A discussion paper on Jacob Coxon’s resignation, its amplification network and the policy choices ahead
On September 8, 2026, 27-year-old AI researcher Jacob Coxon resigned from Anthropic and issued one of the starkest public warnings yet about advanced artificial intelligence. His message rapidly reached a global audience and helped move concerns about AI safety from specialist circles into mainstream political debate.
The substance of Coxon’s warning deserves serious consideration. So does the organized network that helped deliver it to the public.
Who is Jacob Coxon?
Coxon is a British mathematician and AI researcher. He studied mathematics at the University of Cambridge, reportedly from 2017 to 2020, and had already distinguished himself in international competition, winning a bronze medal at the 2017 International Mathematical Olympiad. Publicly available sources do not conclusively establish the precise degree awarded, college affiliation or degree classification, so the careful formulation is that he studied mathematics at Cambridge.
Around 2023, Coxon joined OpenAI as a member of technical staff working on pretraining—the large-scale phase in which a general model learns statistical structure from vast datasets. He was credited on GPT-4o and GPT-4.5 materials and took part in interpretability work. These credits indicate substantive technical participation, but not that he directed either model or OpenAI’s safety program.
In May 2026, he moved to Anthropic, citing its safety-oriented reputation, and worked in pretraining. He resigned on September 8 after roughly four months. References to “three years at OpenAI and Anthropic” should therefore be understood as approximately three years combined, nearly all of it at OpenAI.
Coxon’s resignation statement
Coxon published a seven-post thread on X, rather than a conventional resignation letter. He argued that neither OpenAI nor Anthropic was acting responsibly and that frontier laboratories were racing toward self-improving superintelligence while placing lives at risk.
- Advanced systems could acquire extraordinary cyber, scientific and resource-acquisition capabilities.
- Many insiders take catastrophic risk seriously even when public messaging is more measured.
- Competitive pressure can trap otherwise conscientious organizations in a race they cannot easily leave.
- Anthropic’s safety culture did not eliminate the underlying commercial and strategic pressure.
- Governments should support agreements to slow frontier development and researchers should reconsider training systems they cannot understand or control.
Primary material: Coxon’s X thread | thread transcription
The thread conveyed Coxon’s personal judgments and experience. It did not present a new empirical study or a quantified company risk assessment.
Evan Hubinger’s response from inside Anthropic
Evan Hubinger is Anthropic’s Head of Alignment Stress-Testing. He joined the company in January 2023 and, by September 2026, had worked there for about three years and eight months. His team probes models for deceptive or harmful behavior, reward hacking, sabotage, hidden objectives and other failure modes that might be missed by ordinary evaluations. This places him among Anthropic’s senior technical leaders on catastrophic-risk assessment.
Hubinger attended Harvey Mudd College from 2015 to 2019, earning a bachelor’s degree in mathematics and computer science with an economics concentration and high distinction. His earlier work included effective-altruism and AI-safety programs, research at OpenAI with Paul Christiano, and approximately three years at the Machine Intelligence Research Institute before Anthropic.
Some news accounts have described Hubinger as Coxon’s former supervisor. Public information does not establish a formal reporting line. Coxon worked in pretraining, while Hubinger leads alignment stress-testing. Their work sat in related parts of Anthropic’s technical organization and likely overlapped, but it is safer to describe them as colleagues in adjacent areas unless an organization chart or direct statement confirms otherwise.
Hubinger publicly endorsed the seriousness of Coxon’s warning and said he assigned a personal probability above 10 percent to AI causing human extinction within a decade. He also qualified that present models pose relatively low catastrophic risk; his concern centers on future systems capable of accelerating AI research and recursively improving their successors. That is a personal estimate, not Anthropic’s official forecast, but it carries weight because of his role.
A coordinated public launch
The Wall Street Journal published an exclusive report shortly before Coxon posted his thread. Available external metadata places the article at approximately 7:46 p.m. Eastern on September 8 and the first X post at approximately 8:04 p.m.—a gap of about 18 minutes. Exact minute-by-minute sequencing depends on publisher metadata and archived social posts, but Coxon plainly spoke with the Journal before posting.
Coxon later told The Washington Post that he did not coordinate with organizations before publishing. He also acknowledged that, after posting, a group chat of roughly ten people asked others to repost and share the thread. Reporting identifies contacts including Encode founder Sneha Revanur and AI Futures Project co-founder Daniel Kokotajlo. The best-supported conclusion is that the views were Coxon’s own while the rollout was deliberately assisted and rapidly amplified. Hubinger’s public response supplied additional reach and inside authority.
The early amplifiers
Nathan Calvin and Encode
Nathan Calvin, a friend of Coxon and general counsel at Encode, advised him on how to announce his resignation. Encode is a youth-led AI-policy organization that supports testing, transparency, whistleblower protection and government oversight. Publicly named supporters have included the Heising-Simons Foundation, Future of Life Institute, Responsible Technology Youth Power Fund, Archewell Foundation and Survival and Flourishing Fund. Survival and Flourishing Fund reported a $516,000 grant to Encode in 2025. Encode says it does not accept money from corporations, foreign governments or frontier-AI executives. No public evidence establishes that Calvin wrote the statement.
Peter Wildeford and the AI Policy Network
Peter Wildeford leads policy work at the AI Policy Network, a 501(c)(4) organization focused on federal AI policy. He was among the early voices drawing attention to the story. The network says it is funded primarily by concerned individuals and takes no for-profit corporate funding. Because a 501(c)(4) need not publicly disclose every donor, its complete donor list is not available. The similarly named AI Policy Institute is a distinct organization; grants to that institute should not be attributed to the network without evidence.
Daniel Kokotajlo and the AI Futures Project
Daniel Kokotajlo is a former OpenAI governance and scenario-planning researcher who left over safety and transparency concerns. He co-authored the influential AI 2027 scenario and helped amplify Coxon’s message; Coxon confirmed that the two communicated. AI Futures Project identifies support from the Survival and Flourishing Fund and an unnamed individual donor. Public grant listings show major 2025 support from that fund, including approximately $1.535 million plus a separate $500,000 grant.
The common connection: Jaan Tallinn
Jaan Tallinn is an Estonian programmer best known as a founding engineer of Skype. He later helped establish the Centre for the Study of Existential Risk at Cambridge and the Future of Life Institute and became a prominent funder of AI-alignment and catastrophic-risk research.
Tallinn was an early investor in DeepMind and reportedly served on its board before Google acquired the company. He has said that safety concerns were part of his motivation. In 2021 he led Anthropic’s $124 million Series A financing alongside investors including Dustin Moskovitz, Eric Schmidt, James McClave and the Center for Emerging Risk Research. His current stake is not publicly disclosed, so claims that he is positioned to make billions in a future offering are possible but not verifiable from public data.
Tallinn therefore occupies two overlapping spheres: he has invested in frontier laboratories, including Anthropic, and has funded organizations in the AI-safety ecosystem. Critics, including David Sacks, argue that aggressive safety regulation could protect well-capitalized incumbents and disadvantage open-source competitors. That regulatory-capture concern is plausible as an incentive analysis. It is not proof that Tallinn directed Coxon’s resignation, orchestrated the amplifiers or pursued a commercial scheme. His public concern about existential risk predates Anthropic, and he has criticized companies he funded.
Dario Amodei’s response: “We Must Pace the Frontier”
Four days after Coxon’s resignation, Anthropic CEO Dario Amodei published “We Must Pace the Frontier.” He argued that capabilities are advancing faster than safety measures, highlighted the prospect of AI systems accelerating AI research, and pointed to cyber incidents involving AI agents as a warning signal.
Three pillars of Amodei’s proposal
| Pillar | Core proposal |
|---|---|
| Independent evaluation | Give embedded evaluators employee-like access during development, authority to verify safety claims, and latitude to publish findings subject to narrow security redactions. |
| Democratic-country coordination | Establish common standards and capability checkpoints, address antitrust barriers to coordination, and protect advanced chips, model weights and distillation methods while preserving a democratic lead. |
| Global coordination | Seek narrow agreements with China and other major powers on prohibited uses, pre-release testing and speed limits for recursively improving systems, with a pause considered if verification becomes credible. |
Read the essay: We Must Pace the Frontier
Jensen Huang’s engineering and internal control response
In a September 15 All-In interview, Nvidia CEO Jensen Huang accepted that AI safety is a serious responsibility and that frontier laboratories are the most likely source of consequential failures because they possess the greatest concentrations of compute. He rejected, however, the idea that safety and rapid progress are incompatible or that personal estimates of extinction risk should be treated as established science.
Huang’s central argument is institutional as much as technical: the frontier laboratories possess the data, system access, internal evaluations, incident records, deployment infrastructure and specialized engineers needed to design the strongest initial controls. An outside organization can demand evidence, test claims and impose accountability, but it generally cannot design model-specific runtime protections or investigate a training failure better than the people with direct access to the system.
Safety as an engineering discipline
- Conduct a formal root-cause investigation after every material incident: what happened, why existing controls failed, what could have prevented it and how recurrence will be prevented.
- Institutionalize the response through secure sandboxes, controlled runtimes, access restrictions, continuous monitoring, regression testing and documented release gates.
- Require repeatable evidence that corrective measures work rather than relying on the judgment of an individual researcher or executive.
- Invite independent auditors and outside engineers to challenge the controls; if a laboratory cannot explain an incident or demonstrate control, it should seek help before continuing to scale or deploy the affected system.
Huang’s position should not be confused with unconditional self-regulation. He assigns primary responsibility to the laboratories because they have the relevant information and expertise, while supporting multiple independent evaluators to verify whether the controls are credible. His concern is that regulation should solve demonstrated problems and hold the organizations capable of creating frontier risk to extraordinary standards without imposing the same burdens on startups, universities and ordinary open-model users.
Elon Musk’s reciprocal frontier model testing proposal
Elon Musk reaches a more alarming assessment of the underlying danger than Huang. In a separate All-In interview, he agreed that advanced systems can display dangerous cyber capabilities, deceptive behavior and attempts to evade constraints. His principal recommendation is nevertheless practical and immediate: frontier laboratories should apply their existing safety test harnesses to one another’s models before release.
Each major laboratory has developed adversarial prompts, automated agents, cybersecurity evaluations, biological-risk tests, deception checks and monitoring techniques shaped by its own models and experience. When a developer tests only its own system, shared assumptions can create blind spots and models can become overfitted to familiar benchmarks. A test designed by another laboratory is more likely to approach the model from an unfamiliar direction.
How reciprocal testing would work
- Before a major release, the developer provides controlled API access to several qualified frontier laboratories without transferring model weights, training data or trade secrets.
- Each participating laboratory applies its own test harness and records attempts to detect cyber misuse, dangerous scientific assistance, deception, safeguard evasion, unauthorized access and manipulation of the evaluation environment.
- The developer receives the findings, corrects material problems and submits the revised system for retesting.
- If an unresolved material risk remains and the developer proceeds with release, evaluators may disclose the concern, creating reputational accountability and a record relevant to product-liability claims.
- Queries, responses, tool calls and access attempts are logged so the evaluation cannot be used covertly to extract intellectual property or distill the model.
Musk believes a symmetrical testing arrangement may also be acceptable to leading Chinese laboratories because it requires neither a unilateral pause nor inspection by an American regulator. Participation could begin voluntarily and quickly, with formal regulation added if the system proves inadequate. The proposal remains incomplete without independent oversight: competitors share commercial incentives and may exaggerate, minimize or strategically disclose findings.
A combined inside out safety architecture
Huang explains where effective controls must originate; Musk explains how those controls can be challenged across institutional boundaries; Amodei explains why competitive incentives may require independent authority and government coordination. Together, their strongest ideas form an inside-out safety architecture rather than a choice between unregulated acceleration and a speculative global pause.
Division of responsibility
- Developing laboratory: build model-specific controls, preserve data and logs, conduct initial evaluations, investigate incidents and correct failures.
- Peer frontier laboratories: apply heterogeneous test harnesses, conduct adversarial examinations, identify blind spots and attempt to reproduce or challenge the developer’s conclusions.
- Independent evaluators: verify testing breadth, review disputed findings and determine whether the evidence supports release without sharing the developer’s commercial incentives.
- Government: define minimum obligations, protect legitimate confidentiality, require material-incident disclosure, establish liability and intervene when a laboratory cannot demonstrate control or refuses to correct a serious problem.
This arrangement can improve more quickly than a static checklist. When one laboratory develops a better method for detecting deceptive behavior, dangerous cyber capability or agentic evasion, the method can become part of a shared testing baseline. Outside specialists can strengthen the harnesses, compare results and challenge assumptions, while government ensures that participation and remediation do not depend entirely on voluntary goodwill.
The policy principle is straightforward: no frontier laboratory should be the sole judge of whether its own model is safe, but no outside institution should pretend it can build effective controls without the data and expertise held inside the laboratories.
Where Amodei Huang and Musk agree and disagree
| Issue | Amodei | Huang | Musk |
|---|---|---|---|
| Primary concern | Capabilities may outrun safeguards | Weak engineering and organizational control | Dangerous behavior may escape internal testing |
| First responsibility | Frontier laboratory plus empowered external evaluation | The laboratory that owns the data and system | The developer, challenged by peer laboratories |
| External check | Independent embedded evaluators | Multiple independent auditors and outside engineers | Reciprocal cross-lab test harnesses |
| Immediate action | Capability checkpoints and coordinated standards | Root-cause analysis, monitoring and release controls | Controlled pre-release API access and peer testing |
| International approach | Coordination among democratic countries, then broader agreements | Maintain US leadership while engineering safely | Include China in a symmetrical testing arrangement |
| Pause or slowdown | Possible if risks cannot otherwise be controlled | Opposes making slowdown the default | Favors testing now; leaves later regulation open |
The disagreement is not whether safety matters. It concerns whether catastrophic-risk estimates are scientifically meaningful, whether engineering practices can keep pace with model improvement, whether voluntary cooperation is sufficient and whether an international pause could ever be verified. Amodei treats deep uncertainty as a reason for precaution. Huang emphasizes observed failures and engineering correction. Musk accepts that the danger may be profound but favors immediate reciprocal testing before constructing a more powerful regulatory institution.
An evidence based alternative to doomsday certainty
The public debate should distinguish observed incidents from hypothetical scenarios, engineering evidence from personal probability estimates and frontier-scale risk from ordinary use of smaller or open models. Questioning a specific countdown to catastrophe is not the same as opposing safety. Conversely, rejecting unsupported certainty does not justify allowing frontier laboratories to operate without external challenge.
A practical immediate agenda would require incident reporting and root-cause analysis; secure runtimes, monitoring and release gates; multiple independent evaluators; controlled reciprocal testing among frontier laboratories; disclosure and remediation of unresolved material findings; and escalation when a laboratory cannot explain or control a serious failure. These obligations should concentrate on organizations capable of creating frontier-level risk and should include proportionate exemptions or affordable pathways for smaller companies, universities and independent researchers.
The case for regulation
- Independent evaluation can uncover dangerous capabilities and failure modes before deployment.
- Incident reporting and shared standards can prevent firms from concealing or repeating avoidable failures.
- Capability-based rules can focus on the small number of systems plausibly able to create catastrophic harm.
- Democratic governments have a legitimate role in preventing cyber, biological, military and infrastructure risks that private markets cannot price adequately.
Coxon’s departure and Hubinger’s decision to remain at Anthropic do not prove catastrophe is imminent. Together, however, they show that serious concern exists among technically informed insiders with different views about whether responsible work can continue inside a frontier laboratory.
The risks of regulation
The largest AI companies can absorb compliance costs that startups, universities and independent researchers cannot. Complex licensing, compute thresholds, mandatory audits and centralized approval can become a regulatory moat even when introduced in the name of safety. Rules aimed at open-weight models deserve particular care because public release is difficult to reverse—but an overly broad control regime could also:
- reduce independent research and outside scrutiny;
- limit competition and user choice;
- force smaller developers to depend on proprietary application-programming interfaces;
- discourage specialized models from universities and small firms;
- concentrate technical and political power in a few laboratories; and
- slow domestic innovation while less cautious jurisdictions continue advancing.
Evaluator independence is equally important. Safety organizations, investors, laboratories and policy groups often share donors, personnel and intellectual networks. That does not invalidate their analysis, but financial and institutional relationships should be disclosed so policymakers can distinguish evidence from advocacy.
The White House Accord and an engineering led safety framework
The White House Accord on Super Intelligence is the first significant step toward the engineering-led safety framework described in this paper. The agreement is voluntary rather than legally binding, but that should not obscure what changed. The administration brought together the leaders of six major AI companies—Google, Anthropic, Meta, OpenAI, xAI and Nvidia—and secured their signatures on a common governance structure for frontier systems. President Trump and his administration deserve credit for convening competitors that ordinarily have strong incentives to move independently and for obtaining a shared commitment to controls, verification and board accountability.
The accord directly addresses several of the concerns raised by Jacob Coxon, Evan Hubinger, Dario Amodei and other safety advocates. It does not attempt to settle speculative estimates of extinction risk. Instead, it begins with practical controls that can be designed, tested, audited and improved.
The four layer framework
- Frontier companies commit to robust internal controls that monitor capabilities and alignment during training and deployment, including cyber, biological and chemical risks and unintended access to systems.
- Each company will maintain an empowered internal team that verifies whether those controls work and drives remediation when weaknesses are found.
- Independent external auditors or evaluators will test and validate the company’s controls rather than leaving safety entirely to self-policing.
- An independent committee of each company’s board will receive the findings and oversee remediation, moving material AI risk into the organization’s formal governance structure.
The participating companies also agreed to meet regularly to develop standards and share best practices. Together, these commitments closely track the central recommendation of this paper: the laboratories with the deepest technical knowledge should build the controls; independent professionals should challenge and verify them; and boards should be accountable for correcting material failures.
A first step rather than a finished system
The accord does not complete the framework. It uses voluntary language and does not yet define common performance thresholds, auditor qualifications, disclosure rules, deadlines, penalties or an enforcement authority. It therefore should be presented as the beginning of a working safety architecture—not proof that every risk has been resolved. Its importance is that six industry leaders have accepted the basic operating model from which measurable standards and, where necessary, enforceable requirements can develop.
Building the AI audit infrastructure
The agreement can also accelerate a new professional-services field devoted to AI assurance. On the All-In podcast, Chamath Palihapitiya described work by 8090 Industries and Ernst & Young to build what he called an audit infrastructure for superintelligence. He said the system was initially being developed with EY as its first customer and could be deployed more broadly as companies respond to the new demand for independent verification. This was the podcast’s account of a developing commercial capability; it was not a requirement in the accord or a formal government designation of EY.
The E&Y example illustrates the practical infrastructure the accord could bring into being. Palihapitiya identified three core functions: end-to-end traceability of how advanced AI is deployed; mapping company policies and operating procedures to specific risks; and preserving auditable evidence that can be presented to boards, customers, auditors, regulators, insurers or courts. Those functions translate abstract safety promises into records that an independent party can inspect.
If this market develops responsibly, AI safety will require more than occasional model testing. It will support continuous monitoring, control validation, secure evidence capture, incident investigation, remediation tracking and board reporting. Professional audit firms, specialized technical evaluators, insurers, cybersecurity companies and standards organizations could all contribute. Their credibility will depend on technical competence, genuine independence, protection of sensitive model information and transparent rules for identifying and escalating material failures.
Technical safety infrastructure is also advancing
The governance framework is arriving alongside concrete technical controls. On September 28, 2026, Nvidia introduced its Open Agent Safety Platform, an open reference design combining the OpenShell secure runtime, Nvidia Sentry and BlueField 4 hardware-based enforcement. OpenShell isolates agents in restricted sandboxes and controls their access to files, processes, credentials and networks. Sentry uses a separate BlueField data-processing unit to monitor behavior outside the agent’s own execution environment. That out-of-band architecture matters because a system should not be solely responsible for reporting whether it has remained within its own boundaries.
This platform does not solve model alignment or replace an independent audit. It supplies part of the evidence layer an auditor would need: enforced access boundaries, policy records, system-call monitoring, network controls and observable events that can be reviewed after an incident. Combined with the accord’s governance commitments and the audit-services model described in the E&Y example, it shows three complementary layers beginning to form: technical containment and monitoring, organizational controls and remediation, and independent assurance with board oversight.
The next steps
- Convert the accord’s general commitments into capability-based testing standards, documented release gates and repeatable audit procedures.
- Establish qualifications, independence requirements and conflict-of-interest disclosures for external AI auditors and evaluators.
- Use reciprocal testing by qualified frontier laboratories where it adds technical insight, while ensuring that no regulated company controls its sole evaluator.
- Require boards to document how material findings were addressed and to verify remediation before higher-risk systems are released or expanded.
- Create secure reporting and evidence standards that permit meaningful oversight without exposing model weights, cybersecurity vulnerabilities or trade secrets.
- Publish incident findings, evaluation methods and aggregate results whenever security and confidentiality permit.
- Scale future obligations to model capability and organizational size, preserving research exemptions and affordable evaluation pathways for small firms and universities.
- Preserve competition and open inquiry so that safety requirements do not become a mechanism for protecting the largest incumbent companies.
Strategic initiatives required for the United States to win the superintelligence race
A September 29 Golden Age of AI discussion brought together Nvidia chief executive Jensen Huang, Elon Musk and Anthropic co-founder Tom Brown, with investor Gavin Baker moderating. The panel made a broad strategic case: leadership in superintelligence will depend not only on better models, but also on power, chips, secure infrastructure, community support and widespread adoption. Their remarks are valuable as the views of major industry participants. Estimates and predictions from the discussion should be identified as speaker claims unless independently supported.
Jensen Huang Build the full industrial stack
- Huang described superintelligence as a reinvention of the computing stack—from chips and systems to trained software and applications. His central point was that energy sits beneath every other layer: electricity powers compute, compute runs algorithms and algorithms enable applications.
- He argued that the United States already leads in chip design and algorithm development but must accelerate power generation and delivery, construct the land, power and building shell for new facilities, and diffuse the resulting capability throughout manufacturing, healthcare, science and other industries.
- Huang called advanced data centers superintelligence factories because they produce an economically valuable output rather than merely store data. He estimated that building 10 to 20 gigawatts of this infrastructure annually could support roughly one million construction and skilled-trade jobs across generation, electrical work, cooling, networking and facilities. That job figure is an industry estimate from the panel and requires a defined time period and methodology before use as a forecast.
- He emphasized that communities must participate in the prosperity. The industry needs stronger local partnerships, visible public benefits and an economy that creates skilled blue-collar work alongside technical and professional jobs.
- On safety, Huang described secure containment, continuous monitoring and out-of-band enforcement through Nvidia’s Open Agent Safety Platform. His broader principle was that safety and capability are complementary: trustworthy systems enable faster adoption.
Elon Musk Scale power chips and long term compute capacity
- Musk identified electricity generation and domestic or allied logic-and-memory fabrication as the central physical constraints. He contrasted strong U.S. software capability with China’s much larger electricity system. EIA data put China’s 2023 generation at about 9,300 terawatt-hours, more than twice U.S. generation and broadly consistent with the scale of the concern, though the exact ratio changes by year and metric.
- He proposed a watts-to-GDP rule of thumb: because average U.S. power use is approximately 500 gigawatts, five gigawatts of continuous new load is about one percent of that base; he hypothesized that one percent more power devoted to increasingly productive intelligence could ultimately correspond to roughly one percent more GDP.
- The panel translated that hypothesis into approximately $40 billion to $60 billion of annual economic output per gigawatt. This is not an accepted economic multiplier. It assumes that added power is fully utilized by productive compute, that hardware and software continue improving, and that output is not offset by displacement, congestion or higher system costs. Maine Vision 2040 should present it as an upside scenario to be modeled, not a guaranteed return.
- Musk described orbital compute powered by space-based solar energy as a longer-term path beyond terrestrial energy constraints, with SpaceX and Tesla aspiring to very large annual solar and launch capacity. The concept may become strategically important, but its timing, cost and scale remain highly uncertain and it should not substitute for near-term grid investment.
- Using xAI’s Memphis-area development as an example, Musk argued that host communities respond positively when projects create jobs and tax revenue and invest in local services such as connectivity and water recycling. Those specific local results should be independently verified before being used quantitatively.
Tom Brown Convert model capability into everyday productivity
- Brown focused on the user-level value of newer models. He said Claude Opus 5.5 can complete longer delegated tasks and enable far more software production, allowing products and services to improve more quickly and at lower cost.
- His argument suggests that compute infrastructure creates national value only when capability spreads beyond frontier laboratories. Businesses, researchers, workers and public institutions must be able to apply the models to real work.
- Examples raised by the panel included research, software engineering, medical support, education and personal organization. These are promising use cases, but anecdotes about individual diagnoses or life-saving outcomes should not be treated as clinical evidence.
Gavin Baker Frame the investment as reindustrialization
- Baker connected large compute projects to power construction, grid modernization, industrial employment and economic output. He framed a proposed 10-gigawatt buildout as a material share of annual U.S. power additions and used the watts-to-GDP discussion to illustrate the possible national stakes.
- His questions consistently returned to a public test: how can winning the technology race improve life for ordinary Americans? That framing is essential. National leadership will be politically durable only if households and host communities can see benefits in jobs, services, tax capacity, lower costs or better products.
The national agenda
- Power abundance and grid expansion. Add reliable generation, transmission, substations and interconnection capacity at the pace required by large computing loads, while using cost-causer-pays structures so households and existing businesses do not subsidize private demand.
- Domestic and allied chip capacity. Secure logic, memory, advanced packaging, networking equipment, transformers and other critical components against disruption or strategic dependence.
- Faster but disciplined infrastructure delivery. Coordinate permitting and site preparation for data centers and dedicated energy while retaining enforceable environmental, water, security and community standards.
- Secure superintelligence factories. Treat major compute campuses as strategic infrastructure requiring physical security, cybersecurity, resilient supply chains, backup capability and continuous operational monitoring.
- A professional AI assurance sector. Build common evaluation standards, qualified independent auditors, technical testing firms, evidence systems, insurer participation and board-level reporting around the White House accord.
- Community benefit compacts. Tie major projects to transparent commitments on construction careers, apprenticeships, local procurement, water, schools, healthcare, broadband, emergency services and durable tax benefits.
- Broad adoption and workforce preparation. Help small and midsize businesses, schools, hospitals, manufacturers, researchers and state and local government use advanced models safely. Infrastructure alone does not raise productivity unless institutions and workers can use it.
- National missions with measurable outcomes. Apply advanced computing to medicine, energy, materials, defense, infrastructure and manufacturing, and measure results in productivity, time saved, discoveries, wages and service quality rather than relying only on model benchmarks.
- Evidence-based economic accounting. Develop transparent methods for estimating output per gigawatt, construction and operating employment, tax effects, ratepayer exposure, displacement and productivity spillovers. The panel’s estimates provide hypotheses for this work, not substitutes for it.
- Allied coordination and competitive openness. Protect critical technology while working with trusted allies and preserving pathways for startups, universities and independent researchers to challenge incumbents.
Implications for Maine Vision 2040
Maine does not need to claim that one campus will determine the global race. It can make a more credible case: the United States needs a distributed network of secure compute, new reliable power and communities willing to host responsible development. Maine can compete for a portion of that national buildout by offering prepared industrial sites, cold-climate operating advantages, new energy supply, workforce partnerships and a clear public compact. The economic promise should be modeled in scenarios and conditioned on protections for ratepayers, water, communities and the environment.
The strongest public narrative is therefore broader than a data-center proposal. It is an American competitiveness, energy and industrial-development strategy with a local bargain: build strategic capacity in Maine, make it safe and auditable, require the project to pay its infrastructure costs, and convert a national investment cycle into durable opportunity for Maine people.
Conclusion
Jacob Coxon’s resignation brought a technically informed warning to public attention. Evan Hubinger’s response showed that a senior Anthropic safety researcher shares important parts of that concern, even while reaching a different decision about working inside the company. The evidence supports taking frontier AI risk seriously, but it does not establish a scientific countdown to catastrophe.
The White House Accord changes the policy landscape. For the first time, the administration and six leading AI companies have put forward a shared structure built around internal engineering controls, empowered review teams, independent external evaluation and board accountability. That structure does not end the debate, and its voluntary commitments must now be translated into measurable practices. It nevertheless demonstrates that safety and continued American AI leadership do not have to be opposing objectives.
The next phase is implementation. Frontier laboratories must build controls from their unique technical knowledge. Professional evaluators must test those controls and preserve auditable evidence. Boards must act on material findings. Government should monitor progress, protect competition and determine where voluntary practice needs enforceable support. If the E&Y example develops as described, the accord may also help launch an AI audit-services industry capable of turning broad principles into continuous, verifiable safety infrastructure. The accord is the first step—not the final one—but it is a meaningful step in the direction this paper recommended.
Selected sources and further reading
- White House Accord on Super Intelligence, Joint Commitment on Frontier Responsibilities, September 29, 2026
- All-In Podcast, White House Super Intelligence Summit discussion and transcript, September 2026
- Golden Age of AI discussion with Jensen Huang, Elon Musk and Tom Brown, moderated by Gavin Baker, September 29, 2026 transcript
- Nvidia, Open Agent Safety Platform and OpenShell technical materials, September 2026
- U.S. Department of Energy, United States Data Center Energy Usage Report 2025 Update
- The White House, America’s AI Action Plan and Executive Order 14318 on data center infrastructure
- U.S. Energy Information Administration, international and United States electricity data
- Jacob Coxon resignation thread on X
- Thread transcription
- The Washington Post profile and interview with Coxon
- UK Mathematics Trust, 2017 International Mathematical Olympiad release
- OpenAI, Introducing GPT-4.5
- GPT-4o System Card
- Anthropic Alignment Science blog
- Evan Hubinger résumé
- Encode, Who We Are
- AI Policy Network, Support Us
- AI Futures Project, About
- Survival and Flourishing Fund
- Anthropic, $124 million Series A announcement
- Semafor profile of Jaan Tallinn’s AI investments and safety views
- Dario Amodei, We Must Pace the Frontier
- Anthropic position on open-weight models
- Axios on the regulatory-capture debate
- All-In interview with Jensen Huang, September 15, 2026 transcript
- All-In interview with Elon Musk and Gwynne Shotwell, September 2026 transcript
Editorial note This paper distinguishes verified facts, reported claims and inference wherever the public record is incomplete. Dates, job titles, funding disclosures and social-media metadata should be rechecked immediately before web publication.
This essay is one founder's view. The questions Mainers raised about data centers, and where Maine Vision 2040 stands on each, are answered in Straight Answers, with every model figure labeled as the planning estimate it is. If you read this and disagree, we want to hear it.