Version One's instruction is right and almost universally ignored: seed supply first, because there is zero motivation for customers without inventory. But the instruction is incomplete, and the missing half is what actually kills companies. Seed supply first, and seed only as much as your first contract will keep busy.
Answer the utilisation question before you recruit anyone
Andrew Chen's post-mortem on the "Uber for X" wave is entirely supply-side. Rideshare worked because a driver could fill 20–50 hours a week with continuous demand. Valet, car-wash, massage and cleaning had demand that was infrequent and spiky, so workers could not stay productively employed — and those companies paid the same $300+ per supplier acquisition cost with none of the volume to amortise it: "they can never dig out of that hole, and stay unprofitable forever." His summary is four words: supply side is king.
The question to answer before writing a line of code is not "is there demand?" It is: can one worker fill a week here, at a rate they will accept, indefinitely? If the honest answer is "for the first three months of the first contract," your supply cost is a recurring re-acquisition cost, not a one-off — and you should model it as a monthly line, not as CAC.
The decay path runs one way. Cannot keep the worker busy → cannot pay them enough → they churn → the ones who stay have no alternative → quality falls → buyers leave. Mechanical Turk is the fully-decayed end state: registered workers passed 500,000 by 2011, but by 2018 only about 2,000 were active simultaneously, and a study of 3.8M tasks found a median $2/hr with only 4% of workers above $7.25/hr. A pool you do not invest in decays into a thin, adversarial residue that is still technically a network.
Single-player mode is worth roughly ten times the alternative
Across the 100 largest marketplaces, single-player mode — the product is useful before a second side exists — accounted for 34% of solutions at roughly 10:1 revenue-to-funding. Filling empty seats accounted for 33% at roughly 1:1. See Which side you build first for the full argument. What matters operationally is that you can often manufacture a single-player wedge after the fact.
Handshake is the cleanest case in the atlas and nobody calls it that. It spent a decade selling campus recruiting software and accumulated 17–20M students, 1,600+ institutions and roughly 500,000 PhDs as a side effect. Launching AI data work in January 2025 on top of that graph took it from $5–10M to about $1B of gross annualised revenue by April 2026 (Dealroom). Scale and Mercor had been recruiting off Handshake; it disintermediated its own customers.
Mercor made the funnel the product. The 20-minute structured AI interview with up to three retakes was a standalone recruiting artefact before it was a supply pipe; by February 2025 it had evaluated 468,000+ applicants. The interview does the vetting that would otherwise be a headcount line.
Uber is running fill-empty-seats at a scale nobody can match. Uber AI Solutions routes labelling tasks to drivers and couriers in downtime, from a 12-city India pilot in September 2025 (CIO) — marginal recruiting cost near zero, aimed at the commodity tier already being automated away.
Paraform changed which side is scarce. Hired, Vettery and Triplebyte expensively acquired candidates, a side that exits the market the moment it succeeds. Paraform's scarce side is recruiters: durable, repeat, better with use.
Toptal and Surge AI refused to buy either side. Toptal raised $1.4M in 2012 and nothing since; Surge bootstrapped past $1B of revenue by 2024 — basis unstated, almost certainly gross — with roughly 130 employees, paying contractors well and letting labs come to it. Both took longer; both are still founder-owned.
Sourcing channels and what they actually cost
There is no published CAC benchmark for a 2026 crowd. What the record gives you is the cost floor of running one properly.
| Line | Cost | Note |
|---|---|---|
| Supplier acquisition, paid channels | $300+ per supplier | The "Uber for X" figure; treat as the order of magnitude, not the price |
| Identity verification | $1–3 per worker one-off [UNVERIFIED] | Persona starts at $250/month on a 12-month minimum, priced per successful verification, no charge on drop-off |
| Gold-standard / honeypot tasks | 5–15% of task volume [UNVERIFIED] | Pure overhead, and non-negotiable — see Who is actually on the other end |
| Trust, safety and quality ops headcount | 1 per 200–500 active contributors [UNVERIFIED] | The ratio that determines whether quality survives scale |
| Contractor tooling | Deel from $49/contractor/month | $2.94M a year at 5,000 contributors, before a single payout. Per-head pricing is built for staffing, not crowds |
| Mass payouts | Tipalti mass payments from $249/month plus per-transaction | Includes the W-8/W-9 and 1099/1042-S workflows, which is the actual reason to buy it |
The organic channels compound and nobody budgets them. An owned graph (Handshake), a standalone screening artefact (Mercor's interview), an existing earner base (Uber), referral inside a professional community — all near-zero marginal cost with a better quality distribution than paid. Use the paid channel to prove the funnel works, not to scale on.
Whatever channel you pick, publish the pay rate. In Voice, speech and low-resource language data the contributor pay is documented down to the dollar across the whole category because recruiting is marketing — advertised rates from $6/hr to $150 per recorded hour. That transparency is a supply weapon and an information leak simultaneously: the moment your rate is public, so is your spread.
Quality tiers, and building them before you need them
Buyers do not want one crowd; they want a ladder, and they will pay differently for each rung. Design three from the start:
- Open crowd. Document and liveness identity verification, watchlist and sanctions screening, gold-standard task injection, behavioural analytics — keystroke and timing distributions, paste-event detection, device/IP/ASN clustering for multi-accounting, inter-annotator agreement drift.
- VDI-only. Data never leaves a controlled enclave. Screenshot suppression, no bulk export, per-worker watermarking and canary tokens.
- Secure facility. No BYOD, no personal devices, no phones, clean desk, badged access, background checks proportionate to tier, and named-worker approval lists for the highest tier. This is a capex line, not a policy.
Assume contamination is already present at the open tier. On Mechanical Turk, 33–46% of crowd workers used LLMs on an abstract-summarisation task, detected by combining keystroke analysis with synthetic-text classification (Veselovsky, Ribeiro & West) — in 2023. For a lab buying human preference data, undetected LLM contamination silently converts the product into model self-distillation: the worst defect you can ship.
The retention problem when wages fall
This is the part of supply operations that has no clean answer, and the sector has run the experiment in public.
Mercor moved Meta "Musen" workers to a replacement project at $16/hr, down from $21 (Forbes). Outlier's baseline reportedly fell from $40–50 to $15–20 [WEAK — community source]. Handshake's "Project HH" paid 20–50% of earned amounts in May 2026 [WEAK]. And Appen — whose accounts are audited — disclosed crowd NPS falling from 33 to 22 while the business was otherwise stabilising.
Three consequences, in the order they arrive.
Adverse selection, immediately. The contributors with alternatives leave first. Your median quality falls before your headcount does, which means the metric that moves first is inter-annotator agreement, not attrition.
Reputation cost, within a quarter. These pools are shared across platforms and they talk in public. Pay cuts, deactivations and withheld pay are the dominant complaint themes on every community tracker in the sector. That is a cost of goods payable next quarter as higher acquisition cost and lower funnel quality.
Litigation, within a year. Non-payment and deactivation are now a distinct legal surface — see You pay weekly, they pay in sixty days and The law is about to arrive — and from 2 December 2026 Art. 11 of the EU Platform Work Directive requires written reasons for any refusal to pay an EU-resident contributor.
The one thing not to do: employ your way out of it. Managed by Q made cleaners W-2 employees with stock options, a 401(k) and health insurance. Quality and retention improved. WeWork bought it in April 2019 and sold it eleven months later for $25M — about 11% of the purchase price. Every point of quality bought with employment is a point of gross margin lost, and gross margin is what the exit is priced on (What the public market pays for labour).
The workable answer is narrower: constrain the wedge hard enough that the contributors you have stay busy, pay at the top of the band for the tier you actually need, pay fast, and never fund the receivable out of the crowd. When to stop has the numbers at which to stop.