Selling Data to Frontier Labs: The Data Generation Industry

The framework

THESIS: Data-vendor revenue is real but non-recurring (e.g. one vendor: $30M first quarter, $100M open POs, ~$150M by December, <12 people). The market keeps replacing its own product — each model generation graduates past the data that trained it.

WHAT IS SOLD (6 products, one label): HOURS — labeling/ratings, the crowd era product, status: dying. JUDGMENT — doctors, lawyers, physicists at $100–500/hr, status: peak demand. WORLDS — simulated software + rubrics, $20k–$300k per environment, status: the current wave. VERDICTS — benchmarks, evals, red teams, status: becoming compliance. BODIES — gloves, rigs, capture ops; robots need to watch hands, status: supply-starved. RIGHTS — licensed archives, text into an annuity, status: renegotiating up. Footer: marketplaces pass 60–70% of billings to experts; offshore delivery keeps 80%+; same invoice, different business.

MARKET MAP (July 2026): SELL EXPERT HUMAN DATA — Surge AI, Scale AI, Mercor, Handshake AI, micro1, Turing, Invisible, Toloka, AfterQuery, Evaratus, Deccan AI, Prolific, Rise Data Labs. BUILD TRAINING ENVIRONMENTS — Mechanize, Fleet AI, Deeptune*, Sepal AI*, Datacurve, Bespoke Labs, Proximal, HUD, Plato, Matrices, Chakra Labs, Refresh, Vmax, Andromede, AIChamp, Habitat, Halluminate, Collinear, Veris AI, BenchFlow, Huzzle Labs, Calaveras, Idler, Originator, Hillclimb, Phinity, Quesma, Aviro, Haladir, Theta, Preference Model, Verita, Vetto, General Reasoning, Good Start Labs. SCORE THE MODELS — LMArena, METR, Epoch AI, Vals AI, Patronus AI, Haize Labs, Gray Swan, Andon Labs, Artificial Analysis, Trajectory Labs, FutureSearch. COLLECT PHYSICAL/VIDEO/AUDIO DATA — Encord, David AI, Human Archive, Cortex, Objectways, Pronto, Sieve. PAY PEOPLE FOR THEIR DATA + LICENSED RIGHTS — Kled, Luel, Reddit, News Corp, Wiley, Shutterstock, Axel Springer. GENERATE SYNTHETIC DATA — Gretel*, MOSTLY AI, Datology, Rockfish, Tonic.ai, Synthesis AI. RUN THE INFRASTRUCTURE — Prime Intellect, Modal, E2B, Daytona, Runloop, Morph, Cua. THE LABELING INCUMBENTS — Appen, Innodata, TELUS International, Snorkel AI, Labelbox, Argilla*, Sama, iMerit. (* = acquired)

WHY LABS PAY: no durable capability lead, treadmill economics — vendor revenue is 'a tax on nobody winning'; every invoice is a confession about what models still can't do.

PRICING & CLOCK: simple website replica ~$20k (a lab bought hundreds, once); high-fidelity enterprise clone ~$300k; tasks $200–$2k; exclusivity 4–5x. Tasks discarded once models pass ~70% — product depreciates by succeeding, guaranteeing repeat orders but nothing annuitizes. Buyers are commoditizing vendors (Anthropic $1B+ environments across a dozen vendors; OpenAI building internal data platform; xAI cutting annotators for specialist tutors).

PRECEDENTS: self-driving data companies absorbed after 2016–2021; Appen, once $4B, lost a Google contract in Jan 2024, stock down 95%+; pharma kept trials outsourced and it compounded. THE LAW: 'Whatever a machine can verify, machines will learn without you' — verification is the scarcity.

LADDER: 2023 crowd ratings (thumbs up/down at scale, models outgrew the raters) → 2024 expert data (credentialed judgment, commoditizing at the edges) → 2025–26 environments (worlds + rubrics + verifiers, discarded at ~70% pass) → NEXT: experience (agents + fleets generating their own training data).

WHAT THE CASH SHOULD BUY: Own one hard domain (200 best experts in one verification-resistant field) — COMPOUNDS. Go physical, full stack (capture ops + in-house industry experts, exclusive datasets) — COMPOUNDS. Sell environments up the stack (labs today, enterprises tomorrow) — COMPOUNDS. Build the trust rails (provenance, neutrality covenants, benchmarks, certifications) — COMPOUNDS. Enter enterprise workflows — IF RUN SEPARATELY. Buy the next rung (~18-month seed-runway acquisitions) — SMART WITH A MAP. Buy GPUs/host training — ONLY IF IT FEEDS PRODUCT. Generalist expansion (giants' lane) — AVOID. Moderate raise, big multiple — AVOID. 'PO income is financing, not the business.'

What it's for

• Evaluate startup opportunities and competitive moats within the AI training data and RL environment sector
• Source specialized data vendors across expert judgment, benchmarks, simulated worlds, and robotics capture
• Structure pricing, margin targets, and delivery contracts when selling datasets or evaluation suites to AI labs
• Assess venture investment risks in AI data vendors facing model graduation obsolescence

Key points

• Six-product taxonomy categorizing AI data assets: hours, judgment, worlds, verdicts, bodies, and rights
• Comprehensive market map categorizing dozens of data generation, eval, synthetic, and environment startups
• Economic breakdown of synthetic environment pricing, task costs, and gross margin profiles
• The Data Treadmill framework explaining why data products depreciate upon achieving high model pass rates
• Strategic playbook detailing which business models compound versus which lead to rapid commoditization

Pricing

Free

Details

Type

Framework

Category

Business & Money

Topic

AI Industry Analysis

Open source

No

Platforms

Web

For

Founders, venture capitalists, and data strategists working in the generative AI ecosystem.

Tags

ai data, rlhf, frontier models, market map, data strategy, synthetic data

Why it was saved

• @viks_rum, 16 July 2026: Full mental model of a fast-moving, little-discussed market: product taxonomy, vendor map, pricing, risk structure and strategy options.