Which side you build first

The chicken-and-egg problem has been studied empirically, once, properly. Eli Chait went through how the 100 largest marketplaces solved it and counted the strategies. Three answers cover most of the field, and one of them is an order of magnitude cheaper than the others.

Version One's marketplace guide supplies the sequencing instruction that goes with it — seed supply first, because "there's zero motivation for customers without inventory" — plus the expectations nobody sets: marketplaces take 3+ years to reach traction against 6–9 months for SaaS, and liquidity means something specific (30–60% visit-to-purchase conversion), not a rising GMV line.

#The question that actually kills these companies

Supply utilisation. Not demand.

Andrew Chen's post-mortem on the "Uber for X" wave is the clearest statement of it, and its diagnosis is entirely supply-side: rideshare works because a driver can fill 20–50 hours a week with continuous demand, whereas valet, car-wash, massage and cleaning have "demand that is often infrequent and there's spikes at a few points in the day," so the worker cannot stay productively employed. Those companies paid the same $300+ per supplier acquisition cost and the same labour costs without ever generating the volume to amortise them: "they can never dig out of that hole, and stay unprofitable forever." His conclusion is four words: supply side is king.

The mechanism is a loop, and it runs in one direction only. Cannot keep the worker busy → cannot pay them enough → they churn → the ones who stay are the ones with no alternative → quality falls → buyers leave. Amazon 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, with 15,000–30,000 US workers completing at least one task monthly. Wage studies across that period run from ~$1/hr (2013) to a 2018 study of 3.8M tasks finding a median $2/hr with only 4% of workers above $7.25/hr. A supply pool you do not invest in does not stay the same size. It decays into a thin, adversarial, low-quality residue that is still technically a network.

So 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 would accept, indefinitely? If the honest answer is "for the first three months of the first contract," you are building a project business with a recruiting cost, and the recruiting cost recurs.

That question has teeth in this atlas. In Expert data for frontier labs, where average rates run $81–$100+/hour, a doctor or ex-banker will tolerate a gap of days, not months. When Meta cancelled Mercor's "Musen" project weeks after a $10B round, workers were moved to a replacement at $16/hr, down from $21 (Forbes). Community trackers of Scale's Outlier report empty queues as one of the top complaint themes, alongside deactivations and payment problems [WEAK — community-data analysis, not press]. Utilisation is not an operational nicety in this sector; it is the thing the supply side talks about on Reddit, and those posts are a cost of goods.

#How the companies here actually did it

Handshake is the cleanest single-player case in the atlas, and nobody calls it that. It spent a decade building campus recruiting software: 17–20M students, 1,600+ institutions, roughly 500,000 PhDs and 3M advanced-degree holders already in-network. The AI-data business then launched in January 2025 on top of supply it had already paid for once — and went from $5–10M gross ARR to ~$1B gross annualised by April 2026, the fastest zero-to-$1B in the sector (Dealroom; Lenny's interview with Garrett Lord). The detail that makes the point: Scale AI and Mercor had been recruiting PhD annotators off Handshake. It disintermediated its own customers with a supply graph it had accumulated as a side effect of a different business.

Mercor built the funnel as the product. It started as an AI-interviewer recruiting marketplace and pivoted into training data — which means the 20-minute structured video interview, with up to three retakes, was a standalone 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 recruiting 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 their downtime — a 12-city India pilot from September 2025 and a US pilot with selected drivers (CIO). Its addressable supply is claimed at "8M+ global earners" [WEAK]. Marginal recruiting cost near zero, aimed squarely at the commodity tier — which is the tier already being automated away, so the strategy and the target may be mismatched.

Paraform changed which side is scarce. Hired, Vettery and Triplebyte all built marketplaces where the expensively-acquired side was candidates — a side that exits the market the moment it succeeds. Paraform's scarce side is recruiters, who are durable, repeat, and get better with use. That single swap is the strongest structural argument in Contingency recruiting marketplaces, and it is a cold-start decision, not a growth decision.

Toptal and Surge AI show what happens when you refuse to buy either side. Toptal raised $1.4M in 2012 and nothing since; every venture-funded competitor from its cohort is dead, absorbed or silent. Surge bootstrapped to more than $1B of revenue by 2024 (basis not stated by any source; almost certainly gross) — ahead of Scale's $870M, which had raised $1.6B — with roughly 130 full-time employees, by paying contractors well and letting labs come to it rather than building a sales organisation. Neither company subsidised a side. Both took longer. Both still exist and are still owned by their founders.

#Constrain the thimble

Every canonical success launched inside a constraint tight enough to make liquidity achievable at trivial absolute scale: Etsy's handmade rule, Poshmark's six months on fashion-obsessed users, Uber black cars in one city, Airbnb during one conference week. Tavel's formulation — "if you pick the right thimble, you can win the ocean, but you'll never win by going after the ocean first" — is the operating instruction, and it is also a supply-utilisation instruction in disguise. A narrow enough wedge means the few suppliers you have are busy.

#The question to answer before you build

Write down the answers to these three, in this order, before anything else:

1. What does side one get on the day nobody else is here? If the answer is "nothing," you have chosen the 1:1 path and should budget accordingly. See Building the supply side.
2. Can a single supplier fill a week, every week, at a rate they will accept? Compute it in hours, not in headline pay rates. If the work is spiky, your supply cost is a recurring re-acquisition cost, not a one-off.
3. Which side is scarce, and does it leave when it succeeds? Build the durable side first. Candidates leave; recruiters stay. Annotators churn; institutional supply graphs do not.

Then, and only then, ask whether there is demand. There nearly always is. That is not the part that kills you.