Prometheus: A $41B Mark, $18.2B Raised, and Not Yet a Product

Executive Summary

Prometheus is a physical-AI company founded by Jeff Bezos and Vik Bajaj. It builds what Bezos calls an Artificial General Engineer (AGE): a set of AI tools intended to compress the cycle from product idea to manufacturable physical object. As of October 21, 2026 it has raised approximately $18.2 billion in primary capital at a $41 billion post-money valuation, employs roughly 150 people, and has not shipped a product.

Three facts frame the investment case, and they point in different directions.

First, the talent is unusually strong and unusually specific. The founding team includes Kamyar Azizzadenesheli, a co-developer of the Fourier Neural Operator and lead author of FourCastNet, who lists himself as having initiated the founding team in spring 2025. The founding advisers include Ashish Vaswani and Jakob Uszkoreit, the co-authors of "Attention Is All You Need." The acquired company General Agents brought Sherjil Ozair and William Guss, the creators of the Ace computer agent. For a company whose thesis is that physical simulation needs a different architecture class than language modelling, hiring the people who built that architecture class is a meaningful, verifiable signal about intent.

Second, the capital position is the strongest in the category. The $12 billion Series B at a $41 billion post-money valuation closed on June 11, 2026, and is confirmed by an unrelated third party: Powerlaw Corp's SEC Schedule of Investments as of June 30, 2026 lists both the $12 billion round and the $41 billion post-money figure. An EquiDeFi special purpose vehicle announced on August 10, 2026 to invest indirectly in the same Series B preferred shares states $29 billion pre-money and $41 billion post-money, and opened subscription on August 12 with a $10,000 minimum. On the arithmetic, the launch round's roughly $30 billion pre-money and the Series B's $29 billion pre-money are nearly identical marks on the same asset eight months apart, which tells you the Series B was priced close to the launch round rather than at a step-change premium.

Third, there is no product, no customer and no revenue. The company website at prometheus.ai is a splash page containing only the word "Prometheus," a privacy policy link, the line "Copyright 2026 Prometheus Industries, Inc." and a trademark notice. There is no API, no documentation, no demo, no pricing, no published benchmark and no named paying customer. On CNBC on June 11, 2026 Bezos said only that "we will have early product rollouts" and that "it's too early to lay out a timetable for that."

Our view: a $41 billion mark on an unbuilt product, funded by an industrial roll-up plan that is itself unbuilt. The thesis is coherent and unusually legible: physical AI's bottleneck is not compute or architecture but access to high-fidelity manufacturing and experimental data that cannot be scraped from the internet, so the founders intend to manufacture that data the way they intend to manufacture physical goods, by owning industrial assets. Bezos confirmed the intent directly on CNBC ("we may buy parts of companies and so on who could benefit from this technology ... that plan is still in work") but confirmed no vehicle, no size, no close and no target. An investor buying Prometheus today is underwriting a $18.2 billion option on a data strategy whose acquisition mechanism has not been funded and whose delivery mechanism has not been shipped.

Evidence Tiers

Every number below carries a source, a date and a tier.

• OFFICIAL DISCLOSURE - the company's own statements to a named reporter, its own website, a government commercial register, or a third party's regulatory filing describing its own investment in the company.
• MEDIA REPORT - Reuters, Bloomberg, Financial Times, Wall Street Journal, New York Times, Axios, CNBC, TechCrunch, GeekWire, WIRED, Forbes, Fox News, plus Chinese-language outlets and commercial databases, all of which are labelled as reporting rather than disclosure.
• UNVERIFIED - asserted somewhere but not supported by a primary source or by two independent ones. Such claims are recorded and labelled rather than dropped or silently converted into fact.

Section 11 lists every conflict we found between sources, including the ones that undercut the headline narrative. We have not reconciled them away.

1. Founding, Legal Structure and Governance

1.1 The founding timeline is a stack of different dates, not a single date

• Late 2024 - preparation. Bezos said on CNBC on June 12, 2026: "I started out in late '24 with Vik as an investor in Prometheus ... and as I saw what we were doing, I became so impressed by what was happening and the potential, that I decided I couldn't sit on the sidelines and I needed to jump in with both feet." (OFFICIAL DISCLOSURE, CNBC, 2026-06-12)
• Early June 2025 - Bajaj approaches third parties. Reuters documents that over dinner at a restaurant in greater Los Angeles in late 2024, Bajaj discussed a collaboration with Anima Anandkumar and Benedikt Jenik, and later discussed an offer letter titled "Project Prometheus" with them. (MEDIA REPORT, Reuters, 2026-08-25)
• Spring 2025 - founding team formed. Kamyar Azizzadenesheli's own page states: "Kamyar initiated the founding team of Prometheus, starting Spring 2025." (OFFICIAL DISCLOSURE, personal academic page, accessed 2026-10-21)
• 2025-06-02 - Swiss legal entity registered. Prometheus Industries GmbH, UID CHE-308.691.679, registry number CH-020.4.088.310-3, registered in Switzerland on June 2, 2025. The entity was originally named FL2024-008 Schweiz GmbH. (OFFICIAL DISCLOSURE, Swiss commercial register / FOSC)
• 2025-11-17 - public reveal. The New York Times published "Jeff Bezos Creates A.I. Start-Up Where He Will Be Co-Chief Executive," naming Bezos and Bajaj as co-chief executives and reporting $6.2 billion of launch funding. (MEDIA REPORT, New York Times, 2025-11-17)
• 2026-06-11 - first public company statements. Bezos and Bajaj jointly interviewed by CNBC's David Faber from Prometheus headquarters in San Francisco, publicly announcing the $12 billion Series B. (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)
• 2026-09-24 - Swiss entity renamed. FL2024-008 Schweiz GmbH becomes Prometheus Industries GmbH, with a shareholder change recorded. The US parent Prometheus Industries, Inc. holds 200 shares at CHF 100 each against registered capital of CHF 20,000. (OFFICIAL DISCLOSURE, Swiss commercial register / FOSC)

Two dates in this list are commonly reported as "the founding date" and neither is clean. The Swiss entity was originally named FL2024-008 Schweiz GmbH, an identifier that maps to Foresite Labs, Bajaj's venture studio, which suggests the Swiss entity began life as a Foresite Labs subsidiary rather than as a purpose-built AI company. Separately, WIRED-derived reporting says Prometheus was "operating in stealth mode since early 2024," which conflicts with Bezos's own account of joining as a founding investor in late 2024 and with Azizzadenesheli's spring 2025 founding-team date. The US parent's Certificate of Formation date remains UNVERIFIED: a name search on EDGAR did not locate a Form D filing for Prometheus Industries, Inc. as of the research date. That is a statement about the search, not a claim that no filing exists.

1.2 Entity structure

• US parent: Prometheus Industries, Inc. Named as the copyright holder on the company website and confirmed by the Swiss register as the shareholder of the GmbH. Incorporation date UNVERIFIED.
• Swiss subsidiary: Prometheus Industries GmbH, a GmbH under Swiss law, registered June 2, 2025, renamed September 24, 2026, registered address Zollikerstrasse 226. Stated commercial purpose covers software development, data analysis, machine learning, AI and related research. (OFFICIAL DISCLOSURE, Swiss commercial register)
• The $100B acquisition vehicle has no identified legal vehicle. See section 4.

1.3 Vik Bajaj, in full

Bajaj is a Canadian-American scientist and serial founder. The full record matters because his background is the strongest single indicator of what Prometheus intends to build.

• Education. Combined BA/MS in biochemistry, University of Pennsylvania. PhD in physical chemistry, MIT. Postdoctoral research, UC Berkeley.
• Academic and laboratory. Research scientist and principal investigator, Lawrence Berkeley National Laboratory from 2011, seven years; also a principal investigator at UC Berkeley. Anatole Abragam Prize (2012 per his company biography; some profiles date it 2013). R&D 100 Awards in 2011 and 2013. DOE LBL Innovation Grant, 2013.
• Stanford. Adjunct associate professor, Stanford University School of Medicine, from September 2014, in the Department of Radiology's Molecular Imaging Program. CNBC describes him as a professor at Stanford Medicine and as an adjunct professor of radiology.
• Google X and Verily. Co-founded Google Life Sciences in March 2013 as chief scientific officer, alongside a director role at Google X; reported by the New York Times to have worked closely with Sergey Brin on Wing, the program that became Waymo, and Project Baseline. Google Life Sciences was rebranded Verily in 2015; he was founding chief scientific officer until October 2016 and chaired Verily's scientific advisory board. (MEDIA REPORT, New York Times and Wikipedia-sourced profiles)
• Grail. Chief scientific officer, 2016 to 2017; remained on the scientific advisory board until 2021, when Illumina acquired Grail for $8 billion.
• Foresite. Managing director at Foresite Capital from October 2017, focused on companies that personalize medicine to individual genetic code. Co-founder and CEO of Foresite Labs, an AI-biotech venture studio, from 2018. His professional profile reportedly records the Foresite Capital role as running through March 2026, which is consistent with a full-time commitment to Prometheus from early 2026. (MEDIA REPORT)
• Xaira Therapeutics. Co-founder, director and interim president from January 2023; San Francisco AI drug discovery company.
• Boards and public service. Genomics England board from 2021; Quantum-Si board from June 2022; advisor to the US Department of Defense Defense Science Board's Task Force on Biology as of 2022; UC Berkeley College of Chemistry advisory board 2015 to 2021; affiliate scientist, LBNL and UC Berkeley.
• At Prometheus. Co-founder and co-CEO from November 2025. CEO of Foresite Labs concurrently.
• How Bezos and Bajaj connected. Through Bezos's investments in biotech companies Bajaj helped start or run, including Grail and Xaira Therapeutics. (MEDIA REPORT, WIRED, 2025-11-26)

1.4 Is Bezos co-CEO? Yes, and the operating model is unusual

Bezos is confirmed as co-founder and co-CEO by the New York Times (2025-11-17), CNBC (2026-06-11), Fox News (2026-10-07, which identifies him as "co-founder and co-CEO of Prometheus") and GeekWire (2026-06-11). It is his first operating CEO role since leaving Amazon in 2021. CNBC asked specifically about the division of labour, and Bezos answered:

"Bezos and Bajaj ... 'don't really divide responsibilities' as co-CEOs ... We're talking to each other multiple times a day ... We're both involved in all of the decisions really. It's not like it's, 'OK. You handle this half of the company and I'll handle this half.' We're tight at the hip. We're both jacks of all trades."

On time allocation he said "Prometheus is the bulk of my time. I'm also spending a lot of time on Blue Origin. I'm spending a lot of time on AI at Amazon." He called the role "Type 2 fun" and "a grind, but it's a good grind." (OFFICIAL DISCLOSURE, CNBC, 2026-06-12)

1.5 What Dave Limp's board role entails

Dave Limp, chief executive of Blue Origin, joined the Prometheus board. The Wall Street Journal reported this on March 19, 2026, in its coverage of the $100 billion fund. (MEDIA REPORT)

What is not disclosed: committee assignments, whether he chairs anything, whether he holds an investment-committee seat, the scope of his authority, or whether his role carries any day-to-day operating responsibility. Any claim that Limp manages Prometheus's daily operations is UNVERIFIED and contradicts Bezos's description of the co-CEO arrangement. The defensible statement is narrow: Limp is a director of a company whose adjacent aerospace business is simultaneously described by Bezos as a case study for a potential Prometheus customer.

1.6 Founding advisers and the founding team

• Ashish Vaswani and Jakob Uszkoreit, co-authors of "Attention Is All You Need" (2017), are founding advisers to Prometheus while running their own startups. (MEDIA REPORT, WIRED and Forbes, 2025-11 and 2026-03)
• Kamyar Azizzadenesheli is a founding team member. He co-developed the Fourier Neural Operator, led development of FourCastNet for high-resolution global weather modelling, co-authored Physics-Informed Neural Operator (PINO) and U-NO, created signSGD, and built Neural-Fly for adaptive drone control. He was a senior research scientist at NVIDIA from June 2022 to March 2025, an assistant professor of computer science at Purdue from 2020 to 2022, a postdoctoral scholar at Caltech, and did his PhD at UC Irvine from 2015 to 2020 in Anima Anandkumar's MEGA Data Lab. (OFFICIAL DISCLOSURE, his own page and NVIDIA Research profile)
• Sherjil Ozair (formerly DeepMind and Tesla) and William Guss (formerly OpenAI) joined via the General Agents acquisition, as did the leader of General Agents' India training team. (MEDIA REPORT, WIRED, 2025-11-26)
• Recruiting profile. The company has recruited from OpenAI, Google DeepMind, Meta, NVIDIA and xAI. (MEDIA REPORT)

2. Offices and Headcount

2.1 Locations

• Headquarters: San Francisco. CNBC conducted its interview "from Prometheus headquarters in San Francisco"; GeekWire describes the company as "based in San Francisco with teams in London and Zurich." (OFFICIAL DISCLOSURE via CNBC, 2026-06-11)
• San Francisco office: 101 Mission Street, approximately 30,000 square feet. (MEDIA REPORT)
• Alternate San Francisco corporate address: 2261 Market Street, Suite 710, listed by a commercial database. (MEDIA REPORT, database)
• West Oakland: 1960 Mandela Parkway, approximately 100,000 square feet, with roughly 8,000 amps of power. (MEDIA REPORT)
• London and Zurich. Both are confirmed operating sites. Bajaj said on CNBC "we have 150 people in here in London and Zurich, and those teams continue to grow." (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)
• Swiss registered address: Zollikerstrasse 226. (OFFICIAL DISCLOSURE, Swiss commercial register)

2.2 Headcount over time

• November 2025: close to 100 employees. (MEDIA REPORT, New York Times, 2025-11-17; Crunchbase)
• December 2025: approximately 120. (MEDIA REPORT)
• March 2026: approximately 120. (MEDIA REPORT)
• April 2026: approximately 120. (MEDIA REPORT)
• June 11, 2026: approximately 150. Bezos confirmed "150 now" when Faber asked directly. (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)
• August 2026: some reporting still says "120+". (MEDIA REPORT)
• LinkedIn: company size band 51-200; Founded 2025; tagline "AI for the physical economy". A search-index snapshot showed 191 associated employees and a direct fetch at the same research date showed 209. (MEDIA REPORT, LinkedIn)

Both LinkedIn numbers are platform-associated-account counts, not audited payroll headcount, and they differ from each other between snapshots. A Chinese-language investment outlet reported "several hundred employees" in August 2026; that figure is inconsistent with every other source and we treat it as an overstatement.

The physical footprint is worth reading against the headcount: roughly 130,000 square feet leased across San Francisco and West Oakland, against about 150 people. That is a real-estate commitment roughly 870 square feet per employee, which suggests the company is building for a workforce it has not yet hired rather than for its current one.

3. The Funding Ladder

3.1 Round 1: the launch financing, November 2025

• Amount: $6.2 billion. (MEDIA REPORT, New York Times 2025-11-17; Financial Times 2026-02-26/27; Axios; GeekWire; Crunchbase's November 2025 roundup of December 2, 2025 identifies it as the largest round of that month, in a month when 43% of all global venture funding went to just 14 companies)
• Valuation: approximately $30 billion, and the pre/post distinction is the crux. The Financial Times reported $6.2 billion raised at "about $30 billion", explicitly noting the valuation does not include the new investment. Silicon UK restated the figure as "$30 billion pre-money." Read that way, the implied post-money is approximately $36.2 billion, which is our arithmetic and not a company-stated figure. Several commercial databases label the same ~$30 billion as post-money or as "current valuation"; that reading is inconsistent with the FT's exclusion note and should not be adopted. (MEDIA REPORT, Financial Times, 2026-02-26; the $36.2 billion post-money is derived)
• Round name: inconsistent across sources. The same $6.2 billion is called launch funding, seed, and Series A. Axios and GeekWire call it a Series A; the New York Times calls it launch funding; Crunchbase files it as a Series A. (MEDIA REPORT, conflicting)
• Bezos was the largest backer of the $6.2 billion Series A. (MEDIA REPORT, Axios, 2026-06-11)
• Bezos's own stake is not disclosed. Forbes has estimated roughly 20% and roughly $8.2 billion of value; that is a media estimate, not a company disclosure. (MEDIA REPORT, Forbes)

3.2 Round 2: the Series B, April and June 2026

• April 2026 snapshot: approximately $10 billion at approximately $38 billion post-money. Bloomberg reported the round had closed, with JPMorgan Chase and BlackRock among the participants and no lead investor. Business Insider, reporting while the round was still in progress, gave roughly $10 billion at a post-money valuation of roughly $38 billion and called it the company's first raise since the launch round. (MEDIA REPORT, Bloomberg and Business Insider, April 2026)
• June 11, 2026 final: $12 billion at a $41 billion valuation. Announced by the company and confirmed by CNBC, Axios, GeekWire, TechCrunch, and by Powerlaw Corp's own Schedule of Investments as of June 30, 2026, which lists both the $12 billion Series B and the $41 billion post-money figure. (OFFICIAL DISCLOSURE for Powerlaw's own filing; MEDIA REPORT for the announcements)
• Pre-money: $29 billion, post-money: $41 billion. Stated directly in the EquiDeFi SPV announcement of August 10, 2026, which offers indirect exposure to the Series B preferred shares and states the $29 billion pre-money and $41 billion post-money explicitly. (OFFICIAL DISCLOSURE, EquiDeFi, 2026-08-10)
• Investors: Bezos, JPMorgan Chase, BlackRock, Goldman Sachs, DST Global, Arch Venture Partners. (MEDIA REPORT, Axios, GeekWire, TechCrunch, 2026-06-11) On CNBC, Bezos confirmed he participated in the new round as well as in the launch round.
• Lead investor: UNVERIFIED. Bloomberg explicitly said the April round had no lead. Some outlets describe JPMorgan and BlackRock as co-leads of the June round; no announcement from Prometheus or from any investor naming a lead was located. A commercial database lists the lead as Bezos, which conflicts with the no-lead reporting. Recorded as a discrepancy in section 11 rather than resolved.

3.3 Were $10 billion and $12 billion separate rounds? No - one round, upsized

This is the single most consequential question in the funding record, and the arithmetic settles it.

• CNBC's own wording was that the company was "raising its Series B round to $12 billion" - an upsize, not a new round.
• TechCrunch described the June event as the company's second fundraising, following the $6.2 billion launch raise. A $12 billion round in June would be the third only if the April $10 billion were a separate round; TechCrunch's count and CNBC's upsize language agree that it was not.
• Arithmetic: $38 billion minus $10 billion leaves a $28 billion pre-money. $41 billion minus $12 billion leaves a $29 billion pre-money. The pre-money moved by about $1 billion, not by $13 billion. Independent rounds eight weeks apart on the same company at the same stage do not price at a $1 billion pre-money difference.
• If $10 billion and $12 billion were genuinely two separate rounds, the company would have raised $28.2 billion in Series B alone, which no source claims.

Conclusion: $10 billion at $38 billion (April) and $12 billion at $41 billion (June) are two reporting snapshots of the same Series B, which was upsized between them. The April figure is not retracted; both numbers were reported.

3.4 Cumulative capital: $18.2 billion

• $18.2 billion is the best-supported figure. $6.2 billion plus $12 billion. This is the number carried in Powerlaw Corp's Schedule of Investments as of June 30, 2026, describing its own holdings, and in the EquiDeFi SPV documentation. (OFFICIAL DISCLOSURE, Powerlaw SEC filing and EquiDeFi, 2026-08)
• $16.2 billion is the April 2026 interim cumulative figure ($6.2 billion plus $10 billion), reported at the time the round was smaller. (MEDIA REPORT)
• $28.2 billion appears in several commercial databases that file $10 billion and $12 billion as two separate rounds. It double-counts the same capital. (MEDIA REPORT, and it conflicts with the SEC-schedule-backed figure)

3.5 Use of proceeds

Bezos was explicit that compute dominates. On the CNBC live blog: "The latest funding round will help Prometheus acquire more compute ... That is a big chunk of the funding we've raised. And one of the reasons we've had to raise a significant amount of funding is because ... what we're doing is very compute intensive and we need to, you know, create that data." On capital intensity: "there's a lot of investment required in a company, because you do have to build these big datasets, which is expensive, and you do have to do a lot of compute, which is expensive. So this is a capital-intensive start-up. There's no question about that." (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)

On an IPO: "You know, I guess it's just too early to think about that." (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)

On the size of the company itself: when asked how large it might become, Bezos said "No, I don't think - it'll be more - that's more of an output ... We have certain needs on the data side, on the engineering side, on the research side, and so on we'll keep growing the company as we kind of evolve into that." (OFFICIAL DISCLOSURE, CNBC transcript, 2026-06-11)

3.6 The EquiDeFi SPV is not a new round

EquiDeFi announced on August 10, 2026 a SPV offering indirect exposure to Prometheus Series B preferred shares, opening subscription on August 12, 2026 with a $10,000 minimum subscription. It is a structured investment channel into an already-closed round. It does not add to Prometheus's primary capital and should not be counted as a third financing event. (OFFICIAL DISCLOSURE, EquiDeFi, 2026-08-10)

4. The $100 Billion Industrial Acquisition Vehicle

This is the part of the story that most often gets reported as settled. It is not.

4.1 What is reported

• The Wall Street Journal reported on March 19, 2026 that Bezos is in early discussions to raise up to $100 billion for a new fund that would acquire manufacturing companies and use AI to accelerate their automation. WSJ reported that Bezos had met with some of the world's largest asset managers, had travelled to the Middle East to discuss the fund with sovereign wealth representatives months earlier, and had recently gone to Singapore to raise money. In investor documents the vehicle is described as a "manufacturing transformation vehicle". Target industries are chipmaking, defense and aerospace. At $100 billion it would match SoftBank's Vision Fund in scale and dwarf most buyout vehicles. WSJ also reported that Prometheus was separately in talks to raise up to $6 billion for its core AI operations. (MEDIA REPORT, WSJ, 2026-03-19)
• The New York Times reported separately that Bezos was in early talks with investors in the Middle East and Southeast Asia to raise as much as $100 billion for a fund that would operate alongside Prometheus, and that both the fund and the startup would sit under the same holding company. (MEDIA REPORT, NYT, 2026-03)
• Prospective backers named in reporting: JPMorgan, through its Security and Resiliency Initiative, a $10 billion fund launched in December and led by former Berkshire Hathaway investment manager Todd Combs; and the Abu Dhabi Investment Authority. (MEDIA REPORT, Forbes and WSJ-derived reporting) Note that Bezos sits on an external advisory council for the JPMorgan initiative, which focuses on defense, aerospace, health care and energy - the same sectors as the proposed vehicle - and JPMorgan is separately a Prometheus Series B investor. (MEDIA REPORT)
• Secondary commentary adds batteries, solar equipment and machine tools as logical extensions. No original reporting names them. Treat as inference, not fact.

4.2 What the CEO actually said - the most important sentence in this section

On CNBC on June 11, 2026, Faber asked directly whether it was a separate vehicle that would do the investing. Bezos's answer, verbatim from the published CNBC transcript:

"We - yes, we may buy parts of companies and so on who could benefit from this technology, and then help them improve their processes, their manufacturing process, or what have you. It's premature to say too much about that, that, you know, that plan is still in work."

Faber then pushed the private-equity framing explicitly, proposing a portfolio of companies into which the technology is deployed. Bezos declined the framing and the disclosure: "I'm not sure I would frame it that way, but it's just a little premature to talk about." (OFFICIAL DISCLOSURE, CNBC transcript, 2026-06-11)

Axios reported that the founders "declined to discuss" the reported $100 billion holding company. (MEDIA REPORT, Axios, 2026-06-11)

4.3 Status as of October 21, 2026

• Has it closed? No evidence. No first close, no final close, no GP entity name, no LP commitments, no final size, no timetable. UNVERIFIED.
• Any deals done? No. No industrial acquisition has been announced by Prometheus, by any reporting outlet, or by any target company, as of the research date. UNVERIFIED.
• Legal structure. Described variously as an affiliated holding company (NYT), an investment fund, and a manufacturing transformation vehicle (WSJ). The NYT's "same holding company" claim and the WSJ's "operate alongside" framing are not the same structure, and Bezos's own answer does not settle it. UNVERIFIED.
• Relationship to Prometheus's own financing. Two entirely separate pools of capital. The $100 billion is not part of the $18.2 billion, and the $12 billion Series B does not demonstrate that the acquisition vehicle has been funded.

This is the crux of the investment case. Bezos has said publicly that physical AI needs proprietary data that does not exist as an internet corpus, and the reported plan to buy industrial companies is the direct answer to that problem. But the mechanism that would solve Prometheus's stated central constraint is itself unfunded and undelivered. An investor in Prometheus is buying a company that has said it cannot proceed without a vehicle that has not been raised.

5. Product: the Artificial General Engineer

5.1 Positioning

• Definition. A set of AI tools that let engineers design and manufacture physical products faster, from bridges and chips to jet engines, medical devices and drug compounds. Bezos's framing to Axios: "The cycle from dream, to manufacturing at rate, to having it out in the world can be very long. For example, if you go to a current jet engine manufacturer and say you want the exact same engine but with 10% more thrust, it could be a 10-year program. Not because they're lazy or bad at their jobs, but because it's so complex. So what we're doing is building a set of tools that will empower engineers to compress that cycle time and make that dream-build loop be 10 times faster or even more." (OFFICIAL DISCLOSURE via Axios, 2026-06-11)
• The CAD analogy. Bezos described AGE as "a very, very modern version of CAD." (OFFICIAL DISCLOSURE, 2026-06-11)
• Bajaj's architecture of the problem. The jet engine takes engineer teams a decade or more to design, prototype and manufacture; what changed is that "the ability to formulate even something as complicated as that, from design to manufacturing, as an end-to-end AI problem." (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)
• Not robotics. Bezos interrupted a robotics framing to clarify that Prometheus has nothing to do with robotics, and that it focuses on optimising pre-production machinery and processes such as prototyping. He acknowledged the tools could later be used to design and manufacture robots and factories, but that is not the starting point. (OFFICIAL DISCLOSURE via TechCrunch and Axios, 2026-06-11)
• CAD is not the victim. Asked whether AutoCAD dies if Prometheus succeeds, Bajaj answered: "No. it's not going to be the death of that. There are products today, big ones, airplanes, jet engines, where CAD systems are the system of record. At the same time, for decades, engineers have worked with tools that dissent for drafting, and for the next generation of invention, there are probably better ways to design things, better ways to interact with designs, better ways to predict whether the designs work in a computer, and better ways to make designs from the beginning that are buildable, that are manufacturable. And that requires us to reimagine that process end to end." (OFFICIAL DISCLOSURE, CNBC transcript, 2026-06-11)
• Market framing. Bajaj: "we are one small company operating in a physical economy, which amounts to 60% of the world's GDP. It's very large, it's $70 trillion the entire physical economy." (OFFICIAL DISCLOSURE, CNBC transcript, 2026-06-11)

5.2 Training data strategy - the actual hard problem

Bajaj stated the constraint plainly when Faber pressed him on it: "you're identifying and creating data sets that don't exist." (OFFICIAL DISCLOSURE, CNBC transcript, 2026-06-11)

• There is no internet of manufacturing data. Axios reported that the founders are not sharing how Prometheus is being trained, "except to acknowledge that there isn't an 'Internet of manufacturing data' they can ingest." (MEDIA REPORT, Axios, 2026-06-11)
• Compute-intensive training and inference, and the data must be created. Bezos: "what we're doing is very compute intensive and we need to, you know, create that data." (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)
• Reported data sources (MEDIA REPORT, mixed primary and secondary): industrial and engineering data; industrial data relating to jet engines and computer chips; real experimental and test data; simulation and synthetic data; and precise 3D models of components such as engines built by engineers near New Delhi.
• Recruitment signals (MEDIA REPORT, from published job descriptions): extraction, parsing, normalisation and de-duplication of CAD data; synthetic data generation using LLM and VLM methods; acquisition of engineering, scientific and multimodal datasets; and data provenance, versioning and reproducibility infrastructure.
• UNVERIFIED and explicitly undisclosed: the model architecture, the network design, the training data mix and volume, the benchmark results, the API surface, and the release date. Bezos said it is "premature" to disclose what the company has built.

5.3 Architecture: what can and cannot be inferred

Prometheus has not disclosed its architecture. What can be stated as fact is the composition of the team, and what follows from it is analysis, not disclosure.

• Fact: the founding team includes the co-developer of the Fourier Neural Operator and the lead author of FourCastNet, a research lineage built specifically for learning mappings between function spaces in the Fourier domain to solve PDEs.
• Fact: the founding advisers include the authors of the Transformer paper.
• Fact: the departed scientific lead, Anandkumar, had argued in a Reuters interview that the Transformer architecture is fundamentally unsuited to modelling the physical world and that neural operators should be used instead.
• Analysis, clearly labelled: the most coherent reading of the public record is that Prometheus intends a two-stack architecture - a Transformer-family general reasoning stack for design intent and constraint negotiation, and a neural-operator stack for physical surrogates. The absence of a disclosed architecture means this remains inference, and it is also possible that the operator lineage was hired for a different reason than the one Anandkumar argued for. Any specific architecture claim should be treated as UNVERIFIED until the company publishes one.

5.4 The Reuters fork: who was offered what, and what it reveals about the data strategy

Reuters published an exclusive on August 25, 2026, based on meeting records and an offer letter seen by its reporters.

• In late 2024, over dinner in greater Los Angeles, Vik Bajaj discussed a collaboration with Anima Anandkumar (Caltech Bren Professor of Computing and Mathematical Sciences; previously a senior director of AI research at NVIDIA and a principal scientist at Amazon Web Services) and Benedikt Jenik, her husband, an AI infrastructure engineer.
• A copy of the offer letter titled "Project Prometheus" proposed Anandkumar as the company's public face, a board member and the owner of its scientific vision, with her and Jenik (as a board observer) taking a 35% stake in the company, a combined annual salary of $1 million doubling to $2 million after three months, and an outline of more than $2 billion in "committed rounds" of financing through Series B from investors including Bezos. (MEDIA REPORT, Reuters, 2026-08-25; the letter's own figures)
• Prometheus declined to comment.
• Anandkumar and Jenik declined and founded Accelerated Understanding Inc, publicly launched August 25, 2026. Their model is built on neural operators rather than the Transformer, and the company says it handled 5 trillion data points in a single prompt, roughly 5 million times what Anthropic and Google's flagship models typically consume. That capability claim has not been independently benchmarked. Nvidia did not reply to Reuters on whether it is backing the company; Anandkumar said Jensen Huang was the one who encouraged her to pursue the idea. (MEDIA REPORT, Reuters, 2026-08-25)
• Anandkumar was named to the TIME100 AI list in the same week as the launch. (MEDIA REPORT, TIME)

Two things follow, and they point in opposite directions.

First, the letter's own numbers are internally inconsistent with what actually happened: it promised more than $2 billion in committed rounds through Series B, while the company ultimately raised $6.2 billion at launch and $12 billion in Series B. The letter was either written at an early, much smaller scale or described a floor rather than a plan.

Second, the fork is not a clean architectural argument about Transformer versus operator learning, because Prometheus already hired the Fourier Neural Operator's co-developer to its founding team in spring 2025, before the letter was discussed. Azizzadenesheli was also Anandkumar's PhD student at UC Irvine from 2015 to 2020 in her MEGA Data Lab. Whatever the terms of the split with Anandkumar were, the operator-learning lineage was already inside the company. The dispute, on the public record, is about who would own the scientific vision, not about whether neural operators would be used.

5.5 Product status, customers, partners and the website

• Status: pre-product. As of October 21, 2026 there is no released product, no API, no developer documentation, no demo, no pricing page, no published benchmark and no disclosed revenue. (Verified directly against the website and all reporting, 2026-10-21)
• Roadmap statement. Bezos: "Well, we will have early product rollouts. It's too early to lay out a timetable for that, so you'll just have to stay tuned. But you will see products roll out." (OFFICIAL DISCLOSURE, CNBC transcript, 2026-06-11)
• Internal performance claim. Reporting on the CNBC interview records Bezos saying the team has beaten traditional methods on certain internal simulations while calling it "a little premature to talk too much" about what has been built, and closing with "This is very early. We're working very hard. This is not a done deal." No external validation exists.
• Named customers: none. No paying customer has been named by the company or by any outlet. Blue Origin is the only named relationship, and Bezos's own framing is that Blue Origin is "a case study for a customer of Prometheus" - a description that does not establish a signed contract, a deployment or a commercial arrangement, and is not counted as a customer here. (OFFICIAL DISCLOSURE via Axios, 2026-06-11)
• Amazon. CNBC's live coverage flagged the possibility that Amazon could work with Prometheus. That is a statement about a possible future relationship, not a contract. (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)
• Formal corporate separation. Bezos: Prometheus "doesn't have corporate ties to either Amazon or Blue Origin ... it deserves a dedicated team that is obsessed with this one thing." (OFFICIAL DISCLOSURE via Axios, 2026-06-11)
• Website. https://www.prometheus.ai/ exists. On October 21, 2026 it renders a splash page with the background image, the single heading "Prometheus", a link labelled "Privacy policy", the line "Copyright 2026 Prometheus Industries, Inc." and "PROMETHEUS is a trademark of Prometheus Industries, Inc." There is no product, no customer logo, no blog post, no documentation and no pricing. (OFFICIAL DISCLOSURE, direct fetch, 2026-10-21) Note that the site was reported as non-existent in May 2026, so the domain and site are recent.
• LinkedIn. Tagline: "AI for the physical economy". Company size band 51-200. Founded 2025. Employee-associated-account counts of 191 and 209 at two snapshots on the research date. (MEDIA REPORT, LinkedIn)

6. Acquisitions

General Agents is the only confirmed acquisition.

• What it does. Founded in 2024 by Sherjil Ozair, formerly of DeepMind and Tesla, and William Guss, formerly an OpenAI research scientist. Its product, Ace, is described as a "realtime computer pilot" - a computer agent that takes over a machine and executes tasks across applications from natural-language prompts, built on a video-language-action (VLA) model architecture. A launch demo shows Ace downloading an image from Google and sending it to a contact over iMessage in under 15 seconds. (MEDIA REPORT, WIRED, 2025-11-26)
• The legal timeline, from Delaware filings. In early June 2025, Bajaj hosted an off-the-record AI dinner at Saison, a two-Michelin-star restaurant in San Francisco, attended by Ozair among others. The following morning Bajaj formed an acquisition entity. Four days later the merger with General Agents closed. (MEDIA REPORT, WIRED, citing Delaware corporate filings)
• So the closing was in June 2025; the public reporting date was November 26, 2025. These are two different facts and sources routinely conflate them. The acquisition was therefore completed roughly five months before the company was publicly revealed, and roughly five months before the launch round was reported.
• Price and terms: not disclosed. WIRED: "Terms of the deal could not be learned." General Agents' investor Mythos Ventures declined to comment. (MEDIA REPORT)
• Post-closing. General Agents' registered address became the San Francisco headquarters of Foresite Labs. General Agents continued to operate and ship new Ace versions as of November 2025, and the leader of its India training team joined Prometheus. After the New York Times story, Guss, Ozair and about three dozen colleagues updated their LinkedIn profiles to Prometheus. (MEDIA REPORT, WIRED)
• Interpretation. The competitive logic is stated by Donely CEO Harsha Abegunasekara: "What General Agents really cracked early on is speed - Ace runs on your computer at lightspeed. We've been working on that for six months and haven't achieved it yet." A VLA-based computer-use agent is directly relevant to systems that must operate inside physical environments and equipment states rather than only in text. (MEDIA REPORT) It does not follow that Prometheus has released a robotics product; the VLA route and a shipped robot are different things.
• Other acquisitions: none. Commercial databases list General Agents and nothing else. UNVERIFIED for any additional transaction.

7. Competitive Landscape

7.1 Companies named in reporting

Media coverage of Prometheus has placed it alongside: Periodic Labs, World Labs, AMI Labs, Physical Intelligence, Atoms (Travis Kalanick's rebranded company, focused on transforming manufacturing with AI), Applied Intuition, PhysicsX, Skild AI and Figure AI. Incumbent industrial software vendors named in the same coverage include Autodesk, Siemens, Dassault Systemes and PTC, with Ansys, CATIA/NX and OpenFOAM appearing in technical accounts of the CAD and simulation toolchain. None of these outlets frames the comparison as a head-to-head product contest, because there is no Prometheus product to contest.

• Periodic Labs is the most frequently cited peer. Its actual co-founders are Liam Fedus and Ekin Dogus Cubuk; several Chinese-language reports name "William Fedus," which is an error. It announced roughly $300 million in seed funding in September 2025 led by a16z, and is building autonomous physical laboratories and an AI scientist. (MEDIA REPORT)
• Accelerated Understanding (Anandkumar and Jenik), launched August 2026, is a direct competitor in physics-first modelling.
• The industrial roll-up contest. Forbes reported that Thrive Capital and General Catalyst have launched vehicles to acquire legacy industrial firms and improve margins through software. The proposed $100 billion Prometheus-linked vehicle would be materially larger. (MEDIA REPORT)

7.2 The conflict-of-interest structure is worth stating plainly

Three facts, all documented, compose into a structural tension rather than a scandal:

1. The proposed acquisition vehicle would buy industrial companies and deploy Prometheus's AI inside them, capturing the productivity gains through ownership rather than through software licences. An affiliated fund that owns portfolio companies has an incentive to select Prometheus's tools for its own assets over a third-party vendor's tools.
2. Blue Origin's CEO sits on Prometheus's board, and Blue Origin is described as a case study for a potential Prometheus customer.
3. Bezos sits on an external advisory council of JPMorgan's Security and Resiliency Initiative, which targets defense, aerospace, health care and energy, and JPMorgan is a Prometheus Series B investor. That initiative and the proposed $100 billion vehicle target overlapping sectors.

Separately, Elon Musk has described Project Prometheus as a "copycat" on X. (MEDIA REPORT, Chinese-language secondary reporting; treated as commentary, not fact)

7.3 No benchmark exists

There is no public benchmark against which Prometheus can be compared to any of these companies on any axis. Any statement about whether AGE outperforms a solver, a neural-operator surrogate or an incumbent CAD tool is currently unsupportable. This is a property of the category and of Prometheus's pre-product status, not a criticism of any individual competitor.

8. Market Context

• Prometheus's launch round was the largest venture round of November 2025. Global venture funding that month was $39.6 billion, up 28% year on year, with 43% of it going to just 14 companies raising $500 million or more - the highest count of mega-rounds in three years. (MEDIA REPORT, Crunchbase, 2025-12-02)
• Physical AI's capital surge. Crunchbase data indicates roughly $47 billion of venture funding into physical AI in H1 2026, close to four times H2 2025. The largest single deal in that period was Waymo's $16 billion Series D at a $126 billion valuation in February 2026. Goldman Sachs raised its humanoid robotics market forecast to $38 billion by 2035 from a prior $6 billion, citing a 40% reduction in component costs. (MEDIA REPORT, via secondary reporting of Crunchbase and Goldman Sachs research)
• The market-size anchor the founders themselves use. Bajaj's $70 trillion physical economy, 60% of world GDP, cited in section 5.1. (OFFICIAL DISCLOSURE, CNBC, 2026-06-11)

9. Risk Register

9.1 A $41 billion mark on an unshipped product

There is no revenue, no customer, no product, no API and no benchmark. At $41 billion against roughly 150 employees, the valuation prices near-flawless execution of a data strategy that has not yet produced a single public artifact. Bezos himself set the bar: "This is very early. We're working very hard. This is not a done deal."

9.2 The data strategy has a funding hole in the middle of it

The stated central constraint is that no internet corpus of manufacturing data exists. The stated answer is to own industrial assets and generate the data. The vehicle that would do that has no disclosed close, no GP, no size and no targets. Either the company finds another way to obtain the data, in which case the $41 billion mark embeds an assumption about a method nobody has described, or it funds the vehicle, in which case the first $18.2 billion is not the last requirement. Either branch is a material capital-intensity risk that the current valuation does not appear to price.

9.3 The physical-to-digital gap

A design that is optimal in simulation still has to survive the last 100 metres of physical reality: machining tolerances, tool wear, process deviation, thermal variation, supplier variation. Prometheus's stated remit is pre-production design and process planning, which is where the value is, and the company has been clear that it is not doing factory automation. The commercial consequence is that the tool must earn trust on the design side while the physical validation loop remains human and slow, which sets a ceiling on how much of a design cycle can actually be compressed in the near term. This is a structural property of the category, not a defect in this particular team.

9.4 Liability and certification friction

AI-generated blueprints for pressure vessels, aircraft components or bridges carry the same legal responsibility structure as human-designed ones. Named human engineers still sign. In the highest-value verticals, which are also the most certification-heavy, adoption will be gated by qualification processes whose duration the company has not disclosed.

9.5 Concentration on a single operating founder with several other demanding roles

Bezos described the co-CEO arrangement as fully shared decision-making, with him spending the bulk of his time on Prometheus but also significant time on Blue Origin and on AI at Amazon. Any portfolio company of an affiliated fund, and any board seat, and any multi-company operating commitment, compounds the same key-person dependency. Blue Origin itself had a launch-pad failure in May 2026 that Bezos described as "a very difficult event, a very bad day for Blue Origin."

9.6 The disclosure vacuum is itself a risk

Until there is a product with a date on it, an externally runnable benchmark, or a disclosed first industrial acquisition, the only information channel is two founders' interviews. That is a thin basis for a $41 billion private mark, and it is why the discrepancies in section 11 could not be resolved by diligence.

9.7 If the fund works, it changes the business

A successful roll-up would make Prometheus a vertically integrated industrial holding company with a software arm, not a software company that sells software. That is a different business with different economics, different capital intensity and a different regulatory profile - defence, aerospace and semiconductor acquisitions trigger national-security review, export controls and foreign-investment scrutiny. The company might end up well; an investor buying a software-company thesis would not own what it bought.

10. Assessment

The technical pedigree is the strongest verifiable asset. Hiring the Fourier Neural Operator's co-developer to the founding team and the Transformer paper's co-authors as founding advisers is a specific and unusual commitment to a physics-first architecture class rather than a generic scaling play, and it is corroborated by the fact that a physics-operator researcher with a competing venture was offered the scientific vision. The capital is stronger still: $18.2 billion from a syndicate of JPMorgan, BlackRock, Goldman Sachs, DST Global and Arch Venture Partners plus the founder's own capital, at a $41 billion mark confirmed by an unrelated SEC schedule of investments, with a further $100 billion fund ambition that is a separate, unfunded, undisclosed vehicle.

The weakness is symmetric and equally verifiable. There is no product. There is no customer. There is no benchmark, no API, no price and no revenue. The website is four lines of text. The stated central constraint on the technology - that the required training corpus does not exist and must be manufactured - is addressed by a plan whose funding vehicle has not closed and whose first transaction has not happened.

That is a coherent, unusually legible thesis with a very long and very expensive proof obligation attached. It is worth watching precisely because the verification events are cheap to track and cheap to fail: a product with a date, a benchmark someone outside the company can run, and the announced close of the industrial roll-up on its first target.

11. Discrepancies

Every conflict we found between sources, preserved rather than reconciled.

1. Launch round size: $6.2B versus $6.3B. $6.2B is supported by the New York Times, Financial Times, Axios, GeekWire, TechCrunch, CNBC and Crunchbase's own roundup. A $6.3B figure appears in some aggregators. The weight of evidence is on $6.2B; $6.3B is recorded here rather than dropped.
2. Launch valuation pre-money versus post-money. The Financial Times stated roughly $30 billion and explicitly excluded the new investment; Silicon UK restated it as "$30B pre-money". Several databases label the same ~$30B as post-money or as "current valuation". Our implied ~$36.2B post-money is arithmetic, not a disclosed number.
3. Series B: $10B at $38B versus $12B at $41B. Resolved as one round upsized (see section 3.3), supported by CNBC's "raising its Series B round to $12 billion", TechCrunch's "second fundraising", and the near-identical pre-money arithmetic. The April figures were reported and not retracted.
4. Cumulative capital: $16.2B versus $18.2B versus $28.2B. $16.2B is the April interim figure. $18.2B is supported by Powerlaw's SEC Schedule of Investments and EquiDeFi's SPV documentation. $28.2B is a double count found in several commercial databases that treat $10B and $12B as separate rounds.
5. Lead investor. Bloomberg said the April round had no lead. Some June coverage calls JPMorgan and BlackRock co-leads. One database lists the lead as Bezos. No lead announcement from Prometheus or from any investor was located. UNVERIFIED.
6. Round naming. The November 2025 financing is variously launch funding, seed, and Series A. The June round is Series B by the company's own usage.
7. Announcement date of the Series B: June 11, June 12, June 13, or July 2026. June 11 is the US announcement date from the CNBC interview and the GeekWire same-morning piece. June 12 is the date of CNBC's follow-up feature and of many Chinese-language reprints. June 13 appears in aggregator trackers. July appears in some databases and in Chinese encyclopedia entries, and reflects Powerlaw's retrospective confirmation rather than a new event.
8. The $100B fund: $100 billion versus "tens of billions". Some FT-derived coverage refers to "tens of billions of dollars or more" for a holding company; WSJ and NYT reporting gives the up-to-$100 billion figure for the fund. The two numbers may describe different scopes.
9. The fund's legal structure. NYT: fund and startup under the same holding company. WSJ: a separate "manufacturing transformation vehicle". Other coverage: an affiliated fund operating alongside Prometheus. Bezos's own CNBC answer confirms only intent. UNVERIFIED.
10. The fund's status. WSJ and NYT reported early talks in March 2026. Secondary articles later described the effort as actively advancing. No close, no GP, no LP list, no first deal. UNVERIFIED.
11. Founding date. Stealth since early 2024 (WIRED-derived) versus Bezos joining as founding investor in late 2024 versus a spring 2025 founding team versus 2025-06-02 Swiss registration versus 2025-11-17 public reveal versus LinkedIn "Founded 2025". All are individually sourced and they are not the same claim.
12. The Swiss entity's original name. FL2024-008 Schweiz GmbH indicates an origin inside Foresite Labs' Swiss entity structure, which complicates the common description of Prometheus as a purpose-built AI company founded in 2025.
13. Offices. 101 Mission Street (about 30,000 sq ft), 2261 Market Street, Suite 710 (database corporate office), 1960 Mandela Parkway in West Oakland (about 100,000 sq ft, roughly 8,000 amps), plus London and Zurich, plus the Swiss registered address Zollikerstrasse 226. All are sourced; they are not necessarily the same category of site.
14. Headcount. 100, 120, 120+, 150, 191, 209, and one report of "several hundred". The 191 and 209 are different LinkedIn snapshots and are associated-account counts, not payroll headcount.
15. Website existence. Reported as non-existent in May 2026; present and live on October 21, 2026 as a splash page only.
16. General Agents timing. Legal close in early June 2025 per Delaware filings; public reporting November 26, 2025.
17. General Agents stealth start. WIRED-derived reporting says Prometheus had been in stealth since early 2024, which conflicts with the late 2024 founding-investor account and the spring 2025 founding-team date.
18. Blue Origin's status. "A case study for a customer of Prometheus" is the founder's phrase. It is not equivalent to a signed customer, a deployment or a contract, and it is not counted as one.
19. Periodic Labs' co-founder. Several Chinese-language reports write "William Fedus". The co-founder is Liam Fedus. The incorrect name is recorded here because it is still circulating.
20. Dave Limp's remit. Board membership is confirmed by WSJ. Any suggestion of day-to-day management authority is UNVERIFIED and is inconsistent with Bezos's account of the co-CEO arrangement.
21. Bezos's ownership. Roughly 20% and roughly $8.2 billion of value are Forbes estimates. No stake percentage has been disclosed by the company.
22. The Anandkumar offer letter. More than $2 billion in committed rounds through Series B, a 35% stake for two people, and a $1M-to-$2M salary - figures inconsistent with the $6.2B and $12B rounds the company actually raised. Either the letter was written against a much smaller plan or described a floor.
23. Anandkumar's departure reason. Reuters reports a scientific disagreement about Transformer versus neural operators. Prometheus declined to comment. The disagreement's precise terms are UNVERIFIED, and the presence of the FNO co-developer in the founding team since spring 2025 complicates any reading of it as an operator-versus-Transformer split.
24. Anandkumar's capability claim. Five trillion data points in a single prompt, roughly 5 million times the flagship models' typical context, is a company claim reported by Reuters and explicitly not independently benchmarked.
25. Anatole Abragam Prize year. His company biography says 2012; some profiles say 2013.
26. Bajaj's Foresite Capital tenure. Some profiles still describe him as managing director; his professional profile reportedly records the role as running through March 2026.

12. Sources and Evidence Ledger

Primary and regulatory:

• Prometheus Industries, Inc., https://www.prometheus.ai/, fetched 2026-10-21. OFFICIAL DISCLOSURE.
• Swiss commercial register / FOSC: Prometheus Industries GmbH, UID CHE-308.691.679, registry CH-020.4.088.310-3, registered 2025-06-02, renamed 2026-09-24. OFFICIAL DISCLOSURE.
• Powerlaw Corp., SEC Schedule of Investments as of 2026-06-30, and its Business Wire announcement. OFFICIAL DISCLOSURE for its own holdings.
• EquiDeFi, Prometheus AI SPV announcement, 2026-08-10; subscription opened 2026-08-12. OFFICIAL DISCLOSURE.
• CNBC Exclusive transcript, Prometheus co-founders and co-CEOs Jeff Bezos and Vik Bajaj with David Faber, "Squawk on the Street", 2026-06-11. OFFICIAL DISCLOSURE.
• CNBC live blog, "Bezos opens up about AI startup Prometheus after $12 billion raise", 2026-06-11. OFFICIAL DISCLOSURE.
• CNBC, "Why Jeff Bezos says he's embraced being a CEO again", 2026-06-12. OFFICIAL DISCLOSURE.
• Kamyar Azizzadenesheli, personal academic page and NVIDIA Research profile. OFFICIAL DISCLOSURE.

Reporting:

• The New York Times, "Jeff Bezos Creates A.I. Start-Up Where He Will Be Co-Chief Executive", 2025-11-17.
• Financial Times, launch financing and the approximately $30 billion valuation, 2026-02-26/27.
• The Wall Street Journal, $100 billion fund, Dave Limp board appointment, up-to-$6 billion company raise, 2026-03-19.
• The New York Times, $100 billion fund alongside Prometheus, 2026-03.
• Reuters, $100 billion fund report, 2026-03-19.
• Reuters exclusive, Anandkumar and Jenik and the "Project Prometheus" offer letter, 2026-08-25.
• Bloomberg, approximately $10 billion round closed at approximately $38 billion with no lead, April 2026.
• Business Insider, approximately $10 billion in progress at approximately $38 billion post-money, April 2026.
• Axios, "Prometheus, the industrial AI startup from Jeff Bezos, is now worth $41 billion", 2026-06-11.
• CNBC, Series B announcement and coverage, 2026-06-11 and 2026-06-12.
• TechCrunch, "Jeff Bezos's Prometheus raises $12B to build an 'artificial general engineer' for the physical world", 2026-06-11.
• GeekWire, "Bezos' AI startup Prometheus raises $12B at $41B valuation, and the CEOs explain what they're doing", 2026-06-11.
• WIRED, "Jeff Bezos' New AI Venture Quietly Acquired an Agentic Computing Startup", 2025-11-26, based on Delaware corporate filings.
• Forbes, $100 billion fund and Abu Dhabi Investment Authority as a prospective backer, 2026-03-19.
• Fox News, Bret Baier interview with Jeff Bezos, aired 2026-10-07, recorded at Blue Origin in Cape Canaveral.
• Crunchbase, "Startup funding continued on a tear in November as megarounds hit 3-year high", 2025-12-02.
• Oxford Economics and Goldman Sachs physical-AI market data as reported by secondary sources.
• LinkedIn company page, "Prometheus", accessed 2026-10-21.
• Silicon UK, launch valuation restated as pre-money.
• Chinese-language sources: Xinhuawang, Pengpai, Jiemian, China Finance News Network, TechOrange, Baidu Baike, chinaventure, aifunding.me, InforCapital, Silicon Valley Investclub, Newzino - all treated as MEDIA REPORT and flagged where they conflict with primary sources.

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of October 21, 2026; please refer to official company announcements and filings. All amounts are in US dollars unless stated otherwise; the source-conflict list is set out in Section 11.