Capability Gap

← All topics

One customer is a binary event

Above 15% of revenue a customer is a coin-flip, above 25% you are a division of that customer — and when your buyers compete with each other, neutrality is the product you are actually selling.

high confidence8 minupdated 2026-08-29concentration · buyers · risk · ai labs

Customer concentration is the only risk in this atlas with a clean base rate, a live natural experiment, and a survivor's blueprint all in the public record. It is also the one every private company in the sector is currently running hotter than the company that died of it.

So what

Any single customer above ~15% of revenue should be modelled as a binary event, not a growth line. Above 25%, you are a division of that customer with a different cap table. Concentration risk is not linear in the percentage. It is a cliff, and the cliff moves closer when your buyers compete with each other.

The base rate: Appen

Appen is the only company in this space that has run the full cycle in public, and the shape of it should be printed on the wall of every board meeting in the sector.

Figure
Peak market cap, August 2020US$4.3B
Revenue concentration at peak80% from five clients — Microsoft, Apple, Meta, Google, Amazon
Google contract terminated, 22 Jan 2024~US$83M, roughly a third of remaining revenue
Share price reaction, same day–40% to –41%
Revenue, FY2021 → FY2025 (USD)$447.3M → $388.3M → $273.8M → $235.2M → $232.7M
Drawdown from peak97%
Market cap, 28 Aug 2026A$330M

Sources: The Verge's December 2025 feature for the peak and the 97% drawdown; stockanalysis.com for the revenue series and current market cap; NBC for the Alphabet termination. Note the currency bases differ — the revenue series is USD, the market cap AUD. Never mix the two series.

The lesson is not that Appen executed badly. FY2025 was a decent year: operating revenue $230.8M, underlying EBITDA up 250% to $12.2M, gross margin 40.3% — a better gross margin than Mercor runs today. The collapse was caused by buyer concentration meeting a change in training technique. And Appen's top five are still 74.3% of revenue, up from 67.3%. Concentration does not necessarily fall as you shrink; it can rise.

One layer down the same supply chain, Sama issued redundancy notices to 1,108 employees at its Nairobi delivery centre in April 2026 after Meta terminated a single content and annotation contract (TechCabal). One buyer decision, a thousand jobs, weeks.

Where the current cohort sits

The private set is running the same exposure, and in several cases worse.

CompanyConcentrationSource
Mercor~91% of H1 2026 revenue from AI foundation-model companies, OpenAI and Anthropic dominantThe Information, via BigGo
TaskUstop 10 = 58%, top 20 = 71%, Meta alone 26% (up from 22%)TaskUs FY2025 10-K
Innodatalargest customer 37% of Q2 2026 (from 56% in Q1), second 34%, top two 71%StockTitan
Appentop five 74.3%Appen AR2025
Concentrixfive largest ≈ 19%, 2,000+ clientsConcentrix FY2025 10-K
GLGtop ten 19.0%, no client above 6%, 90% of revenue recurringGLG S-1, Oct 2021

Mercor's founder, Brendan Foody, has publicly compared the company's concentration to Nvidia's — four customers at 61% of revenue — as a defence of it [WEAK — the comparison appears in the research record without a retrievable primary link]. It is a revealing analogy. Nvidia's four customers each need Nvidia more than Nvidia needs any one of them, and Nvidia has no substitute. A data vendor at 91% has neither condition.

Not verified

The GLG "no client above 6%" figure is cited to the October 2021 S-1 in one research file and flagged as unsourced in the contradictions register, which also records the S-1 as unread. The 19.0% top-ten figure and the 90%-recurring figure come from the same document. Treat the 6% as directionally right and worth verifying against the filing before it is quoted anywhere consequential.

Neutrality is the product

The sharpest fact in this atlas is that Meta's $14.3B was simultaneously the largest financing and the largest customer-loss event in the category's history.

On 12 June 2025 Meta bought a 49% non-voting stake in Scale AI at a ~$29B valuation and took founder-CEO Alexandr Wang to run its superintelligence effort (NYT; Wikipedia reports the figure as $14.8B — the two circulate interchangeably). What followed, within weeks: Google cut ties, having spent ~$150M with Scale in 2024 with ~$200M planned for 2025; OpenAI departed; Microsoft pulled back; xAI began exploring alternatives (Sacra; Computerworld). In July 2025 Scale cut 200 FTEs (~14%) and roughly 500 contractor relationships.

Then the punchline. By August 2025, Meta's own researchers reportedly rated Scale's data low quality, and Meta itself was routing work to Surge AI and Mercor (TechCrunch). The buyer that destroyed the vendor's neutrality did not even end up consuming the output it had paid $14.3B for a claim on.

Read

If your buyers compete with each other, neutrality is the product. Anything that compromises it — equity, exclusivity, a shared executive, a board seat — destroys revenue outside that buyer faster than it grows revenue inside it. This is why Google's approach to Mechanize was reportedly structured as a technology licence plus hiring select staff at $1.5B+, not an acquisition (TNW). The structure is the lesson.

What the survivors did differently

Set the two spread books beside each other. GLG: top ten at 19.0%, no client above 6%, 90% of revenue on subscription, 22 of the top 25 clients retained for over a decade, 2,700+ clients, segment contribution margin above 70%. Concentrix: five largest at 19%, 2,000+ clients, 160+ Fortune Global 500 brands, largest-client relationships averaging 16 years.

Neither of those is an accident of growth. They are policies. Six things separate them from the casualty list:

  1. Contract structure over transaction structure. GLG converted a per-call spot business into account-level subscriptions. That turns revenue visibility from a weakness into a strength — the same move ADP made in payroll.
  2. Deliberate buyer fragmentation as a rule, not an outcome. "No client above 6%" is a decision you make while turning down revenue.
  3. Multi-year embedded relationships instead of project wins. Sixteen-year averages are not sold, they are installed.
  4. Never take equity or strategic investment from one buyer in a competitive buyer set.
  5. Sell into a function, not a project. Functions have budgets that renew; projects have budgets that end.
  6. Own the compliance obligation. GLG's chaperoning, ADP's payroll-tax liability. This is the only moat in the whole model that gets stronger as the buyer gets more sophisticated — see Getting cut out.

And the honest coda: diversification prevents the cliff, it does not create a multiple. Concentrix trades at 0.70x EV/revenue with a $1.32B net loss. GLG carried $959.9M of debt into an IPO it withdrew in March 2022. GLG is the best-designed version of this business anyone has built and it still could not get out at a good price. Spread your book because the alternative is Appen, not because it will re-rate you.

The in-housing clock

Concentration and in-housing are the same risk viewed from two ends. The trigger is predictable: a buyer in-houses when your function becomes (a) large enough to matter in their P&L, (b) strategic enough to be a differentiator, and (c) legible enough to hire for. All three are already true at the commodity end of Expert data for frontier labs — OpenAI hired 100+ ex-bankers directly, xAI cut 500 generalists to build a specialist team in-house. Track the three conditions per customer. When all three go true, the clock has started, and the only question left is whether your book is wide enough to survive it.

Linked from
vertical

Expert data for frontier labs

Labs buy throughput of credentialled human labour — annotation, preference data, reasoning traces, RL environments — and pay nine figures for it at a staffing margin.

company

Mercor

The fastest-growing broker of expert human labour to AI labs — $2B gross run rate on a 33% gross margin and 91% revenue from foundation-model companies.

page

What the model actually is

Buyers with money but not capability, sellers who are fragmented, and a middleman who organises them and keeps a big cut. Four conditions have to hold at once — and calling the result a marketplace is how it gets mis-priced by a factor of ten.

page

How to evaluate one of these

Six scores, one to five, higher always better for the operator — what each is trying to capture, where each misleads, and the three disqualifiers that override any total.

playbook

Pricing and the contract

The buyer's alternative is nearly always to hire someone, and increasingly an offshore someone at half the cost. Price against that — then decide gross or net deliberately, because the same facts that book you gross make you an employer.

vertical

Paid creators, clipping and UGC ad ops

Pay-per-view clip campaigns and creator-made ad assets. The take is thin, the cheapness claim breaks for exactly the buyer with money, and the evidence base is the weakest in the atlas.

company

Scale AI

Took $14.3B from Meta for 49% of itself and lost most of its frontier-lab book within weeks — the cleanest natural experiment in why neutrality is the product.

page

The capital register

Every company the sweep found, with what it raised, what it was marked at, what it earns and whether that revenue is gross or net. Plus the exits, the failures and the absence of a public bear case.

playbook

The first ninety days

A sequence: prove buyer urgency with money before building supply, build only what the first contract consumes, and measure seven things at day 90 — one of which decides everything.

topic

What the public market pays for labour

Below ~40% gross margin the revenue multiple is capped near 1.6x, permanently. Accenture proves the ceiling; Fiverr proves that clearing it is necessary and not sufficient.

vertical

Adversarial evals and red-team crowds

Labs pay five and six figures per model to be attacked; the attackers are paid in prize money. The widest spread and the scarcest supply in the sweep.

page

Evidence register

Not a bibliography — a graded list of the claims the atlas leans on, what each one holds up, where it came from and how much weight it will take.

topic

How much money is actually in the buyer pool

There are two buyers, not one: about 10–20 labs signing six-to-nine-figure contracts, and several thousand startups buying at roughly $19K. Quoting the $510B headline as a TAM confuses them.

playbook

Operating from Europe, selling to the US

The entity can be European. The crowd should not be — from 2 December 2026 an EU-resident contributor is a presumptive employee and the anti-fraud stack that works becomes unlawful.

company

micro1

5x'd in eight months to a $500M gross run rate, and is the only vendor claiming 80–90% gross margin on anything — by reselling the same dataset more than once.

vertical

Robotics teleoperation and physical-world data

Buyers pay $50–200/hour for demonstrations, operators get $25–50. The rig and the floor space are what make this defensible — and what the commodity end lost.

page

How this was built

Eight research passes, about 54,000 words of notes, a writing pass and a verification pass — built with a search budget that ran out partway through, which shaped what is here and what is missing.

company

Turing

A remote-developer staffing firm that repositioned as a lab data provider — profitable, $2.2B marked, and with no hard revenue number since 2024.

vertical

Expert networks

$200 to the expert, $800–900 to the client, held for four decades. The oldest version of this model has the most durable rake in the atlas — and may not sell to AI labs at all.

company

Invisible Technologies

An operations-as-a-service business that discloses a real profit figure — $134M revenue, $15M EBITDA — and is marked at 15x for it.

company

Appen

The only pure-play with audited numbers, and a 97% drawdown from peak — the base rate for what a concentrated data vendor is worth when one hyperscaler leaves.

topic

You pay weekly, they pay in sixty days

At a 25% gross margin on net-60 terms, roughly 11% of annual revenue is permanently trapped in the gap — and it has to be funded again every time you grow. The tempting fix is to fund it out of the crowd, which is how you acquire a docket.

vertical

Forward-deployed engineering

Distyl at $1.8B on ~$31M ARR is the purest test of services-as-software. Palantir's 82% gross margin is the only proven escape from the labour multiple — and it works because the humans install a licence rather than being the product.

company

Mechanize

$9.1M of seed capital, ~35 people, no disclosed revenue, and reported Google talks at $1.5B+ — the licence-and-hire template for selling environments instead of hours.

company

Prolific

The only company in the vertical that publishes its take rate — 42.8% on top of participant pay — and the best public anchor for what this spread actually is.

vertical

Compute and capacity brokerage

The largest new budget line in the economy, brokered at the thinnest spread in the atlas — because compute is the one supply here that has a public price index and, from October 2026, futures.

vertical

Data-centre labour and site brokerage

A 439,000-worker shortage, a 30% wage premium, and $14.2M a month burning on a delayed 60MW build — against a supply that has no price index and cannot be delivered from a laptop.

company

Gray Swan AI

15,000 red-teamers paid in prize money; revenue from the software their attacks train. The widest inferred spread in the atlas, on the thinnest evidence.

company

Distyl AI

$1.8B on an estimated $31M ARR — 58x, the highest multiple in the register, paid for 111 people doing implementation work.

company

GLG

The largest expert network, a 70–80% take held for decades, and a market share that halved anyway. Its 2021 S-1 is the highest-value document in the atlas — quoted in one research file and recorded as unopened in another.