Appen's own annual report shows crowd NPS falling from 33 to 22 while the business was otherwise stabilising. A reputation collapse in the recruiting funnel is a cost of goods, payable next quarter, in higher acquisition cost and lower quality — see Which side you build first.
Do not fund working capital out of the crowd. It is the cheapest money on the day you take it and the most expensive money in the sector eighteen months later — a class action, a discovery process, and a supply base that has moved to a competitor who paid on time.
One last connection. Whether this problem is yours at all depends on whether you book gross or net. An agent — Upwork's marketplace, where "talent and clients negotiate and agree upon the scope and the price directly with each other" — largely escapes it. A principal, who sets the spec, directs the work, invoices the lab and carries rework risk, owns it entirely, because the crowd payments are that principal's cost of goods. The same facts that put you on the gross side of GMV is not revenue are the facts that put the funding gap on your balance sheet. Choose deliberately, and price the 11%.