The capital register

Sixty-seven rows, one table — a few of them cover a cohort rather than a single company. The columns that matter are the last three: whether the revenue figure is gross or net, what multiple that implies, and whether anyone has tested the mark with cash.

Read the table with three warnings in hand. First, most of these companies disclose nothing — "not disclosed" is the modal entry and it is honest reporting, not laziness. Second, where a third-party estimator is the only source, the figure is marked [WEAK] and should not be used; see the evidence register for why. Third, a multiple computed on gross billings and a multiple computed on net revenue differ by five to ten times with no change in the underlying business — GMV is not revenue is the page that explains it and this table is where the damage shows up.

#The register

Multiple = last valuation ÷ the revenue in the adjacent column. Where a company books gross, the net multiple is given in brackets where net is knowable.

#Implied multiples, and where gross has been mistaken for net

Set the anchors before reading anything above. Appen trades at ~0.9x revenue (stockanalysis). Innodata trades at ~6.2x on 40% gross margins (stockanalysis). Those are the only two companies in this table doing something close to what the private set does, with audited numbers, in a market that can see inside them. Everything above 6x is a wager that the private companies are structurally different from the public ones — see What the public market pays for labour for why the public market caps a labour business near 1.6x and what the three exceptions did.

Mercor is the worked example. At the $10B mark on ~$760M of annualised gross, the headline is a defensible-sounding 13x. Contractors take 60–70% of that top line (Sacra), so the money Mercor keeps is roughly $230–300M and the real multiple is about 38x. At the reported $20B talks on a $2.0B gross run-rate, the headline is ~10x and the net figure is ~29x. Both numbers are true. Only one of them is a multiple on money the company owns.

The same arithmetic runs through the table:

• Handshake looks like the cheapest large name in the sector at 3.2x — and it is cheap, but by about half as much as it looks, because $1.10B gross is ~$450M net (Sacra). The real reason it screens cheap is that the $3.5B mark predates every dollar of AI revenue. Nobody has re-priced it.
• micro1's headline $500M gross run-rate is 2.5–3.3x the $150–200M net number that matters. The same TechCrunch sentence also says it retains 60–70% of gross, which cannot both be true; see What we could not establish.
• Turing's model is explicitly "a staffing spread — Turing manages the engagement end-to-end, pays developers on fixed terms, and earns margin on the difference" (Sacra). Reporting that spread at $300M "annualised revenue" and marking it at $2.2B produces a 7.3x headline that would be materially higher on retained margin.
• Prolific is honest by construction. It charges a 42.8% platform fee on top of participant rewards (Prolific pricing), so what it keeps is separable, visible and published. It is the only company in the vertical that does this.

The one figure in this document that should be retired: the "92x ShopMy" ratio. It does not exist. GMV to net is ~12.5x; valuation to net is 18.8x; valuation to GMV is under 1.5x. A multiple computed on GMV is always lower than one computed on net, and any figure claiming otherwise has an arithmetic error in it.

#The exits, and what buyers actually paid

There is no natural strategic acquirer that pays a software price for a human-supply business. The buyers fall into four groups and only one of them pays big.

(a) Labs and hyperscalers buying capability, structured as licence-and-hire. This is where the money is, and it is not paid for revenue.

• Meta / Scale, June 2025: $14.3B for a 49% non-voting stake at $29B, plus Alexandr Wang and a research team (NYT). Not an acquisition — a stake plus an acqui-hire, structured to avoid merger review. (The price circulates as both $14.3B and $14.8B; the atlas uses $14.3B and flags the discrepancy.)
• Google / Mechanize, in talks August 2026: $1.5B+ to hire some of 35 staff and take a non-exclusive licence (TNW). Precedents cited: Windsurf ($2.4B, Jul 2025) and Character.AI (2024). Mechanize's last priced round was $9.1M at $500M.

The pattern: the value is in the judgement of about thirty-five people, not in a revenue line. These deals price talent and IP. No multiple was involved and none can be inferred.

(b) BPO and IT-services strategics buying capacity. They pay services prices.

• EXL / iMerit, June 2026: up to $310M, closed August 2026. BII exited. This is the single most useful exit price in the register, because iMerit is a real operating data business and $310M is what a disciplined strategic paid for one.
• Uber / Segments.ai, October 2025: undisclosed. Uber's intent was distribution — routing labelling work to idle drivers.

(c) Holdcos and consultancies buying creator supply. They pay agency prices and they do not disclose.

• Publicis / Captiv8, May 2025: undisclosed. ContentGrip reports $175M; every other source says terms were not disclosed. Publicis had already bought Influential at ~$500M.
• Accenture Song / Whalar, announced 8 June 2026, closed 31 July 2026: undisclosed. Only the 170-person agency (Accenture newsroom). Accenture's stock fell on the day.
• Later / Mavely, January 2025: $250M cash from Nu Skin — the one clean disclosed price in creator M&A, and a strategic paid it.

(d) Staffing incumbents mopping up failed marketplaces. They pay scrap prices.

• Adecco / Vettery, 2018: $100M.
• Vettery / Hired, November 2020: undisclosed. Hired had raised $132.7M and been marked near $500M. By July 2024 the brand had been erased into Adecco's LHH (TechCrunch).
• Karat / Triplebyte, March 2023: undisclosed. Only the assessment products; the Magnet talent network was shut down (TechCrunch).

(e) Consolidation inside the sector, all undisclosed except one: Handshake bought Cleanlab (Jan 2026); Invisible agreed to buy WeCP (Mar 2026); Mercor bought Sepal AI (Feb 2026) and Deeptune (Jul 2026); Toptal bought YouTeam (Jan 2025); Andela bought Woven (Jan 2026); CreatorIQ bought Tribe Dynamics (2021); AlphaSense bought Tegus for $930M (2024).

#The failures, and why nobody printed a down round

Appen is the base rate, and the only company here that has run a full cycle in public.

• Peak: market cap above the equivalent of US$4.3B in August 2020.
• Concentration: at the peak, 80% of revenue came from five clients — Microsoft, Apple, Meta, Google, Amazon. (Appen's own FY2025 report puts top-5 concentration at 74.3%, up from 67.3%; that is a different fact about a different year and the two are routinely conflated.)
• Revenue, USD (stockanalysis): FY2021 $447.3M → FY2022 $388.3M → FY2023 $273.8M → FY2024 $235.2M → FY2025 $232.7M. A 48% decline over four years.
• Losses: $28.5M net income (2021) → –$239.1M (2022) → –$118.1M (2023) → –$20.0M (2024) → –$21.8M (2025).
• The trigger: on 22 January 2024 Alphabet terminated a contract worth ~US$83M — roughly a third of remaining revenue. The shares fell 40–41% in a day.
• Where it landed: A$330M market cap on 28 August 2026, a 97% drawdown from peak, at ~0.9x revenue. (That 97% is The Verge's figure and is carried as reported rather than as computed: this register's own two numbers — a ~US$4.3B peak and an A$330M cap — give about 92% if both are put in Australian dollars and about 95% if both are put in US dollars, so the exact digit depends on which currency series you use. See the standing warning against mixing the USD and AUD series in the table above.) The business has stabilised and is growing again on an AI-data tailwind. The market still prices it as labour services.

The lesson the private cohort has not absorbed: Appen's collapse was not bad execution on labelling. It was buyer concentration plus a change in training technique. Both conditions are present, in more extreme form, across the current set — Mercor reportedly derives ~91% of H1 2026 revenue from AI foundation-model companies. See One customer is a binary event.

#Investor behaviour, and the missing bear case

Who keeps writing the cheques. Felicis is the most concentrated bet in the category, having led both Mercor's $100M Series B at $2B (Feb 2025) and its $350M Series C at $10B (Oct 2025) — a 5x step-up in eight months, doubled down on. Sequoia led Juicebox's Series A and separately published the category's defining thesis. General Catalyst led Haize Labs' seed at $100M on a pre-revenue red-teaming company. Coatue appears in both eras: $200M into Handshake's Series F and $100M of Braintrust tokens. Bezos Expeditions led Toloka. Tether put $200M into Whop. Nvidia is reportedly weighing an investment in Mercor at $20B, which would make the largest supplier of AI compute an investor in the largest supplier of AI labour. Khazanah Nasional led Turing; Vanara Capital (a TPG spin-out) led Invisible; Wellington led Encord.

This is not classic early-stage venture. It is crossover, sovereign, corporate-strategic and now crypto-treasury capital, buying revenue growth at scale. See Selling through the investor.

The bull thesis is explicit and named. Sequoia's "Services: The New Software" (Julien Bek, 5 March 2026) argues that "the next $1T company will be a software company masquerading as a services firm"; that "for every dollar spent on software, six are spent on services"; that as models cross a competency threshold on intelligence work (rule-based) as against judgement work, companies that sell the work itself beat companies that sell the tool; and that "autopilots" selling outcomes capture the labour budget while "copilots" selling tools do not. It sizes recruitment at $200B+, consulting at $300–400B, IT services at $100B+. Mercor and Juicebox are named in it. Foundation Capital has run an adjacent "service-as-software" line since April 2024.

The load-bearing assumption is precise: that AI compresses the labour cost of delivery fast enough for gross margin to migrate from services levels (30–40%) to software levels (70–85%) before the multiple has to be justified. If that migration happens, 58x on Distyl is early. If it does not, it is 58x on a consultancy. What the public market pays for labour and What better models do to each layer are where the atlas argues both sides.

#The bootstrapped counterexamples

Two companies in this register refute the capital thesis directly.

Surge AI bootstrapped from 2019, profitable since launch, past $1B of revenue in 2024 (basis unstated, almost certainly gross) — ahead of Scale's $870M, which had raised $1.6B. Roughly 130 full-time employees against ~50,000 contractors: about $9M of billings per employee. Edwin Chen reportedly still owns ~75%. It opened to outside capital only in 2025, largely for secondary liquidity — and the resulting round is the single biggest unresolved number in this register.

Toptal raised $1.4M in 2012 and nothing since, has never disclosed a valuation, is still independent fourteen years later, is acquisitive, and won a jury verdict in January 2025 (Sacra) against a financier found to have orchestrated a plot against it. Every VC-funded competitor from its cohort — Hired, Vettery, Triplebyte, Andela, Braintrust, A.Team — is dead, absorbed or silent.

The mechanism generalises. Human-supply businesses convert cash into gross margin at a predictable rate and rarely need capital for anything except growth-stage working capital. Venture money here buys speed of headcount and speed of client acquisition, at the price of a valuation only a narrow future can justify.