Gonçalo Esteves / building capable systems Field Notes →

The Vault / Companies

Company

Mercor

a marketplace that bills AI labs, by the hour, for the judgement of human domain experts (doctors, lawyers, bankers, engineers) and pays those experts as independent contractors. The experts write the evaluations, rubrics and training tasks that frontier models learn from. Founde

status · living added 27 Sept 2026 updated 16 Nov 2026 from moonshots 162 · moonshots 175 · moonshots 181 · moonshots 190 · moonshots 204 · moonshots 217 · moonshots 228 · moonshots 232 · moonshots 248 · moonshots 267 · moonshots 273 · moonshots 282 · Moonshots #283 · Moonshots #289 · Moonshots #293 organizationoperator-lifeai see it in the graph →

What it is: a marketplace that bills AI labs, by the hour, for the judgement of human domain experts (doctors, lawyers, bankers, engineers) and pays those experts as independent contractors. The experts write the evaluations, rubrics and training tasks that frontier models learn from. Founded in January 2023 by Brendan Foody, Adarsh Hiremath and Surya Midha. Private; it files nothing publicly, so every number below is company-reported or press-reported, and marked as such.

Updated 2026-11-16. Refreshed weekly (see How this page stays current at the bottom).

This week (2026-11-16, window 11–16 Nov). No new round, acquisition or court ruling read. The week after the Series D brought three research posts, a paid call for benchmark tasks, a refreshed newsroom and a revenue number that sits only in the CEO's X post. Run-rate, his claim only. Foody's X post announcing the round (10 Nov, 15:52 UTC per two outlets) says "@mercor scaled from $1M to $3B in run-rate revenue in 27 months, the fastest in history to $3B". Four outlets quote it identically. Neither the BusinessWire release nor the blog post gives a revenue figure. The release was read this week through a Yahoo Finance copy, which closes last week's 403 gap: it names no lead, repeats "more than five million experts" and states no revenue, payout or headcount. The post itself was not read here ⚠️. It is a gross run-rate, unaudited, with no period stated. Two Indian outlets add that Foody cited $250M gross sales in the latest month and more than 100 customers, more than double a year earlier. Where he said that is not stated ⚠️. Newsroom caught up, partly: the valuation stat now reads $20B (it read $10B on 10 Nov), still above the old investor line "Benchmark, Felicis, General Catalyst & Menlo Ventures". The page lists 700+ employees "across San Francisco, New York, and London" (400+ on 4 Nov) and $5M "paid out to the expert network every day", while the home page still says $4M+ ⚠️. The careers page now has the heading "$20 billion valuation" and still says "We are a profitable Series C company". Three new research posts: (1) 11 Nov, "What workers want to hand to AI, and what they want to keep" (Isaac Robinson, Aden Barton, Aurash Vatan): of "roughly 9,000 task ratings, 14% were 'none needed' and 27% 'essential'". Lawyers rated tasks essential most often (38%), software developers least (19%). An applicant extension drew "over 2,000 workers". No data release is stated. (2) 12 Nov, "Introducing OWE-Bench for Long-Horizon Engineering Projects" (Jacob Cole, Cindy Yang, Aravind Selvan, Daniel Babalola, Bertie Vidgen), a coding benchmark of 26 long-horizon engineering and research tasks. It is fully public on Hugging Face, hidden test sets included, under Apache-2.0 per the dataset card (some third-party files carry other licences; read via summary ⚠️). Top score: Claude Fable 5.2 at 54.6%, using about 211M tokens per attempt, then GPT-6 Astra 44.2% and Claude Opus 5.5 41.6%. Exploit attempts appeared in 39 of 2,912 trials. The post gives 54.6% both overall and for the machine-learning domain, a possible copy error ⚠️. The leaderboard did not render, so these figures come from the post only. The site itself warns that now the held-out data is public, a model trained on it is no longer tested on unseen cases. (3) 13 Nov, "How frontier models miss risk in routine enterprise work" (Rich Willats, Bertie Vidgen, Edward Hu): frontier models acted unsafely in 2 of 180 runs when the risk was stated, and in 108 of 180 (60%) when it was only implied. The spread runs from Opus 5.5 at 22% to GPT-6.1 Sol at 83%. The post ends with "apply for safety research funding at Mercor", with no amount or deadline. The AI Safety Fund Grants listing still shows "$0". All scores are company-reported on Mercor's own benchmarks and not reproduced. ★ New and open: OWE-Bench's contribute page pays $5,000 "per accepted task" "for each task accepted into version 2 of OpenWorldEngineer". Submission is a web form. The page states no eligibility, no deadline and no IP terms, and the blog post does not mention the payment. mercor.com home: $91/hr ($90), 752k roles created (731k), $4M+ daily payouts, 5M+ experts. The counter artefacts persist ("$91/hr$/hr", "752kk") ⚠️. PLE: ple-bench.com re-read: unchanged at 9 Nov 2026 – 5 Jan 2027 and $250K / $100K / $50K. Eligibility is still not stated, and the Kaggle rules were not read ⚠️. Grants: both $5M AI Research Fund listings are still on /careers. The deadline is 27 Nov, per job-board copies. A Research Grants Coordinator role ($140K–180K) is now listed. Link Ventures: Mercor · 2023 · LV3 · Active. This again comes from a summary of a truncated render, not verbatim ⚠️. Dockets: White v. Mercor.io has an order entered 12 Nov 2026 (ECF 28) on Mercor's motions to compel arbitration, dismiss, stay and strike. Docket trackers list it by title only, and the outcome was not read and no report of it was found ⚠️. Breach consolidation: still stayed for mediation, no ruling found. Scale v. Mercor: no sign of reopening. Spirit data: trackers give the sale hearing as 17 Nov (Dkt. 1779), and Google has agreed to tighter privacy terms on its winning bid. Mercor remains the $7.5M alternate bidder, and no sale order has been entered ⚠️. Moonshots: EP310 (11 Nov) and EP311 (13 Nov) were swept for all nine spellings plus "Foody", in transcripts and raw caption files: zero hits. Not re-read this week: the 60 Minutes segment, the X post itself, PACER, Foody's CNBC Squawk Box interview (reported, no transcript found).

Previous week (2026-11-10, window 5–10 Nov). ★ The round closed. Mercor announced a $750M Series D at a $20B valuation on 2026-11-10 in a blog post by Brendan Foody ("Organizing human intelligence for the entire economy"). The post says: "The round included an employee tender along with participation from numerous investors, including Benchmark, Felicis, Fundomo, funds managed by affiliates of Apollo, G Squared, General Catalyst, Glade Brook Capital, NVIDIA, Salesforce, Sapphire Ventures, StepStone Group, XTX Ventures, and Zoom." No lead investor is named. The size of the employee tender is not stated, so how much of the $750M is new money for the company is unknown. Reuters (same day) and The Next Web report the same figures. Link Ventures is not among the named participants. The list says "including", so that is not evidence either way about its participation. Mercor says the money goes to "infrastructure and research for evals and post-training and in building more environments where agents can learn complex workflows", sold to enterprises as well as the labs. The post gives no revenue, payout or headcount figure. It repeats "more than five million experts in our network" and says "We already spend 3x as much on inference as on employee salaries" (company statement, no amounts). It also projects "something closer to double-digit GDP growth over the next five years" if that ratio spreads through the economy, which is a forecast, not a measurement. Mercor's own pages have not caught up: the newsroom stat still reads $10B and the careers page still says "We are a profitable Series C company", while the home page features the Series D. New research post (2026-11-05): "Playing well with others: Putting AI in the org chart" (Victor Barres, Aden Barton, with Nicolai J. Foss and Henrik M. Nielsen of Copenhagen Business School) introduces a "preliminary" taxonomy of organizational difficulty for AI benchmarks. It reports that fewer than a quarter of roughly 500 benchmarks (522 in a figure caption) claim any organizational element, and that only 16% report performance on one separately. The paper is arXiv 2611.03494, "Harder Than the Task: From Technical to Organizational Difficulty in AI Benchmarking", submitted 2 Nov 2026 under CC BY 4.0. It continues the line of the Benchmark Map and TaskNet: research that maps the gaps Mercor sells into. No open call is attached. mercor.com home: $90/hr (unchanged), 731k roles created (716k), $4M+ daily payouts. The body still says 5M+ experts, and the counter artefacts persist ("$90/hr$/hr", "731kk", "$4M+$M+") ⚠️. PLE competition: ple-bench.com re-read 10 Nov. It now gives the window as 9 Nov 2026 – 5 Jan 2027 and prizes of $250K / $100K / $50K, with submissions on Kaggle and a starter harness on GitHub. Neither page states eligibility, and the Kaggle rules were not read ⚠️. Grants: both $5M AI Research Fund listings are still on /careers, with the deadline of 27 Nov unchanged per job-board copies. AI Safety Fund Grants still shows "$0". Link Ventures: about page verbatim: Diamandis "Partner, Link-XPV", Blundin "Founder and General Partner". The Mercor card (2023 · LV3 · Active) and Wissner-Gross "Venture Advisor" came this week from a summary of a truncated render, not verbatim ⚠️. Dockets: no ruling found in White v. Mercor.io, the breach consolidation or Scale v. Mercor. Web search only; PACER and CourtListener not read ⚠️. Spirit data: no sale-approval outcome found; Mercor is still reported as the $7.5M backup bidder ⚠️. Moonshots: EP308 (published 4 Nov) and EP309 (published 7 Nov) were swept for all nine spellings plus "Foody", in transcripts and raw caption files: zero hits. The panel did not mention the round, which came after both episodes. Not re-read this week: the 60 Minutes segment, the BusinessWire release (403), PACER.

Earlier (2026-11-04, window 29 Oct–4 Nov). No announced round, acquisition or court ruling found. The newsroom still says $10B and 400+ employees and its newest dated item is still 8 Oct; the careers page still says "We are a profitable Series C company". Two new blog posts. (1) "Offline Evaluations in Practice: Improving Mercor's SRE Agent" (2026-10-29, seven named authors): two internal alert-response agents, Medic and Momo, on the same unnamed model; Medic scores 0.79 vs 0.55 on a five-incident evaluation suite the post itself calls "a direction, not a final measurement"; cost per alert $7.18 → $2.26; weekly cost $8,295 → $345 after retiring Momo. It names no partner — no 1Password, no Codex — and none of the Forbes 13 Oct figures (300+ incidents, 95% accuracy, 57% faster remediation) appear in it, so the Forbes item stays unconfirmed; the post's own numbers also don't reconcile ($7.18 old cost per alert vs its stated operating costs of $1.86/$2.06 per alert) ⚠️. (2) "Introducing Physical-Engineering's Last Exam (PLE) with Google Cloud AI Research" (2026-11-03, Bennett, Haegler Pilnik, Mittal): a benchmark of 85 long-horizon engineering tasks (mechanical 30, electrical 27, civil 18, structural 10; 44 expert-authored, 41 built by Google) running on Archipelago, Mercor's open-source environment framework; top score Claude Opus 5.5 at 48.3% on the 56-task hidden set (~27M tokens and 63 minutes per task; "14 frontier models across seven agent harnesses"). Google released the 29 public tasks plus grading tools, and an open competition runs at ple-bench.com (read 4 Nov: "Submissions: Open", 3 scored submissions per participant, GitHub sign-in required, no deadline, eligibility rules or prize amounts published — "Competition details to be announced"; contact ple-bench@google.com; the blog's "Thirteen Google-sponsored models can be run for free" does not appear on the site ⚠️). All scores are company-reported on Mercor's own benchmark, not reproduced. Second Google tie in this record: Google outbid Mercor at the Spirit auction in August; it now co-builds and hosts Mercor's benchmark. mercor.com home: the stat strip now reads $90/hr average contracted rate ($94 on 10-29), 716k roles created (670k) and $4M+ daily payouts; the Experts label still shows no number and the body still says 5M+; the counter artefacts persist ("$90/hr$/hr", "716kk", now also "$4M+$M+") ⚠️. New listings include "Chemicals Safety for Redteaming" at $65–75/hr and a Retail Trader whose rate renders ambiguously as "$400311 recently hired" ⚠️. Round: still talks, still unannounced; one more secondary (a Yahoo Finance copy) dates The Information's Nvidia piece to 2026-08-19, supporting August over the 22 Oct author-page date; read the report as August ⚠️. Link Ventures: unchanged — Mercor · 2023 · LV3 · Active (this week read from a truncated render's summary, not verbatim ⚠️); Diamandis "Partner, Link-XPV" and Blundin "Founder and General Partner" verbatim; Wissner-Gross "Venture Advisor" (summary only ⚠️). Dockets: no ruling found on Mercor's motion to compel arbitration in White v. Mercor.io (a Law360 headline covers the motion only); nothing new on the data-breach consolidation; no sign of Scale v. Mercor reopening. Web only; PACER and CourtListener not read ⚠️. Spirit data (context): reported hearing dates conflict (19 Aug → 9 Sep → 14 Oct; last week's 6 Nov came from a tracker) and no outcome was found; the privacy ombudsman recommended approving the Google sale, micro1 filed a competing bid, and Mercor remains the $7.5M backup bidder ⚠️. Grants: both $5M AI Research Fund listings still on /careers (deadline 27 Nov per the job-board copies; the careers page itself shows no deadline text); AI Safety Fund Grants still "$0". A widely re-surfaced contractor pay-cut story (5,000+ cut on a Meta project, replacement work at $16/hr vs $21) dates to November 2025, not this window. Moonshots: EP307 (31 Oct) swept for all nine spellings plus "Foody": zero hits. Not re-read this week: the 60 Minutes segment, PACER, the TaskNet paper.

Earlier (2026-10-29, window 23–29 Oct). No announced round, acquisition or court ruling found. The newsroom still says $10B and its newest dated item is still 8 Oct; the careers page now also says "We are a profitable Series C company" (company statement, no period or figure given). Round: The Information ran "Nvidia Discusses Funding Its AI Data Supplier Mercor at a $20 Billion Valuation" (Phoebe Liu, Julia Hornstein); its author page dates it 2026-10-22, but trackers place the same report in August, so the date is unresolved ⚠️. The story is paywalled and was not read. Secondary copies say General Catalyst is in talks to lead, that Nvidia paid Mercor "tens of millions of dollars" for data last quarter, and that Nvidia used Mercor data for its two latest Nemotron models ⚠️. Still talks, not a close. New company research: blog post "Introducing TaskNet: A New Taxonomy of Work" (2026-10-27, Isaac Robinson) describes a working paper by researchers from Stanford, Mercor, the O*NET Resource Center and NBER covering six US occupations. It reports 79 tasks validated by O*NET analysts that have no O*NET counterpart (5 of them created or reshaped by AI adoption), a rewrite of 21% of task statements after analysts reviewed 229 candidates, and coverage of 89.3% of tasks in a holdout of 1,184 reports from 57 workers, against 78.6% for O*NET. It values the new tasks at ~$140B a year in time-adjusted wages and says Mercor has run "over 3 million interviews" since 2023. All of this is company-reported and was not reproduced. Like the 15 Oct Benchmark Map, it is research that maps the gaps Mercor sells into. Data supply: on 23 Oct Foody posted on X that Mercor will pay "up to $100K for every company you refer that licenses its data", claiming "> 500 companies" partnered so far, and asked for introductions to firms with 100+ employees in code, gaming, manufacturing, healthcare or professional services. Two outlets quote the post identically; the post itself was not read ⚠️. One of them says Mercor's referral page (read 25 Oct) claims $50M+ paid to companies for data and $2M+ paid to referrers in 2026 ⚠️ secondary. Office: Mercor has leased 25,550 sq ft on the 77th floor of One World Trade Center, Lower Manhattan (REBusinessOnline, 2026-10-23; trade press, term and rent not disclosed, not confirmed by Mercor ⚠️). mercor.com home: the stat strip now reads $94/hr average contracted rate (it was $100 on 10-23 and $111 on 10-17), 670k roles created (600k) and $4M+ daily payouts. The Experts label shows no number; the page body still says 5M+. The same counter artefacts appear ("$94/hr$/hr", "670kk") ⚠️. The newsroom now lists 400+ employees. Correction, disputes: Scale AI v. Mercor (N.D. Cal. 3:25-cv-07402) is listed below as open, but three docket mirrors quote entry #30, a "STIPULATION of Voluntary Dismissal with Prejudice filed by Scale AI" on 2026-01-02, with the case terminated 5 Jan. CourtListener and Justia returned 403, so this was not read at the docket ⚠️. Secondary trackers also give the breach suits as consolidated in In re Mercor.io Corporation Data Breach Litigation (N.D. Cal. 3:26-cv-02821, 19 May), stayed for mediation since 21 Jul, and show no ruling on Mercor's 9 Jun motion to compel arbitration in White v. Mercor.io ⚠️. Spirit Airlines data (context, not new this week): press citing the bankruptcy record says Mercor was the $7.5M backup bidder to Google's $10M winning bid at the 14 Aug auction for Spirit's internal archive. The court-appointed privacy ombudsman said she would need more on Mercor's 2026 security incident before assessing it as a buyer, and the sale hearing was moved to 6 Nov (Travel Weekly via a tracker ⚠️). AI Research Fund: both $5M listings are still on /careers, and the deadline is unchanged at 27 Nov. AI Safety Fund Grants still shows "$0" and its posting still did not render. Link Ventures: unchanged (Mercor · 2023 · LV3 · Active; Diamandis "Partner, Link-XPV"; Wissner-Gross "Venture Advisor"; Blundin "Founder and General Partner"). Moonshots: EP305 (23 Oct) and EP306 (28 Oct) were swept for all nine spellings plus "Foody": zero hits. Not re-read this week: the 60 Minutes segment, PACER.

Earlier (2026-10-23, window 17–23 Oct). No new round, acquisition, product line or court ruling found; the $20B round is still unannounced and the newsroom still says $10B. New and open: the Mercor AI Research Fund. Two grant tracks are listed at mercor.com/careers, each labelled $5M: Benchmarks (first posted 2026-09-23, so it was already open during last week's refresh, which said nothing was open; that line was wrong) and Economics of AI (posted 2026-10-20). The terms below come from two job-board copies of Mercor's own posting (General Catalyst's and Emcap's boards), which agree; Mercor's posting page did not render its text. Budget tiers are under $50K, $200K and $500K. Applicants must be based in the US or UK: university and nonprofit researchers, public benefit corporations, and startups with fewer than 20 people that do not sell AI training data. The award goes to the institution. The process starts with a 1–3 page expression of interest. Applications are due 27 Nov 2026, 11:59 pm PT, with notification by 18 Dec. The grantee keeps ownership, but Mercor gets a non-exclusive licence, commercial use included. Outputs must be published under an open licence that permits commercial use. For benchmark projects, Mercor gets early access and the exclusive right to build and host a private held-out test set. Read with the 15 Oct Benchmark Map, this means Mercor is paying outside researchers to build the benchmarks its own map says are missing, and it keeps the hidden test set. Whether the two "$5M" labels are one pool or two is not stated. A third listing, AI Safety Fund Grants, shows "$0" and its terms were not read. mercor.com home: the stat strip now reads an average contracted rate of $100/hr (it was $111 on 10-17), 600k roles created (538k), $4M+ daily payouts and 5M+ experts; the 30k+ figure no longer appears there. The rendered text carries counter artefacts ("$100/hr$/hr", "600kk"), so the $100 may be an animated counter's starting value rather than a new average ⚠️. New listings shown include an oncologist at $350–450/hr. No blog post after 15 Oct and no newsroom item after 8 Oct. Missed last week, not re-read: Forbes (2026-10-13, Tim Keary) reports that Mercor and 1Password built an SRE agent with Codex, tested on 300+ past incidents, with 95% accuracy and a 57% lower median time to remediate, per 1Password's CTO. The article returned 403, so this rests on a search snippet ⚠️. Link Ventures: portfolio unchanged (Mercor · 2023 · LV3 · Active). The about page now lists Diamandis as "Partner, Link-XPV"; on 29 Sep it read "Co-Founder and Partner", and the date of the change is unknown. Wissner-Gross is still "Venture Advisor". Moonshots: EP303 (17 Oct) and EP304 (21 Oct) were swept for all nine spellings with zero hits (Raor only as a whole word, since it otherwise matches inside "extraordinary"). The 23 Oct watch found nothing newer. Dockets: no new filing or ruling found by web search; CourtListener returned 403 and PACER was not read ⚠️. Not re-read: the 60 Minutes segment.

Earlier (2026-10-17). No new round, acquisition, product line or court ruling found. mercor.com home and newsroom re-read: valuation still $10B (newsroom's press list still ends at the Oct 2025 Series C), payouts $4M+ a day, home 30k+ experts and 538k "roles created", average contracted rate now $111/hr (was $109 on 10-11, $108 on 10-05; the figure moves week to week and nothing explains it). Newsroom still says 5M+ experts and "more than $95 an hour" (FT), so the site's counts remain inconsistent and unaudited. Two new research items, both company-reported and not reproduced: (1) a Benchmark Map (mercor.com/research/labor-index/benchmark-map, undated on the page; dated 2026-10-15 only by Mercor's X post and secondary coverage) that ties 29 benchmarks to the Labor Department's O*NET taxonomy and finds coverage for 208 of 923 occupations and 924 of 18,797 tasks, with the headline "90% of American working hours are spent on tasks with little or no coverage from existing benchmarks"; the page itself calls the results "provisional", says the benchmark list "is not a complete or perfectly representative sample" and that an incomplete list "will necessarily bias down estimates", so the 90% is a floor on uncovered work by Mercor's own selection rules, and Mercor sells the benchmarks it finds missing. (2) Blog "Small, targeted data, measurable gains" (2026-10-15), no named author: Mercor post-trained NVIDIA's Nemotron 3.5 Super on "a few hundred tasks" per domain from its APEX environments, with Harvey, Cognition and Ramp credited; reported gains are APEX-SWE Pass@1 26.6% → 34.7%, APEX-Accounting mean rubric 34.6% → 41.9%, legal slice Pass@1 19.8% → 25.0%. No weights, data or scripts are stated as released, and no NVIDIA collaboration is claimed (the post says the work was done "ahead of its release"). A team post (2026-10-12, Derek Shimozawa, ex-Airbnb payments, joining Payments Platform Engineering) says weekly disbursements run at "30 times the scale" of the earlier system; no base given. Link Ventures portfolio re-read: Mercor · 2023 · LV3 · Active, unchanged; Wissner-Gross still appears as a "Link technical advisor". Moonshots: transcripts of 3, 5, 7 and 15 Oct 2026 swept for all nine spellings: zero hits (swept, not read; the 5 Oct caption is machine-translated). Dockets: no new filing or ruling found by web search (White v. Mercor.io docket last mirrored 30 Jun; Scale v. Mercor nothing new); a Currey v. Mercor Global, Inc. (M.D. Fla., default entered 20 Jul) names a different entity and is not verified as this company, so not added. The $20B round is still unannounced; nothing dated October found. Programmes: the Research Fellowship closed (aggregators give a 26 Sep 2026 deadline), the 2025 Graduate & Postdoc Fellowship has no 2026 cycle found, so nothing open to enter. Not re-read this week: the 60 Minutes segment, PACER.

Earlier (2026-10-11). No new round, acquisition or lawsuit ruling found. mercor.com home and newsroom re-read: valuation still $10B (the newsroom's press list still ends at the Oct 2025 Series C), payouts $4M+ a day, average contracted rate $109/hr (back from $108 on 10-05), home says 30k+ experts while the newsroom page still says 5M+ and "more than $95 an hour" (FT), so the site's counts stay inconsistent and unaudited. One new company item: blog post "Why Mercor is building a research team" (2026-10-08), credited to Mercor, no named author. It names five researchers (Edward J. Hu, Minyang Tian, Victor Barres, Hardik Bhatnagar, Bertie Vidgen), says the team will expand the APEX benchmarks (consulting, investment banking, corporate law, accounting, software engineering), work on training-data quality and post-training, and mentions collaborators from Anthropic, Ai2, Bridgewater and D. E. Shaw. It claims post-training lifted Qwen3.5-397B's Pass@1 on APEX-Agents from 16% to 27%, with a guide, training script, weights and traces (company-reported, not reproduced; Mercor writes the benchmark it reports on). No fellowship or funding programme in the post; a "Mercor Research Fellowship" job listing exists on a third-party board but is marked closed (30 Sep 2026), so nothing to apply to. Link Ventures portfolio re-read: Mercor · 2023 · LV3 · Active, unchanged. The $20B round is still unannounced; web search found no October 2026 coverage of it. Not re-read this week: the 60 Minutes segment, dockets, new Moonshots transcripts (EP299+).

On Moonshots with Peter Diamandis it is the panel's standing example of "how an individual gets paid in the AI economy". A caption sweep finds it in 20 of the 91 transcripts on disk (sweep of 2026-09-29), under nine spellings (Mercor, Meror, Merkor, Merkore, Merur, Merore, Mercorps, mercur.com, Raor). The four episodes since EP298 (3, 5, 7, 15 Oct 2026) were swept on 2026-10-17 with zero hits; EP303 (17 Oct) and EP304 (21 Oct) were swept on 2026-10-23 with zero hits; EP305 (23 Oct) and EP306 (28 Oct) on 2026-10-29, zero hits; EP307 (31 Oct) on 2026-11-04, zero hits (transcript and raw caption file both); EP308 (4 Nov) and EP309 (7 Nov) on 2026-11-10, zero hits (transcripts and raw caption files); EP310 (11 Nov) and EP311 (13 Nov) on 2026-11-16, zero hits (transcripts and raw caption files).


The business model

Who pays. Frontier AI labs, overwhelmingly. Mercor names OpenAI, Anthropic and Meta as customers (TechCrunch, 2025-10-29) and says it works with "the top five AI labs". The Information, citing internal documents, reported that roughly 91% of first-half 2026 revenue came from OpenAI and Anthropic (via AI Weekly, 2026-07-21). ⚠️ Reported, not disclosed by the company.

What they buy. Human expertise that cannot be scraped. In Foody's words: the labs' customers "don't want to give them data to automate large portions of their value chains, so they need to hire contractors who previously worked at those companies" (TechCrunch, 2025-10-29). In practice:

  • Human data and evaluations: experts grade model output and write the rubrics that score it.
  • Benchmarks: the APEX family (AI Productivity Index; paper arXiv 2509.25721, Sep 2025), which measures models on professional tasks. It is research and a sales document at once. Since 2026-10-15 a Benchmark Map (29 benchmarks against O*NET's 923 occupations) argues that 90% of US working hours are not covered by existing benchmarks; Mercor's own page calls it provisional. Since Sep–Oct 2026 an AI Research Fund (Benchmarks; Economics of AI) pays outside US/UK researchers up to $500K per project, and for benchmark projects Mercor keeps early access and the exclusive right to a private held-out test set (job-board copies of Mercor's posting, read 2026-10-23). Since 2026-11-03, PLE (Physical-Engineering's Last Exam, built with Google's cloud AI research team): 85 physical-engineering tasks, 29 public and 56 hidden, with an open harness-and-model competition on Kaggle running 9 Nov 2026 – 5 Jan 2027 for $250K / $100K / $50K prizes (ple-bench.com, read 2026-11-10; eligibility not stated there). Since 2026-11-05, a taxonomy of organizational difficulty for benchmarks (arXiv 2611.03494, with Copenhagen Business School). Since 2026-11-12, OWE-Bench (OpenWorldEngineer): 26 long-horizon engineering tasks released in full, hidden tests included, under Apache-2.0. Its contribute page pays $5,000 per task accepted into version 2 (read 2026-11-16; no eligibility, deadline or IP terms stated).
  • Post-training as a service (2026-10-08 and 2026-10-15 posts): Mercor reports fine-tuning open models (Qwen3.5-397B, NVIDIA Nemotron 3.5 Super) on a few hundred APEX tasks per domain and scores the result on APEX. All gains are company-reported, scored on Mercor's own benchmarks, not reproduced.
  • RL environments: simulated workplaces where agents practise and get scored. Mercor bought Deeptune, an environment builder, on 2026-07-09, having been its customer and, per Foody, an early investor (Mercor blog).
  • Newer lines on mercor.com (2026-09-29): Data monetization (companies sell anonymised workflow data), Enterprise AI (agents deployed with forward-deployed engineers) and a robotics data line.

How it charges. An "hourly finder's fee and matching rate" on top of what the expert is paid (TechCrunch, 2025-09-09). Revenue is booked gross: Foody confirmed the reported figure "includes the total amount that customers pay Mercor for services before its contractors receive their portion" (same source). That single accounting choice is why the headline numbers are so large.

Take rate: not published by Mercor. Press reports put the contractors' share at 60–70% of the gross figure (The Information / The Next Web, 2026). One analyst's arithmetic puts net revenue at roughly $600M on the $2B gross run-rate (Value Add VC, Trace Cohen, 2026-07-19). ⚠️ Third-party estimate, unaudited.

A cross-check from Mercor's own numbers (arithmetic, not a disclosed figure): Mercor's site says it pays out more than $4M a day. Annualised, that is about $1.46B a year to experts. Against a $2B+ gross run-rate, experts would receive about three-quarters and Mercor keep about a quarter, which sits in the same range as the press reports. The two numbers come from different dates, so read this as a bracket, not a measurement.

How the model changed.

when what Mercor was selling source
2023 an AI interviewer and hiring marketplace: vetting overseas software engineers and placing them with US companies seed coverage; Link Ventures' portfolio card still describes it this way
Sep 2024 "an AI hiring platform that enables top tech companies and leading AI labs to hire experts" Series A release, PR Newswire
Feb 2025 most revenue now from AI labs; $75M annual run-rate TechCrunch, 2025-02-20
Oct 2025 "We connect human expertise with leading AI labs and enterprises"; paying experts $1.5M+ a day Mercor blog, 2025-10-27; TechCrunch, 2025-10-29
Jul 2026 RL environments (Deeptune acquisition) Mercor blog, 2026-07-09
Sep 2026 "a leading AI data company": human data, APEX, data monetization, enterprise agents mercor.com
Nov 2026 the lab toolkit (evals, post-training, environments) sold to "big tech, developers of open models, application layer companies, and the Fortune 500" Series D post, Mercor blog, 2026-11-10

The hiring marketplace did not disappear. It became the supply side of a data business.

The valuation, dated

date round amount post-money valuation lead source · type
2023 (announced Jan 2024) Seed $3.6M not disclosed General Catalyst (with Soma Capital, Link Ventures and others) FinSMEs, Jan 2024 · secondary
2024-09-18/19 Series A $30M $250M Benchmark Mercor release, PR Newswire · primary
2025-02-20 Series B $100M $2B Felicis (with Benchmark, General Catalyst, DST Global, Menlo) TechCrunch · secondary
2025-10-27 Series C $350M $10B Felicis (with Benchmark, General Catalyst, Robinhood Ventures) Mercor blog · primary
2026-07-09 reported talks, ~$500M — ~$20B not named at first; later reports said General Catalyst in talks to lead, Nvidia in talks to join Bloomberg via TechCrunch, Forbes, PYMNTS; The Information (paywalled, not read) · secondary
2026-11-10 Series D $750M (includes an employee tender of undisclosed size) $20B none named; participants include Benchmark, Felicis, Fundomo, Apollo affiliates, G Squared, General Catalyst, Glade Brook, NVIDIA, Salesforce, Sapphire, StepStone, XTX Ventures, Zoom Mercor blog (Foody), 2026-11-10 · primary; Reuters, TNW same day · secondary

The $20B round is now announced (2026-11-10). It came in at $750M rather than the ~$500M first reported in July, and the announcement names no lead, so the reports that General Catalyst would lead are not borne out. Mercor's newsroom stat still read $10B on 2026-11-10 and reads $20B on 2026-11-16.

Revenue, as reported (all gross, all company-reported or from internal documents): $75M run-rate (Feb 2025) · ~$450M (Sep 2025, sources) · ~$500M (Oct 2025) · crossed $1B early 2026 (Mercor blog, 2026-05-08) · more than $2B in June 2026 (The Information; announced by Foody) · $614M gross revenue booked in H1 2026 (The Information) · $3B run-rate (Foody on X, 2026-11-10: "$1M to $3B in run-rate revenue in 27 months"; quoted identically by four outlets, the post not read here, and not in the release or the blog post ⚠️). Profit: $6M in H1 2025 (TechCrunch sources); nothing more recent is public.

"The fastest appreciating company in the history of the world"

Said on Moonshots #293 0:45:59 (2026-09-22). Read: Off. The record shows one of the fastest climbs on record, not a provable first:

  • Mercor: $250M (Sep 2024) to $10B (Oct 2025), 40× in about 13 months.
  • Anysphere (Cursor), over overlapping dates: $400M (Aug 2024) to $29.3B (Nov 2025), about 73× in about 15 months (TechCrunch 2025-06-05; CNBC 2025-11-13).
  • On moonshots 162 (2025-04-04) the same speaker framed it more carefully: reaching a multi-billion valuation within two years of founding "has happened about five, six times in the history of the world", sourced by him to a chatbot and flagged as "maybe not totally reliable".
  • The frequently quoted "~80× in two years" now has its round: $250M (Sep 2024) to $20B (announced 2026-11-10) is 80× in about 26 months. Say "80× in a little over two years". The superlative still doesn't follow, because Cursor's climb was steeper over a shorter window.

Who holds a position

Found, and disclosed on tape, though not on the episode where Mercor was sold hardest.

person what the public record shows on tape
Dave Blundin / Link Ventures Yes. Link Ventures' portfolio page lists Mercor · year invested 2023 · fund LV3 · status Active (read 2026-09-29). Seed coverage names Link Ventures among the participants. Stake size not public. Not among the participants named in the 2026-11-10 Series D announcement, which lists investors as "including", so neither participation nor its absence is established. A Foody testimonial sits on Link's homepage. "one of the investments we made" moonshots 175 0:28:24 · "from the day we invested at the seed stage" moonshots 181 0:03:09 · "through our portfolio company" moonshots 248 0:50:58 · "every investment we've made in a company like Mercor" moonshots 282 2:11:26
Peter Diamandis Affiliated, exposure not established. Listed by Link Ventures as "Partner, Link-XPV" (read 2026-10-23; it read "Co-Founder and Partner, Link-XPV" on 2026-09-29). Link's own page puts Mercor in LV3, a different fund. Whether he participates in LV3 is not public. "full disclosure… I'm a partner with Dave in… Link Exponential Ventures", followed by Mercor as "one of two $2 billion success stories in the portfolio" moonshots 162 0:15:06 · "this is one of your key investments at Link XPV" moonshots 181 0:47:46 (this conflicts with Link's LV3 label)
Alex Wissner-Gross Affiliated with the investor, no Mercor stake found. Listed by Link Ventures as a Venture Advisor. none found
Salim Ismail None found. none found

What was searched: Link Ventures' portfolio, about and home pages; Mercor's release and newsroom investor lists; round coverage from seed to Series C; web searches for each panelist's name with "Mercor"; and a caption sweep of 91 Moonshots transcripts for every spelling above.

Other ties on tape: Mercor calling Link looking for teams to acquire ("I got calls from both Merur and from Orin… looking for acquisition targets", Moonshots #283 0:49:56) and offering to buy human data (moonshots 273 0:46:07, Moonshots #289 0:21:24). Link Ventures is also a Moonshots sponsor (pre-roll).

On Moonshots #293 itself the position is not restated. That episode contains the superlative, the "walk exactly in the footsteps of Mercorps" prescription and the advice to "do something on Mercor", and zero uses of "portfolio company" or "full disclosure". The interest was disclosed earlier, repeatedly, and those earlier disclosures are to the speakers' credit. It was not restated on the day the company was pitched to callers. → Moonshots: the disclosures ledger

The reversal, in their own words

The panel uses Mercor to argue both sides of the operator's main question: where does the value go?

episode how Mercor is used verbatim
moonshots 204 2025-10-30 a new class of worker on old payment rails "a company 2 years old getting close to a billion dollars of revenue… unleashed a whole new class of employee" 0:29:00
🔴 moonshots 232 2026-02-19 the case against the platform "Meror has tens of thousands of people in India doing work that benefits Merore that is actually for open AI. The fraction of all value created that flows back to the mother ship is just a massive fraction of that value chain" 0:55:34
moonshots 267 2026-06-29 Chinese labs substituting it out with models "why does Mercor have a billion-dollar revenue? Cuz they get all the experts in… So they're substituting out Mercor and these other guys" 1:23:14
Moonshots #289 2026-09-13 a slow buyer of archives (a caller) "Merore and Scale have both expressed interest but it's been weeks and nothing is closing" 0:20:52
★ Moonshots #293 2026-09-22 the earnings prescription and the company to imitate "do something on Mercor plus do something on Fiverr" 0:29:43 · "I would walk exactly in the footsteps of Mercorps" 0:46:30

Between those two Moonshots #293 lines the same speaker says "almost all value is going to accrete to capital gains through ownership… and not to payroll" (0:30:48). A marketplace pays its experts for output (payroll, in effect); the ownership sits with the marketplace and its investors. Both statements can be true, and they point in different directions. The gross-to-net gap above is the same disagreement expressed as an accounting line.

On tape vs reality

Three reads. Close means near enough to repeat with the correction. Off means don't repeat it. Aspiration: investigate means a forward claim worth tracking, not a fact yet.

on tape reality read
"fastest appreciating company in the history of the world" Moonshots #293 0:45:59 40× in ~13 months; Cursor ~73× in ~15 months over overlapping dates Off. Say "one of the fastest valuation climbs on record"
"unleashed 100,000 people" Moonshots #293 0:46:30; "50 to 100,000 individual actors" (2026-09-09) Mercor: more than 30,000 weekly active contractors (blog, 2026-05-08); "30k+ experts" on site; "5M+" profiles Off. Use Mercor's 30,000 active and say it is company-reported
"Mercor is now at a $10 billion valuation" moonshots 190 2025-08-27 0:29:37 talks reported 2025-09-09; closed 2025-10-27 Off at the time. Early by two months
"close to a billion dollars of revenue" moonshots 204 2025-10-30 ~$500M gross run-rate that week (TechCrunch, 2025-10-29) Off. About half, and gross
"a billion dollar revenue run rate now" moonshots 228 2026-02-09 Mercor: crossed $1B run-rate early 2026 Close. Say "gross run-rate"
"Mor is worth 40 billion or 20 now, 40 by the end of the year" moonshots 282 2026-08-21 2:11:38 at the time: last closed $10B, $20B only in talks; $20B announced 2026-11-10 Off at the time on $20B (early by about 12 weeks) · Aspiration: investigate on $40B by year-end, an investor's forecast
"$2 billion… 100 million of revenue doing 20 million on the bottom line" moonshots 175 2025-05-30 0:28:34 $2B matches Feb 2025; $75M run-rate reported Feb 2025; only public profit figure is $6M in H1 2025 Close on $2B · unverified on the profit
"$20 million a week of value" since seed moonshots 181 0:03:19 seed valuation undisclosed; $0→$2B in ~17 months is ~$27M/week Close as arithmetic, not a measurement
"idea to… $10 billion valuation two years" moonshots 262 1:56:29 founded Jan 2023, $10B Oct 2025: ~2 years 9 months Close. Say "under three years"
sell old code or records "for a million, 2 million dollars" to Mercor moonshots 273 0:46:38 Mercor now lists a Data monetization product; no public price Aspiration: investigate
Mercor "almost entirely operating in India" moonshots 190 Mercor pays contractors "across dozens of countries"; no public split Unverified. Don't repeat

Disputes on the record

  • Scale AI v. Mercor (filed 2025-09-03, N.D. Cal. 3:25-cv-07402): Scale alleged a former employee took customer documents to Mercor. Mercor co-founder Surya Midha said Mercor has "no interest in any of Scale's trade secrets". Closed: docket mirrors quote a stipulation of voluntary dismissal with prejudice filed by Scale on 2026-01-02, case terminated 5 Jan. No finding either way. (Earlier refreshes of this page called it open; that was wrong. ⚠️ Mirrors only; the docket itself returned 403 on 2026-10-29.)
  • Security incident, late March 2026: a supply-chain attack through the LiteLLM open-source package. Meta paused work with Mercor (reported, early April 2026). Mercor's update of 2026-06-25 says the investigation is complete, a "very limited subset" of experts had sensitive data affected, and "all frontier labs have increased their work with us over the last few months". Class actions followed, consolidated as In re Mercor.io Corporation Data Breach Litigation (N.D. Cal. 3:26-cv-02821) and stayed for mediation since 2026-07-21 (secondary trackers ⚠️). Allegations, not findings.
  • White v. Mercor.io (N.D. Tex., filed 2026-05-08): alleges experts were misclassified as independent contractors. Allegations, not findings. Mercor moved to compel individual arbitration (2026-06-09; amended motion 2026-07-09). Docket trackers list an order on the motions to compel, dismiss, stay and strike entered 2026-11-12 (ECF 28) by title only. What the court decided was not read and no report of it was found as of 2026-11-16 ⚠️.

Re-disclosure check — EP298 (2026-10-02)

Dave praises Brendan Foody on air ("Raor", 2:19:47, Moonshots #298): "a an obviously good person and so successful at such a young age". No Link Ventures / "portfolio company" restatement; the pre-roll names Link Ventures at 0:01:05, 2 h 18 earlier and about nothing in particular. Same direction as EP293: Mercor is cited approvingly with the tie held in a different episode. ⚠️ Caption spelling variant "Raor" is new to the sweep list (add to Meror/Merkor/Merur).

What an operator can take from it

If you are thinking of working on it as an expert:

  • The route is public: a profile at mercor.com, AI assessments and an AI interview, then offers matched to listed roles. Listings on 2026-09-29 ran from about $60 to $250 an hour (on 2026-10-17 the home page showed Spanish tutors at $60–65/hr and clinicians at $120–150/hr); Mercor states an average contracted rate of $91/hour as rendered on 2026-11-16 ($90 on 11-04 and 11-10, $94 on 10-29, $100 on 10-23, $111 on 10-17, $108–$109 the weeks before; the renders show counter artefacts, so treat any single reading with care ⚠️) and weekly pay. Company-reported.
  • It is contract work: no guaranteed hours, projects end, and you handle your own taxes and benefits. The pay is for judgement you already have, which is the point and also the ceiling.
  • The best-paid listings ask for specific, verifiable experience (a licensed clinician, a securities analyst, a named codebase migration), not general "AI skills".

If you run a small company, what the model teaches:

  1. Sell the scarce input to whoever has the budget. Mercor did not invent demand. It noticed the labs would pay for expert judgement and built the fastest way to supply it.
  2. Know your own gross-to-net. A marketplace that books gross revenue looks four times larger than the business it keeps. Price, hire and plan on the net.
  3. Concentration is the risk that comes with the speed. A reported ~91% from two customers is a growth engine and a single point of failure at once.
  4. Climb the ladder from where you already are. Interviewer → talent pool → data → benchmarks → environments → enterprise. Each step reused the asset built by the one before.
  5. Trust is the product. One compromised open-source dependency was enough for a major customer to pause. For any company that holds other people's data, security is part of the offer.

⚠️ Unverified or open

  • The $20B round was announced on 2026-11-10 (Mercor blog), and NVIDIA is named as a participant. Still not disclosed: a lead, the size of the employee tender (and so the primary capital raised), and any revenue figure. The BusinessWire release was read through a Yahoo Finance copy on 2026-11-16; it states none of these.
  • $3B run-rate, $250M latest-month gross sales, 100+ customers (Nov 2026): the $3B is Foody's X post as quoted by four outlets (post not read). The $250M and 100+ come from two Indian outlets, and where he said them is not stated. All are gross, unaudited, with no period given.
  • Daily payouts, conflicting (2026-11-16): the newsroom says $5M a day, the home page $4M+. Headcount is 700+ on the newsroom (400+ on 4 Nov).
  • White v. Mercor.io order of 2026-11-12 (ECF 28): listed by title only; outcome unread.
  • OWE-Bench (2026-11-12): scores are from the post only (the leaderboard did not render), and the 54.6% appears both overall and for one domain. The $5,000-per-task call states no eligibility, deadline or IP terms. The Apache-2.0 licence was read through a summary of the dataset card.
  • Spirit sale hearing 17 Nov (Dkt. 1779): from docket trackers; no sale order entered.
  • "3x as much on inference as on employee salaries" and the double-digit-GDP projection (Series D post): company statement and forecast, with no amounts given.
  • Take rate and net revenue: not published by Mercor; the 60–70% and ~$600M figures are press and analyst estimates.
  • Customer concentration (~91%): one outlet, internal documents.
  • Stake sizes for Link Ventures, and whether any panelist beyond Link's LV3 participants is exposed: not public.
  • The caption sweep counts candidate episodes by regex; hits were read in context and attribution is inferred from hand-offs (auto-captions carry no speaker labels). Episodes through 2026-11-13 (EP311) are swept; the 2026-10-05 fireside has a machine-translated caption, so a miss there is weak evidence.
  • CBS 60 Minutes segment on Foody (reported 2026-10-05 via Axios): 100,000+ freelancers, about 400 employees, $2B annualised run-rate. Segment not watched; the 100,000 figure conflicts with Mercor's own 30,000+ weekly active. ⚠️ Unverified.
  • Benchmark Map and post-training gains (Oct 2026): company-reported, scored on Mercor's own benchmarks, not reproduced; the map's page is undated and its 2026-10-15 date rests on Mercor's X post.
  • AI Research Fund terms (Oct 2026): the listing titles and "$5M" labels were read at mercor.com/careers. Deadline, tiers, eligibility and licence terms come from two job-board copies of the posting (General Catalyst, Emcap), because the posting page did not render. Whether there is one $5M pool or two is not stated.
  • Forbes, 2026-10-13 (Mercor × 1Password SRE agent; 95% accuracy, 57% faster remediation): article returned 403, so the figures come from a search snippet. Not read. Mercor's own SRE post (2026-10-29) names no partner and contains none of these figures, so they remain unconfirmed by the company.
  • Nvidia round detail (The Information, paywalled): General Catalyst leading, Nvidia's "tens of millions" in data spend, Nemotron training on Mercor data: secondary copies only. On the date, a Yahoo Finance copy gives 2026-08-19, agreeing with the trackers; read the report as August, and the 22 Oct author-page date as unexplained.
  • PLE (2026-11-03): all scores company-reported on a benchmark Mercor and Google built and host; not reproduced. Window (9 Nov 2026 – 5 Jan 2027) and prizes ($250K/$100K/$50K) published by 2026-11-09; eligibility still not stated on ple-bench.com and the Kaggle rules not read (2026-11-10); the blog's "free runs for thirteen Google-sponsored models" is not stated on ple-bench.com. The SRE post's cost figures ($7.18 vs $1.86/$2.06 per alert) do not reconcile internally.
  • Foody's $100K data-referral offer (X, 2026-10-23) and the referral page's "$50M+ paid to companies": quoted by two outlets, neither the post nor the page read here.
  • One World Trade Center lease (25,550 sq ft, 77th floor): trade press only, not confirmed by Mercor.
  • "Profitable" (careers page, read 2026-10-29): company statement, no period or figure.
  • TaskNet (2026-10-27): working-paper figures are company-reported and not reproduced; the paper itself was not read.
  • Spirit Airlines backup bid ($7.5M): press citing the bankruptcy record; hearing date (6 Nov) from a tracker citing Travel Weekly.
  • Currey v. Mercor Global, Inc. (M.D. Fla., 2026): different entity name; not verified as this company, not counted among its disputes.

How this page stays current

A weekly pull (Mondays). Each step names the source and what to diff:

  1. mercor.com (home, /experts, /newsroom, blog, /careers for grants and fellowships): diff the stated valuation, daily payouts, average rate, expert and employee counts; log any new round, acquisition or product line.
  2. News, last 7 days ("Mercor" + raises / valuation / lawsuit / acquires in TechCrunch, Bloomberg, The Information, Forbes): a round enters the dated table only when Mercor announces it or it is reported as closed.
  3. Link Ventures portfolio and about pages: diff Mercor's fund and status, and the panelists' listed roles.
  4. Moonshots: sweep new transcripts for every spelling listed at the top (nine, including Raor, matched as a whole word only); add each hit to On tape vs reality with a read, and log a zero-hit sweep as swept, bounded by episode dates.
  5. Dockets (White v. Mercor.io, N.D. Tex. 6:26-cv-00201; In re Mercor.io Corporation Data Breach Litigation, N.D. Cal. 3:26-cv-02821; Scale v. Mercor is closed, check only for a reopening), then bump updated: and the Updated line. Anything not re-read at its source stays marked ⚠️.

Related

Moonshots with Peter Diamandis · Moonshots: the disclosures ledger · Dave Blundin · Peter Diamandis · Alexander Wissner-Gross · Salim Ismail · project through line matchmaking

Reading rule. Anything marked unverified or on-air is a claim made on the podcast and has not been checked against a primary source yet. It stays marked until it has. Sourced numbers link to where they came from.