Miju LabsEvery competitive brief in this lane repeats one number: Design Arena is doing $60M ARR. It is the number that makes the arena route look like it has already won. It should not be used.
The claim's origin is a single quote from CEO Grace Li: "the site is currently generating $60 million in ARR, solidifying its position as a key source of human-led evaluation data for the AI industry" (Yahoo Finance). TechCrunch characterises the revenue as frontier labs paying for enterprise access to human evaluation data, using the platform as "a source of endless instant feedback for their media-generating models" (TechCrunch, 3 Aug 2026).
The $60M is self-reported, unaudited and uncorroborated. Dealroom's independent write-up of the same round makes no mention of any ARR figure and states that valuation was not disclosed (Dealroom). AI Weekly flags the self-reporting explicitly and notes that nothing is disclosed about "the shape of Intelligence's contracts, retention on those 5.3 million voters" (AI Weekly). [UNVERIFIED]
Why it is almost certainly not what it sounds like
One: there is no consumer monetisation at all. The voting product is free. So the entire $60M would have to be enterprise data-access contracts, sold by a company that was a group of Harvard friends founding "a few weeks before graduation" in 2025 (YC launch).
Two: it would not be recurring even if it were real. The structurally identical business next door prices by consumption — Arena charges on "volume of battles, votes, and prompts consumed", and its CEO says plainly that the revenue is not recurring (Sacra; TechCrunch, 29 Jun 2026). Calling project-shaped data-access revenue "ARR" is the category error the whole sector is committing; GMV is not revenue is the general case.
Three, and decisively: the round contradicts the number. The seed was $7.9M, announced ~3 August 2026, led by Index Ventures, with Conviction (Sarah Guo, Mike Vernal), A*, Valkyrie and YC (TechCrunch; intelligence.ai/careers). A company genuinely at $60M of annual revenue does not raise $7.9M at seed from Index. It raises a $50–150M growth round, at a valuation it announces. Dealroom notes only that the round "sits in the 95th percentile for seed deals in US enterprise software" — which is the correct compliment for a strong seed, and an absurd one for a company with $60M of revenue.
The ratio check points the same way. $60M across 5.3M users is about $11.32 per user, against Arena — 10M monthly visitors, 82M cumulative votes, 28 staff, $250M raised — at $100M annualised (Sacra). Design Arena would be earning 60% of Arena's revenue on half the users, a fraction of the votes and 3% of the capital.
Treat Design Arena as a well-funded seed-stage arena with real consumer traction and no verified revenue. The $7.9M is the number to plan against. If the client's deck contains "$60M ARR", it will be challenged by the first investor who reads the Dealroom note, and the client will lose the room over someone else's press quote.
What is genuinely real, and it is not small
- 5.3M users worldwide as of August 2026 (TechCrunch); the company's own careers page claims 6M+ across 190+ countries (intelligence.ai/careers). Growth was 47,000 users four weeks after launch, then 1.4M, then 2M+, then 5.3M (YC launch; Grokipedia).
- 40+ leaderboards — website, UI component, game, data visualisation, 3D asset, mobile app, logo, SVG, image, slide, video, image editing, agentic game dev and more (designarena.ai). Prompt goes to several anonymous models, user picks A or B, results aggregate to Elo via Bradley–Terry.
- A product no expert panel can build. Intelligence can cut preference by continent and by time, the published example being that "web dashboards in Asia tend to have a more maximalist design style" (Yahoo Finance). Geographic and temporal taste drift needs millions of raters. Thirty designers cannot produce it at any price.
- The company has already left design: it is publishing evaluations of Thinking Machines, Muse Spark 1.2, Fal H3 Max, Grok and Ideogram 4.0, and positions itself as "building the benchmarks and arenas that shape the future of intelligence" (intelligence.ai/careers). Design was its beachhead too.
Corporate identity is a trap worth getting right: the product is Design Arena, the company is Intelligence ("© 2026 Design Arena by Intelligence"), and both the YC launch and Grokipedia name the parent Arcada Labs. Read it as Arcada Labs renamed to Intelligence [WEAK] on the mechanism. The short profile at Design Arena carries the same reading.
The real argument: this is Pick-a-Pic, commercialised
The client's whole thesis is that credentialed judgement is worth a large multiple of anonymous judgement. There is a well-cited paper that says the opposite, and Design Arena is that paper with a cap table.
Pick-a-Pic collected 500K+ (later 1M+) preference examples from "real, intrinsically-motivated users, rather than paid crowd workers" recruited off Twitter, Discord and Reddit (arXiv 2305.01569). Its conclusion is stated flatly:
"The superhuman performance of PickScore showcases the importance of using real users as ground truth rather than expert annotators when collecting human preferences." — Pick-a-Pic (arXiv 2305.01569)
The measurement behind it: the expert baseline scored 68.0%, PickScore, trained on ordinary users, 70.5% (arXiv 2305.01569). Experts lost to the crowd on the crowd's own test.
The strongest version of the answer — and the client needs to be able to give it without flinching — is that the Pick-a-Pic result is not about expertise, it is about whose preference you declared to be the target. The ground truth in that experiment was aggregated user preference. Testing whether experts predict aggregated user preference, and finding they do so 2.5 points worse than a model fitted on it, measures agreement with the crowd. It does not measure whether the crowd's preference is the thing a buyer wants to optimise.
Three consequences follow, and each is a separable product:
- When the target is user preference, buy the arena. Cheaper, larger, drift-aware. There is no honest specialist pitch against 5.3M votes for a model whose success metric is average user delight.
- When the target is a professional standard the crowd cannot see, the crowd is the wrong instrument. Typography, grid discipline, hierarchy, accessibility contrast, brand-system adherence, print production correctness — these have right answers that a two-second A/B click cannot express. HPS v2's own authors concede the point about their raters: "they are a group of average annotators rather than experts in the field" (arXiv 2306.09341).
- A vote is not a reason. Arena and Design Arena capture a click. They do not capture why, and a lab debugging a failure mode needs the why. That is the argument Eight units and one cost anchor builds the product line on, and the honest boundary of it is set out in The oracle problem.
Design Arena's existence does not disprove the specialist thesis; it prices it. Anonymous preference at 5.3M-user scale is now free and funded. Anything the client sells must be something a click cannot express — a rationale, a rubric axis, a named rater, or a recorded process.
The systemic version of this problem, with the vote counts, is next door. The generalists' version — Contra and Mercor selling the same critique at 100× the price — is in The generalists in creative.
Design Arena's cumulative vote count, HQ, headcount and contract shapes are all undisclosed, as is whether the $60M is bookings, recognised revenue, or a quarter annualised. No independent corroboration of any revenue figure exists.