Memo 03
Halluminate — a run rate built on inventory that expires
A run rate built on inventory that expires.
1. Thesis
Halluminate builds finance simulations and expert-written tasks that AI labs use to train agents, and it raised a $30M Series A on a company-reported mid-eight-figure run rate P1R2. The run rate annualizes one quarter of delivered work, and the CEO says environment complexity must roughly double every six to eight months, so much of that revenue may have to be re-earned on harder inventory R1F30inference. The company is worth watching if that treadmill turns into a repeatable process that labs can't cheaply copy.
Confidence: 4/10 that it becomes a durable, high-margin business. Demand looks real by the company's account; revenue quality and training impact aren't public.
2. Variant perception
- Consensus thinks: The story is the score and the speed. Coverage led with AI topping out at 51% on finance work and a nine-person team at a mid-eight-figure run rate P4P37. Its lead investor argues that specialization "will really be key" M60.
- I think: The useful question is what a dollar of that run rate is worth a year later. Two things decide it, and neither is public: how much goes to outside experts, and how much of the work repeats R2inference.
- Why I might be right: Wu's own mechanism sets the pace. Doubling every six to eight months means about 2.8x to 4x the complexity each year C2estimate. Epoch AI reports tasks may be retired near a 70% pass rate M20. Expert pay is $150–250 an hour for PE and banking roles P60. At Mercor, contractors keep 60–70% of billings per TNW, which would leave 30–40 cents per dollar at that rate M31C3estimate. Halluminate's basis may differ, and I found no statement of it R2.
An illustration, not a forecast C4estimate: a dollar billed today, multiplied by what's kept after experts and by how much of the work repeats next year, is worth about 9 cents (35% kept, 25% repeats), 18 cents (70%, 25%), 26 cents (35%, 75%) or 53 cents (70%, 75%). Expansion at the same labs could lift the repeat share above 100%, rising task prices could offset the treadmill, and labs may keep several vendors hypothesis. Both inputs are my assumptions. To keep 50 cents next year, repeat work would have to reach 143% at the 35% case, which means expansion, or 71% at the 70% case C5estimate.
3. The company in one paragraph
Halluminate was founded in 2024 by Jerry Wu (CEO) and Wyatt Marshall (CTO) and went through YC S25 F2F26F1. It sells reinforcement-learning environments and expert-built tasks to frontier labs, starting with financial services; enterprise "could come later" P7F29. Funding: a $30M Series A led by Oak HC/FT on 2026-10-01, with YC, Orange Collective, FT Partners and Heavybit, for $38.5M in total P1F34. The roughly $8.5M before the A is arithmetic; its stage and dates are not in public sources F38. Fortune counts nine people P4. Valuation, gross margin and customer names are not in public sources.
4. Why now
- RL became a major data category. Scale reported nearly half of its new data-training projects involve environments P69. Anthropic leaders reportedly discussed more than $1B on environments over a year, a discussion rather than a budget M12.
- Finance became a lab product line. OpenAI's Project Mercury was reported in October 2025, with more than 100 former bankers M64. Anthropic launched finance agents on 2026-05-05 M63.
- Vendor neutrality opened. After Meta's investment in Scale, Google and OpenAI reportedly left it M66.
- Caveat. OpenAI's Sherwin Wu said in 2025 he was "short" on these startups P66.
5. Product
| Item | Score | Key evidence |
|---|---|---|
| Problem | 3 | Labs reportedly spend; agents score poorly across benchmarks P65M67 |
| Solution | 2 | No public result showing finance training uplift R17 |
| Why now | 4 | Environments now a large category P69 |
| Moat | 2 | Claimed process IP, not yet observable P8 |
| Distribution | 3 | A handful of buyers; no sales postings P7P62 |
| PMF signals | 3 | Company-reported revenue and customers; no lab has confirmed R4R8 |
| Competition | 2 | Mercor, AfterQuery and lab in-house teams P67M45M27 |
The benchmark is the best public window into the product. Westworld Finance Diligence Bench has 88 problems across one acquisition process, in seven categories, run inside a desktop with hundreds of steps P18P19. Practicing deal professionals authored and reviewed the tasks; their number and names aren't stated P21. It was authored, run and QA-adjudicated in-house P20P29. In the sample task, agentic checks carry 95% of the weight P23. The post states no human baseline or confidence interval and doesn't name the grading model P24.
The published scores need a label, and they have an ordinary explanation.
| Source | Models | Top score |
|---|---|---|
| NeurIPS 2026 workshop listing (undated) | 7 | about 0.497 ±0.030 P33 |
| Fortune, 2026-10-01 | 7 | 51% P6 |
| RL List directory, 2026-10-05 | 8 | 0.59 R16 |
| Company post, read 2026-11-06 | 9 | 0.61 F23 |
These aren't four company statements. Claude Opus 5.5 was announced on 2026-09-22 and Gemini 4 Argon was generally available on 2026-10-13, so a rising top reflects added models R9R10R14. I found no same-model score that changed, and 0.51 sits inside the listing's ±0.030 R14R15. What remains is a documentation gap: the post displays 2026-08-11, contains later models, and has no changelog R13. A reader can't match a number to a run. That's easy to fix, and I'd ask the company about it.
Openness is a fair comparison, within limits. The post links no tasks or verifiers as of 2026-11-06, and the accompanying paper sat behind a verification page, so this is provisional P32R17R19. Vals and Mercor also keep scored sets private, and publish a subset and harness R20R21. BankerToolBench publishes all 100 tasks R22. Deal data carries a confidentiality reason P40. Most benchmark builders also sell data, agents or evals, so this isn't specific to Halluminate M68M70M71. Scores don't port across benchmarks: Argon gets 0.61 here and 82.4% on APEX F23R20.
Moat. The realistic candidates are process and cornered data. Wu calls producing ever-harder complexity the company's IP P8, and a posting seeks to license real-world financial data P63. Exclusivity isn't in public sources. Switching costs look low for project-style work inference.
PMF. The revenue basis is quarterly revenue from delivered, paid work, per Wu R1. The only outside signal is an investor: Oak HC/FT lists the company in its portfolio, without an amount P9. The run-rate claim isn't implausible: Epoch reports contracts often run six to seven figures a quarter R25.
6. Founders & team
| Item | Score | Key evidence |
|---|---|---|
| Founder-market fit | 3 | Bank agent launch, self-reported; customer evals F3F13 |
| Execution speed | 4 | Dated research releases and the A within 16 months F10F23F24 |
| Clarity of thought | 3 | Specific mechanism; states study limits F30F22 |
| Talent magnetism | 3 | Researchers from labs as angels, unnamed F25 |
| Track record | 2 | No prior exit found in accessible sources F64 |
| Team completeness | 3 | No sales role among visible postings F58 |
| Integrity of claims | 3 | Round matches across sources; revenue unverifiable F33R4 |
Wu says he led product and research at Capital One Labs F3. Marshall worked in data infrastructure at MediaWallah F6. Buy-side or deal experience of either founder is not in public sources; deal professionals write the tasks P21. LinkedIn was blocked for this check, so this is incomplete.
The choice to note: with four of the top five closed-source US labs reportedly paying, Wu is "deliberately concentrating on a small group of frontier model labs", and "Wu sees that focus as a strength" R5F29. It's a clean way to learn fast, and it leaves revenue with few buyers inference.
A detail in the team's favor: in July the company published a study where a simulator-only model reached 0.6% on real Best Buy, against 23.4% for a sim and real mix, while noting its small sample F22P44. It's one model and one site pair, so it says little about finance, but publishing it is a useful habit inference.
7. Market & competitive structure
Size estimate. External lab spend on expert data and environments is about $4.8–5.5B annualized, from disclosed vendor billings at 80–91% lab share M77. Environments alone are about $0.5–2B M78. The finance slice, at an assumed 10–25%, is about $100–250M M79. That share has no public source. I don't compute Halluminate's share: the inputs are assumed and partly circular R23R24.
| Player | Type | Disclosed |
|---|---|---|
| Mercor | Direct, buying Deeptune | $2B gross run rate; $10B last confirmed valuation M30M34P67 |
| AfterQuery | Direct, finance overlap | $3.2B reported valuation M45 |
| Handshake | Publishes BankerToolBench | About $1.1B gross, Sacra estimate M42 |
| Lab in-house | Substitute | About 3,000 Meta engineers moved to this work, per SemiAnalysis M27 |
| Hire experts directly | Substitute | Mercury runs through third-party suppliers M65 |
Labs can build this inference. Finance is less exposed than coding because labs don't employ deal teams, though Mercury shows experts can be bought without an environment vendor M65inference. Google hired Mechanize staff with a licence, final terms undisclosed M48.
8. Traction signals (verifiable only)
| Signal | Date | Source |
|---|---|---|
| Series A announced by company; Oak lists it in portfolio | 2026-10-01 | P1P9 |
| 88-task benchmark published | 2026-08-11 | P18 |
| Listed as a poster at a NeurIPS 2026 workshop; review process not stated | Dec 11 2026 | P33P34 |
| Ten open postings, San Francisco | 2026-07 to 2026-10 | P62 |
Revenue and customers aren't on this list: they're company statements R4R8.
9. What has to be true
- Net revenue is close to gross, or margin holds after expert pay.
- Work repeats: multi-year or expanding contracts, rather than one-off projects.
- Labs see measurable finance gains after training, on tasks held out from it.
- Tooling lets a small team keep up with 2.8x to 4x complexity a year [C2].
- Finance stays a real share of lab spend while labs build in-house.
10. Bear case & pre-mortem
Bear case. Halluminate may be a services business with software-company optics, selling a depreciating asset to buyers who can replace it hypothesis. Revenue annualizes one quarter of project work, tasks may retire, and depth comes from outside experts R1M20P60. If one lab in-houses, about a quarter of revenue could leave within a contract cycle, assuming an even split across four labs hypothesis. I found no public source showing a finance environment improving a lab's model. In that case, $38.5M buys time rather than a moat.
It's 2029 and Halluminate is dead. - Market: Finance proves a thin slice, or budgets move to cheaper data. Early sign: reported task prices and contract sizes R25. - Execution: The complexity treadmill outruns a small team, or a data incident costs a customer, as Meta's pause did at Mercor M37. Early signs: the paper and task release, a SOC 2 or trust page P34P71. - Competition: Mercor with Deeptune or AfterQuery bundles environments with expert supply. Early sign: a finance subset in APEX, or another vendor acqui-hire R20M48.
11. Key risks & what would retire them
| Risk | Severity | What would retire it | When we'd know |
|---|---|---|---|
| Revenue quality | High | Disclosed net revenue or margin | Next raise |
| Concentration | High | A named customer; a fifth buyer or enterprise | 12–18 months |
| No proof of uplift | High | Published uplift on held-out tasks | After the paper |
| In-housing | Medium | Lab renewals despite internal teams | First renewals |
| Documentation gap | Low | Versioned results, or task subset release | Around Dec 11 |
12. What I'd ask the founders
- What share of billings goes to experts? Is the run rate gross or net?
- What share of this quarter's revenue comes from customers who bought last year's inventory?
- Which models and run produced 51%? Will the post get a changelog?
- Will a subset of tasks and a harness ship with the paper?
- Have you seen a lab's finance score rise after training, on tasks it didn't train on?
- How does your tooling hold cost per task as complexity doubles?
- What rights do you hold over licensed deal data?
13. Verdict & kill criteria
A real company with a credible team, and an open question about what kind of business it is. I wouldn't call this a software margin story yet, and I found nothing showing the reported numbers are wrong. Evidence of repeat work and training gains would move me up quickly.
Kill criteria: - A customer lab announces in-house finance environments and ends vendor work. - Reported net revenue is a small fraction of the run rate. - A rival publishes open finance tasks that labs adopt as the standard.
Falsifiable prediction: By 2027-12-31, neither Halluminate nor a frontier lab will publish a result showing a finance environment raised a model's score on held-out tasks. If one does, I'm wrong.
Sources (86)
- P1Company post: $30M Series A led by Oak HC/FT; YC, Orange Collective, FT Partners, Heavybit participated; total raised $38.5M · 2026-10-01
halluminate.ai/blog/series-a - R2Company post: "Grown from $0 to a mid-eight-figure revenue run rate while remaining strongly profitable"; no figure, gross or net basis, or audit stated · 2026-10-01
halluminate.ai/blog/series-a - R1Fortune attributes the mid-eight-figure run rate to CEO Jerry Wu ("based on quarterly revenue from work already delivered and paid for"; profitable, "he said"); the article mentions no documents or records reviewed · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - F30Fortune: Wu's "Moore's law of environments"; environment complexity needs to roughly double every six to eight months · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - P4Fortune describes Halluminate as a nine-person San Francisco startup; reports $30M Series A led by Oak HC/FT and $38.5M total; lists YC, Orange Collective, Heavybit and individual lab researchers per Wu · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - P37TechTimes FAQ: 51% not independently validated by a third party or peer-reviewed; benchmark published by the company selling related environments · 2026-10-01
www.techtimes.com/articles/328407/20261001/finance-ai-tops-out-51-due- - M60Fortune quote, Matt Streisfeld (Oak HC/FT): "testing work and specialization will really be key."; Wu uses the term "Moore's law of environments" · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - M20Epoch AI: tasks may be retired near 70% pass rate; one startup employee says RL requests are perhaps 10-20x benchmarking requests · 2026-01-12
epoch.ai/gradient-updates/state-of-rl-envs - P60Expert Network: PE and IB $150-250 per hour; other roles $100-200; median monthly earnings $3,800, top 25% average $30,000+ (September 2026; method not stated); 1099 only · 2026-09
halluminate.ai/experts - M31TNW: contractors receive 60-70% of billings; TNW estimates Mercor net revenue at $600M-$800M · 2026-07-10
thenextweb.com/news/mercor-20-billion-valuation-deeptune-acquisition - F2Wu: "We left our jobs last year" · 2025-08-11
news.ycombinator.com/item?id=44865290 - F26Fortune: nine-person San Francisco company founded 2024; CEO Jerry Wu, CTO Wyatt Marshall · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - F1Launch HN post by wujerry2000 presents Halluminate (YC S25); signed by Jerry and Wyatt · 2025-08-11
news.ycombinator.com/item?id=44865290 - P7Fortune via Yahoo: company intentionally focused on frontier labs; enterprise customers and other industries not a current priority · 2026-10-01
finance.yahoo.com/technology/ai/articles/exclusive-nine-person-hallumi - F29Fortune: "deliberately concentrating on a small group of frontier model labs"; enterprise "could come later"; "Wu sees that focus as a strength." · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - F34Company post: "This brings our total capital raised up to $38.5 million." · 2026-10-01
halluminate.ai/blog/series-a - F38Pre-Series A funding of about $8.5M is arithmetic ($38.5M minus $30M); neither primary source states its stage, date or composition · 2026-10-01
halluminate.ai/blog/series-a - P69Fortune via Yahoo: Scale AI reported nearly half of its new data-training projects involve RL environments · 2026-10-01
finance.yahoo.com/technology/ai/articles/exclusive-nine-person-hallumi - M12TechCrunch citing The Information: Anthropic leaders discussed spending more than $1B on RL environments over the next year · 2025-09-21
techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-t - M64Bloomberg via Fortune: OpenAI's Project Mercury has more than 100 ex-investment bankers paid $150/hour, about one Excel model per week · 2025-10-21
fortune.com/2025/10/21/openai-entry-level-banking-finance-jobs-automat - M63Fortune: on 2026-05-05 Anthropic announced about 10 finance agents, including pitchbooks, credit memos and underwriting, plus Microsoft 365 integration · 2026-05-05
fortune.com/2026/05/05/anthropic-wall-street-financial-services-agents - M66TechCrunch: Meta invested $14B in Scale; Google and OpenAI reportedly dropped Scale as a data provider · 2025-09-21
techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-t - P66Same article: OpenAI's Sherwin Wu says he is "short" on RL environment startups; Ross Taylor warns about scaling difficulty and reward hacking; Surge reportedly $1.2B revenue; Mercor valued $10B · 2025-09-21
techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-t - P65TechCrunch: Anthropic leaders reportedly discussed $1B+ on RL environments per The Information; a16z's Jennifer Li says all big labs build in-house · 2025-09-21
techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-t - M67BankerToolBench abstract: built with 502 bankers; best of nine models, GPT-5.4, fails nearly half of rubric criteria; 0% of its outputs rated client-ready · 2026-04-13
arxiv.org/abs/2604.11304 - R17Benchmark post: "We will release an accompanying paper soon"; no statement on releasing tasks, verifiers or code; no confidence intervals or human baseline; grading model not named · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P8Fortune via Yahoo: Wu says environment complexity must roughly double every six to eight months; calls ability to keep producing that complexity the company's IP; plans to turn benchmark failure modes into RL environments · 2026-10-01
finance.yahoo.com/technology/ai/articles/exclusive-nine-person-hallumi - P62Ten open postings, all San Francisco on-site: engineering, research, finance researcher, FDE, data, people, recruiter, chief of staff, project lead; no sales title; platform posting seeks software that 10-100x's RL environment quality and throughput · 2026-07-15 to 2026-10-06
api.ashbyhq.com/posting-api/job-board/halluminate - R4The run rate is confirmed only as a company statement; all sources found trace to the company; the amount is not verifiable from public sources · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - R8Four-of-five-labs is confirmed only as a company statement; no lab has confirmed; which labs is not in public sources · 2026-10-01
halluminate.ai/blog/series-a - P67Mercor announces acquisition of Deeptune; cites $43M Series A led by a16z; Mercor already a customer · 2026-07-09
mercor.com/blog/mercor-to-acquire-deeptune - M45TechCrunch citing Forbes: AfterQuery reportedly valued at $3.2B, up from $300M at its April $30M Series A; $100M run rate as of April; customers include Nvidia, Legora, Motif · 2026-09-01
techcrunch.com/2026/09/01/afterquery-reportedly-becomes-y-combinators- - M27SemiAnalysis: about 3,000 Meta engineers moved to making RL tasks and environments full-time after a late-May restructuring · 2026-07-09
newsletter.semianalysis.com/p/the-future-of-meta-superintelligence - P18Blog: Westworld Finance Diligence Bench has 88 problems on a full acquisition diligence process, anonymized data from real private transactions, dynamic desktop environments, hundreds of steps · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P19Blog categories: modeling 14, underwriting 13, quality of earnings 9, data preparation 11, pitch materials 15, closing 4, diligence review 22 · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P21Blog: every task authored and reviewed by practicing finance deal professionals; number and names of experts not stated · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P20Blog authors: Alina Hyk, Robert Alward, Victoria Knapp Perez, Wyatt Marshall (Marshall is CTO per Fortune via Yahoo) · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P29Blog QA: six automated post-run checks, mostly by claude-opus-5; red-team solver; claude-sonnet-5 judges hollow high scores; a core-team member adjudicates all findings · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P23Blog sample task: three deterministic checks carry about 5% of weight, seven agentic checks 95%; no fixed ratio across tasks · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P24Blog text states no human baseline, no confidence interval or standard error, and does not name the model grading agentic verifiers · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P33NeurIPS 2026 workshop poster listing by same authors: seven models, fourteen configurations, peak mean about 0.497 (+/-0.030); code execution improves every model by +0.002 to +0.149; instruction loss 42.4% of long-horizon failures; no code or data links on the listing · n/d
neurips.cc/virtual/2026/160087 - P6Fortune: benchmark gave seven frontier models 88 tasks; highest average score 51% · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - R16RL List, last updated 2026-10-05: "Eight models are scored on two harnesses"; top 0.59 (Opus 5.5) from "the current blog table"; it reads the company table and is not a separate result · 2026-10-05
www.rl-list.com/vendors/halluminate - F23Finance Diligence Bench post: authors Alina Hyk, Robert Alward, Victoria Knapp Perez, Wyatt Marshall; nine models, 88 problems; top mean 0.61 (Gemini 4 Argon, native CLI harness), 0.51 in CU + TU harness · 2026-08-11
halluminate.ai/blog/due-diligence-bench - R9Anthropic announced Claude Opus 5.5 on 2026-09-22 · 2026-09-22
www.anthropic.com/news/claude-opus-5-5 - R10Google post on Gemini 4 Argon is dated 2026-10-13 and says it is "now globally available"; search-result text shows a limited announcement on 2026-09-30, not read on a live page · 2026-10-13
blog.google/innovation-and-ai/models-and-research/gemini-models/introd - R14Opus 5.5 and Argon were not generally available on 2026-08-11, so the live table was revised after its displayed date; 0.59 and then 0.61 as the top follow from added models, with no evidence that any same model's score changed · 2026-08-11
halluminate.ai/blog/due-diligence-bench - R15Workshop listing: mean scores "top out at $0.497 \pm 0.030$"; the stated interval, 0.467 to 0.527, contains 0.51, so 51% and 0.497 need not be different results; link not confirmed · n/d
neurips.cc/virtual/2026/160087 - R13Benchmark post says Gemini 4 Argon's trajectory runs "were added later" and that the "earlier v1 run" has models that "differ from those in the results above"; no changelog · 2026-08-11
halluminate.ai/blog/due-diligence-bench - P32Blog: "We will release an accompanying paper soon"; no GitHub, arXiv, Hugging Face or leaderboard link for this benchmark in the post · 2026-08-11
halluminate.ai/blog/due-diligence-bench - R19OpenReview page for the paper returned a browser verification page; not bypassed; paper unread · n/d
openreview.net/forum?id=DTef39ShHg - R20Mercor APEX-Agents leaderboard: "The full task set stays private so that models can't be trained on it"; open subset and harness; Gemini 4 Carbon 84.2%, Argon 82.4% · n/d
www.mercor.com/apex/apex-agents-leaderboard/ - R21Vals Finance Agent v2.1: 449-sample test set private "to prevent overfitting"; 27 public samples and harness open; Opus 5.5 top at 67.85% · 2026-11-05
www.vals.ai/benchmarks/fabv2-1 - R22BankerToolBench README: tasks.jsonl carries prompts and rubrics for 100 tasks; data CC-BY-4.0, code Apache-2.0 · n/d
github.com/Handshake-AI-Research/bankertoolbench - P40DealTrace repo README: files are a fictional mock, not the paper's dataset; results "are not comparable"; deal packages confidential and not distributed · n/d
github.com/Halluminate/DealTrace/blob/main/models/README.md - M68BankerToolBench full text: 100 tasks; GPT-5.4 about 16% Pass@1; authors at Handshake AI, one also at McGill · 2026-04-13
arxiv.org/html/2604.11304 - M70APEX-Agents v1: Pass@1 is the share of tasks meeting all criteria; investment banking n=160, top 27.3%; authors at Mercor · 2026-01-20
arxiv.org/html/2601.14242v1 - M71BigFinanceBench: 928 questions; top rubric score 58.8%; top answer accuracy 44.3%; authors at Rogo and OpenAI; 50-question public subset · 2026-06-02
arxiv.org/html/2606.03829 - P63Data Platform Lead posting ($295-375K base): "Source, license, and purchase real-world financial data and artifacts"; first hire on data team · 2026-08-17
api.ashbyhq.com/posting-api/job-board/halluminate - P9Oak HC/FT portfolio page lists Halluminate as Active, CEO Jerry Wu, "Reinforcement-learning environments for financial services"; no amount or date · n/d
www.oakhcft.com/portfolio - R25Epoch AI: "Contract sizes are often six to seven figures per quarter"; article does not address vendors per lab and gives no finance versus coding share · 2026-01-12
epoch.ai/gradient-updates/state-of-rl-envs - F3Wu: "I previously led product/research at Capital One Labs, where I launched one of the first AI agents in banking." (self-reported) · 2025-08-11
news.ycombinator.com/item?id=44865290 - F13Wu: "faced these problems first-hand while building evals for our customers" · 2025-08-11
news.ycombinator.com/item?id=44865290 - F10Web Bench post: authors Suchintan Singh, Shuchang Zheng, Wyatt Marshall, Jerry Wu; open-sourced with Skyvern; about 5,750 tasks, about 2,454 public; best write-task success 46.6% · 2025-05-30
halluminate.ai/blog/benchmark - F24Fortune (syndicated on Yahoo Finance): $30M Series A led by Oak HC/FT; total funding $38.5M · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - F22Sim-to-real study by Victoria Knapp Perez: sim-only model 0.6% on real Best Buy, real-only 15.3%, 50/50 mix 23.4%; small sample, one site pair · 2026-07-27
halluminate.ai/blog/sim-to-real-webagents - F25Fortune: existing investors Y Combinator, Orange Collective, Heavybit participated, plus individual researchers from Anthropic, OpenAI, Meta (per Wu; unnamed) · 2026-10-01
finance.yahoo.com/technology/ai/articles/exclusive-nine-person-hallumi - F64No prior exit or earlier venture-backed company by either founder found in accessible sources · 2025-08-11
news.ycombinator.com/item?id=44865290 - F58Aggregator lists San Francisco roles including Data Platform Lead, MTS Research / Post-Training, MTS Finance Researcher, Chief of Staff, People Platform Lead; no sales role among the 8 of 10 visible · 2026-11
www.fastaijobs.com/companies/halluminate - F33Company post by Wu: $30M Series A led by Oak HC/FT, participation from YCombinator, Orange Collective, FT Partners, Heavybit · 2026-10-01
halluminate.ai/blog/series-a - F6AI Engineer bio: Marshall studied CS and philosophy at Cornell (Milstein Program); at MediaWallah from December 2022 on data infrastructure · n/d
www.ai.engineer/speakers/wyatt-marshall - R5Fortune: "According to Wu, four of the top five closed-source U.S. AI labs are paying customers"; no lab named; researchers from Anthropic, OpenAI and Meta are named only as individual investors · 2026-10-01
fortune.com/2026/10/01/halluminate-raises-30-million-series-a-oakhc-ft - P44Sim-to-real study releases Hugging Face dataset, models and W&B report; one model, one site pair · 2026-07-27
halluminate.ai/blog/sim-to-real-webagents - M77Disclosed vendor billings sum to about $6.0B: Mercor 2.0 + Handshake AI 0.95 + Surge 1.2 + Scale 1.5 + Fleet, Mechanize, AfterQuery 0.3 + Halluminate 0.05; at 80-91% lab share, about $4.8-5.5B · 2026-11-06
derived (see claim) - M78External RL environment spend about $0.5-2B annualized, center about $1B; buyer-side and supply-side builds in market.md · 2026-11-06
derived (see claim) - M79Finance RL environment slice about $100-250M a year at an assumed 10-25% share of about $1B; share has no public source · 2026-11-06
derived (see claim) - R23Share sensitivity on market.md inputs: lowest $50M / ($2B × 25%) = 10%; highest $50M / ($0.5B × 10%) = 100%; with $30M-$70M run rate (P57), 6% to 140%; in the base buyer-side case share = 1 / (10 × finance share), so the revenue figure cancels because the $12.5M contract is itself $50M ÷ 4 · 2026-11-06
derived (see claim) - R24The 20-50% share is not verifiable: finance share, vendors per lab and the $50M midpoint are assumptions, and the numerator may include non-finance or expert-data revenue · 2026-11-06
derived (see claim) - M30TNW: CEO Brendan Foody says Mercor run rate passed $2B and doubled in four months; figure is gross billings · 2026-07-10
thenextweb.com/news/mercor-20-billion-valuation-deeptune-acquisition - M34Sacra: Mercor raised a $350M Series C at a $10B valuation in October 2025, led by Felicis · undated
sacra.com/c/mercor/ - M42Sacra estimate, April 2026: Handshake about $1.1B annualized gross, about $450M net, about $950M gross from AI data · 2026-04
sacra.com/research/handshake-indeed-for-data-labelers/ - M65OpenAI via the same article: "Experts are recruited, managed and compensated by third party suppliers." · 2025-10-21
fortune.com/2025/10/21/openai-entry-level-banking-finance-jobs-automat - M48RL-List relaying Business Insider: talks reported 2026-08-05 above $1.5B including a non-exclusive licence; on 2026-09-11 reported complete, cofounder Besiroglu at DeepMind, more than a dozen staff at Google; final terms undisclosed; Mechanize still operates · undated
www.rl-list.com/vendors/mechanize - P34Workshop "Agentic AI Benchmark and Application for Enterprise Tasks", Dec 11 2026; organizers include Graham Neubig; page does not state review process or archival status; Halluminate not named · n/d
neurips.cc/virtual/2026/workshop/137523 - M37TNW: after a March supply-chain attack on LiteLLM exposed Mercor data, Meta paused work with Mercor indefinitely; Mercor calls the impact "very limited"; OpenAI and Anthropic remained customers · 2026-07-10
thenextweb.com/news/mercor-20-billion-valuation-deeptune-acquisition - P71RL List: no SOC 2, security page or trust center found as of 2026-10-05 · 2026-10-05
www.rl-list.com/vendors/halluminate
Disclosure. Not investment advice. This memo is based solely on public information as of 2026-11-06. Disclosure: I'm a student at Minerva University, not a professional investor or analyst. I write these memos to learn how to evaluate companies, in public. I hold no position in Halluminate or its competitors, have no relationship with the company or its investors, and no one paid for this memo. It was researched with Claude. Anthropic, which makes Claude, has models scored in this benchmark, and may or may not be one of the unnamed customers; I don't know. Halluminate was not contacted before this draft. I found no evidence that any figure Halluminate reports is inaccurate. I will get things wrong; corrections are welcome at aldiyar@nur.capital.