Market & Buyers
Guido's wiki (4/4-convergence, B226) partially reverses this page's buyer
priority: "Researchers are spec authors and advisors, not buyers. Actual
procurement happens at frontier labs (Meta ~$150M/yr with Surge; Google
~$100M/yr) and publishers' AI initiatives." The one startup conversation on
record (Hugo, 2026-03-20) deprioritized human data (budget-constrained,
synthetic-first). "Corpus-less" is also wrong for parts of the named list
(State Space Labs holds ~100M hrs with engine state). Revised posture:
work the live grounded threads first — Roblox Foundation AI ($350–425K
first-batch anchor, Q3 refresh), FACEIT (revenue-split impasse to revisit),
GDM partnerships (joint publisher licensing openness) — treat the
corpus-less-startup lane as real but unproven, and treat this page's claims
about Decart/Odyssey/Luma/Runway/Worldmodeldata/Grunt as net-new intel
pending a wiki ingress batch. Full detail:
ec-workspace/research/world-model-data-flywheel/alignment-review-vault-addendum-2026-07-08.md.
Condensed from Deliverable 1 of the research pass
(ec-workspace/research/world-model-data-flywheel/01-market-map.md), which
carries full citations and per-claim evidence grades. Figures below were
adversarially verified against primary sources; corrections are noted where a
commonly-quoted number is wrong.
The market in one paragraph
There is now a real, priced market for gameplay data — but almost nobody demonstrably licenses third-party gameplay data yet. The observed behaviors are scraping (NVIDIA, Runway, apparently Sora), first-party capture (Microsoft/Xbox, Wayve, Tesla), free open datasets (Decart's Oasis trained on OpenAI's MIT-licensed VPT corpus), and vertical integration (Medal became a lab rather than sell). The first true marketplaces appeared only in mid-2026, and a paid player-fleet identical in shape to Playroll's is already live (Grunt Games). The market exists, is months old, and is being entered from three directions at once: player fleets, publisher-side marketplaces, and input reconstruction from free video.
Timeline — how the market got priced
| When | Event | Why it matters |
|---|---|---|
| 2022 | OpenAI VPT: contractors at $20/hr, ~4,500 hrs ≈ $90K, ~$160K project total; 1,962 labeled hours trained an IDM that labeled 70,000 scraped hours | The founding precedent and the founding warning: labeled-hour demand saturates (leverage 35:1) |
| Feb 2025 | Microsoft Muse/WHAM in Nature: 1.6B params on ~500K Bleeding Edge sessions (≈61K hour-equivalents — the oft-quoted "500K hours" is wrong), 300×180 @ 10 Hz, consent via EULA | Platform owners can EULA their way to first-party data — but only first-party |
| Late 2024 → Oct 2025 | OpenAI reportedly offers $500M for Medal (figure single-source, The Information; rejected offer company-confirmed); Medal spins out General Intuition ($133.7M seed) | The market's blockbuster datapoint — and it was a failed acquisition, not a license |
| May 2026 | Origin Lab $8M seed (Lightspeed): publishers sell rights-cleared game-world data; 20+ exclusive publisher partnerships | The publisher-side channel exists; unlicensed sellers now compete with rightsholders |
| Jun 2026 | General Intuition $320M Series A at $2.3B (Khosla; Bezos Expeditions, Schmidt); reported ~$40M 2025 revenue via early-access arrangements (thinly sourced) | Vertical integration is what the best corpus chose; "early access" is the revenue template |
| Jul 2026 | Worldmodeldata £7M seed (Iona Star): licenses gameplay/engine data from Unreal/Unity developers, targets 1M hours | Investors now price the exact gap Playroll targets — from the developer side |
| Ongoing | Grunt Games (Genmo) live beta: pays gamers per hour for game-window video + K/M inputs, per-game bounties | Playroll's recipe, replicated by a mid-size lab in months — the moat clock is running |
Buyer categories
| Category | Named orgs | Appetite for Playroll-shaped data |
|---|---|---|
| Frontier labs + publisher AI initiatives | Meta ( | Where procurement actually happens (Guido's wiki, 4/4 convergence B226: "researchers are spec authors and advisors, not buyers"). Episodic-but-huge; enter via bespoke collection (VPT/SIMA pattern) and the named live threads below |
| Named live EC threads | Roblox Foundation AI (parked Q2, refresh Q3; $8–10/hr commodity anchor), FACEIT (30M CS hrs/mo; revenue-split impasse to revisit; wants a $/yr case), GDM partnerships (rights-clean mandatory; open to joint publisher licensing) | Work these first — they exist, have numbers, and have named contacts in the wiki |
| World-model startups | Decart (~$4B), World Labs, Odyssey, Luma ($900M C), Runway, AMI Labs ($1B+) | Real but unproven — the one EC conversation on record (Hugo) deprioritized human data (budget-constrained, synthetic-first); several aren't corpus-less (State Space Labs: ~100M hrs w/ engine state). Steady-but-small budgets; pursue, don't build the pipeline around them |
| Gameplay-native labs | General Intuition | Competitor, not buyer — but proves the price |
| Robotics / embodied AI | NVIDIA GR00T, Physical Intelligence, Figure, Tesla | Low today — no game footage in any major model's documented data mix. 3–5 yr option (Technical Evidence §C) |
| Game industry | Publishers (via Origin-Lab-style channels), Tencent, NetEase | Emerging; publisher-partnered channels favored — see Products & Publishers |
Pricing comparables
No public transaction prices action-aligned gameplay per hour. Every figure is an adjacent proxy — label all extrapolation as such in any pitch.
| Comparable | Price | Reading |
|---|---|---|
| VPT contractor gameplay | $20/hr production cost | The cost floor |
| YouTube-creator footage (unlabeled) | $1–4/min = $60–240/hr of footage | The raw-footage band |
| Adobe purpose-shot video | $2.62/min avg, up to $7.25/min | Purpose-shot premium over archive |
| Teleoperation (fully loaded) | ~$118–200/hr (vendor estimate) | The action-labeled ceiling from robotics |
| Tesla Optimus mocap wage | $25.25–48/hr | What embodied-data labor pays |
| xAI "Video Games Tutor" | $45–100/hr | What game-domain AI labor pays |
| Shutterstock ↔ labs | $25–50M initial deals | Archive-scale licensing, images |
| Reddit ↔ Google | ~$60M/yr | Archive-scale licensing, text |
| OpenAI → Medal | $500M (rejected) | Archive-scale M&A ceiling for gameplay |
Working band for Playroll: between raw-footage licensing ($60–240/hr) and teleop ($118–200/hr), with $20/hr as the cost floor. The way to move within the band is the coverage delta: a priced quest delivers a measured improvement, not hours (Engine Architecture §4).
What buyers actually require
- Fidelity is not the premium axis. WHAM shipped on 300×180 @ 10 Hz; VPT on 360p @ 20 Hz; Genie 1 at 10 fps. Over-investing in 4K capture is wasted spend.
- Alignment precision is. Millisecond-aligned multimodal streams (video, K/M/controller, mic, metadata) are the emerging reference spec (PLAICraft). This matches Quests as a Data Product axis C: the moat is sync precision, not hours.
- License cleanliness is priced. Post-Bartz ($1.5B settlement), buyers' counsel demand provenance. Gameplay carries two copyright layers — the publisher's game content and the player's recording — so player consent alone may not clear a dataset for sophisticated buyers. DeepMind's SIMA and Origin Lab both went publisher-first. See Risk & Compliance.
- EU AI Act Art. 53 forces GPAI buyers to publish training-content summaries (fines from Aug 2026). Whatever we sell must survive public disclosure of its provenance — which is an argument for our consent ledger, and against anything gray.
Where the demand thesis is weakest
- No proven licensing market for player-recorded gameplay. The blockbuster datapoint was a failed acquisition. Origin Lab and Worldmodeldata are seed-stage and source from publishers/developers, not player fleets.
- Label demand may saturate. The VPT/V-JEPA recipe (small labeled seed unlocks huge unlabeled corpus; 62 robot-hours rode on 1M video-hours) means per-buyer volume may be thousands of hours, not millions. What's actually priced is breadth across games/genres/input regimes, directability, and provenance — which is exactly what Medal's implied valuation says is priced, and exactly what a passive archive can't deliver.
- The biggest buyers are structurally self-sufficient (YouTube, Xbox, lifetime-per-day scraping). The realistic buyer list is the corpus-less startups — real, well-funded, but a handful of counterparties.
- Input-alignment scarcity is eroding. NVIDIA's NitroGen reconstructs gamepad inputs from on-screen overlays at usable fidelity across 1,000+ games. KBM-precise streams (high-DPI mouse aim) and mic-aligned intent resist reconstruction longest — bias collection accordingly.
- The rights stack is the most likely lab-counsel objection — see Risk & Compliance for why and what defuses it.
Consequences for strategy
- Work the named live threads first (Roblox Q3 refresh, FACEIT, GDM joint- licensing opening); enter frontier-lab procurement via bespoke collection (VPT/SIMA pattern); treat the startup lane as pipeline, not plan-of-record — and answer Hugo's synthetic-data objection with the founding experiment's delta, since "why not synthetic?" is the recorded kill scenario.
- Sell coverage and directability, not hours. Price against the delta.
- Provenance is a feature buyers pay for, not compliance overhead — build the consent ledger into the product from day one.
- Keep KBM-heavy and voice-intent-heavy titles central to the catalogue: they hold value longest against reconstruction.