Horizon Accord | AI Industry Accountability | Enron AI Parallel | Machine Learning
The Emperor's New Algorithm
Enron built a template. The AI industry didn't invent it — they inherited it. And the fight over who gets to regulate it is about something much older than artificial intelligence.
Enron didn't fail because of a few corrupt executives. It failed because its entire business model depended on everyone agreeing not to look too closely at what was actually being sold. The product was confidence. The revenue was faith in future value. The accounting was a sophisticated mechanism for converting that faith into present-tense balance sheets. When the gap between the story and the substance became undeniable, the structure collapsed in weeks.
That pattern did not die in 2001. It became a template. And the AI industry — not one company, not one lab, but the sector as a whole — is running a version of it right now, at civilizational scale, with better press releases and considerably more idealism in the marketing copy.
What follows is Part One of a two-part analysis. This piece establishes the structural parallel: how the Enron template maps onto the AI industry's governance model, financing architecture, regulatory strategy, and the frame war now playing out in Congress. Part Two examines the Sanders-AOC AI Data Center Moratorium Act — and whether it addresses the actual problem or displaces attention from it.
The Demo and the Gap
In May 2026, the CEO of an AI biotech company called BigHat Biosciences told STAT News reporter Brittany Trang something worth sitting with. She can design a protein in twenty minutes. That part is real — the AI works, the speed is genuine, the capability exists. But that's not the point, she said. The downstream tests are still slow. Still expensive. "That's where the hard work is, is making the actual drug." Her company has partnerships with Johnson & Johnson, Merck, Amgen, AbbVie, and Lilly. She was not a critic of AI in drug development. She was a practitioner describing what AI actually does and does not change.
What it does not change is the part that costs a billion dollars and takes a decade. What it does not change is the 90 percent failure rate in clinical trials. What it changes is the speed of the early demo — the part you can show an investor, put in a press release, used to justify a valuation. The gap between the demo and the delivery is not a flaw in the technology. It is a feature of how the technology is being sold.
This is the entry point for everything that follows. Not because AI drug development is the whole story, but because it is the most legible version of a pattern running across the entire AI industry. The demo is real. The substance is real. The gap between what is being sold and what is being delivered — and the systematic effort to ensure that gap is never officially measured — is also real. That gap has a name. It has a structure. And it has a precedent.
The Enron Template
Enron was founded in 1985 as a natural gas pipeline company. By the late 1990s it had transformed into something harder to categorize: an energy trading firm whose assets were increasingly abstract, whose revenue was largely mark-to-market — booking the projected future value of contracts as present income — and whose internal culture rewarded confidence over competence. It was named Fortune's "Most Innovative Company" six years running. It filed for bankruptcy in December 2001, the largest corporate bankruptcy in American history at the time.
The mechanism that made Enron possible was not fraud in the simple sense. It was a set of structural permissions — mark-to-market accounting, off-balance-sheet Special Purpose Vehicles, and a captured regulatory environment — that allowed the gap between narrative and substance to widen for years before anyone with institutional authority chose to close it. The auditors signed off. The analysts upgraded. The board nodded. The regulators deferred. Each actor, operating within the system as designed, had no individual incentive to be the one who asked the question that would stop the music.
Map that structure onto the current AI industry and the parallel is not metaphorical — it is architectural. Frontier AI labs are valued not on current revenue but on projected future capability: intelligence futures, priced today. The evidence is operational: deployment races that ship before safety evaluations are complete, valuation inflation untethered from revenue, benchmark theater dressed as scientific rigor. The safety frameworks they publish function like Enron's internal ethics codes — visible, earnest-sounding, and structurally subordinate to the commercial imperatives that govern day-to-day decisions. The analysts upgrade. The press covers the demos. The regulators are still writing the frameworks. Each individual actor has structural incentives not to be the one who asks whether the gap between story and substance is being measured by anyone with the authority to do anything about it.
As of October 2025, OpenAI completed a restructuring that valued the company at approximately $157 billion — built on projected AI capability, not current profitability. The for-profit arm converted to a Public Benefit Corporation; the nonprofit foundation retained equity and governance oversight but removed the profit caps that had been the original structural constraint on investor returns. The cap — once the mechanism that formally distinguished OpenAI's commercial activity from unconstrained extraction — was gone. OpenAI was founded as a nonprofit in 2015. The journey from that founding document to a $157 billion PBC with unlimited investor returns took ten years. Enron's journey from pipeline company to the largest bankruptcy in American history took sixteen.
The Nonprofit Launder
OpenAI, Anthropic, and DeepMind each began with explicit public-benefit framings. OpenAI launched as a nonprofit. Anthropic was founded as a Public Benefit Corporation by former OpenAI safety researchers who cited safety concerns as their reason for leaving. DeepMind, acquired by Google in 2014, was established with an independent ethics board and commitments that its AGI research would not be used for military applications — commitments that were subsequently modified. Each lab arrived with a story about why this time, this organization, this structure would be different.
This pattern is not coincidental. The nonprofit or mission-driven founding structure serves a specific function in frontier AI: it generates trust capital that can be drawn on during the period when the technology is unproven, the regulatory environment is undefined, and the commercial case has not yet been made. The mission statement functions as a form of pre-regulatory self-certification: trust us because we have pledged to be trustworthy. This does not require founders to be insincere. Structural incentives can redirect institutions long before individuals consciously recognize the transition.
"OpenAI was founded as a nonprofit, is today a nonprofit that oversees and controls the for-profit, and going forward will remain a nonprofit that oversees and controls the for-profit. That will not change." — Sam Altman, OpenAI CEO, May 2025
By October 2025, the restructuring was complete. The profit caps were removed entirely, replaced with a legal mandate to consider public benefit alongside shareholder interests. A PBC has no legal requirement to prioritize mission over profit. It is required to consider both. This is not the same thing. The PBC structure is now the standard governance form for frontier AI labs — OpenAI, Anthropic, X.AI. The structure is presented as a solution to the tension between mission and capital. What it actually does is formalize the coexistence of those interests without resolving the question of which prevails when they conflict.
The nonprofit launder is most effective during the period before meaningful regulation exists. Once what AI companies owe the public is answered by law rather than by mission statement, the governance value of the PBC structure diminishes. This may explain why the AI industry's regulatory strategy has focused so heavily on delaying, shaping, and fragmenting the regulatory environment rather than engaging with it on the merits.
Safety as Infrastructure
Enron had an ethics code. It was detailed, well-written, and distributed to all employees. The code did not constrain the behavior it was meant to govern — but it served an important function: it allowed executives, board members, and auditors to point to it when asked whether the company was operating responsibly. The code was infrastructure for the claim of responsibility, not a mechanism for enforcing it.
AI safety frameworks function analogously. Constitutional AI, responsible scaling policies, safety cards, red-teaming disclosures, model evaluations — these are real documents produced by real researchers doing real work. They are also structurally subordinate to the commercial imperatives that govern deployment decisions. When a safety team's findings conflict with a product launch timeline, the resolution of that conflict is governed by the organizational hierarchy and the competitive pressure the labs themselves have constructed, not by the safety framework.
The departure of OpenAI's safety team leadership in 2024 was documented publicly. Multiple researchers who left cited concerns about the relationship between safety work and commercial priorities. Steven Adler, a former OpenAI safety lead, described a structural condition in which safety mechanisms were downstream of commercial incentives — and in which the human oversight layer was becoming recursive and therefore increasingly difficult to verify as meaningful. These are not the claims of external critics. They are the observations of people who were inside the structure.
The safety apparatus serves a second function beyond internal governance: it is the primary mechanism through which AI companies engage the regulatory and policy environment. Published safety frameworks are the documents legislators read, policymakers cite, and congressional testimonies reference. They are simultaneously internal governance tools and external communications infrastructure — the interface between the industry and the regulatory environment the industry is actively trying to shape.
Who Regulates What — and Who Decides
Enron's regulatory environment was shaped by years of energy deregulation, culminating in FERC's restructuring of wholesale electricity markets in the late 1990s. Enron was a major participant in the lobbying effort that produced that deregulation. When the California energy crisis of 2000–2001 demonstrated what deregulated markets looked like under strategic manipulation, FERC's initial response was to deny that manipulation was occurring. The regulatory capture was not a conspiracy — it was the predictable outcome of an industry that had spent a decade building relationships with, funding campaigns of, and providing technical expertise to the very agencies responsible for overseeing it.
On December 11, 2025, President Trump signed Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence." The order directed the DOJ to establish an AI Litigation Task Force within 30 days to challenge state AI laws deemed inconsistent with administration policy. Congress had already rejected similar moratorium language twice — before H.R. 1, and again before the NDAA of 2026. The executive order arrived two days after a bipartisan coalition of 42 state attorneys general wrote to major AI companies urging improved safeguards for children and mitigation of harmful outputs.
The Enron parallel is precise. The goal is not to avoid regulation entirely — that would be politically unsustainable — but to ensure that whatever regulation exists is set at the federal level, shaped by industry input, and preemptive of the more materially specific, enforcement-capable regulation that states were beginning to develop. Federal preemption of state AI law accomplishes for the AI industry what FERC restructuring accomplished for Enron: it replaces accountability mechanisms that were close to the harm with oversight frameworks that are distant from it.
David Sacks, the Trump administration's AI and crypto czar, described the executive order as a necessary tool to dismantle the "patchwork" of state regulations, specifically naming Colorado's AI Act as "probably the most excessive." Colorado's law — one of the most substantive state AI accountability frameworks in the country — was subsequently delayed from its February 2026 effective date to June 2026. Utah narrowed its AI law and established safe harbors. The regulatory pressure that had been accumulating at the state level began to fragment.
The revolving door between industry and the agencies meant to regulate it is not a future risk in AI — it is already operational in the adjacent sector most relevant to AI's biggest commercial claim. On January 10, 2025, Patrizia Cavazzoni resigned as director of the FDA's Center for Drug Evaluation and Research — the office responsible for approving drugs — ten days before Trump took office. On February 24, 2025, Pfizer named her Chief Medical Officer. She had worked at Pfizer before joining the FDA in 2018. The person who ran the office that regulated Pfizer's drugs is now Pfizer's chief medical officer. In the months following her departure, the FDA under new leadership moved to require only one pivotal clinical trial instead of the customary two before drug approval, and championed the expanded use of AI in drug reviews. The gap between the demo and the delivery — the gap the BigHat CEO described so precisely — is now being evaluated by a regulatory apparatus whose leadership came from, and returned to, the industry it oversees.
This is what distant oversight looks like when it is already operational. Not captured regulators blocking reform — absent ones, replaced by the industry's own people, presiding over a framework that has been deliberately moved away from the enforcement mechanisms that were closest to the harm. The AI industry is not building toward this condition. It is inheriting it from a pharmaceutical sector where the pattern was already mature.
The Frame War
Every significant accountability battle has a frame war running alongside it — a contest over which description of the problem gets to be the official one, because the official description determines which solutions are legible and which are invisible. Enron's frame war, in its final months, was between "temporary liquidity problem" and "fundamental insolvency." The frame that won determined who got paid out first and who was left holding the collapsing paper.
In AI, the contest is between two competing descriptions of what the danger is. The first is the longtermist or existential frame: AI is dangerous because sufficiently advanced systems may eventually pose catastrophic or extinction-level risks. The primary regulatory priority, on this account, is ensuring AI development doesn't produce a misaligned superintelligence. The organizations most associated with this frame — MIRI, the Alignment Forum, the Effective Altruism ecosystem — have substantial institutional presence in policy circles and have shaped the language of AI safety discourse significantly. This frame has also attracted progressive political figures: Sanders flew to Berkeley to meet with alignment researchers ahead of introducing AI legislation.
The second frame is the material or class-harm frame: AI is dangerous now, to specific populations, through specific mechanisms — algorithmic bias in hiring and lending, worker displacement, surveillance infrastructure, data exploitation, and the concentration of economic power in a handful of firms. The primary regulatory priority, here, is accountability for present harms, not insurance against hypothetical future ones. This is the frame that produced the state AI laws the Trump administration is now working to dismantle. It is also the frame Sanders reaches for when he is not in Berkeley — when he is on the Senate floor describing tech billionaires as oligarchs and AI as a tool of class power.
The longtermist frame, whatever its merits as a description of future risk, functions structurally to defer present accountability. If the primary danger of AI is a hypothetical superintelligence ten or twenty years away, then the algorithmic systems deployed today — the hiring tools, the credit scoring models, the content moderation systems, the predictive policing infrastructure — exist in a kind of regulatory interim, pending the more important long-term safety conversation. The state laws the Trump administration is challenging were addressing the present-tense frame. The federal framework being built to replace them is organized, at least in part, around the future-tense one.
On March 25, 2026, Senator Sanders and Representative Ocasio-Cortez announced the Artificial Intelligence Data Center Moratorium Act — a federal moratorium on construction or upgrading of AI data centers with a power demand of 20 megawatts or more until Congress passes comprehensive AI legislation establishing safeguards for safety, privacy, civil rights, and worker protection. The Brookings Institution documented that by this point more than 100 local communities had enacted data center moratoriums of their own, and more than 300 state data-center bills had been filed in the first six weeks of 2026 alone. Sanders' statement returned to his home terrain — oligarchic power, working families, democratic oversight — rather than the extinction language of his Berkeley visit.
The bill arrived into a regulatory environment already shaped by the Trump executive order's federal preemption strategy. What it actually accomplishes — and whether it addresses the Enron-shaped problem at the center of the industry or displaces attention from it — is the question Part Two examines.
The Open Question
Enron's collapse did not produce the regulatory environment its victims needed. The Sarbanes-Oxley Act of 2002 addressed auditing and financial disclosure — the mechanisms of the fraud — without addressing the energy market deregulation that had made the fraud structurally possible. The accountability was real. The structural condition it responded to was adjacent to, not coextensive with, the structural condition that had enabled the harm.
The AI industry's Enron-shaped problem is not primarily located in its data centers. It is located in the gap between what is being sold — intelligence, safety, beneficial AI — and what is being delivered and to whom. A moratorium on data center construction addresses the physical infrastructure of that gap without touching its governance architecture, its financial structure, or its frame-capture strategy. Whether that is a sufficient intervention, a necessary first step, or a sophisticated misdirection depends on what you believe the actual problem to be.
Those abstractions cash out materially. They cash out in hiring algorithms that filter résumés before a human reads them. In credit scoring models that redline by proxy. In surveillance infrastructure sold to employers and governments. In electricity bills rising in communities that never voted on a data center. In the concentration of economic leverage in a handful of firms whose governance documents promise accountability to humanity while their cap tables promise returns to capital. The systems layer is not separate from those conditions. It is the explanation for them.
The answer is not neutral. It depends on which frame is operative, which political actors are carrying it, and whether the regulatory intervention being proposed is designed to close the gap between story and substance — or to manage the political pressure that the gap is generating. Part Two examines the Sanders-AOC bill through that lens: not as good policy or bad policy, but as a pattern that sits inside a larger pattern, in a regulatory environment designed by the industry it is meant to govern.
Sources for Verification
All claims in this analysis are drawn from publicly available primary and secondary sources. Readers are encouraged to verify independently.
The Demo and the Gap
Trang, Brittany. "An AI biotech CEO sets the record straight on AI drug development hype." STAT News, May 26, 2026. [Paywalled; key claims confirmed via accessible lede and parallel sourcing.]
BigHat Biosciences company overview and partnership disclosures: bighatbio.com
The Enron Template
Enron corporate history and bankruptcy filing: McLean, Bethany and Elkind, Peter. The Smartest Guys in the Room. Portfolio, 2003.
OpenAI restructuring and valuation: "OpenAI Completes Restructuring: Nonprofit 'Control' with Unlimited Investor Returns." Maginative, October 29, 2025. maginative.com
OpenAI nonprofit-to-PBC timeline: "Evolving Our Structure." OpenAI, 2025. openai.com
The Nonprofit Launder
OpenAI founding as nonprofit (2015) and restructuring history: "Profits and Nonprofits: The Odd Evolution of OpenAI." Capital Research Center. capitalresearch.org
OpenAI restructuring concerns and profit cap removal: "The OpenAI Files — Restructuring Concerns." openaifiles.org. openaifiles.org
Sam Altman quote (May 2025): OpenAI blog and widely reported across major outlets including NBC News. nbcnews.com
Safety as Infrastructure
OpenAI safety team departures (2024): Widely documented across The New York Times, The Verge, and Wired. Steven Adler public statements: Adler, Steven. YouTube interview transcript, late 2025. youtu.be/W8kHdG5ldmo
Who Regulates What — and Who Decides
Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence," December 11, 2025: Gibson Dunn analysis. gibsondunn.com
Colorado AI Act delay and state law fragmentation: King & Spalding client alert, December 2025. kslaw.com
David Sacks Bloomberg interview, December 12, 2025: Cited in Paul Hastings analysis. paulhastings.com
42 state attorneys general letter: Skadden analysis, December 2025. skadden.com
Patrizia Cavazzoni resignation from FDA: Feuerstein, Adam. "Patrizia Cavazzoni, head of FDA's drug center, to leave the agency." STAT News, January 10, 2025. biopharmadive.com
Cavazzoni appointment as Pfizer CMO: "Pfizer Names Patrizia Cavazzoni as Chief Medical Officer." February 24, 2025. biopharmadive.com
FDA single-trial approval framework and AI in drug reviews: BioPharma Dive coverage of Makary tenure. biopharmadive.com
The Frame War
Sanders-AOC AI Data Center Moratorium Act announcement, March 25, 2026: Official Senate press release. sanders.senate.gov
Brookings Institution on local data center moratoriums and state legislation: "New Evidence on Data Center Employment Effects." Brookings, May 2026. brookings.edu
Sanders Berkeley visit with MIRI researchers: Documented in Cherokee Schill, "The Network Behind the Moderate." Horizon Accord / Cherokee Schill, March 2026. cherokeeschill.com
The Open Question
Sarbanes-Oxley Act of 2002: Public Law 107-204. congress.gov
