The working description this atlas was built to test is three lines long: a buyer with money but not capability; a seller side that is fragmented; and a middleman who organises the sellers and takes a big cut. Every vertical in the atlas is a test of whether that mechanism holds in a particular market, and the answer is different in each one for reasons that are structural rather than a matter of execution.
The mechanism deserves a sharper definition, because the loose one is doing damage:
An operator assembles a supply that cannot be bought as a unit, specifies and inspects an outcome the buyer could not specify or inspect alone, carries the liability for that outcome, and charges the buyer for the assembly rather than for the discovery. Every word is load-bearing. Assembly, not matching. Outcome, not hours. Liability, not introduction. Take out any one of them and you are describing a different, worse business.
It is not a marketplace, and that is not a quibble
The word "marketplace" carries an implicit claim about network effects, and NfX's operational test is the one to apply: does the platform provide a better experience to customer n+1,000 than to customer n, directly as a function of adding 1,000 more participants? For a labour middleman with a large but finite and broadly interchangeable supply pool, the honest answer is no. Past the point where a buyer can always get a qualified person inside their tolerance window, additional supply adds roughly zero buyer value while adding cost. NfX has a name for this shape — the asymptotic marketplace — and rates it more vulnerable than a classic marketplace, not less.
That is the single most under-appreciated structural fact in this model, and it is why "we have a million experts" is a slide rather than a moat. GLG has more than a million registered experts and watched its share of the expert-network market fall from 51% to 24%.
Calling it a marketplace mis-prices it three ways at once, and all three are visible in the current register.
It invites a net-revenue multiple on a gross-revenue business. A marketplace books a fee; an assembler books the buyer's whole cheque and pays the crowd out of it. Gross versus net moves the apparent revenue multiple by 5–10x with no change in the economics — C.H. Robinson trades at 1.14x on gross and about 13x restated to net. See GMV is not revenue.
It implies a take rate the form cannot support. Marketplace rakes cluster where discovery is the product. An assembler's spread is a services spread: Mercor's leaked gross margin was 27% in 2025 rising to 33% in Q2 2026 (The Information, via BigGo) — below Robert Half's 39.0% contract-staffing margin and below Appen at 40.3%. See What a rake can actually be.
It promises defensibility that does not exist. The only anti-leakage mechanisms that survive a sophisticated buyer are liability transfer and workflow embedding. Ratings, escrow and non-circumvention clauses are speed bumps — Getting cut out sets out why, and Upwork's own filing admits circumvention is "difficult or impossible to measure" two decades in.
The taxonomy that actually predicts the P&L is in Marketplace, staffing firm, BPO or agency. Two questions generate all of it: how much of the value chain you own, and who signs the worker's contract.
The four conditions
The model works only when all four hold simultaneously. Three out of four is not 75% of a business; it is a different business.
1. A buyer with capital, urgency and no capability. Not one or two of those. Capital without urgency produces a procurement process; urgency without capital produces a pilot that never converts; both without a capability gap produces a buyer who hires instead.
Where it fails: Design, video and content production. The budget is real but shallow — Series A/B companies spending $50k–$300k a year across creative and content — and the capability gap is imaginary, because the buyer's alternative is a designer whose salary they can look up, and generative tooling is dragging that salary down. A 70–88% inferred spread on that footing scores avoid.
2. A supply that is genuinely fragmented and hard to assemble. Hard is the asset. If money alone buys the supply, money alone buys your competitor the same supply.
Where it fails: Outbound and GTM-as-a-service. The scarcity that made Clay-operator agencies expensive in 2024 was gone by 2026, and the spread survives on operational competence rather than on anything anyone would call a barrier. The general rule sits in Andrew Chen's post-mortem on the "Uber for X" wave: those companies paid $300+ per supplier in acquisition cost, the same as everyone else, and never generated the volume to amortise it.
3. An outcome that can be specified and inspected. Someone has to be able to write down what "done" means and check it. Inspection is what lets you take delivery risk, and delivery risk is what you are actually charging for.
Where it fails: Community, campus and events. Developer distribution is sold as an outcome and approved on mindshare rather than pipeline; the deliverable resists specification, which is precisely why no intermediary layer has formed and why almost no economics have ever been published in the category. Inspection failure inside an otherwise-good vertical is just as fatal: Inc obtained internal logs showing Scale's Google programme was plagued by spam for eleven months, with contributors using ChatGPT despite a prohibition and faking advanced degrees (Inc). See Who is actually on the other end.
4. A reason the two sides cannot easily meet without you. Compliance, indemnity, capital, or a workflow the buyer runs their business inside. Not a contract clause. Not a relationship.
Where it fails: Paid creators, clipping and UGC ad ops. A brand and a clipper who have worked together once can transact by direct message forever after, and the marketplace take of 9–18% is priced accordingly — it is a payments fee wearing a marketplace costume. There is a second, subtler failure mode in Compute and capacity brokerage: once a supply becomes fungible and acquires a public price index, the two sides meet at the index and the middleman earns volume times a tiny fee. Indexation is disintermediation with better manners.
The verticals where all four hold are the short list: Adversarial evals and red-team crowds, where the supply cannot be recruited by job advertisement and the crowd's output is monetised as software rather than resold as hours; and Data-centre labour and site brokerage, where the unit is not fungible, not indexed and cannot be delivered from a laptop.
Note what "all four hold" does and does not mean. It is a claim about market structure, and structure is the part you can reason about from outside. Only Adversarial evals and red-team crowds carries a build stance on it; Data-centre labour and site brokerage sits on watch, because the four conditions there are argued entirely from the buyer's side and no intermediary in that vertical has published a single figure. A market that should support a middleman and a market observed to support one are different findings, and the second is the one worth money.
What to call it
Four names are in circulation and each one smuggles in a claim.
"Marketplace" claims network effects the form does not have. "Staffing" is closer to the P&L but understates the specification and inspection work, and implies the buyer manages the worker. "Managed marketplace" is a fundraising term whose entire function is to avoid answering who signs the labour contract. "Services-as-software" is a statement about the future written in the present tense — a target gross margin of 60–85% described as though it were a current one.
The honest name is the one that points at the scarce act. Call it the assembly business, and the operator an assembler.
The name earns its place because it tells you where the margin comes from and when it ends. The margin is paid for assembly — sourcing, vetting, specification, inspection, liability and payment across jurisdictions — and not for discovery, which is the cheapest function to replicate and the first one search, SEO and now language models attack. It ends when the supply becomes assemblable by anyone with an ad budget, or indexable by anyone with a price feed.
The pattern is old, and the record is not flattering
Nothing above is new. Expert networks have run it for four decades at a 70–80% take — a $200/hour expert rate against $800–900 billed to the client — and the product was never the expert; it was recruitment speed plus a chaperone the fund cannot legally do without. Staffing has run it since the 1940s, and Robert Half's own filings contain the cleanest statement of the whole thesis: contract staffing, which sells hours, at a 39.0% gross margin; permanent placement, which sells a matched outcome, at 99.8% — same company, same salespeople, same buyers (RHI FY2025 10-K). Advertising agencies ran it on a media commission that collapsed into fees. Freight brokerage runs it at ~8.5% on $17.0B of gross billings, licence and bond included, because discovery is most of the job and discovery is a price.
Then the largest run of all: the 2000s offshore BPO wave. Infosys at $20.3B, Cognizant at $21.6B and Accenture at $73.1B are the successful outcome of the biggest labour-arbitrage trade in history, and forty years of flawless execution buys ~30–33% gross margins and 1.4–2.3x EV/revenue. See What the public market pays for labour.
What is actually new in 2026
Three things, and only three.
The buyer's capital intensity. OpenAI is projected to burn ~$27B in 2026 against a ~$19B R&D compute budget; Anthropic reached a $65B annualised run rate by July 2026 and has discussed spending over $1B a year on RL environments alone (Sacra; Epoch AI). Megadeals of $100M+ took 87.5% of all venture capital in H1 2026, and OpenAI and Anthropic together took 43% of every venture dollar raised globally in the half (PitchBook; Crunchbase). No previous version of this model faced a buyer whose cheque was that large relative to the vendor. That is why One customer is a binary event has stopped being a risk factor and become the governing variable.
The speed of the budget. Handshake AI went from $5–10M to roughly $1B of gross annualised revenue in about fifteen months; Mercor doubled gross revenue in four months; micro1 grew 5x in eight. Version One's benchmark for a marketplace to reach traction is 3+ years. Contracts of six to seven figures per quarter are normal at the labs, and exclusive RL-environment deals price at 4–5x non-exclusive (Epoch AI) — the buyer paying for denial of the asset to rivals. Budgets now appear and are spent inside a quarter, which compresses Which side you build first and inflates what a first ninety days can prove.
Models are changing what "capability" means. ICONIQ's finding is the most thesis-relevant sentence in the benchmark literature: as AI products move from beta to scale, "talent's share of cost falls while inference rises" (ICONIQ). Seventy-eight per cent of AI companies are rethinking staffing, a third are planning smaller teams, and half are scaling forward-deployed-engineer models. The buyer's stated intent is to substitute purchased capacity for headcount — which is the assembler's tailwind. The same force is deleting the bottom of the market: Appen's global segment fell 21.1% to $127.9M in FY2025, and Sama issued redundancy notices to 1,108 Nairobi employees in April 2026. See What better models do to each layer.
What is not new is the margin. Forty years of this model has never produced a durable spread from assembling people, and the best-funded version of it in 2026 is running at 27–33%. The framework in How to evaluate one of these exists to work out, before you commit, whether a particular market is one of the rare ones where it can.