THE FIRST AGE OF AI: MONEY, MYTH, POWER, AND THE PRIVATIZATION OF INTELLIGENCE
An archival investigative report on the financialization, institutionalization, and political economy of artificial intelligence, 2010–August 15, 2026
Scope. This report treats 2010–August 15, 2026 as the primary historical window, while separating events that have already occurred from commitments, forecasts, and projects extending into 2027–2030 and beyond. “Privatization of intelligence” is used analytically: it does not mean that corporations literally own intelligence. It refers to the concentration of control over frontier models, model weights, compute, data-center capacity, distribution channels, intellectual property, safety decisions, and access within private legal and financial structures.
Standard of proof. Documented financial relationships are distinguished from interpretation. Political access is distinguished from regulatory capture. Conflicts of interest are distinguished from corruption. Forecast error is distinguished from deception. Nothing reviewed for this report establishes that Leopold Aschenbrenner or the institutions examined here committed fraud; where public evidence is incomplete, that is stated explicitly.
Executive summary
The archival record supports a significant but more precise thesis than “AI was a scam.” Between 2010 and 2026, artificial intelligence underwent a remarkable transformation from an academic and research frontier into a financial asset class, strategic infrastructure project, geopolitical doctrine, philanthropic cause, policy field, and cultural mythology. Capital did not merely finance AI after technical progress occurred. Capital increasingly became an input into what counted as frontier AI: obtaining the largest clusters, attracting researchers, constructing data centers, securing electricity, and financing years of losses became prerequisites for participation at the frontier. Stanford’s 2026 AI Index estimates that global corporate AI investment reached $581.7 billion in 2025, while private AI investment reached $344.7 billion; U.S. private AI investment alone was $285.9 billion, more than twenty-three times China’s private-investment figure.
The resulting industry is not adequately described as conventional venture capital. The frontier increasingly operates through a circular capital-and-compute system. Cloud companies invest billions in model developers; developers commit large portions of that capital back to the cloud providers through compute purchases; providers receive equity, revenue-sharing rights, technical information, product integration opportunities and, in some cases, exclusivity or consultation rights. That is not speculation: a 2025 FTC investigation based partly on non-public company documents identified these characteristics in the Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic relationships and warned of switching costs, access-to-input concerns, and possible lock-in.
This makes compute ownership and finance central to the privatization question. OpenAI moved from a nonprofit announced in 2015 with $1 billion of commitments to a capped-profit structure in 2019 because it said AGI would require billions of dollars, then to a public-benefit-corporation structure controlled by its nonprofit foundation. Microsoft’s post-recapitalization interest was valued at approximately $135 billion in October 2025. In March 2026 OpenAI announced $122 billion of new funding involving Amazon, Nvidia, SoftBank, Microsoft, a16z, D. E. Shaw Ventures, MGX, TPG, BlackRock-affiliated funds, Blackstone, Sequoia, Temasek, UC Investments and many others.
Anthropic developed an apparently different institutional identity—explicit emphasis on AI safety, public-benefit status and its Long-Term Benefit Trust—but converged financially toward a similar capital intensity. Its Trust was designed to become capable of selecting a majority of its board over time and to remain financially disinterested; that governance experiment is materially different from ordinary shareholder governance. Yet by May 2026 Anthropic had raised $65 billion at a $965 billion post-money valuation, following a $30 billion round at $380 billion only three months earlier, while Amazon, Google, Microsoft and Nvidia all became important capital or compute partners. Anthropic confidentially submitted a draft S-1 for an IPO on June 1, 2026.
The archival record also establishes that AI safety was not simply an internal laboratory discipline. It became an institution-building project. Open Philanthropy—renamed Coefficient Giving—explicitly described its AI-governance strategy as including both research and “practice and influence”; it funded or supported work across Georgetown’s CSET, RAND, CNAS, CSIS, the Wilson Center, FHI/GovAI, 80,000 Hours, the Future of Life Institute and OpenAI, among others. Its current program describes hundreds of grants for governance, technical safety, capacity building and professional networks, and in July 2026 it appointed an inaugural managing director of public policy to lead U.S. AI-policy work and government relations.
That is important because forecasting created institutions before the forecasts could be resolved. In 2016 Holden Karnofsky argued that it was appropriate to act as though there were at least a 10% probability of transformative AI within twenty years while explicitly emphasizing very high uncertainty. That forecast is not due to resolve until 2036, but it contributed to a philanthropic rationale for creating an AI-risk ecosystem years earlier. In 2024 Leopold Aschenbrenner argued that “AGI by 2027” was strikingly plausible and forecast a transition toward enormous compute clusters and trillions of dollars of industrial investment. Dario Amodei argued that what he calls “powerful AI” could arrive as early as 2026. These claims cannot simply be scored as correct or incorrect today because horizons remain open and definitions are often flexible.
That produces a central finding of this investigation:
The First Age of AI frequently converted forecasts into present-day institutional authority long before those forecasts became empirically resolvable.
The mechanism operated across philanthropy, investment and politics. A sufficiently consequential future scenario can rationally justify current action even when probability is low. But this creates a structural asymmetry: capital, organizations, careers and regulatory influence are created today, while falsification may arrive years later—or become impossible because the prediction was never operationalized sufficiently to score.
The 2026 collapse of Situational Awareness provides the clearest case where AI prediction, investment authority and leveraged financial exposure became visibly intertwined. Aschenbrenner’s June 2024 essays made explicit technological forecasts; he then established an AI-focused hedge fund. Its SEC Form 13F reported $13.677 billion in public-equity holdings across 42 positions as of March 31, 2026. Reuters reported that the fund returned approximately 439% from January through June 2026, before its portfolio value fell 67% in July, forcing the unwinding of most of a roughly $16 billion public-equities book and the removal of all leverage.
Reuters subsequently reported that Jane Street suffered a roughly $15 billion July hit from a combination of its exposure to Situational Awareness and other technology positions, and that the Situational investment itself ended roughly flat for Jane Street in 2026 after having grown enormously earlier. Reuters Breakingviews reported, citing the Wall Street Journal rather than primary filings, that Situational had leveraged parts of its stock portfolio approximately three to four times. Because the precise leverage ratio has not been independently established through public SEC disclosures reviewed here, this report treats “three-to-four-times leverage” as reported rather than independently verified.
This is evidence of extreme concentration and risk, not evidence of fraud. The stronger investigative question is how a technological thesis acquired enough financial credibility to support that scale of leverage and interconnected exposure.
The crypto relationship is similarly real but must not be exaggerated. FTX’s Future Fund funded longtermist and AI-safety-related work before FTX collapsed; archived records and Reuters reporting indicate a broad grantmaking program that had already distributed or committed substantial sums. Separately, Sam Altman co-founded Worldcoin/World, whose own materials explicitly connect digital identity and a global financial network to a possible AI/UBI future. MGX, which appears in major OpenAI and Anthropic financing rounds, also invested $2 billion in Binance using stablecoin. These are meaningful overlaps between crypto capital, AI ideology and AI investment, but they do not demonstrate that token issuance was a major financing mechanism for OpenAI or Anthropic.
Likewise, SPACs offered an AI financialization channel for companies such as BigBear.ai and SoundHound during the 2021–2022 SPAC era, but frontier-model developers largely relied instead on private equity, strategic hyperscaler investment, sovereign capital and increasingly infrastructure debt.
The evidence for political access is strong. Frontier executives testified before Congress; laboratories participated directly in White House processes; governments built policy around frontier-model categories; and the United States explicitly shifted toward an AI-leadership and infrastructure doctrine, while the EU created special obligations for the most computationally intensive general-purpose models. The evidence for regulatory capture, however, is more ambiguous. Government bodies have simultaneously challenged market concentration: the FTC investigated cloud-lab relationships, while NTIA explicitly considered whether open-weight models could broaden participation.
The most defensible conclusion is therefore neither that the system is innocent nor that it is a coordinated conspiracy. It is that technical uncertainty, concentrated private capital, strategic infrastructure, philanthropic agenda-setting and government dependence on scarce expertise developed together, creating recurring conflicts between public-interest claims and private accumulation.
Scope. This report treats 2010–August 15, 2026 as the primary historical window, while separating events that have already occurred from commitments, forecasts, and projects extending into 2027–2030 and beyond. “Privatization of intelligence” is used analytically: it does not mean that corporations literally own intelligence. It refers to the concentration of control over frontier models, model weights, compute, data-center capacity, distribution channels, intellectual property, safety decisions, and access within private legal and financial structures.
Standard of proof. Documented financial relationships are distinguished from interpretation. Political access is distinguished from regulatory capture. Conflicts of interest are distinguished from corruption. Forecast error is distinguished from deception. Nothing reviewed for this report establishes that Leopold Aschenbrenner or the institutions examined here committed fraud; where public evidence is incomplete, that is stated explicitly.
Executive summary
The archival record supports a significant but more precise thesis than “AI was a scam.” Between 2010 and 2026, artificial intelligence underwent a remarkable transformation from an academic and research frontier into a financial asset class, strategic infrastructure project, geopolitical doctrine, philanthropic cause, policy field, and cultural mythology. Capital did not merely finance AI after technical progress occurred. Capital increasingly became an input into what counted as frontier AI: obtaining the largest clusters, attracting researchers, constructing data centers, securing electricity, and financing years of losses became prerequisites for participation at the frontier. Stanford’s 2026 AI Index estimates that global corporate AI investment reached $581.7 billion in 2025, while private AI investment reached $344.7 billion; U.S. private AI investment alone was $285.9 billion, more than twenty-three times China’s private-investment figure.
The resulting industry is not adequately described as conventional venture capital. The frontier increasingly operates through a circular capital-and-compute system. Cloud companies invest billions in model developers; developers commit large portions of that capital back to the cloud providers through compute purchases; providers receive equity, revenue-sharing rights, technical information, product integration opportunities and, in some cases, exclusivity or consultation rights. That is not speculation: a 2025 FTC investigation based partly on non-public company documents identified these characteristics in the Microsoft–OpenAI, Amazon–Anthropic, and Google–Anthropic relationships and warned of switching costs, access-to-input concerns, and possible lock-in.
This makes compute ownership and finance central to the privatization question. OpenAI moved from a nonprofit announced in 2015 with $1 billion of commitments to a capped-profit structure in 2019 because it said AGI would require billions of dollars, then to a public-benefit-corporation structure controlled by its nonprofit foundation. Microsoft’s post-recapitalization interest was valued at approximately $135 billion in October 2025. In March 2026 OpenAI announced $122 billion of new funding involving Amazon, Nvidia, SoftBank, Microsoft, a16z, D. E. Shaw Ventures, MGX, TPG, BlackRock-affiliated funds, Blackstone, Sequoia, Temasek, UC Investments and many others.
Anthropic developed an apparently different institutional identity—explicit emphasis on AI safety, public-benefit status and its Long-Term Benefit Trust—but converged financially toward a similar capital intensity. Its Trust was designed to become capable of selecting a majority of its board over time and to remain financially disinterested; that governance experiment is materially different from ordinary shareholder governance. Yet by May 2026 Anthropic had raised $65 billion at a $965 billion post-money valuation, following a $30 billion round at $380 billion only three months earlier, while Amazon, Google, Microsoft and Nvidia all became important capital or compute partners. Anthropic confidentially submitted a draft S-1 for an IPO on June 1, 2026.
The archival record also establishes that AI safety was not simply an internal laboratory discipline. It became an institution-building project. Open Philanthropy—renamed Coefficient Giving—explicitly described its AI-governance strategy as including both research and “practice and influence”; it funded or supported work across Georgetown’s CSET, RAND, CNAS, CSIS, the Wilson Center, FHI/GovAI, 80,000 Hours, the Future of Life Institute and OpenAI, among others. Its current program describes hundreds of grants for governance, technical safety, capacity building and professional networks, and in July 2026 it appointed an inaugural managing director of public policy to lead U.S. AI-policy work and government relations.
That is important because forecasting created institutions before the forecasts could be resolved. In 2016 Holden Karnofsky argued that it was appropriate to act as though there were at least a 10% probability of transformative AI within twenty years while explicitly emphasizing very high uncertainty. That forecast is not due to resolve until 2036, but it contributed to a philanthropic rationale for creating an AI-risk ecosystem years earlier. In 2024 Leopold Aschenbrenner argued that “AGI by 2027” was strikingly plausible and forecast a transition toward enormous compute clusters and trillions of dollars of industrial investment. Dario Amodei argued that what he calls “powerful AI” could arrive as early as 2026. These claims cannot simply be scored as correct or incorrect today because horizons remain open and definitions are often flexible.
That produces a central finding of this investigation:
The First Age of AI frequently converted forecasts into present-day institutional authority long before those forecasts became empirically resolvable.
The mechanism operated across philanthropy, investment and politics. A sufficiently consequential future scenario can rationally justify current action even when probability is low. But this creates a structural asymmetry: capital, organizations, careers and regulatory influence are created today, while falsification may arrive years later—or become impossible because the prediction was never operationalized sufficiently to score.
The 2026 collapse of Situational Awareness provides the clearest case where AI prediction, investment authority and leveraged financial exposure became visibly intertwined. Aschenbrenner’s June 2024 essays made explicit technological forecasts; he then established an AI-focused hedge fund. Its SEC Form 13F reported $13.677 billion in public-equity holdings across 42 positions as of March 31, 2026. Reuters reported that the fund returned approximately 439% from January through June 2026, before its portfolio value fell 67% in July, forcing the unwinding of most of a roughly $16 billion public-equities book and the removal of all leverage.
Reuters subsequently reported that Jane Street suffered a roughly $15 billion July hit from a combination of its exposure to Situational Awareness and other technology positions, and that the Situational investment itself ended roughly flat for Jane Street in 2026 after having grown enormously earlier. Reuters Breakingviews reported, citing the Wall Street Journal rather than primary filings, that Situational had leveraged parts of its stock portfolio approximately three to four times. Because the precise leverage ratio has not been independently established through public SEC disclosures reviewed here, this report treats “three-to-four-times leverage” as reported rather than independently verified.
This is evidence of extreme concentration and risk, not evidence of fraud. The stronger investigative question is how a technological thesis acquired enough financial credibility to support that scale of leverage and interconnected exposure.
The crypto relationship is similarly real but must not be exaggerated. FTX’s Future Fund funded longtermist and AI-safety-related work before FTX collapsed; archived records and Reuters reporting indicate a broad grantmaking program that had already distributed or committed substantial sums. Separately, Sam Altman co-founded Worldcoin/World, whose own materials explicitly connect digital identity and a global financial network to a possible AI/UBI future. MGX, which appears in major OpenAI and Anthropic financing rounds, also invested $2 billion in Binance using stablecoin. These are meaningful overlaps between crypto capital, AI ideology and AI investment, but they do not demonstrate that token issuance was a major financing mechanism for OpenAI or Anthropic.
Likewise, SPACs offered an AI financialization channel for companies such as BigBear.ai and SoundHound during the 2021–2022 SPAC era, but frontier-model developers largely relied instead on private equity, strategic hyperscaler investment, sovereign capital and increasingly infrastructure debt.
The evidence for political access is strong. Frontier executives testified before Congress; laboratories participated directly in White House processes; governments built policy around frontier-model categories; and the United States explicitly shifted toward an AI-leadership and infrastructure doctrine, while the EU created special obligations for the most computationally intensive general-purpose models. The evidence for regulatory capture, however, is more ambiguous. Government bodies have simultaneously challenged market concentration: the FTC investigated cloud-lab relationships, while NTIA explicitly considered whether open-weight models could broaden participation.
The most defensible conclusion is therefore neither that the system is innocent nor that it is a coordinated conspiracy. It is that technical uncertainty, concentrated private capital, strategic infrastructure, philanthropic agenda-setting and government dependence on scarce expertise developed together, creating recurring conflicts between public-interest claims and private accumulation.
Archive and chronology of the first age
The historical pattern is easier to see when separated into phases. The first phase, roughly 2010–2016, converted machine intelligence from a research ambition into an institutional mission. DeepMind began in 2010 explicitly around general AI; the AlexNet breakthrough in 2012 demonstrated the power of deep learning at scale; OpenAI launched in 2015 as a nonprofit explicitly promising broad human benefit; and by 2016 philanthropic institutions were treating transformative AI as a sufficiently plausible future risk to justify a new governance field.
The second phase, roughly 2017–2021, revealed that “intelligence” at the frontier would be capital-intensive. OpenAI later said that by 2017 it had concluded that building AGI would require vastly more compute and billions of dollars per year. In 2019 it created a capped-profit entity, then entered a $1 billion Microsoft partnership centered on Azure supercomputing. OpenAI itself estimated in 2018 that training compute used by the largest AI systems had risen more than 300,000-fold since 2012, with a roughly 3.4-month doubling time over the period it studied.
The third phase, roughly 2022–2024, turned generative AI into a mass cultural event and safety into mass politics. Generative-AI private investment reached approximately $25.2 billion in 2023, nearly nine times 2022’s level. In 2023 the Future of Life Institute’s pause letter and the Center for AI Safety’s extinction-risk statement converted technical disagreements about model trajectories into highly visible political language; the latter was signed by figures including Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman, Dario Amodei and Ilya Sutskever.
The fourth phase, 2025–2026, is infrastructure financialization. Models became inseparable from multi-gigawatt power commitments, data-center debt, cloud contracts, sovereign investment, hyperscaler capex and capital markets. Stanford estimates global corporate AI investment more than doubled in 2025 to $581.7 billion. OpenAI’s Stargate announcement described an intention to invest $500 billion over four years, beginning with $100 billion, explicitly coupling AI infrastructure with U.S. strategic leadership. Anthropic separately announced $50 billion in U.S. infrastructure and multi-gigawatt Google/Broadcom arrangements.
The historical pattern is easier to see when separated into phases. The first phase, roughly 2010–2016, converted machine intelligence from a research ambition into an institutional mission. DeepMind began in 2010 explicitly around general AI; the AlexNet breakthrough in 2012 demonstrated the power of deep learning at scale; OpenAI launched in 2015 as a nonprofit explicitly promising broad human benefit; and by 2016 philanthropic institutions were treating transformative AI as a sufficiently plausible future risk to justify a new governance field.
The second phase, roughly 2017–2021, revealed that “intelligence” at the frontier would be capital-intensive. OpenAI later said that by 2017 it had concluded that building AGI would require vastly more compute and billions of dollars per year. In 2019 it created a capped-profit entity, then entered a $1 billion Microsoft partnership centered on Azure supercomputing. OpenAI itself estimated in 2018 that training compute used by the largest AI systems had risen more than 300,000-fold since 2012, with a roughly 3.4-month doubling time over the period it studied.
The third phase, roughly 2022–2024, turned generative AI into a mass cultural event and safety into mass politics. Generative-AI private investment reached approximately $25.2 billion in 2023, nearly nine times 2022’s level. In 2023 the Future of Life Institute’s pause letter and the Center for AI Safety’s extinction-risk statement converted technical disagreements about model trajectories into highly visible political language; the latter was signed by figures including Geoffrey Hinton, Yoshua Bengio, Demis Hassabis, Sam Altman, Dario Amodei and Ilya Sutskever.
The fourth phase, 2025–2026, is infrastructure financialization. Models became inseparable from multi-gigawatt power commitments, data-center debt, cloud contracts, sovereign investment, hyperscaler capex and capital markets. Stanford estimates global corporate AI investment more than doubled in 2025 to $581.7 billion. OpenAI’s Stargate announcement described an intention to invest $500 billion over four years, beginning with $100 billion, explicitly coupling AI infrastructure with U.S. strategic leadership. Anthropic separately announced $50 billion in U.S. infrastructure and multi-gigawatt Google/Broadcom arrangements.
The chronology can be represented as four transitions rather than a simple story of technical progress:
The crucial historical change is visible in the units themselves: the early archive talks about papers, benchmarks and researchers; the late archive increasingly talks about billions of dollars, gigawatts, debt facilities, sovereign funds, data-center contracts and national leadership.
The financialization and privatization of intelligence
The most important financial mechanism was not traditional venture capital in isolation. It was the emergence of a vertically interdependent AI balance sheet connecting model laboratories, cloud providers, chip companies, data-center specialists, banks, sovereign wealth, venture funds, hedge funds and eventually public markets.
The FTC’s 2025 findings are unusually important because they move the analysis beyond public-relations announcements. Based partly on material compelled from the companies, FTC staff found that large cloud providers obtained significant equity and sometimes revenue-sharing interests in AI developers; consultation, control or exclusivity rights existed to varying degrees; developers committed significant portions of investment proceeds to purchase cloud services from the investing provider; and partnerships gave cloud firms access to compute, IP and certain business or training information.
That creates a feedback loop:
The financialization and privatization of intelligence
The most important financial mechanism was not traditional venture capital in isolation. It was the emergence of a vertically interdependent AI balance sheet connecting model laboratories, cloud providers, chip companies, data-center specialists, banks, sovereign wealth, venture funds, hedge funds and eventually public markets.
The FTC’s 2025 findings are unusually important because they move the analysis beyond public-relations announcements. Based partly on material compelled from the companies, FTC staff found that large cloud providers obtained significant equity and sometimes revenue-sharing interests in AI developers; consultation, control or exclusivity rights existed to varying degrees; developers committed significant portions of investment proceeds to purchase cloud services from the investing provider; and partnerships gave cloud firms access to compute, IP and certain business or training information.
That creates a feedback loop:
This structure means that saying “Microsoft invested X in OpenAI” or “Amazon invested Y in Anthropic” can be misleading if interpreted like a passive stock purchase. Part of the economic relationship is a closed or semi-closed commercial loop in which the investee is simultaneously a major purchaser of the investor’s infrastructure. The FTC specifically highlighted this feature.
A second loop emerged around infrastructure. Core
Weave illustrates it. Its SEC filings report revenue increasing from $229 million in 2023 to $1.9 billion in 2024 and $5.1 billion in 2025, while the company remained loss-making, recording a $1.2 billion net loss in 2025. Interest expense reached approximately $1.229 billion, up $868 million year over year, and cash investment in property and equipment was approximately $10.3 billion in 2025. The same filing records an OpenAI commitment for up to approximately $6.5 billion through May 2031 under later orders.
CoreWeave is therefore a useful demonstration that the AI boom is not merely equity speculation. It is being translated into fixed assets and debt obligations. At the same time, those assets derive their economic value from future expectations of model demand, creating a financial bridge from an uncertain technological future into present-day collateral and borrowing.
The scale is expanding further. Reuters reported in August 2026 that Nvidia was working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on arrangements targeting more than $500 billion of third-party capital for AI infrastructure, with Nvidia potentially providing substantial backstop support. This is the point at which “AI” ceases to be an ordinary technology sector and begins behaving like an infrastructure-finance complex.
A simple view of the investment acceleration illustrates the change:
A second loop emerged around infrastructure. Core
Weave illustrates it. Its SEC filings report revenue increasing from $229 million in 2023 to $1.9 billion in 2024 and $5.1 billion in 2025, while the company remained loss-making, recording a $1.2 billion net loss in 2025. Interest expense reached approximately $1.229 billion, up $868 million year over year, and cash investment in property and equipment was approximately $10.3 billion in 2025. The same filing records an OpenAI commitment for up to approximately $6.5 billion through May 2031 under later orders.
CoreWeave is therefore a useful demonstration that the AI boom is not merely equity speculation. It is being translated into fixed assets and debt obligations. At the same time, those assets derive their economic value from future expectations of model demand, creating a financial bridge from an uncertain technological future into present-day collateral and borrowing.
The scale is expanding further. Reuters reported in August 2026 that Nvidia was working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on arrangements targeting more than $500 billion of third-party capital for AI infrastructure, with Nvidia potentially providing substantial backstop support. This is the point at which “AI” ceases to be an ordinary technology sector and begins behaving like an infrastructure-finance complex.
A simple view of the investment acceleration illustrates the change:
Stanford reports global corporate AI investment of approximately $252.3 billion in 2024 and $581.7 billion in 2025. These figures aggregate more than frontier-model financing alone and should not be mistaken for spending exclusively on generative models.
The public financial record can be summarized as follows.
The public financial record can be summarized as follows.
There is a further distinction between money raised, valuation, contractual commitments and revenue. They are not interchangeable. A $965 billion private-company valuation is not $965 billion of cash. A $500 billion infrastructure announcement is not $500 billion already spent. A cloud contract scheduled through 2031 is not current-year revenue. A hedge fund’s gross market exposure is not its net asset value. Failing to maintain these distinctions is one of the principal ways technological narratives can become financially misleading even without anyone making an explicitly false statement.
The same discipline applies to “returns.” Situational’s reported 439% first-half 2026 rise and subsequent 67% July loss cannot be combined by simple subtraction. A portfolio gaining 439% grows to 5.39 times its starting level; a subsequent 67% decline reduces that enlarged amount to approximately 1.78 times the original value before accounting for subscriptions, withdrawals, fees, changing gross exposure or other portfolio effects. Reuters reported the fund remained approximately 80% up for 2026 after July, which is consistent with the non-linear arithmetic of large gains and losses.
This episode illustrates why extraordinary percentage returns can themselves become a marketing-like epistemic signal: the market success of investments associated with a worldview can make the worldview appear empirically validated. But stock-price appreciation demonstrates that investors repriced assets; it does not, by itself, validate a forecast about AGI.
Institutional networks, narratives, predictions, and political power
The archival record does not reveal one secret organization directing AI. It reveals something more ordinary and, in some respects, more consequential: a dense elite network in which capital providers, technical researchers, philanthropic organizations, policy institutes, cloud companies, safety advocates and governments repeatedly intersect.
That matters because independence is not binary. Two institutions may be legally independent while sharing funders, board members, intellectual frameworks, talent pipelines or strategic assumptions. These relationships should neither be ignored nor converted automatically into conspiracy.
The same discipline applies to “returns.” Situational’s reported 439% first-half 2026 rise and subsequent 67% July loss cannot be combined by simple subtraction. A portfolio gaining 439% grows to 5.39 times its starting level; a subsequent 67% decline reduces that enlarged amount to approximately 1.78 times the original value before accounting for subscriptions, withdrawals, fees, changing gross exposure or other portfolio effects. Reuters reported the fund remained approximately 80% up for 2026 after July, which is consistent with the non-linear arithmetic of large gains and losses.
This episode illustrates why extraordinary percentage returns can themselves become a marketing-like epistemic signal: the market success of investments associated with a worldview can make the worldview appear empirically validated. But stock-price appreciation demonstrates that investors repriced assets; it does not, by itself, validate a forecast about AGI.
Institutional networks, narratives, predictions, and political power
The archival record does not reveal one secret organization directing AI. It reveals something more ordinary and, in some respects, more consequential: a dense elite network in which capital providers, technical researchers, philanthropic organizations, policy institutes, cloud companies, safety advocates and governments repeatedly intersect.
That matters because independence is not binary. Two institutions may be legally independent while sharing funders, board members, intellectual frameworks, talent pipelines or strategic assumptions. These relationships should neither be ignored nor converted automatically into conspiracy.
The institutional map looks less like a pyramid than an interlocking network:
The diagram represents documented categories of ties, not common control. In particular, capital relationships should not be interpreted as proof of ideological coordination, and grants should not be interpreted as commands.
Myth in this report does not mean “falsehood.” In the anthropological sense, myths are organizing stories that tell institutions what historical moment they inhabit and therefore what behavior is justified. Four such stories became especially powerful.
The first was AGI imminence: the possibility that a qualitative break in intelligence was only years away. Open Phil’s 2016 analysis was actually cautious in wording—it explicitly said uncertainty was extremely high—but the institutional implication was immediate: build a field now. Aschenbrenner later made a substantially more aggressive 2027 argument. Amodei described powerful AI as potentially arriving as early as 2026.
The second was existential risk. By 2023 the Center for AI Safety could gather leading laboratory executives and AI scientists behind a statement placing extinction risk from AI among global catastrophic priorities, while the Future of Life Institute mobilized a public letter calling for a six-month pause on systems more powerful than GPT-4. Whether one agrees with these claims or not, the documents demonstrate that the rhetoric was institutionally consequential.
The third was national competition. AI increasingly ceased to be described merely as software and became a national strategic resource. Aschenbrenner’s essays explicitly framed frontier AI through U.S.–China competition; OpenAI’s Stargate announcement framed infrastructure in terms of American leadership; the White House’s 2025 policy emphasized sustaining U.S. AI dominance and removing barriers to development.
The fourth was benefit to all humanity. OpenAI, Anthropic and many safety organizations use genuinely universal language—benefiting humanity, long-term benefit, responsible scaling, democratizing benefits. The unresolved historical question is not whether this language was sincere. Sincerity cannot be reliably determined from filings. The question is how universal-benefit rhetoric interacted with ownership structures in which a small number of private organizations controlled frontier systems.
That tension is particularly visible in OpenAI’s trajectory. Its 2015 launch announced $1 billion of funding commitments, but OpenAI later said the nonprofit actually raised less than $45 million from Elon Musk and more than $90 million from other donors. The 2019 capped-profit conversion was justified by the extraordinary expected cost of AGI; by 2024 OpenAI said it had originally estimated that building AGI would require capital “on the order of $10B.” By 2026 a single funding round was more than ten times that early estimate.
This should make historians cautious about retrospective inevitability. The capital intensity of contemporary AI was not simply a timeless fact. It was produced jointly by technical scaling choices, competitive racing, cloud economics, model size, investor willingness, organizational strategy and demand growth.
The measurable-predictions archive reinforces that caution.
Myth in this report does not mean “falsehood.” In the anthropological sense, myths are organizing stories that tell institutions what historical moment they inhabit and therefore what behavior is justified. Four such stories became especially powerful.
The first was AGI imminence: the possibility that a qualitative break in intelligence was only years away. Open Phil’s 2016 analysis was actually cautious in wording—it explicitly said uncertainty was extremely high—but the institutional implication was immediate: build a field now. Aschenbrenner later made a substantially more aggressive 2027 argument. Amodei described powerful AI as potentially arriving as early as 2026.
The second was existential risk. By 2023 the Center for AI Safety could gather leading laboratory executives and AI scientists behind a statement placing extinction risk from AI among global catastrophic priorities, while the Future of Life Institute mobilized a public letter calling for a six-month pause on systems more powerful than GPT-4. Whether one agrees with these claims or not, the documents demonstrate that the rhetoric was institutionally consequential.
The third was national competition. AI increasingly ceased to be described merely as software and became a national strategic resource. Aschenbrenner’s essays explicitly framed frontier AI through U.S.–China competition; OpenAI’s Stargate announcement framed infrastructure in terms of American leadership; the White House’s 2025 policy emphasized sustaining U.S. AI dominance and removing barriers to development.
The fourth was benefit to all humanity. OpenAI, Anthropic and many safety organizations use genuinely universal language—benefiting humanity, long-term benefit, responsible scaling, democratizing benefits. The unresolved historical question is not whether this language was sincere. Sincerity cannot be reliably determined from filings. The question is how universal-benefit rhetoric interacted with ownership structures in which a small number of private organizations controlled frontier systems.
That tension is particularly visible in OpenAI’s trajectory. Its 2015 launch announced $1 billion of funding commitments, but OpenAI later said the nonprofit actually raised less than $45 million from Elon Musk and more than $90 million from other donors. The 2019 capped-profit conversion was justified by the extraordinary expected cost of AGI; by 2024 OpenAI said it had originally estimated that building AGI would require capital “on the order of $10B.” By 2026 a single funding round was more than ten times that early estimate.
This should make historians cautious about retrospective inevitability. The capital intensity of contemporary AI was not simply a timeless fact. It was produced jointly by technical scaling choices, competitive racing, cloud economics, model size, investor willingness, organizational strategy and demand growth.
The measurable-predictions archive reinforces that caution.
The methodological problem is striking: the most culturally influential AI forecasts are often difficult to score. Terms such as AGI, transformative AI, powerful AI, superintelligence and human-level performance are not interchangeable; probabilities differ from deadlines; “could arrive” is not the same proposition as “will arrive”; a prediction of investment is not a prediction of capability.
A serious archive should therefore resist a Nostradamus-style hit/miss game. Instead, it should preserve each claim as originally stated, with its definition, probability, date, horizon and measurable criteria, then prevent retrospective rewriting.
Safety rhetoric deserves the same analytical discipline. It can create genuine social value and simultaneously create competitive advantages. A frontier laboratory that has extensive safety teams, evaluation infrastructure, lobbying capacity and compliance departments may be better positioned to satisfy sophisticated regulation than a small entrant. The EU AI Act, for example, presumes systemic-risk status for GPAI models above a training-compute threshold of (10^{25}) FLOP, while recognizing that only a handful of companies currently develop such models.
That does not imply the threshold was created to protect incumbents. The same threshold may be a rational attempt to concentrate regulatory scrutiny where risk is greatest. The dual effect is what matters:
A safety rule can simultaneously reduce danger and increase fixed costs of entry.
The archival question should therefore be: Who proposed a rule, what evidence justified it, who bears its compliance cost, and who gains competitive advantage from it? Intent need not be assumed.
Political access presents a similar distinction. Sam Altman’s 2023 Senate testimony is direct evidence that frontier executives obtained high-level legislative access. Subsequent White House strategies explicitly relied on private-sector frontier expertise and infrastructure. Yet the FTC simultaneously investigated those same firms’ market relationships. The evidence therefore describes interdependence and asymmetric access, not a simple story in which the state has been completely captured.
Academic work increasingly supports the structural concern. Research on “concentrating intelligence” examines how foundation-model scaling can drive concentrated market structure, while governance scholarship has documented the dominance of large developed-country actors and risk-management framings in emerging AI institutions. These studies do not prove conspiracy; they strengthen the case that concentration is an economic and governance property worth treating independently of individual motives.
Case studies: where money, myth, and power intersect
Situational Awareness: prophecy becomes a portfolio. The importance of Leopold Aschenbrenner’s case is not that his essays prove or disprove AGI. It is the unusually direct connection between a public technological worldview and a large financial vehicle. His 2024 series argued that AGI around 2027 was plausible, that algorithmic progress plus compute scaling would produce a major capability jump, and that industrial mobilization around enormous clusters would follow. An SEC Form D dated September 23, 2024 identifies Situational Awareness Partners LP as a hedge fund and Aschenbrenner as a related executive person.
By March 31, 2026 its SEC 13F disclosed 42 listed positions worth $13.6767 billion. Reuters later reported that the fund had about $20 billion under management before its July crisis and that public holdings included Broadcom, Intel and CoreWeave, alongside a private Anthropic investment. SEC beneficial-ownership filings independently confirm Aschenbrenner’s control relationships and Carl Shulman’s co-portfolio-management role.
The first half of 2026 then produced a spectacular validation narrative: Reuters reported a 439% year-to-date gain through June. But in July the portfolio value fell 67%; most of the public-equity book was sold, much of it to Citadel, and the firm removed all leverage. In the investor letter seen by Reuters, Aschenbrenner acknowledged that the fund had come too close to permanent capital impairment and that it had let investors down that month.
This is precisely why the case should not be reduced to personal ridicule. It is an unusually clear experiment in epistemic financialization. A thesis about the future of intelligence generated financial authority; strong early returns increased the authority of the thesis; that authority supported larger exposure; and when the trade reversed, financial leverage created forced selling independent of whether the long-term AI thesis was ultimately correct.
The most important unanswered questions are not criminal ones. Public disclosures do not reveal the complete investor base, fee structure, month-by-month gross and net exposure, derivatives book, financing terms, collateral haircuts, counterparty concentrations or internal risk limits. Those are necessary for a genuine forensic evaluation. Reuters provides some of this picture through sources, but not enough to conclude fraud.
OpenAI and Microsoft: from public-benefit promise to capital-compute symbiosis. OpenAI is the paradigmatic privatization case because the transformation is documented by the organization itself. The 2015 launch presented a nonprofit intended to ensure that AI would benefit humanity and announced $1 billion of commitments. OpenAI later disclosed that actual nonprofit contributions were far below the headline commitment figure.
By 2019 OpenAI concluded that nonprofit financing could not provide the scale required and created a capped-profit company. Microsoft simultaneously supplied a $1 billion investment and Azure supercomputing partnership. By 2023 Azure was described as OpenAI’s exclusive cloud provider, powering all OpenAI workloads. Microsoft’s 2025 annual report described reciprocal revenue sharing, IP rights, API exclusivity and a right of first refusal on new capacity needs.
The 2025 recapitalization loosened aspects of that relationship while maintaining major strategic ties. Microsoft said its stake after recapitalization was worth approximately $135 billion and represented roughly 27% on an as-converted diluted basis. OpenAI remained controlled by its nonprofit Foundation through OpenAI Group PBC.
The analytical point is not that nonprofit language was necessarily fraudulent. The point is that the institution discovered that pursuing its universal mission required participation in exactly the private capital markets, hyperscale infrastructure and strategic partnerships from which control questions arise.
Its March 2026 $122 billion round makes that contradiction structural rather than rhetorical. The capital list spans hyperscalers, chip firms, sovereign-linked money, leading venture firms, traditional asset managers, private-equity firms and a major public university investment office. Intelligence had become an investable asset held indirectly across a remarkably broad segment of global institutional capital.
Anthropic: safety governance inside hyper-capitalization. Anthropic is an especially important test because it cannot easily be dismissed as a company that ignored institutional safety. Its Long-Term Benefit Trust was explicitly engineered as a financially disinterested governance body eventually capable of selecting and removing a majority of directors. Anthropic itself calls the arrangement an experiment and has not presented it as a proven template.
Yet its financing trajectory demonstrates a fundamental constraint: responsible governance does not eliminate the economics of frontier scaling. Amazon’s initial commitments reached $8 billion, followed in April 2026 by another $5 billion investment with up to $20 billion more possible. Google-related TPU arrangements reached tens of billions in expected value and multi-gigawatt capacity. Microsoft and Nvidia subsequently announced investment commitments of up to $5 billion and $10 billion.
Then private valuations accelerated from $380 billion in February 2026 to $965 billion in May. Anthropic reported annualized run-rate revenue above $47 billion by May, but run-rate revenue is an extrapolation from current business activity, not audited full-year realized revenue. Its June 2026 confidential draft S-1 means the safety institution is now potentially transitioning into public-market governance as well.
The case therefore refutes a simplistic binary in which “safety” and “capitalism” are opposites. A company can sincerely develop elaborate safety institutions while simultaneously becoming one of the most highly valued private companies ever created. The relevant investigative question becomes whether governance safeguards retain power when they conflict materially with capital, competitive or geopolitical incentives. The LTBT experiment is too young for the archive to answer that conclusively.
Open Philanthropy, Effective Altruism, and FTX: privately financed epistemic infrastructure. The Open Phil/Coefficient story is not principally one of corporate profit. It concerns who finances the production of ideas that governments later encounter as expertise.
Open Phil’s own 2020 account says it sought to support both AI-governance research and AI-governance “practice and influence.” Its historical portfolio included grants related to Georgetown CSET, the Wilson Center, CNAS, RAND, CSIS, FHI/GovAI, FLI, 80,000 Hours and OpenAI. Its present program describes hundreds of grants covering governance, technical safety, capacity building and professional networks.
This is neither clandestine nor inherently improper. Philanthropy has always built academic and policy fields. What makes the AI case historically significant is that a small private donor ecosystem helped build policy capacity around a technology whose future it simultaneously forecast as extraordinarily consequential.
The 2016 Karnofsky essay is revealing precisely because of its humility. It said there was no basis for confidence about timelines, acknowledged historical false AI projections, and nevertheless concluded that a ≥10% chance over twenty years was high enough to warrant substantial action. That is a coherent expected-value argument. But it creates a governance question: a privately selected probability can justify the creation of institutions that later acquire public influence even before society reaches consensus about the underlying premise.
The FTX episode exposes another vulnerability. The Future Fund’s crypto-derived resources funded longtermist and AI-risk-related work, then abruptly disappeared when FTX collapsed. Reuters reported that its program had ambitions to spend between $100 million and $1 billion and that archived materials showed extensive grants and investments before the collapse. Archived Future Fund records include AI-safety projects.
The relevant lesson is not “AI safety was a crypto scam.” It is almost the reverse: research institutions can become economically dependent on fortunes whose origins are epistemically unrelated to the validity of the research being financed. The collapse of the donor does not prove the grantee’s work false; the donor’s wealth nevertheless influences what research exists.
CoreWeave and the infrastructure complex: intelligence becomes collateral. CoreWeave demonstrates the transition from software mythology into hard financial obligations. Its rapid revenue growth was accompanied by billions in GPU and infrastructure investment, substantial losses, expensive debt and enormous customer commitments. Its original S-1 reported major revenue dependence on Microsoft; later filings show OpenAI becoming an important future customer.
The important innovation here is financial. A GPU cluster has physical existence and resale value, but its profitability depends on future demand for computation. Lenders, infrastructure funds and investors are therefore underwriting an implicit forecast about AI usage.
Aschenbrenner’s “trillion-dollar cluster” language can consequently be understood in two ways. It is a technological prediction, but it is also a blueprint for an asset class. Once enough investors believe compute scarcity will persist, the prediction can mobilize debt and equity that cause additional clusters to be built—partially making the infrastructure component of the forecast self-fulfilling.
This does not mean that capability forecasts become self-fulfilling in the same way. Capital can build more GPUs; it cannot guarantee that additional GPUs produce AGI. That distinction is fundamental.
Democratization, accountability, and what an alternative would require“Democratization of intelligence” cannot mean only giving everyone access to a chatbot owned and remotely governed by the same three or four companies. That would democratize consumption of intelligence, not control over it.
A stronger definition has at least five dimensions: access to capable models; ability to inspect or independently evaluate them; ability to modify or locally govern them; access to sufficient compute to build alternatives; and legal/political ability to contest decisions made through AI systems.
Open-weight models therefore matter, but they are not enough. The U.S. NTIA concluded in 2024 that widely available model weights can broaden access for researchers, small companies and developers and recommended monitoring rather than immediately restricting open models. Yet the same NTIA analysis observed that open weights alone were unlikely to radically alter the frontier foundation-model market because compute and other resource constraints remain formidable.
The archival evidence suggests that democratization should consequently target the entire stack, not merely licensing.
Public and shared compute infrastructure. Universities, independent laboratories, nonprofit organizations and small firms need meaningful access to accelerator clusters. When frontier computation requires tens or hundreds of millions of dollars—or eventually billions—formal permission to download model weights cannot equalize capacity. Government-backed compute commons, transparent allocation rules and cross-university facilities would create independent technical centers able to verify frontier claims rather than relying entirely on company access. The U.S. government itself has acknowledged that universities cannot presently match the supercomputing resources available to frontier firms.
Interoperability and cloud portability. The FTC’s finding that cloud/model partnerships can increase technical and contractual switching costs suggests a straightforward accountability agenda: require transparent portability terms for qualifying frontier providers, prevent investment contracts from foreclosing multi-cloud competition where antitrust law permits intervention, and scrutinize arrangements in which invested capital is contractually recycled into the investor’s own cloud services.
A forecast registry. AI governance needs something analogous to a financial prospectus for historically consequential technological claims. When laboratory executives, funds or policy organizations make predictions used to justify billions of dollars or sweeping regulation, an independent archive should record: exact wording, date, probability, operational definition, horizon, measurable threshold, subsequent revisions and final resolution. Forecasts could remain speculative; what disappears is the ability to move the goalposts silently.
Such a registry would have made the distinction between “AGI by 2027 is plausible,” “powerful AI could arrive as early as 2026,” and “≥10% chance of transformative AI by 2036” immediately visible. Those statements have profoundly different epistemic content.
Financial disclosure proportional to systemic narrative influence. Hedge funds already file selected information, but Form 13F does not reveal short positions, many derivatives, real-time leverage, private holdings or financing terms. Situational Awareness’s Q1 filing told the public that its listed long positions were worth $13.677 billion; it did not tell observers enough to reconstruct the risk architecture that produced July’s forced deleveraging. Regulators should examine whether exceptionally large concentrated thematic funds require more timely aggregate leverage and counterparty reporting to supervisors—not necessarily to the public at position level, where disclosure itself can create market-instability risks.
Separate philanthropic agenda-setting from undisclosed policy influence. Private philanthropy has a legitimate role in financing neglected research. But grant databases should make funding provenance, subgrants, board overlaps, fellowships, government secondments and major donor dependencies machine-readable. Coefficient/Open Phil already publishes substantial grant information; that is a useful starting point. The objective is not to prohibit private ideas from reaching government. It is to let the public see the network through which expertise was produced.
Conflict-of-interest disclosure for AI governance. Participants in governmental advisory panels, standards bodies and publicly funded AI-safety processes should disclose material equity, carried-interest, venture-fund, consulting, cloud-contract and major philanthropic relationships involving companies affected by their recommendations. This matters because the same institutions increasingly recur as investors, model suppliers and governance experts. The dense cross-investment network documented above makes ordinary disclosure rules more important, not less.
Safety regulation with competition-impact analysis. Frontier safety obligations may be justified, particularly for genuinely catastrophic-risk capabilities. But every major fixed-cost rule should receive a parallel assessment: can only incumbent companies afford compliance; does the rule implicitly require access to proprietary benchmarks; does it lock in a compute-centric architecture; does it disadvantage open alternatives without evidence that openness creates the relevant marginal risk? The EU’s systemic-risk framework and NTIA’s open-weight work show two complementary approaches that can be evaluated together rather than treated ideologically.
Independent evaluations rather than company-defined intelligence. The organization selling a model should not be the sole authority defining whether that model is “AGI,” “PhD-level,” “safe,” “aligned,” “superhuman” or economically transformative. Benchmark selection itself affects valuation and policy narratives. Independent public-interest evaluation institutes should preserve model versions, prompts, scaffolding, benchmark contamination controls and economic-task data so historical claims remain reproducible.
Beneficial public ownership of infrastructure. The alternative to purely private AI need not be nationalizing AI companies. Public pension funds, universities and governments already provide capital to the sector directly and indirectly. Public investment could more deliberately acquire rights to compute capacity, open research outputs, interoperability, public-interest licensing or revenue participation rather than simply socializing infrastructure costs while privatizing the resulting intellectual property.
Antitrust should focus on the stack rather than model count. Ten model developers do not constitute meaningful competition if all depend on two chip architectures, three cloud providers and the same financing channels. The FTC’s work correctly emphasizes inputs, switching costs and cross-layer information. Academic work on foundation-model concentration similarly suggests that scale economics can change market structure before conventional monopoly metrics make concentration obvious.
Preserve open weights where risk permits. NTIA’s finding that open models can broaden access is important because democratization requires the possibility of intelligence operating outside remote corporate control. This does not imply publishing every frontier model regardless of demonstrated biological, cyber or autonomous-action risk. A defensible policy asks for empirical evidence of marginal risk before restricting openness, and uses the least concentrated governance mechanism capable of mitigating that risk.
Audit announced investment separately from realized investment. Stargate’s $500 billion figure, Anthropic’s infrastructure plans and multi-gigawatt announcements should be tracked through permits, construction, energized capacity and actual capital expenditures. The archive should not allow announced dollars to become historical facts before they are spent.
Above all, democratization requires rejecting a category error that repeatedly appears in the First Age of AI:
Technical intelligence does not automatically confer epistemic authority, financial wisdom, political legitimacy, or moral authority.
A laboratory can build an extraordinarily capable model and make poor macroeconomic forecasts. A hedge fund can earn extraordinary returns without proving its theory of intelligence. A safety researcher can identify a genuine risk without being entitled to determine public policy. A billionaire can finance important science without acquiring special insight into humanity’s future. A government can legitimately protect national security without turning private corporate success into synonymous national success.
This separation of competencies is the institutional foundation that an era of abundant intelligence will require.
Evidence appendix and methodological assessment
The sources below are prioritized according to evidentiary weight. Tier A means primary filings, government findings or first-party corporate documents directly establishing the relevant fact. Tier B means high-quality independent reporting, generally Reuters and comparable investigative journalism. Tier C means academic interpretation or contextual analysis. Forecast essays are treated as primary evidence of what the author predicted, not evidence that the prediction is correct.
A serious archive should therefore resist a Nostradamus-style hit/miss game. Instead, it should preserve each claim as originally stated, with its definition, probability, date, horizon and measurable criteria, then prevent retrospective rewriting.
Safety rhetoric deserves the same analytical discipline. It can create genuine social value and simultaneously create competitive advantages. A frontier laboratory that has extensive safety teams, evaluation infrastructure, lobbying capacity and compliance departments may be better positioned to satisfy sophisticated regulation than a small entrant. The EU AI Act, for example, presumes systemic-risk status for GPAI models above a training-compute threshold of (10^{25}) FLOP, while recognizing that only a handful of companies currently develop such models.
That does not imply the threshold was created to protect incumbents. The same threshold may be a rational attempt to concentrate regulatory scrutiny where risk is greatest. The dual effect is what matters:
A safety rule can simultaneously reduce danger and increase fixed costs of entry.
The archival question should therefore be: Who proposed a rule, what evidence justified it, who bears its compliance cost, and who gains competitive advantage from it? Intent need not be assumed.
Political access presents a similar distinction. Sam Altman’s 2023 Senate testimony is direct evidence that frontier executives obtained high-level legislative access. Subsequent White House strategies explicitly relied on private-sector frontier expertise and infrastructure. Yet the FTC simultaneously investigated those same firms’ market relationships. The evidence therefore describes interdependence and asymmetric access, not a simple story in which the state has been completely captured.
Academic work increasingly supports the structural concern. Research on “concentrating intelligence” examines how foundation-model scaling can drive concentrated market structure, while governance scholarship has documented the dominance of large developed-country actors and risk-management framings in emerging AI institutions. These studies do not prove conspiracy; they strengthen the case that concentration is an economic and governance property worth treating independently of individual motives.
Case studies: where money, myth, and power intersect
Situational Awareness: prophecy becomes a portfolio. The importance of Leopold Aschenbrenner’s case is not that his essays prove or disprove AGI. It is the unusually direct connection between a public technological worldview and a large financial vehicle. His 2024 series argued that AGI around 2027 was plausible, that algorithmic progress plus compute scaling would produce a major capability jump, and that industrial mobilization around enormous clusters would follow. An SEC Form D dated September 23, 2024 identifies Situational Awareness Partners LP as a hedge fund and Aschenbrenner as a related executive person.
By March 31, 2026 its SEC 13F disclosed 42 listed positions worth $13.6767 billion. Reuters later reported that the fund had about $20 billion under management before its July crisis and that public holdings included Broadcom, Intel and CoreWeave, alongside a private Anthropic investment. SEC beneficial-ownership filings independently confirm Aschenbrenner’s control relationships and Carl Shulman’s co-portfolio-management role.
The first half of 2026 then produced a spectacular validation narrative: Reuters reported a 439% year-to-date gain through June. But in July the portfolio value fell 67%; most of the public-equity book was sold, much of it to Citadel, and the firm removed all leverage. In the investor letter seen by Reuters, Aschenbrenner acknowledged that the fund had come too close to permanent capital impairment and that it had let investors down that month.
This is precisely why the case should not be reduced to personal ridicule. It is an unusually clear experiment in epistemic financialization. A thesis about the future of intelligence generated financial authority; strong early returns increased the authority of the thesis; that authority supported larger exposure; and when the trade reversed, financial leverage created forced selling independent of whether the long-term AI thesis was ultimately correct.
The most important unanswered questions are not criminal ones. Public disclosures do not reveal the complete investor base, fee structure, month-by-month gross and net exposure, derivatives book, financing terms, collateral haircuts, counterparty concentrations or internal risk limits. Those are necessary for a genuine forensic evaluation. Reuters provides some of this picture through sources, but not enough to conclude fraud.
OpenAI and Microsoft: from public-benefit promise to capital-compute symbiosis. OpenAI is the paradigmatic privatization case because the transformation is documented by the organization itself. The 2015 launch presented a nonprofit intended to ensure that AI would benefit humanity and announced $1 billion of commitments. OpenAI later disclosed that actual nonprofit contributions were far below the headline commitment figure.
By 2019 OpenAI concluded that nonprofit financing could not provide the scale required and created a capped-profit company. Microsoft simultaneously supplied a $1 billion investment and Azure supercomputing partnership. By 2023 Azure was described as OpenAI’s exclusive cloud provider, powering all OpenAI workloads. Microsoft’s 2025 annual report described reciprocal revenue sharing, IP rights, API exclusivity and a right of first refusal on new capacity needs.
The 2025 recapitalization loosened aspects of that relationship while maintaining major strategic ties. Microsoft said its stake after recapitalization was worth approximately $135 billion and represented roughly 27% on an as-converted diluted basis. OpenAI remained controlled by its nonprofit Foundation through OpenAI Group PBC.
The analytical point is not that nonprofit language was necessarily fraudulent. The point is that the institution discovered that pursuing its universal mission required participation in exactly the private capital markets, hyperscale infrastructure and strategic partnerships from which control questions arise.
Its March 2026 $122 billion round makes that contradiction structural rather than rhetorical. The capital list spans hyperscalers, chip firms, sovereign-linked money, leading venture firms, traditional asset managers, private-equity firms and a major public university investment office. Intelligence had become an investable asset held indirectly across a remarkably broad segment of global institutional capital.
Anthropic: safety governance inside hyper-capitalization. Anthropic is an especially important test because it cannot easily be dismissed as a company that ignored institutional safety. Its Long-Term Benefit Trust was explicitly engineered as a financially disinterested governance body eventually capable of selecting and removing a majority of directors. Anthropic itself calls the arrangement an experiment and has not presented it as a proven template.
Yet its financing trajectory demonstrates a fundamental constraint: responsible governance does not eliminate the economics of frontier scaling. Amazon’s initial commitments reached $8 billion, followed in April 2026 by another $5 billion investment with up to $20 billion more possible. Google-related TPU arrangements reached tens of billions in expected value and multi-gigawatt capacity. Microsoft and Nvidia subsequently announced investment commitments of up to $5 billion and $10 billion.
Then private valuations accelerated from $380 billion in February 2026 to $965 billion in May. Anthropic reported annualized run-rate revenue above $47 billion by May, but run-rate revenue is an extrapolation from current business activity, not audited full-year realized revenue. Its June 2026 confidential draft S-1 means the safety institution is now potentially transitioning into public-market governance as well.
The case therefore refutes a simplistic binary in which “safety” and “capitalism” are opposites. A company can sincerely develop elaborate safety institutions while simultaneously becoming one of the most highly valued private companies ever created. The relevant investigative question becomes whether governance safeguards retain power when they conflict materially with capital, competitive or geopolitical incentives. The LTBT experiment is too young for the archive to answer that conclusively.
Open Philanthropy, Effective Altruism, and FTX: privately financed epistemic infrastructure. The Open Phil/Coefficient story is not principally one of corporate profit. It concerns who finances the production of ideas that governments later encounter as expertise.
Open Phil’s own 2020 account says it sought to support both AI-governance research and AI-governance “practice and influence.” Its historical portfolio included grants related to Georgetown CSET, the Wilson Center, CNAS, RAND, CSIS, FHI/GovAI, FLI, 80,000 Hours and OpenAI. Its present program describes hundreds of grants covering governance, technical safety, capacity building and professional networks.
This is neither clandestine nor inherently improper. Philanthropy has always built academic and policy fields. What makes the AI case historically significant is that a small private donor ecosystem helped build policy capacity around a technology whose future it simultaneously forecast as extraordinarily consequential.
The 2016 Karnofsky essay is revealing precisely because of its humility. It said there was no basis for confidence about timelines, acknowledged historical false AI projections, and nevertheless concluded that a ≥10% chance over twenty years was high enough to warrant substantial action. That is a coherent expected-value argument. But it creates a governance question: a privately selected probability can justify the creation of institutions that later acquire public influence even before society reaches consensus about the underlying premise.
The FTX episode exposes another vulnerability. The Future Fund’s crypto-derived resources funded longtermist and AI-risk-related work, then abruptly disappeared when FTX collapsed. Reuters reported that its program had ambitions to spend between $100 million and $1 billion and that archived materials showed extensive grants and investments before the collapse. Archived Future Fund records include AI-safety projects.
The relevant lesson is not “AI safety was a crypto scam.” It is almost the reverse: research institutions can become economically dependent on fortunes whose origins are epistemically unrelated to the validity of the research being financed. The collapse of the donor does not prove the grantee’s work false; the donor’s wealth nevertheless influences what research exists.
CoreWeave and the infrastructure complex: intelligence becomes collateral. CoreWeave demonstrates the transition from software mythology into hard financial obligations. Its rapid revenue growth was accompanied by billions in GPU and infrastructure investment, substantial losses, expensive debt and enormous customer commitments. Its original S-1 reported major revenue dependence on Microsoft; later filings show OpenAI becoming an important future customer.
The important innovation here is financial. A GPU cluster has physical existence and resale value, but its profitability depends on future demand for computation. Lenders, infrastructure funds and investors are therefore underwriting an implicit forecast about AI usage.
Aschenbrenner’s “trillion-dollar cluster” language can consequently be understood in two ways. It is a technological prediction, but it is also a blueprint for an asset class. Once enough investors believe compute scarcity will persist, the prediction can mobilize debt and equity that cause additional clusters to be built—partially making the infrastructure component of the forecast self-fulfilling.
This does not mean that capability forecasts become self-fulfilling in the same way. Capital can build more GPUs; it cannot guarantee that additional GPUs produce AGI. That distinction is fundamental.
Democratization, accountability, and what an alternative would require“Democratization of intelligence” cannot mean only giving everyone access to a chatbot owned and remotely governed by the same three or four companies. That would democratize consumption of intelligence, not control over it.
A stronger definition has at least five dimensions: access to capable models; ability to inspect or independently evaluate them; ability to modify or locally govern them; access to sufficient compute to build alternatives; and legal/political ability to contest decisions made through AI systems.
Open-weight models therefore matter, but they are not enough. The U.S. NTIA concluded in 2024 that widely available model weights can broaden access for researchers, small companies and developers and recommended monitoring rather than immediately restricting open models. Yet the same NTIA analysis observed that open weights alone were unlikely to radically alter the frontier foundation-model market because compute and other resource constraints remain formidable.
The archival evidence suggests that democratization should consequently target the entire stack, not merely licensing.
Public and shared compute infrastructure. Universities, independent laboratories, nonprofit organizations and small firms need meaningful access to accelerator clusters. When frontier computation requires tens or hundreds of millions of dollars—or eventually billions—formal permission to download model weights cannot equalize capacity. Government-backed compute commons, transparent allocation rules and cross-university facilities would create independent technical centers able to verify frontier claims rather than relying entirely on company access. The U.S. government itself has acknowledged that universities cannot presently match the supercomputing resources available to frontier firms.
Interoperability and cloud portability. The FTC’s finding that cloud/model partnerships can increase technical and contractual switching costs suggests a straightforward accountability agenda: require transparent portability terms for qualifying frontier providers, prevent investment contracts from foreclosing multi-cloud competition where antitrust law permits intervention, and scrutinize arrangements in which invested capital is contractually recycled into the investor’s own cloud services.
A forecast registry. AI governance needs something analogous to a financial prospectus for historically consequential technological claims. When laboratory executives, funds or policy organizations make predictions used to justify billions of dollars or sweeping regulation, an independent archive should record: exact wording, date, probability, operational definition, horizon, measurable threshold, subsequent revisions and final resolution. Forecasts could remain speculative; what disappears is the ability to move the goalposts silently.
Such a registry would have made the distinction between “AGI by 2027 is plausible,” “powerful AI could arrive as early as 2026,” and “≥10% chance of transformative AI by 2036” immediately visible. Those statements have profoundly different epistemic content.
Financial disclosure proportional to systemic narrative influence. Hedge funds already file selected information, but Form 13F does not reveal short positions, many derivatives, real-time leverage, private holdings or financing terms. Situational Awareness’s Q1 filing told the public that its listed long positions were worth $13.677 billion; it did not tell observers enough to reconstruct the risk architecture that produced July’s forced deleveraging. Regulators should examine whether exceptionally large concentrated thematic funds require more timely aggregate leverage and counterparty reporting to supervisors—not necessarily to the public at position level, where disclosure itself can create market-instability risks.
Separate philanthropic agenda-setting from undisclosed policy influence. Private philanthropy has a legitimate role in financing neglected research. But grant databases should make funding provenance, subgrants, board overlaps, fellowships, government secondments and major donor dependencies machine-readable. Coefficient/Open Phil already publishes substantial grant information; that is a useful starting point. The objective is not to prohibit private ideas from reaching government. It is to let the public see the network through which expertise was produced.
Conflict-of-interest disclosure for AI governance. Participants in governmental advisory panels, standards bodies and publicly funded AI-safety processes should disclose material equity, carried-interest, venture-fund, consulting, cloud-contract and major philanthropic relationships involving companies affected by their recommendations. This matters because the same institutions increasingly recur as investors, model suppliers and governance experts. The dense cross-investment network documented above makes ordinary disclosure rules more important, not less.
Safety regulation with competition-impact analysis. Frontier safety obligations may be justified, particularly for genuinely catastrophic-risk capabilities. But every major fixed-cost rule should receive a parallel assessment: can only incumbent companies afford compliance; does the rule implicitly require access to proprietary benchmarks; does it lock in a compute-centric architecture; does it disadvantage open alternatives without evidence that openness creates the relevant marginal risk? The EU’s systemic-risk framework and NTIA’s open-weight work show two complementary approaches that can be evaluated together rather than treated ideologically.
Independent evaluations rather than company-defined intelligence. The organization selling a model should not be the sole authority defining whether that model is “AGI,” “PhD-level,” “safe,” “aligned,” “superhuman” or economically transformative. Benchmark selection itself affects valuation and policy narratives. Independent public-interest evaluation institutes should preserve model versions, prompts, scaffolding, benchmark contamination controls and economic-task data so historical claims remain reproducible.
Beneficial public ownership of infrastructure. The alternative to purely private AI need not be nationalizing AI companies. Public pension funds, universities and governments already provide capital to the sector directly and indirectly. Public investment could more deliberately acquire rights to compute capacity, open research outputs, interoperability, public-interest licensing or revenue participation rather than simply socializing infrastructure costs while privatizing the resulting intellectual property.
Antitrust should focus on the stack rather than model count. Ten model developers do not constitute meaningful competition if all depend on two chip architectures, three cloud providers and the same financing channels. The FTC’s work correctly emphasizes inputs, switching costs and cross-layer information. Academic work on foundation-model concentration similarly suggests that scale economics can change market structure before conventional monopoly metrics make concentration obvious.
Preserve open weights where risk permits. NTIA’s finding that open models can broaden access is important because democratization requires the possibility of intelligence operating outside remote corporate control. This does not imply publishing every frontier model regardless of demonstrated biological, cyber or autonomous-action risk. A defensible policy asks for empirical evidence of marginal risk before restricting openness, and uses the least concentrated governance mechanism capable of mitigating that risk.
Audit announced investment separately from realized investment. Stargate’s $500 billion figure, Anthropic’s infrastructure plans and multi-gigawatt announcements should be tracked through permits, construction, energized capacity and actual capital expenditures. The archive should not allow announced dollars to become historical facts before they are spent.
Above all, democratization requires rejecting a category error that repeatedly appears in the First Age of AI:
Technical intelligence does not automatically confer epistemic authority, financial wisdom, political legitimacy, or moral authority.
A laboratory can build an extraordinarily capable model and make poor macroeconomic forecasts. A hedge fund can earn extraordinary returns without proving its theory of intelligence. A safety researcher can identify a genuine risk without being entitled to determine public policy. A billionaire can finance important science without acquiring special insight into humanity’s future. A government can legitimately protect national security without turning private corporate success into synonymous national success.
This separation of competencies is the institutional foundation that an era of abundant intelligence will require.
Evidence appendix and methodological assessment
The sources below are prioritized according to evidentiary weight. Tier A means primary filings, government findings or first-party corporate documents directly establishing the relevant fact. Tier B means high-quality independent reporting, generally Reuters and comparable investigative journalism. Tier C means academic interpretation or contextual analysis. Forecast essays are treated as primary evidence of what the author predicted, not evidence that the prediction is correct.
The evidence base has several important holes.
First, private fund letters are mostly unavailable. Reuters obtained portions of Situational Awareness communications, but a complete archival analysis would require original investor letters from inception, audited returns, administrator records, prime-broker statements and offering documents. Without these, claims about intentional deception would be irresponsible.
Second, private frontier-company financing remains opaque. Announced round size and valuation are visible; liquidation preferences, side letters, investor protections, compute-purchase dependencies, secondary transactions and derivative arrangements are often not.
Third, the policy network is only partially observable. Published grant databases and board lists are unusually useful, but private meetings, informal advice, draft-policy circulation and donor conversations are difficult to reconstruct. This limits any strong claim of regulatory capture.
Fourth, media amplification has not been quantitatively measured here. The report establishes high-profile statements, major press coverage and investor attention, but it does not include a complete corpus analysis of newspapers, television, podcasts or social media. A publication-grade follow-on investigation should measure repetition of specific AGI dates, extinction claims, valuation narratives and “race” metaphors over time.
Fifth, forecast performance requires prospective rather than retrospective scoring. The archive should freeze claims now. Aschenbrenner’s 2027 prediction cannot responsibly be classified as failed in August 2026; Karnofsky’s 2036 probability is even further from resolution; Amodei’s “as early as 2026” formulation is not equivalent to a deadline.
The overarching archival conclusion is therefore narrower—and stronger—than an accusation of fraud:
During the First Age of AI, society allowed claims about a radically uncertain future form of intelligence to become present-day financial assets, corporate valuations, philanthropic priorities, infrastructure commitments, policy categories and sources of political authority.
Some of those claims may ultimately prove remarkably accurate. Others may look naïve in retrospect. That uncertainty is not a reason to dismiss the technology. It is precisely why the institutions surrounding it require scrutiny.
The historical danger is not merely that someone might make a bad AGI forecast.
It is that forecasting intelligence can become a way of acquiring power over intelligence before the forecast has been tested.
The deepest democratization principle that follows from the archive is consequently not “everyone must agree that AI is safe,” nor “everyone must receive the same model,” nor even “all AI must be open source.”
It is this:
No corporation, investor, government, laboratory, philanthropic network, safety movement, ideology, or individual should acquire unquestionable authority merely by claiming privileged knowledge of where intelligence is going.
The appropriate answer to intelligence is not another mythology of certainty.
It is pluralism, independent verification, transparent finance, contestable governance, open inquiry, distributed technical capacity, and permanent permission to ask who benefits from the future being predicted.
First, private fund letters are mostly unavailable. Reuters obtained portions of Situational Awareness communications, but a complete archival analysis would require original investor letters from inception, audited returns, administrator records, prime-broker statements and offering documents. Without these, claims about intentional deception would be irresponsible.
Second, private frontier-company financing remains opaque. Announced round size and valuation are visible; liquidation preferences, side letters, investor protections, compute-purchase dependencies, secondary transactions and derivative arrangements are often not.
Third, the policy network is only partially observable. Published grant databases and board lists are unusually useful, but private meetings, informal advice, draft-policy circulation and donor conversations are difficult to reconstruct. This limits any strong claim of regulatory capture.
Fourth, media amplification has not been quantitatively measured here. The report establishes high-profile statements, major press coverage and investor attention, but it does not include a complete corpus analysis of newspapers, television, podcasts or social media. A publication-grade follow-on investigation should measure repetition of specific AGI dates, extinction claims, valuation narratives and “race” metaphors over time.
Fifth, forecast performance requires prospective rather than retrospective scoring. The archive should freeze claims now. Aschenbrenner’s 2027 prediction cannot responsibly be classified as failed in August 2026; Karnofsky’s 2036 probability is even further from resolution; Amodei’s “as early as 2026” formulation is not equivalent to a deadline.
The overarching archival conclusion is therefore narrower—and stronger—than an accusation of fraud:
During the First Age of AI, society allowed claims about a radically uncertain future form of intelligence to become present-day financial assets, corporate valuations, philanthropic priorities, infrastructure commitments, policy categories and sources of political authority.
Some of those claims may ultimately prove remarkably accurate. Others may look naïve in retrospect. That uncertainty is not a reason to dismiss the technology. It is precisely why the institutions surrounding it require scrutiny.
The historical danger is not merely that someone might make a bad AGI forecast.
It is that forecasting intelligence can become a way of acquiring power over intelligence before the forecast has been tested.
The deepest democratization principle that follows from the archive is consequently not “everyone must agree that AI is safe,” nor “everyone must receive the same model,” nor even “all AI must be open source.”
It is this:
No corporation, investor, government, laboratory, philanthropic network, safety movement, ideology, or individual should acquire unquestionable authority merely by claiming privileged knowledge of where intelligence is going.
The appropriate answer to intelligence is not another mythology of certainty.
It is pluralism, independent verification, transparent finance, contestable governance, open inquiry, distributed technical capacity, and permanent permission to ask who benefits from the future being predicted.