Competition & Moat
Condensed from Deliverable 1/§competitive of the research pass
(ec-workspace/research/world-model-data-flywheel/). "Could not find public
info" entries are themselves signal — they mark either strategic openings or
undisclosed capabilities.
The competitor table
| Who | Captures | Real inputs? | Consent for AI? | Directable fleet? | Scale |
|---|---|---|---|---|---|
| Medal / General Intuition | Highlight clips + semantic actions ("player moved forward") | No — raw keys explicitly never stored (their public privacy stance) | Opt-out by default; disclosure criticized by community | No — organic clipping only | 15M+ MAU (self-reported), 2.5B clips/yr; $454M raised, $2.3B |
| Grunt Games (Genmo) | Full game-window video + K/M input logs | Yes | Explicit opt-in, paid | Yes — per-game bounties | Live beta; scale undisclosed (small) |
| Overwolf | Overlay telemetry, game events, per-app capture | Partial (app events) | No AI-data program found | No | 113M MAU network; anti-cheat whitelists |
| NVIDIA | ShadowPlay local; GFN server frames (latent); NitroGen = scraped video + reconstructed gamepad | Reconstructed | None disclosed for user data | No | 40K hrs / 1,000+ games (NitroGen) |
| Microsoft/Xbox | First-party server-side visuals + controller actions | Yes (server-recorded) | Via game EULA (first-party only) | No | 1B+ pairs, one game (Bleeding Edge) |
| Valve demos / esports replays | Game-state (positions, angles, buttons) — no pixels, no device inputs | Game-state only | Public demos; no licensing program | No | e.g. ESTA: 1,558 pro matches |
| Origin Lab | Publisher-side engine/world data marketplace | n/a (engine capture) | Rights-cleared by design | n/a | $8M seed; 20+ publishers |
| Worldmodeldata | Developer-licensed gameplay/engine data | Unstated | Licensed by design | Unclear | £7M seed; 1M-hr target |
| Twitch/YouTube (scraped) | Video only | No (reconstructable) | None — ToS-violating, IP-murky | No | Unbounded |
What is actually scarce (ranked)
- Consent / license cleanliness. Medal's is opt-out and reputationally dented; scraped video is unsellable to provenance-sensitive buyers; EULA consent works only for your own games. Explicit, paid, opt-in consent with a clean chain of title is the most defensible single attribute — but note: it's a practice, not a barrier. Anyone can adopt it; few have.
- Directability. Only Grunt's bounties resemble our quests. Buyers increasingly want targeted distributions (specific games, scenarios, failure modes, long-horizon sessions) that passive archives structurally can't deliver. Strongest differentiator vs Medal specifically.
- Input alignment — scarce but eroding. NitroGen reconstructs gamepad inputs from overlays; VPT-style IDMs reconstruct from pixels. Reconstruction is weakest for high-DPI KBM (FPS aim) — so KBM-precise streams retain value longest. Medal's no-raw-keys stance is self-imposed and reversible.
- Cross-game breadth — least scarce; Medal spans tens of thousands of games, NitroGen 1,000+ from free video.
- Mic/voice-aligned intent — genuinely rare (only PLAICraft, at academic scale). Underappreciated: this is Quests as a Data Product axis A (intentionality), the measured +28% ECoT lever, and it's near-impossible to reconstruct from video. Arguably our most durable data-side asset.
Fast-follow clock — how quickly each threat materializes
Medal is the highest-probability, highest-impact threat. Technically weeks (they built input translation deliberately — the capture layer exists); realistically 6–18 months via a separate opt-in paid tier, because reversing "raw key data is never stored or transmitted" after a community trust episode is a scandal they must engineer around. Their consent blemish (opt-out default, late disclosure of the General Intuition relationship) is our positioning wedge for exactly that window.
Honest moat assessment
Defensible for ~12 months
Being the only supplier of consented, KBM-precise, mic-inclusive, cross-game, on-demand-directable sessions. Medal won't store raw inputs; Grunt is sub-scale; the marketplaces source from developers, not player fleets. Speed to signed buyer contracts matters more than corpus size — General Intuition's early-access revenue shows budgets exist now, and exclusivity + audit-trail terms lock in early buyers.
Mostly not defensible at 3 years — unless converted
None of the ingredients (capture client, quests, opt-in consent) is patentable or structurally exclusive. What can still be defensible at 3 years:
- Longitudinal player histories — the same player across games, skill curves, thousands of hours of per-player context. Late entrants cannot backfill time. Accrues automatically; costs nothing extra.
- The provenance/QC standard — becoming what enterprise and AI-Act- exposed buyers audit against. Scraped and opt-out data can never match it.
- Exclusive publisher partnerships pairing server-side ground truth with client-side inputs — the one dataset shape nobody (including Microsoft, outside first-party) has cross-publisher. See Products & Publishers.
- Quest-ops know-how — QC, fraud detection, distribution targeting, bounty pricing. Operational, compounding, invisible to competitors.
The fleet is a head start. The moat has to be built out of contracts, history depth, publisher partnerships, and trust — not out of the capture client.
Positioning (adopted 2026-07-08, Sestini's own line): "We are Hugging Face; General Intuition is OpenAI." Every serious competitor hoards (GI's corpus is closed; Origin Lab and Worldmodeldata license privately; scraped corpora can't be shown). An EA researcher's unprompted read: no open platform for this data exists, and the data itself does not exist in the open — what exists is closed (SIMA), arbitrary, or proprietary. The open alternative is an unoccupied position with a built-in growth loop (open datasets → papers → replication pulls the field onto the platform), and it converts the openness the moat analysis treated as a practice into a structural differentiator the closed incumbents cannot copy without destroying their own model. Scope per the open/owned split (Products & Publishers).
Anti-cheat: barrier and risk, symmetrically
Kernel anti-cheats (Vanguard, BattlEye, EAC) treat input-hooking software as keylogger-shaped; per-game whitelisting relationships take years (Overwolf's are a real asset). This is a moderate barrier to new entrants — but no barrier to the two most dangerous fast-followers (Medal and Overwolf already hold those relationships), and a standing supply risk to us in Vanguard-class titles. Mitigations in Risk & Compliance #6.
Open datasets — why buyers can't just use them
Open data is deep in exactly one game (Minecraft: VPT, MineDojo, PLAICraft) and shallow-or-input-free everywhere else (ESTA has no pixels; Matrix-Game's corpus is synthetic/scripted; GamePhysics is tiny). A buyer wanting cross-game, input-aligned, commercially licensable human data cannot assemble it from public sources today. That gap is the market — and it is being raced at from three directions (Medal's semantic-action archive, Grunt's paid fleet, the publisher marketplaces), which is the argument for moving fast rather than perfecting the corpus first (Roadmap).