Mercor was marked at $10B in October 2025 on roughly $760M of annualised gross billings. That is 13x — a rich but recognisable growth multiple. But contractors take 60–70% of that top line (Sacra), so the money Mercor actually keeps was roughly $230–300M, and the multiple on that is about 38x.
Both numbers are true. Only one of them is a multiple on money the company owns. Every serious mistake in this sector starts here.
Gross versus net changes the apparent revenue multiple by 5–10x with no change whatsoever in the underlying economics. C.H. Robinson books gross and trades at 1.14x EV/revenue; restate to net revenue and the same business is at ~13x — a headline software multiple, for freight brokerage. Never compare a multiple across a gross-booker and a net-booker.
The accounting rule underneath
Whether you may book the buyer's whole payment or only your cut is not a marketing choice. Under the revenue-recognition standard's principal-versus-agent test (ASC 606), it turns on control: if you control the service before it transfers to the customer — you set the price, you carry the delivery risk, the worker is contractually yours — you are the principal and you book gross. If you arrange for someone else to provide it, you are an agent and you book only your fee.
Upwork runs both models inside one 10-K, which makes it the cleanest teaching example in the record. Its marketplace revenue is $682.9M, a net fee representing an 18.7% take on $4.028B of gross services volume. Blended across all of Upwork's revenue — including the parts where Upwork stands between the client and the delivered work rather than beside it — the company books $787.8M, and the blended take on the same $4.028B of GSV is 19.6% (UPWK FY2025 10-K).
One company. One filing. Two revenue bases, roughly 5x apart in what they imply about size. Nothing about the economics changes between them.
The register, restated
| Company | Headline number | What it actually keeps | Multiple on gross | Multiple on net |
|---|---|---|---|---|
| Mercor | $2.0B gross annualised run rate (Jun 2026); $614M gross H1 2026 | 27–33% gross margin (leaked) | 13x at the $10B mark | ~38x |
| Handshake AI | $1.10B gross annualised (Apr 2026) | ~$450M net after contractor payouts | 3.2x | ~7.8x |
| micro1 | $500M gross run rate (Aug 2026) | $150–200M net (TechCrunch) | ~1x on a stale $500M mark | 2.5–3.3x the gross headline |
| ShopMy | ">$1B GMV" (2025) | $80M net revenue | <1.5x GMV | 18.8x net |
| Whop | $2.67B cumulative GMV (Feb 2026) | $142M net annualised (Oct 2025) | ~11x net | |
| Later (+ Mavely) | "$2.4B annual GMV run rate" (Dec 2025) | Not disclosed | — | — |
| Prolific | $350M annualised (Apr 2026) | Books its 42.8% platform fee — this figure is already close to net | — | — |
Sources: Sacra on Mercor, Handshake and Whop; TechCrunch, 20 Aug 2026 on micro1; Sacra on ShopMy; Later's own release via Google News; Sacra on Prolific.
Two things fall out of that table.
Handshake looks cheap at 3.2x and is not. The 3.2x is a multiple on gross billings against a $3.5B valuation mark that predates all of the AI revenue. On net it is ~7.8x — still the cheapest large name in the register, but for a different reason than the headline suggests: the mark is stale, not the business.
Prolific is the only company in the vertical whose take is separable by construction. It charges a 42.8% platform fee on top of participant rewards, so the fee is visible on the invoice rather than buried in a spread. That is why it is worth more as evidence than companies ten times its size.
Four ways the confusion actually propagates
These are not hypotheticals. All four are live in the source record this atlas was built from.
1. Two different numbers that look identical. Mercor's $614M for H1 2026 is gross revenue. Net on a $2B gross run-rate at a 33% margin is coincidentally also about $600–660M. A reader who sees "$614M" and a reader who computes "about $650M net" arrive at the same figure meaning completely different things, and the difference is roughly a factor of three in what the company is worth. Never collapse them.
2. A ratio that cannot exist. The claim in circulation is that ShopMy's GMV is "92x" its net revenue. It is not: >$1B of GMV against $80M of net revenue is about 12.5x. The real ratios are GMV/net ≈ 12.5x, valuation/net = 18.8x and valuation/GMV < 1.5x. The structural check is one line long — a multiple on GMV must always be lower than a multiple on net revenue, because GMV is the bigger denominator. Any figure that violates that ordering is arithmetic, not analysis. (This error appears in the atlas's own research notes; it is recorded here rather than quietly repaired.)
3. A take rate that does not reconcile with the revenue. ShopMy publishes 2.9% direct and ~3.9% subaffiliate. But $80M of net revenue on >$1B of GMV implies a blended take of roughly 8% — two to three times the published rates. Either GMV is understated, the published rates omit a fee, or the net figure includes something that is not affiliate commission. Nobody has said which. See ShopMy.
4. A take rate that does not reconcile with the volume. Whop's $142M annualised net at a ~5.5% blended take implies about $2.6B of annualised GMV — but the company's cumulative lifetime GMV was $2.67B by February 2026. The two statements can only both be true if essentially all of Whop's lifetime volume occurred in its most recent year, which is possible for a fast-growing business and is never actually claimed. As it stands the numbers describe two different companies.
In creator commerce, the lead investors have generally underwritten net revenue while the press has reported GMV. ShopMy's $1.5B mark on $80M of net is a rational-if-rich 18.8x; a reader anchoring on ">$1B GMV" would compute 1.5x and conclude it was cheap. The damage is done in secondary markets and in founder comparisons — someone benchmarking their own company against "ShopMy at 1.5x" is off by more than an order of magnitude.
How to read a press number in ten seconds
- Find the noun. GMV, GSV, GOV, GBV, "gross run rate", "annualised billings", "payouts to creators" — every one of these is somebody else's money passing through. "Revenue", "net revenue", "ARR" may be the company's own. Nothing else is.
- If the noun is missing, assume gross. In this sector the omission is the tell. Mercor's $2B "run-rate" milestone was covered almost universally without the word gross until TechCrunch and Sacra started using it.
- Ask who pays the worker. If the company pays the human out of the number quoted, the number is gross. If the human is paid by the buyer and the company invoices a fee, it is net. This one question resolves most cases without any accounting.
- Check the ordering. Valuation ÷ GMV must be smaller than valuation ÷ net revenue. If a comparison violates that, one of the inputs is wrong.
- Restate before you compare. A 3.2x and a 38x in the same table are usually the same kind of business measured two different ways.
Where the record is thin
Scale, Surge, Turing, Snorkel, Labelbox, Sama, Toloka and Later all report or leak gross and have never published a net figure. Every net multiple in this atlas for those companies is therefore a range built on a reported or inferred contractor share, not a disclosure. Scale AI's own revenue basis is unreconciled between $870M (2024), an estimated ~$2B for 2025, and company guidance to "surpass $1B" in 2026 — figures that cannot all be on the same basis.
The two exceptions worth their weight: Prolific, which charges a separable platform fee, and Appen, which employs or contracts its crowd and books the work — its US$230.8M FY2025 at a 40.3% gross margin is audited, and it is the only number in the vertical that has been through an auditor at all. That is a large part of why Appen at ~0.9x revenue is the anchor the whole private set is priced against. See What the public market pays for labour.