Key judgements
Our assessment. Confidence reflects the quality of the evidence behind each judgement, not the probability of an event.
- An "Anthropic mafia" is likely and has already begun: two start-ups with former Anthropic researchers among their founders raised a combined $680m in 2026, before the listing.Basis: Parts 4 to 6; Figures 4, 5 and 6ModerateModerate confidenceSourced facts, but the judgement rests on our reading of how the pieces fit
- On an illustrative reading of US data from 1992 to 2006, Anthropic could produce roughly 120, perhaps up to 400, alumni in senior start-up roles by 2029: dozens of companies, but only a handful of frontier-scale labs.Basis: Parts 5 and 7; Figure 7LowLow confidenceA reasoned view from incomplete evidence; plausible, not established
- Most alumni ventures will stay close to AI models and software, with robotics, materials and biology a minority of perhaps a quarter, as among OpenAI's alumni.Basis: Part 5; Figure 6ModerateModerate confidenceSourced facts, but the judgement rests on our reading of how the pieces fit
- A larger share of the windfall will go to philanthropy than in earlier mafias, given the founders' 80% pledges (about $100bn at the midpoint) and the company's matching.Basis: Part 6ModerateModerate confidenceSourced facts, but the judgement rests on our reading of how the pieces fit
- For France, a single giant listing would help, but repeated exits and the policies that enable them matter more than any one company or any one election.Basis: Part 8; Figure 8LowLow confidenceA reasoned view from incomplete evidence; plausible, not established
Contents9 parts, 8 figures
Part 1The millionaire factory
The most consequential number in Anthropic's prospectus may be a headcount. If the listing goes as planned, about 4,250 of the company's 5,300 employees would become millionaires on paper, according to an analysis by Hill, a San Francisco investment platform that trades stakes in private companies, first reported in early November by the British newspaper The Observer and since relayed by CNBC-TV18 and Xataka. Roughly 50 early employees could each pass $100m, and at least 1,200 more could hold over $10m. The seven co-founders are not counted.
The estimate rests on a large assumption: a valuation of $1.5tn. On 10 November Anthropic set a range of $70 to $78 a share for 825m Class A shares, to trade on Nasdaq under the ticker ANT. Pricing is expected on 18 November and the first trade on 19 November, according to Yahoo Finance and Decrypt. The range was unchanged as of 16 November, with orders close to seven times the shares on offer, Bloomberg reported, according to The Next Web; on 11 November the Financial Times reported that bankers were weighing a price above $78. On the roughly 13.09bn shares outstanding after the offering, that range values Anthropic at $916bn to $1.02tn, the "about $1tn" in circulation. Counting options and restricted stock units, the fully diluted value at the top of the range is $1.40tn. Le Détroit's calculation: on that count of about 18bn shares, and if Hill used the same basis, its $1.5tn means about $83 a share, 7% above the top of the range. The two headline valuations are closer than they look, because they measure different things.
Crypto traders have already put a price on it. On Hyperliquid, a decentralised exchange, pre-IPO perpetual contracts let traders bet on Anthropic's value before the listing; they are leveraged derivatives settled against the eventual price, not shares. The largest by open interest, Entropy's io:ANTH, is priced on Anthropic's total market value, according to Coin Metrics data cited by TechFlow; TokenPost equated a mark of about 1,651 on 11 November with a valuation of about $1.64tn. On the morning of 17 November it stood at about 1,660, down about 6% in 24 hours, or roughly $1.65tn. Le Détroit's calculation: on about 18bn fully diluted shares, that is about $92 a share, some 18% above the top of the range and above Hill's $1.5tn assumption. A rival contract quoted per Class A share, Trade.xyz's xyz:ANT, traded between about $97 and $106 on 10 and 11 November, according to TechFlow and PANews. Thin, leveraged markets can overshoot; Wednesday's pricing will test them.
Hill has not published its full method. According to Xataka's account of The Observer's article, it sorted staff into bands by the equity in their pay packages and, for early staff, founding grants, and presented the results as minimums at $1.5tn. The figures are gross: before tax, before lock-ups and before any move in the share price. Hill puts employees' combined potential gains at about $270bn. Le Détroit's calculation: that is about $51m a head on average, an average that flatters, since on Hill's own numbers about 1,050 employees, a fifth of the staff, would not reach $1m.
What those people do next is the subject of this article. The idea we test is that Anthropic is about to seed what Silicon Valley calls a "mafia", after the alumni networks of PayPal and SpaceX: thousands of people who know how to train frontier models from first principles gain financial freedom and, within about three years, leave to found ambitious labs in robotics, materials science and biology. We also ask what a French equivalent would need.
Choose a scenario; hover or tap a band
Bands at Hill's $1.5tn valuation; the co-founders are excluded. The other views rescale the same holdings: the people do not change, the sums do. Hill has not published its method. Values are on paper, before tax.
Part 2What the prospectus says
Behind the wealth estimate is a business whose growth explains the valuation and whose spending explains the risk. According to the prospectus, as reported by TechCrunch, Decrypt and The Tech Hacker, revenue rose from $41m in 2023 to $386m in 2024 and $4.6bn in 2025, then reached $33.2bn in the first nine months of 2026, $17bn of it in the third quarter. Anthropic put its annualised run-rate at about $81bn at the end of October. Headcount nearly tripled in a year, and trailing revenue per employee passed $6.6m, Yahoo Finance reported, citing the prospectus; by Yahoo's calculation Nvidia's is about $7.2m.
The losses look alarming and mostly are not what they seem. The net loss for the nine months was $104bn, but $94.3bn of it was a non-cash charge reflecting the rising value of convertible notes and similar instruments; the operating loss was about $1.26bn, Decrypt reported. The real commitment is compute: about $900.9bn of spending commitments, of which about $70.7bn falls due within twelve months, against cash and short-term investments of nearly $132bn. The offering would raise $57.75bn to $64.35bn, and seven suppliers, AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC, have agreed to buy $26bn of non-voting Class N shares at the IPO price, Google and Nvidia about $10bn each.
The governance is built so that money does not run the company. A single Class F share, held by an LLC of the seven co-founders, carries 50.1% of the vote on most matters, and a single Class T share lets the Long-Term Benefit Trust, whose voting trustees include Ben Bernanke and Richard Fontaine, elect a majority of the board, Reuters and PPC Land reported. Amazon and Google hold non-voting stock, and Google's stake is capped at 14.99%, according to Fortune. "Our capital structure will concentrate voting power with the LTBT and the Founder LLC," the filing warns.
That design matters for the "mafia" question. Anthropic's employees will be owners with money but little say over the company. Those who want to steer something of their own will have to build it elsewhere.
Part 3Paper, not cash
A paper fortune becomes spendable slowly, and it can shrink on the way. Anthropic's lock-up is unusual: instead of a single 180-day expiry, shares are released in stages, through more than 20 windows over about a year, Yahoo Finance reported. Certain non-founder directors, officers and employees may sell up to 10% of their covered shares from the first trading day, a further 35% after full-year 2026 results and the rest at the earlier of 180 days or first-quarter 2027 results. Co-founders may sell up to 29% at that point and the remainder after a year or third-quarter 2027 results, according to summaries of the filing by Leverage Shares and VIP Wealth Advisors.
Tax arrives first. Many employee awards are restricted stock units that settle when the registration becomes effective, and settled units are taxed as ordinary income, in 2026 for a November listing; Anthropic expects to withhold part of the shares to cover it. The Information reported that the company was considering making staff sell through pre-set trading plans, Forbes noted in September.
History shows how far paper can fall. Facebook listed at $38 in May 2012; when the first lock-up released 271m shares on 16 August, the stock touched $19.69, about half its offer price, and employees' 234m shares became saleable only on 31 October, with the stock near $21. Alibaba's 2014 listing is a closer parallel for a "mafia": about 6,000 current and former employees held stock worth nearly $8bn before the IPO, The New York Times reported, and former staff had already started some 130 internet businesses, more than alumni of any other Chinese company, according to Itjuzi.com data cited by the paper. Le Détroit's calculation: that was about $1.3m a head, against Hill's $51m average for Anthropic, about 38 times more, though the paper noted that the $8bn was only part of the stock granted over the years.
Choose a view; hover or tap a bar, a square or a point
Squares are drawn to scale by area. Alibaba's figure is in 2014 dollars and covers holders at the group and its affiliates; Anthropic's is Hill's gross estimate.
Only prices reported on these dates are drawn; the dotted line between them is not the trading path. Facebook listed on 18 May 2012.
Much of the liquidity has also arrived before the bell. Current and former OpenAI staff sold about $6.6bn of shares at a $500bn valuation in October 2025, less than the $10.3bn on offer, SiliconANGLE and the Associated Press reported. Anthropic ran its own employee tender earlier this year at a $350bn valuation, according to Forbes. And SpaceX, which raised $75bn in June in the biggest IPO on record, was expected by Hill to make more than 4,400 current and former employees millionaires, about 400 of them worth over $100m, Fortune reported. The money is already partly in hand; the listing mostly removes the ceiling on it.
Choose a valuation
Series H valuationThe $965bn valuation Anthropic announced after its May funding round (NPR; TechCrunch).
Bottom of the rangeFully diluted value at $70 a share, about $1.26tn (Le Détroit's calculation from the $1.404tn at $78).
MidpointFully diluted value at the $74 midpoint, about $1.33tn.
Top of the rangeFully diluted value at $78: about $1.404tn, as reported by TechCrunch and CNBC.
Hill's assumptionThe valuation behind the 4,250-millionaire estimate: about $83 a share on a fully diluted count.
Part 4The mafia pattern
Alumni networks become "mafias" when three things meet: a new skill that few people have, a liquidity event that frees them, and a place dense enough for them to find each other. The first case set the template. In 1957 eight engineers left Shockley Semiconductor to found Fairchild Semiconductor, with money from Sherman Fairchild. Gordon Moore and Robert Noyce left to found Intel in 1968; Eugene Kleiner co-founded Kleiner Perkins and a former Fairchild executive, Don Valentine, started Sequoia Capital, both in 1972. By 2014 Endeavor Insight traced more than 2,000 companies to Fairchild's co-founders, including 92 public companies worth about $2.1tn. "It was really the place that got the engineer entrepreneur really moving," Moore said in 2016.
PayPal gave the phenomenon its name, and its story is the closest template for Anthropic's. It began as two start-ups. Confinity, incorporated in December 1998, counted Peter Thiel, Max Levchin, Luke Nosek and Ken Howery among its founders; X.com, a payments company started by Elon Musk, was incorporated in March 1999. The two merged in March 2000, and the combined company took the PayPal name. Mr Musk lost the chief executive's job after the merger, Fortune later reported, and Mr Thiel led the company to the stock market.
The business was brutal, and it taught a rare skill: moving money online without being swamped by fraud. PayPal lost $264.7m between March 1999 and September 2001, according to its prospectus, while cutting transaction losses from 1.08% of payment volume in the first nine months of 2000 to 0.42% a year later, thanks to "risk controls and proprietary technology". By the end of 2001 it had 12.8m accounts and 618 full-time staff, 198 at its Palo Alto headquarters and 418 at a customer-service centre in Omaha. It listed on Nasdaq on 15 February 2002 at $13 a share; the stock closed its first day at $20.09, up nearly 55%, CNNMoney reported. On 8 July eBay, where about 60% of PayPal's business took place, agreed to buy it for about $1.5bn in eBay shares, a deal completed on 3 October.
The windfall was concentrated. The prospectus listed Mr Musk as the largest individual shareholder after the listing, with 11.9%, ahead of Mr Thiel (4.6%) and Mr Levchin (2.9%); Sequoia Capital's funds held 8.9%. Le Détroit's calculation: at the sale price a stake that size was worth roughly $180m, while the whole $1.5bn, shared evenly among the 618 staff, would have come to about $2.4m each.
Most of PayPal's key people then left eBay but stayed in touch, Fortune reported. Mr Musk had run SpaceX since May 2002, and Reid Hoffman started LinkedIn in his living room that year. Within three years of the sale Mr Thiel had co-founded Palantir and, with Mr Nosek and Mr Howery, Founders Fund; Jeremy Stoppelman and Russel Simmons had started Yelp with Mr Levchin's first $1m; and Jawed Karim, Chad Hurley and Steve Chen had founded YouTube, which Google agreed to buy for $1.65bn in 2006. In October 2007 Fortune photographed thirteen of them at Tosca in San Francisco, for a story that made the "PayPal Mafia" nickname famous; the article said the group already used the name for itself. Mr Thiel and Mr Levchin put the value of dozens of PayPal-linked ventures at roughly $30bn. "Today they call upon one another when they need money or advice," Fortune wrote.
16 alumni, 20 companies and funds, 61 sourced links. Choose a view or an alumnus; hover or tap any name for roles and years, and tap a company or fund to select it.
Six hubs (Thiel, Levchin, Hoffman, Rabois, Founders Fund and Sequoia) touch 15 of the 20 companies and funds directly, and 15 of the 16 alumni within two steps.
Funds turned the windfall into a system: Founders Fund alone backed SpaceX, Palantir, Facebook, Affirm and Geni.
Companies founded, led or backed by the alumni, from rockets and electric cars to reviews, video and lending.
Peter Thiel: 6 links beyond PayPal; shares a company or fund with Hoffman, Howery, Levchin, McCormack, Nosek, Rabois, Simmons and Stoppelman; every other connected alumnus is at most 4 steps away.
Ken Howery: 1 link beyond PayPal; shares a company or fund with Nosek and Thiel; every other connected alumnus is at most 6 steps away.
Luke Nosek: 3 links beyond PayPal; shares a company or fund with Howery, Musk and Thiel; every other connected alumnus is at most 6 steps away.
Elon Musk: 2 links beyond PayPal; shares a company or fund with Nosek; every other connected alumnus is at most 7 steps away.
David Sacks: 3 links beyond PayPal; reaches the others through Founders Fund, which put the first money into his Geni; every other connected alumnus is at most 7 steps away.
Premal Shah: 1 link beyond PayPal; no company or fund on this map links him to another alumnus.
Jawed Karim: 1 link beyond PayPal; shares a company or fund with Botha, Chen, Hurley and Rabois; every other connected alumnus is at most 7 steps away.
Chad Hurley: 1 link beyond PayPal; shares a company or fund with Botha, Chen, Karim and Rabois; every other connected alumnus is at most 7 steps away.
Steve Chen: 1 link beyond PayPal; shares a company or fund with Botha, Hurley, Karim and Rabois; every other connected alumnus is at most 7 steps away.
Roelof Botha: 3 links beyond PayPal; shares a company or fund with Chen, Hurley, Karim and Rabois; every other connected alumnus is at most 6 steps away.
Keith Rabois: 5 links beyond PayPal; shares a company or fund with Botha, Chen, Hoffman, Hurley, Karim, Levchin and Thiel; every other connected alumnus is at most 5 steps away.
Reid Hoffman: 3 links beyond PayPal; shares a company or fund with Rabois and Thiel; every other connected alumnus is at most 4 steps away.
Jeremy Stoppelman: 1 link beyond PayPal; shares a company or fund with Levchin, Simmons and Thiel; every other connected alumnus is at most 6 steps away.
Russel Simmons: 1 link beyond PayPal; shares a company or fund with Levchin, Stoppelman and Thiel; every other connected alumnus is at most 6 steps away.
Max Levchin: 3 links beyond PayPal; shares a company or fund with Rabois, Simmons, Stoppelman and Thiel; every other connected alumnus is at most 4 steps away.
Andrew McCormack: 1 link beyond PayPal; shares a company or fund with Thiel; every other connected alumnus is at most 6 steps away.
PayPal links all 16 alumni. Take it away and 15 of them are still connected through the companies and funds they built or backed; only Premal Shah's Kiva stands apart.
Affirm is linked to Levchin and to Founders Fund. Hover or tap it for roles and years.
Craft is linked to Sacks. Hover or tap it for roles and years.
Facebook is linked to Hoffman and Thiel and to Founders Fund. Hover or tap it for roles and years.
Founders Fund is linked to Howery, Nosek and Thiel and to Affirm, Facebook, Geni, Palantir and SpaceX. Hover or tap it for roles and years.
Geni is linked to Sacks and to Founders Fund. Hover or tap it for roles and years.
Gigafund is linked to Nosek and to SpaceX. Hover or tap it for roles and years.
Greylock is linked to Hoffman. Hover or tap it for roles and years.
Kiva is linked to Shah. Hover or tap it for roles and years.
LinkedIn is linked to Hoffman, Rabois and Thiel and to Sequoia. Hover or tap it for roles and years.
Opendoor is linked to Rabois. Hover or tap it for roles and years.
Palantir is linked to Thiel and to Founders Fund. Hover or tap it for roles and years.
Sequoia is linked to Botha and to LinkedIn and YouTube. Hover or tap it for roles and years.
SpaceX is linked to Musk and Nosek and to Founders Fund and Gigafund. Hover or tap it for roles and years.
Square is linked to Botha and Rabois. Hover or tap it for roles and years.
Tesla is linked to Musk. Hover or tap it for roles and years.
Valar is linked to McCormack and Thiel. Hover or tap it for roles and years.
Yammer is linked to Sacks. Hover or tap it for roles and years.
Yelp is linked to Levchin, Simmons, Stoppelman and Thiel. Hover or tap it for roles and years.
YouTube is linked to Botha, Chen, Hurley, Karim and Rabois and to Sequoia. Hover or tap it for roles and years.
The parallel with Anthropic holds in shape, not in size. Both companies gathered people around a scarce skill, fraud control for online payments then, the training of frontier models now. PayPal had 618 full-time staff; Anthropic has 5,300, almost 9 times as many. Le Détroit's calculation: the $1.5bn sale is about 180 times smaller than the $270bn that Anthropic's employees alone may gain on Hill's estimate, and Hill's average of about $51m a head is some 21 times our even split. The timing also fits the three-year horizon we are testing: PayPal's alumni had their key companies and funds within three years of the sale, and most of Anthropic's lock-ups are due to end in 2027. The difference is depth. At PayPal a handful of rich alumni became the investors of the rest; at Anthropic, hundreds could start out with capital of their own.
Frontier AI already has its own lineage. All eight authors of the 2017 Google paper that introduced the transformer, the architecture behind today's large language models, had left the company by July 2023, Bloomberg reported, founding or joining start-ups such as Cohere and Character.AI. Arthur Mensch left Google DeepMind, and Guillaume Lample and Timothée Lacroix left Meta, to found Mistral AI in Paris in 2023. OpenAI's alumni form the densest network so far: TechCrunch counted 18 notable companies they founded, among them Safe Superintelligence, Thinking Machines Lab, Perplexity, Periodic Labs and Anthropic itself, started in 2021 by Dario and Daniela Amodei after they left OpenAI.
Choose a network; hover or tap a name
- Fairchild Semiconductor 1957 eight founders leave ShockleyFairchild Semiconductor, 1957Eight employees left Shockley Semiconductor Laboratory and founded it with money from Sherman Fairchild (Computer History Museum)
- Rheem Semiconductor 1959 the first spin-offRheem Semiconductor, 1959The first Fairchild spin-off (Computer History Museum)
- Intel 1968 Gordon Moore and Robert NoyceIntel, 1968Founded by Moore and Noyce; the other six Fairchild co-founders helped fund it (Rhett Morris, TechCrunch, 2014)
- AMD launched by a former employeeAMDOne Fairchild co-founder financed its launch by a former Fairchild employee (TechCrunch, 2014)
- Kleiner Perkins 1972 Eugene KleinerKleiner Perkins, 1972Co-founded by Fairchild co-founder Eugene Kleiner; has invested in hundreds of companies (TechCrunch, 2014; CHM)
- Google, Symantec among its investmentsKleiner Perkins portfolioNamed by Rhett Morris (Endeavor), TechCrunch, 26 Jul 2014
- Sequoia Capital 1972 Don Valentine, ex-FairchildSequoia Capital, 1972Started by former Fairchild executive Don Valentine (TechCrunch, 2014; CHM)
- Cisco, LinkedIn among its investmentsSequoia portfolioNamed by Rhett Morris (Endeavor), TechCrunch, 26 Jul 2014
- Google and DeepMind the transformer and its authorsGoogleAll eight authors of the 2017 paper "Attention Is All You Need" had left Google by July 2023 (Bloomberg via The Star)
- The transformer's authors 2017 all eight gone by 2023The transformer paperCo-authors founded or joined start-ups including Cohere and Character.AI; Llion Jones, the last to leave, planned his own company (Bloomberg, 11 Jul 2023)
- Cohere models for businessCohereSells large language models to businesses (Bloomberg via The Star)
- Character.AI chatbotsCharacter.AILets users create chatbots that imitate celebrities and historical figures (Bloomberg via The Star)
- Arthur Mensch from Google DeepMindArthur MenschPreviously at Google DeepMind; co-founded Mistral with two former Meta researchers (Le Monde via Wikipedia; The Big Byte)
- Mistral AI 2023 Paris; valued at €21bn in Sept 2026Mistral AIFounded in April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix (Wikipedia, as a lead)Valued at €21bn after a €3bn round led by Samsung in September 2026 (Euronews)
- Behnam Neyshabur Google, then AnthropicBehnam NeyshaburMore than five years at Google DeepMind, then Anthropic from late 2024 to December 2025 (The Decoder; The Next Web)
- Mirendil 2026 AI for researchMirendil$200m seed at a $1bn valuation, announced 24 June 2026 (The Next Web)
- Georges Harik Google's seventh employeeGeorges HarikHelped build Google's first advertising systems (TechCrunch, 20 Jan 2026)
- Humans& 2026 with an ex-Anthropic researcherHumans&$480m seed at $4.48bn; co-founders include Andi Peng, formerly of Anthropic (TechCrunch, 20 Jan 2026)
- The transformer's authors 2017 all eight gone by 2023The transformer paperCo-authors founded or joined start-ups including Cohere and Character.AI; Llion Jones, the last to leave, planned his own company (Bloomberg, 11 Jul 2023)
- OpenAI 18 notable alumni companiesOpenAI's alumniTechCrunch counted 18 notable start-ups founded by OpenAI alumni (20 Feb 2026). A selection, not a census
- Anthropic 2021 Dario and Daniela AmodeiAnthropic, 2021Founded after the Amodeis left OpenAI (TechCrunch, 20 Feb 2026)Valued at $965bn in May 2026; IPO range $70 to $78 a share
- Mirendil 2026 Neyshabur, Mehta: $200m seedMirendilFounders left Anthropic in December 2025; $200m seed at $1bn, a16z and Kleiner Perkins, Nvidia joining (The Next Web, 24 Jun 2026)Targets include biology and materials science (The Decoder)
- Humans& 2026 Andi Peng: $480m seedHumans&Peng worked on reinforcement learning and post-training for Claude 3.5 to 4.5 (TechCrunch, 20 Jan 2026)
- Safe Superintelligence Ilya Sutskever; $32bnSafe SuperintelligenceRaised $2bn; reported valuation $32bn; no product yet (TechCrunch, 20 Feb 2026)
- Thinking Machines Lab Mira Murati; $12bnThinking Machines LabValued at $12bn (TechCrunch, 20 Feb 2026)
- Periodic Labs Liam Fedus; materials, $300m seedPeriodic LabsAI scientists for materials discovery; first goal superconductors (TechCrunch, 20 Oct 2025)
- Perplexity Aravind Srinivas; searchPerplexityReported raise at a $20bn valuation (TechCrunch, 20 Feb 2026)
- Covariant robot foundation modelsCovariantFounded by Pieter Abbeel, Peter Chen and Rocky Duan; Amazon hired the founders in 2024 (TechCrunch)
- Daedalus precision-component factoriesDaedalusFounded by Jonas Schneider; $21m Series A (TechCrunch, 20 Feb 2026)
- Living Carbon engineered plantsLiving CarbonFounded by Maddie Hall; carbon-capturing plants (TechCrunch, 20 Feb 2026)
- Anthropic 2021 Dario and Daniela AmodeiAnthropic, 2021Founded after the Amodeis left OpenAI (TechCrunch, 20 Feb 2026)Valued at $965bn in May 2026; IPO range $70 to $78 a share
Part 5What the evidence says
The pattern is real, but the famous cases are survivors, so the better evidence is statistical. Using US Census records on about 3,400 companies that filed for an IPO between 1992 and 2006, Tania Babina, Paige Ouimet and Rebecca Zarutskie found that 2.3% of employees at a filing company leave within three years to become one of the five best-paid people at a start-up no more than three years old, their proxy for founders and key early staff. Firms that completed the listing saw a sharp rise in that rate (the authors put the jump at "over five percentage points" in one passage and "5.3%" in another), concentrated among high earners, younger workers and high-tech firms. The authors link it to wealth: a windfall makes it easier to bear a start-up's risk or to fund one.
Applied to Anthropic, the arithmetic is an illustration, not a forecast. Le Détroit's calculation: at the study's average rate, about 120 of today's 5,300 staff would hold senior start-up roles within three years. Reading the listing effect as five extra percentage points, and assuming it applies in full to a frontier AI lab, lifts that to about 400; since the average already includes firms that listed, the upper figure is generous. Even the lower one could produce dozens of companies. Whether they would be labs is another matter.
Hover or tap a company or a row
Models and labs5
AnthropicFrontier models; founded 2021AnthropicFrontier models; founded 2021TechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitSafe SuperintelligenceSuperintelligence researchSafe SuperintelligenceSuperintelligence researchTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitThinking Machines LabCustomisable AIThinking Machines LabCustomisable AITechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitxAIGrok; acquired by SpaceXxAIGrok; acquired by SpaceXTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitStem AI (Softmax)Research-focused AIStem AI (Softmax)Research-focused AITechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitSoftware8
PerplexityAI searchPerplexityAI searchTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitAdeptAI tools for workAdeptAI tools for workTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitApplied ComputeEnterprise agentsApplied ComputeEnterprise agentsTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitCrestaAI contact centresCrestaAI contact centresTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitKindoEnterprise chatbotKindoEnterprise chatbotTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitWorktrace AIWorkflow automationWorktrace AIWorkflow automationTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitPilotAccountingPilotAccountingTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitEureka LabsAI teaching assistantsEureka LabsAI teaching assistantsTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitPhysical world5
Periodic LabsMaterials discoveryPeriodic LabsMaterials discoveryTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitCovariantRobot foundation modelsCovariantRobot foundation modelsTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitProsper RoboticsHome robots, LondonProsper RoboticsHome robots, LondonTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitDaedalusPrecision-component factoriesDaedalusPrecision-component factoriesTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitLiving CarbonEngineered plantsLiving CarbonEngineered plantsTechCrunch, The OpenAI mafia: 18 start-ups founded by alumni (20 Feb 2026); sector grouping by Le DétroitLe Détroit's grouping of TechCrunch's 18. "Physical world": robots, factories, materials and plants.
One dot per company. Counts come from different sources and methods: a census for Fairchild, notable companies for OpenAI. Read the shape, not the ranking.
The counter-evidence is just as concrete. Money also goes into houses: across 720 California IPOs from 1993 to 2017, home prices within a mile of the company rose by an average of 3.3% between the filing and the first trade, before any cash arrived, according to Barney Hartman-Glaser, Mark Thibodeau and Jiro Yoshida. Alumni networks also vary hugely in size. BlaBlaCar, France's best-documented "mafia", counted 36 start-ups founded by former staff that had raised more than €115m by 2022, according to its chief executive. And frontier skills transfer unevenly to the physical world: of the 18 OpenAI alumni companies on TechCrunch's list, Le Détroit counts 5 working on robots, factories, materials or plants, against 8 building software and 5 building models.
Periodic Labs shows both the promise and the friction. Founded by Liam Fedus, who led OpenAI's post-training team, and Ekin Dogus Cubuk, a former Google Brain materials researcher, it raised a $300m seed round to automate materials discovery with simulations, robotic labs and language models, starting with superconductors. When TechCrunch reported on it in October 2025, its robots were not yet operational. Mr Fedus put the challenge plainly: "everyone talks about doing science, but in order to do science, you actually have to do science."
Back to contentsPart 6Where the alumni might go
Anthropic's diaspora did not wait for the listing, which suggests that its fuel is reputation more than personal wealth. Behnam Neyshabur and Harsh Mehta joined Anthropic in late 2024, left in December 2025 and in June announced a $200m seed round at a $1bn valuation for Mirendil, co-led by Andreessen Horowitz and Kleiner Perkins, The Next Web reported; the company wants to automate the work of AI researchers, with biology and materials science among its targets, according to The Decoder. In January Humans&, co-founded by Andi Peng, who worked on reinforcement learning and post-training for Claude models, raised $480m at a $4.48bn valuation, TechCrunch reported.
Those numbers change the logic of the "mafia". At PayPal, Mr Thiel's personal cheque mattered. Today a year at a frontier lab can unlock a seed round larger than most employees' paper fortunes, so the windfall matters less as start-up capital than as insurance: it lets people take the risk, and it turns the richest of them into angel investors. That points to a first wave close to Anthropic's own craft, AI that does research, before the harder bets on robots and wet labs.
Anthropic's culture adds a third destination: giving. The company matches employee donations one for one, up to 25% of their equity grants, through a donor-advised fund run by Daffy, HR Brew reported. About 934m shares were reserved for the programme at the end of September, according to the prospectus as reported by CNBC-TV18, which added that just over half of employees had pledged some equity. The matching cost more than $660m in non-cash expense between October 2025 and March 2026, The Information reported. The seven co-founders have pledged 80% of their equity, in what the filing calls "voluntary, revocable expressions of each cofounder's philanthropic intent". Le Détroit's calculation: at the $74 midpoint those pledges would be worth about $100bn, and the separate 5% Public Benefit Fund about $63bn.
Dario Amodei, the chief executive, has framed the stakes himself. "The thing to worry about is a level of wealth concentration that will break society," he wrote in an essay in January. Fortune wrote in August that he had "reportedly expressed concerns" about how the company's financial incentives could affect its ability to keep staff committed to its mission, and, citing Axios, that some candidates were asked how they would feel if the stock went to zero. A culture built on mission could slow departures as much as a windfall speeds them.
Part 7The clock
A three-year horizon fits the calendar of the money better than the calendar of the science. On the reported lock-up terms, most employee stock should be saleable by about May 2027, 180 days after the expected first trade, and the co-founders' stock by about November 2027. A founder who leaves then and raises a seed round within months would run a company about two years old by November 2029.
Two forces pull the other way. Unvested equity keeps people: Anthropic reported about $32.3bn of unrecognised stock-based compensation for restricted stock units, of which $8.2bn relates to awards where employees have met some or all of the service conditions, according to the filing as reported by CNBC-TV18. Le Détroit's calculation: roughly three quarters of that pending compensation is tied to awards on which staff have yet to meet any service condition, so it still has to be earned by staying. And mafias mature over years, not quarters: Intel came eleven years after Fairchild, and the transformer's authors took six years to scatter.
Hover or tap a point for the event and its source
The early signals will be easy to read: seed rounds led by former Anthropic researchers in the first half of 2027, angel investments by early employees once the first windows open, and the first ex-Anthropic company outside software. A Periodic Labs of Anthropic's own would be the clearest sign that the hypothesis is running ahead of schedule.
Back to contentsPart 8The French question
Our starting hypothesis for France was bolder still: a single French AI or deeptech company whose listing turned thousands of engineers into millionaires would do more for the economy, innovation and jobs than any presidential election or government reform, as alumni reinvest, back overlooked talent and set off a chain of hiring, research and industrial sovereignty. The evidence supports the mechanism. It is weaker on the comparison.
France has the talent and is starting to have the companies. Mistral AI, founded in Paris in 2023, was valued at €21bn after a €3bn round led by Samsung in September, Euronews reported, and has more than a thousand staff, according to Maddyness. But the listings that create liquidity are rare. Euronext Paris recorded 32 IPOs in 2021 and just two in 2025, the fewest in ten years, according to EY. Criteo, the Paris ad-tech company, chose Nasdaq in 2013 and raised $251m. In 2025 private AI investment across Europe was $20.9bn, against $285.9bn in the United States, according to Stanford's AI Index; Le Détroit's calculation is that Anthropic's IPO and private placement together could raise more than four times Europe's annual total.
Choose a view; hover or tap a bar
Who owns the upside
Ownership bars on a 0 to 25% scale; tax bars on a 0 to 50% scale.
France's rules on employee equity are not the obstacle. Index Ventures, a venture firm, judged the French BSPCE better than what is available in the US in a handbook first published in 2017. For shares subscribed since 2025, the gain on exercise is taxed at a flat 12.8% for holders with three years at the company and at 30% otherwise, plus social levies now at 18.6%, according to the tax administration; the 2026 finance law extended grants to staff of sub-subsidiaries from January. The gap is in how much employees get: staff own about 10% of late-stage European start-ups, against 20% in the US, and about two-thirds of European options go to executives, Index found, in a handbook whose data date from about 2018. Doctolib, one of the employers French tech workers most want to join, has organised a secondary sale of just under €300m of shares, Maddyness reported.
The objections to the comparison are serious. Policy sets the conditions that a liquidity event needs: the BSPCE regime was amended by two finance laws in two years, and research funding, energy prices, pension savings and labour law are decided by elected governments. Windfalls also concentrate: the California housing evidence suggests the money lands first near the company, which in France would mean Paris. Clément Delangue, Hugging Face's chief executive, reacting to the Anthropic estimate, warned that the concentration of power, capabilities and wealth is the biggest risk in AI. A French "mafia" would need repeated exits, not one.
Back to contentsPart 9Le Détroit's reading
Our reading is that an "Anthropic mafia" is likely, is already under way and will look different from the popular sketch in three ways. It will start in AI itself: on the evidence of OpenAI's alumni and Anthropic's first spinouts, most new companies will build models, research tools and software, with robotics, materials and biology a meaningful minority, perhaps a quarter. It will be paced by lock-ups and vesting more than by ambition: on US evidence from earlier listings, something like 120 alumni in senior start-up roles by the end of 2029, and a few hundred if the listing effect holds in full, enough for dozens of companies but only a handful of frontier-scale labs. And more of its money will go to philanthropy than PayPal's did, because the founders have pledged most of theirs and the company matches its staff's gifts.
What would make it fail? A share price that halves after listing, as Facebook's did, would shrink the cushion that makes risk-taking possible. A frontier race that keeps paying researchers more to stay than any start-up can offer would slow departures. And a mission culture strong enough to hold people inside, which Anthropic is deliberately building, would delay the diaspora even as it shapes where the money eventually goes.
For Europe and France the lesson is narrower than our starting hypothesis, but still sharp. The PayPal and Anthropic stories are about liquidity meeting talent in one place. France now has the talent and some of the companies, but little of the liquidity: a shallow domestic IPO market, smaller employee stakes and few exits large enough to free hundreds of people at once. A Mistral listing that made a few hundred engineers rich would probably do more for French innovation than most single reforms. But whether it happens in Paris, at what tax rate and with whose savings is decided at the ballot box. The choice between a great listing and good policy is a false one: the mafia is the harvest, and policy is the soil.
Back to contentsWhat to watch
- 17 to 19 November: books close, pricing and first trade (expected). The order book was due to close on 17 November (The Next Web). Whether the shares price within, above or below $70 to $78 decides every number in this article; at $1tn fully diluted, each paper fortune would be a third smaller than in Hill's estimate.
- Mid-December 2026 to early February 2027: the first unlocks. Other stockholders get about a quarter of their shares in four steps on days 27, 48, 62 and 76 after the prospectus, then 12.5% after results.
- Early 2027: full-year 2026 results. They release a further 35% for directors, officers and staff, and will show whether the $81bn run-rate held.
- By about 18 May and 19 November 2027: the big releases. The 180-day and 365-day lines (or earlier results) free most staff and then the co-founders. Watch for ex-Anthropic seed rounds and angel deals in between.
- France: the 2027 budget and Mistral. Any change to BSPCE taxation or capital-income levies in the 2027 finance bill, and whether Mistral's next step is a round or a listing, and where.
Wealth figures for Anthropic employees are estimates of paper value before tax, reported by others; Le Détroit has not seen Hill's data. Figures labelled as Le Détroit's calculations are ours. "Mafia" is used here only as the established nickname for alumni networks, after the "PayPal Mafia", and implies no wrongdoing by anyone. Forward-looking items (pricing, trading date, valuation, lock-up dates) are as reported up to 17 November 2026 and may change. Le Détroit has no commercial relationship with Anthropic.

