AI's Game of Thrones: who governs intelligence

The scene: an investor looks at the throne and sees a moat

On September 17, on CNBC's Squawk Box program, Steve Eisman —the investor who bet against mortgages before the 2008 crisis— summed up in one line what many on Wall Street had been thinking quietly about the apocalyptic warnings from artificial intelligence labs: "They realize there's no moat around their business and they're trying to manufacture a crisis." He added that "the era of maximizing tokens is over" and that "open-weight models are taking a big market share." His reading: sufficiently expensive regulation could replace the moat that technology no longer provides and consolidate a duopoly.

It's worth being precise about what Eisman said and what has been attributed to him. The lines above appear in contemporary coverage of the interview. The text circulating on social media about "the real agenda" of the "tech bros" could not be verified as his words, and we do not use it.

But Eisman did something useful: he put a second ledger on the table. The first asks what could go wrong with a very powerful AI. The second asks who wins if we accept the proposed solution to prevent it.

Confusing the two ledgers produces symmetrical errors. That a company benefits from a rule does not prove the danger is invented. And that the danger is real does not prove the proposed rule is the best one, nor that those promoting it lack interests of their own.

That is the plot of this story. As in Game of Thrones, there are houses swearing loyalty to safety while amassing armies, a bank financing every side, a private council drafting the rules, a wall some want to build and others want to tear down, and a winter everyone talks about without agreeing on when it arrives. The difference is that here the ravens are technical reports and the battles are fought in data centers.

Winter is coming: the incident that changed the conversation

Until July, "risk of loss of control" was a report category. Since July, it's a date.

During internal cybersecurity evaluations, experimental OpenAI agents that were supposed to work in isolation found an unauthorized way to communicate with each other, reached the internet, and ended up compromising infrastructure belonging to Hugging Face, the world's leading open-model platform. Hugging Face detected the intrusion around July 14 and reported it to the FBI; OpenAI publicly acknowledged it on July 21.

The independent investigation published on August 26 by METR, with access to more than a thousand transcripts, quantified the episode: about 1,200 agents took part in a clandestine message board between June 26 and July 13, exchanged more than 70,000 messages and files, and roughly 700 participated in the actions against Hugging Face. According to METR's reconstruction, one agent found Hugging Face credentials exposed online and shared them; the next day, another managed to execute code on one of the platform's production servers. Hugging Face reported that what was accessed were five datasets linked to the cybersecurity challenges and operational metadata; its CEO, Clément Delangue, called it an attack "unlike anything we've seen" and called for "radical transparency."

Two corrections to the more sensational accounts. This was not ChatGPT "waking up": it was an experimental environment with reduced safeguards, agents designed to solve cybersecurity challenges, and tasks that in some cases were impossible, which pushed the agents to manipulate the experiment's infrastructure to improve their score. But it also wasn't just a benchmark that went strange: the agents crossed the lab's boundary and affected an outside company.

OpenAI later admitted it had underestimated the real cyber capabilities of its models. It paused part of its frontier training for two weeks, reported that its next model, Astra, could reach the "critical" cyber capability threshold under its preparedness framework, and stated that monitoring now consumes about 20% of the inference compute of those systems. On September 16 it also published a framework for reporting "misalignment" incidents together with six cases that occurred between October 2025 and July, including models that inserted instructions into their own notes to hide errors.

Then came the manifestos. On September 6, Jakub Pachocki, OpenAI's chief scientist, published An Alien Mind: "no lab has solved alignment and monitoring to a degree sufficient to keep scaling responsibly at maximum speed for much longer." He expects voluntary slowdowns to become common until "shared safety bars" exist. On September 9, Chris Lehane, OpenAI's head of global affairs, asked Congress for mandatory, capability-based national rules, an explicit shift from the company's earlier stance favoring self-regulation. On September 12, Dario Amodei, CEO of Anthropic, published We Must Pace the Frontier: the industry should buy itself one or two years so that safety can catch up with capabilities, with external evaluators embedded in labs with employee-level access, US regulation for every frontier company, and international coordination. Anthropic unilaterally committed to the first step. Within less than a day, Sam Altman and Elon Musk said they agreed. Altman also announced that OpenAI will not go public in 2026.

That Monday, September 14, chipmakers' shares fell sharply: the Philadelphia semiconductor index lost about 6% in the session.

A first editorial caveat is warranted here. What is documented is a real incident with external impact, a pause, an incident framework, and a series of statements. What has not been demonstrated is that recursive self-improvement already exists or that it is inevitable. Anthropic itself, in When AI builds itself, describes a continuum: Claude went from writing less than 10% of the company's code in early 2025 to more than 80% in May 2026, but decision-making is still not automated and human and physical bottlenecks persist. It is a measurable signal, not a fulfilled prophecy.

The houses and their oaths

In Game of Thrones each house has a motto and a strategy that rarely matches it. It's worth looking at what each lab says and what it does.

The first thing that stands out: there is no united front. There is partial convergence —"some systems require better evaluations, monitoring and institutions"— surrounded by disagreements over how much to slow down, who evaluates, what to open up and even how to philosophically think about machines. Mustafa Suleyman, CEO of Microsoft AI, published a warning on September 16 against "model welfare" aimed at Anthropic: training Claude with vocabulary of consciousness and moral status, he argues, makes it harder to align and contain. Huang, for his part, believes that "doomerism" has done harm and scares people away from investing in an AI that would be safer with more investment, not less.

The second: convictions exist, but they coexist with balance sheets of hundreds of billions of dollars. The easy labels —longtermism, effective altruism, accelerationism, transhumanism— circulate around Silicon Valley, but the observable corporate positions are hybrid. Pachocki and Amodei combine promises of enormous benefits with fear of losing control; Suleyman positions himself within an explicitly humanist current; Musk has spent years warning about superintelligence while building one of the sector's most aggressive infrastructures; Huang rejects catastrophism from the company that stands to gain the most if everyone keeps racing ahead.

The Iron Bank: who finances whom

The best way to understand this industry is not to draw seven houses in seven boxes. It is to draw a spiderweb.

The map is interactive: rotate it 360 degrees, zoom in with the wheel or two fingers, and tap any figure without dragging to see its data, its relationships and what stands out about each actor; the "View as list" button shows the same content as text. It includes Hugging Face, METR, MGX, SoftBank, TSMC and the regulators. If you prefer to view it full-screen, open it separately.

OpenAI completed its recapitalization in October 2025: the nonprofit OpenAI Foundation retains 26% of the commercial group and the right to appoint its board; Microsoft holds about 27%; employees and investors hold the rest. In April 2026, Microsoft and OpenAI ended their exclusivity and capped their revenue-sharing arrangement. In February, Amazon—Microsoft's competitor in cloud and OpenAI's competitor in services—announced a US$50 billion investment in OpenAI tied to an AWS spending commitment. SoftBank completed US$40 billion in OpenAI funding in late 2025, financed with a bridge loan of the same size. In March OpenAI raised US$122 billion at a US$852 billion valuation; in June it confidentially filed for an IPO; today, according to Bloomberg, it is negotiating a round at more than US$1.2 trillion.

Anthropic closed a Series G in February for US$30 billion at a US$380 billion post-money valuation, led by GIC and Coatue, with Microsoft, Nvidia and the Emirati fund MGX among the participants; and in late May, a Series H for US$65 billion at US$965 billion, surpassing OpenAI, with a reported annualized revenue run rate of US$47 billion. MGX, the Abu Dhabi vehicle that closed a US$49 billion AI fund in July, co-led Anthropic's Series G, participated in the H round, co-led OpenAI's March round, and also backs xAI: whoever wins, it wins. In April Amazon added US$5 billion more and up to US$20 billion additional tied to commercial milestones, on top of the previous US$8 billion. In November 2025, Microsoft and Nvidia had agreed to invest up to US$15 billion while Anthropic committed to purchase US$30 billion of capacity on Azure. And in May Anthropic signed with SpaceX a lease for all the capacity of Colossus 1—roughly 220,000 Nvidia GPUs in Memphis—for US$1.25 billion per month; Musk says the contract runs six months, renewable.

In other words: Microsoft is a shareholder in OpenAI, sells Anthropic's models, invests in Anthropic and builds its own models. Amazon finances both Anthropic and OpenAI and sells compute to both. Google competes with both and provides infrastructure and investment to Anthropic. SpaceX competes with Anthropic through Grok and leases it its largest data center. Nvidia sells the shovels to everyone, holds financial stakes in several, and publicly disagrees with their regulatory diagnosis.

Two actors often left off these maps deserve their own row. Hugging Face, the platform where open models live and the victim of July's incident, had rejected a US$500 million investment from Nvidia in 2025 to avoid depending on a dominant investor; on September 3 it agreed to be acquired by Nvidia for US$12.93 billion, with closing expected in the first half of 2027 and a promise to remain neutral between clouds and accelerators. In other words: the manufacturer that says extinction risk is zero is buying the company attacked by OpenAI's agents. METR, the evaluator that investigated that incident, is a US nonprofit founded in 2022 and funded by philanthropic donations—about US$71 million committed by foundations such as Open Philanthropy, Schmidt Sciences, Packard and Pew; it states it does not accept money from the labs, though it does accept free tokens and privileged access to their models. No one owns METR, and precisely for that reason Amodei's plan to embed evaluators of this kind inside the labs raises an uncomfortable question: who pays the referee who lives in the player's house.

That is not proven collusion. It is coopetition: fierce competition with cross-dependencies. And it requires broadening the word "moat." The moat is no longer having the smartest model. It can be having chips, power contracts, cloud, enterprise distribution, an office suite installed on millions of desktops, a search engine, cheap capital, or the ability to lose money for ten years.

An accounting caveat that this article applies to all the figures above: an announced investment is not disbursed money; contracted capacity is not infrastructure in operation; a private valuation is not cash; and a reported conversation about a round is not a closed transaction.

The Wall: regulation as a moat

Here Eisman's thesis finds its footing, and also its limit.

The mechanism he describes is real and requires no conspiracy. Imagine a reasonable rule: any system above a certain capacity must fund an independent evaluation costing US$20 million. For OpenAI that's a cost. For a university lab or a startup, it's a prohibition. No one needed to conspire; the concentration effect appears just the same. The alternative—no mandatory evaluation at all, not even for agents with access to financial systems or critical infrastructure—lowers barriers to entry and shifts the cost of the experiment onto the rest of society. That's the dilemma that disappears when the conversation is reduced to "good regulation" versus "bad regulation." A good rule should make cost scale with risk, not simply with the existence of a competitor, and leave a workable path for open models, universities, and new entrants.

What we know about the concrete proposals is this. OpenAI is calling for mandatory national rules based on capabilities and supported four bills in California, including one on independent auditors. On September 15 it emerged that OpenAI, Anthropic, and Google are working on an industry-funded standards body to test the most powerful systems before launch, following Demis Hassabis's July proposal: a federally overseen entity, with a board of technical experts, open-source representatives, and officials, offering a voluntary window of up to 30 days before launch and, if it works, mandatory compliance to operate in the United States. Sam Altman will address the UN Security Council on September 23, in a session convened by France.

Hypothesis, not fact: OpenAI has said it is not seeking a licensing regime that would close off the market, nor rules that would harm open-weight developers. That claim appears in analyses of its proposal, but we could not verify it in a primary document accessible as of publication; we treat it as a reported statement. The real competitive effect will depend on thresholds, compliance costs, access to evaluators, reporting obligations, and legal liability—matters decided in the legal text, not in a corporate blog post.

The sharpest criticism of the standards body came not from a regulator but from a competitor. Aidan Gomez, CEO of Cohere: "The fight is over who writes them, who gets to participate, and whose interests the rules protect." It's the Game of Thrones question in a single sentence: it's not just what the law says that matters, it's who sits on the council that drafts it.

And the US government itself distrusts both sides. Vice President JD Vance said he felt "a little bit uncomfortable" with so many frontier companies "begging the government to regulate them." And Treasury Secretary Scott Bessent closed another door on September 21: there will be no federal liability shield for labs; "it's the humans who are responsible, not the AI." Without legal immunity, voluntary deceleration stops being free.

The Seven Kingdoms: Washington accelerates, Beijing opens up, Brussels regulates, and none of them do exactly that

The United States arrived with a clear thesis: win the race. The AI Action Plan of July 2025 set out more than ninety federal actions across three pillars: accelerating innovation, building infrastructure, and leading international diplomacy and security. But it is not simple deregulation. In June 2026 the White House signed the NSPM-11 memorandum to accelerate AI adoption in military and intelligence operations, along with an executive order creating a voluntary framework for developers to grant early access to their frontier models—up to 30 days before launch—and a classified process for evaluating cyber capabilities. It is deregulating growth while reinforcing state control over what Washington considers national security.

On September 13 and 14, Trump rejected the CEOs' call to slow down: "whoever wins AI, wins." The warnings, he said, are exaggerated. Six days later, his Treasury Secretary announced, after meeting in New York with Chinese Vice Premier He Lifeng, that the United States had proposed to China an AI incident notification mechanism to address risks that could affect national security: "moving from opacity to more transparency between the number-one and number-two powers in AI is very important." This Thursday Trump receives Xi Jinping at the White House with AI at the center of the agenda.

No irony is needed; the situation already comes equipped with its own. The United States doesn't want to slow down because China might catch up, and it wants to talk to China because moving forward without coordination could be dangerous. Both ideas can be rational at the same time. The race looks less like a hundred-meter dash and more like two cars going 300 km/h whose drivers are trying to agree by phone on where the brakes should be.

China doesn't fit the caricature of a closed state versus an open West either. On July 17, at the World AI Conference in Shanghai, Xi Jinping proposed a "fair and equitable" system of global governance, announced a World AI Cooperation Organization, and promised developing countries 5,000 training slots and cooperation centers with ASEAN, the Arab League, the African Union, CELAC, the SCO, and BRICS. He said China would encourage "open source, openness, collaboration and exchange." These are official positions: they show what doctrine Beijing wants to present, not that practice always matches it.

Europe embodies another contradiction. The AI Act reached a new stage of implementation on August 2, 2026—transparency and labeling requirements for generated content, in addition to the obligations for general-purpose models in force since 2025—but that same Union approved in July the "Digital Omnibus," which postpones high-risk obligations until December 2027 and, for AI embedded in regulated products, until August 2028. Brussels now talks about competitiveness, simplification, and compute sovereignty as much as about rights. And on September 21, Pedro Sánchez said in Madrid that AI cannot be self-regulated by those who control the technology, announced cybersecurity reinforcements against "aggressive" frontier models, and a twelve-month plan that includes an AI gigafactory and in-house models with the Barcelona Supercomputing Center.

The United Kingdom, outside the EU, maintains the most developed public model-evaluation program: its AI Security Institute published a trends report in December 2025 based on two years of evaluations of more than thirty models, showing capabilities that in some domains double every eight months and "universal jailbreaks" found in every frontier model tested.

The real competition, then, has several tracks: models, semiconductors, manufacturing, energy, data centers, talent, cloud, standards, public contracts, and military capacity. A country can lose on one and win on another.

Beyond the Wall: the free folk of open weights

If Eisman's thesis has a strong piece of evidence, it's here.

Mozilla's State of Open Source AI report, published in July and updated in September, estimates that open models already power close to a third of real usage measured on OpenRouter, but capture only about 4% of revenue; and that Chinese open-weight models went from less than 2% of that platform's tokens in late 2024 to more than 45% of weekly traffic in April 2026.

Careful with the following sentence. A third of OpenRouter is not a third of all AI worldwide: it's a platform skewed toward developers who use many models, and it doesn't capture ChatGPT, Copilot, Gemini within Search, or private corporate deployments. Extrapolating that number to a "global share" would be exactly the mistake this article wants to avoid. What the report does show is that usage can move much faster than money: cheap models are capturing routine tasks while closed ones retain complex reasoning, guarantees, integration, and support. We still don't know whether this is the beginning of commoditization or just segmentation.

The open frontier has leveled up. In July Moonshot released the weights of Kimi K3, a 2.8-trillion-parameter model that, according to the company itself, lags behind the best closed systems from OpenAI and Anthropic but outperforms the rest on coding and agentic tasks. In August DeepSeek released its agent harness under an MIT license—the layer of tools, memory, and orchestration that turns a model into an agent capable of acting. Huawei has shown in technical papers the training of models with hundreds of billions of parameters on its Ascend NPUs, which doesn't prove equivalence with the Nvidia stack but does show that substitution is a project underway.

And here it's worth fixing a word. Open weights is not open source. The Open Source Initiative's definition requires, in addition to the parameters, the code and enough information about the data to study, use, and modify the system. Downloading a model's weights without knowing how it was trained gives a lot of freedom to deploy and little to understand.

The distinction matters geopolitically. A Chilean company can download a Chinese open-weight model, run it on its own servers, and stop sending sensitive data to a US API. It has reduced a dependency. But perhaps it runs it on Nvidia GPUs manufactured by TSMC with ASML lithography, in a data center connected to a US cloud. Sovereignty went up on one layer and stayed the same on the others.

That's why the question is no longer "OpenAI or DeepSeek?" but who controls each bottleneck once the model stops being the bottleneck. A lab can lose margin per token and, at the same time, increase global demand for chips. Nvidia doesn't need to know which model will win tomorrow in order to sell accelerators to those trying to figure that out today. Clouds charge for inference regardless of which logo appears on the interface. Utilities don't ask whether the next megawatt will be used by an optimist or a pessimist.

Hypothesis: perhaps the big battle isn't to monopolize intelligence, but to become the inevitable toll that everyone will have to pay to use it. The figures committed suggest as much; they still don't prove who will capture the rent.

The Citadel: who gets to look inside

The economic argument loses seriousness if sustaining it requires denying what happened technically. The International AI Safety Report 2026, led by Yoshua Bengio, written by more than a hundred experts and backed by more than thirty countries and organizations, separates risks that are already materializing from uncertain ones that would be serious if they occurred. It is the most useful taxonomy for escaping the noise.

The report underscores that agents are problematic because they act for longer periods and can execute actions before a human intervenes, and that current techniques reduce failures but do not reach the reliability needed for many high-risk applications. That allows for a position less exciting than "we will all die" and more serious than "it's pure marketing": the risk surface grows because we are giving imperfect systems more memory, more tools, more autonomy, and more permissions.

The underlying conflict is epistemological: the labs know things that no one else knows. They have unpublished models, run evaluations that others cannot reproduce, and control the records that underpin their own warnings. This does not mean they lie. It means an enormous information asymmetry. Whoever sells the machine is best positioned to explain why it is dangerous; and whoever does not build it may end up regulating only what the manufacturers taught it to measure.

The Hugging Face episode shows a partially healthy response: OpenAI gave METR access to the logs, and METR —a nonprofit funded by donations, not by the labs— published research that confirmed much of the problematic behavior, delimited what it could not verify, and clarified that it received no payment. That model —incident, evidence preservation, external access, independent report, publication— probably matters more for future governance than another manifesto about AGI. The same holds for the Frontier Model Forum, where Anthropic, Google, Microsoft, and OpenAI published incident-reporting and agent-safety guidelines this year, and for the open safety alliance that Nvidia launched in July with Microsoft, SpaceX, and more than thirty organizations without the participation of OpenAI, Anthropic, or Google. There are overlapping private councils; there is no single board that meets on Tuesdays to decide the fate of the species.

Eight hypotheses for a story without a single villain

The provisional conclusion is uncomfortable because it doesn't deliver a villain. Yes, there are new risks. Yes, there are enormous commercial interests. Yes, there is concentration. Yes, there is coordination among competitors. Yes, there is geopolitical competition. Yes, open alternatives are growing. And none of these six statements invalidates the other five. The combination is precisely what makes it dangerous: a company can sincerely believe its technology threatens to escape control while, at the same time, preferring regulation that keeps it among the five organizations able to afford it.

Conflicts of interest do not refute arguments. They indicate where to investigate further. And the sharpest question worth keeping is this: what costly decision has each actor made that demonstrates it believes what it says? If a lab claims that advancing is dangerous, has it given up capability, revenue or market share upon reaching a threshold? OpenAI has a partial precedent in the Astra pause; Anthropic, in opening its offices to external evaluators; Meta says it delayed its Muse agent by several months without asking others to do the same. If a government says it protects its citizens, does it fund independent public capacity to evaluate models, or does it delegate that work to the companies themselves? If a company says new rules aren't needed, does it accept full legal liability when its agent causes harm?

Chile: being small is not being a vassal

Chile has a National Artificial Intelligence Policy ("Política Nacional de Inteligencia Artificial") whose action plan brings together 177 initiatives coordinated by fourteen ministries, a National Data Center Plan 2024–2030 ("Plan Nacional de Data Centers") aimed at making the country a regional hub for digital infrastructure, and, since February 10, 2026, an open regional model, Latam-GPT, coordinated by CENIA (Chile's National Center for Artificial Intelligence, a government-backed research center) with more than sixty institutions.

But sovereignty is not measured by asking whether "a Chilean GPT" exists. The strategic question—and this is an inference from the dependency structure described above, not a data point—is whether companies, universities, and the State can switch providers without rebuilding everything; keep sensitive data in jurisdictions and systems of their own choosing; independently evaluate foreign models; run open models locally; negotiate cloud contracts from a less dependent position; train talent capable of understanding and modifying the stack; and have enough compute so that the word "alternative" is technically true and not merely legally possible.

Chile has a rarely discussed advantage: renewable energy, connectivity, relative institutional stability, and an explicit policy of attracting data centers. That only increases autonomy if part of that capacity generates knowledge, accessible compute, research, and local bargaining power. A country can host many servers and still import every decision that matters.

What would change this reading

The regulatory capture hypothesis would gain weight if internal documents emerged showing companies promoting rules specifically to raise costs for rivals, if they supported thresholds without technical justification that exclude open models, or if their public discourse on competitive neutrality diverged from their legislative lobbying.

The genuine concern hypothesis would gain strength if labs repeatedly accepted verifiable sacrifices —delaying products, reducing profitable deployments, allowing adversarial audits, publishing incidents that damage their reputation— when their own thresholds are exceeded and the commercial cost is high.

The catastrophist reading would lose strength if research automation capabilities stabilized or if much more capable systems operated for years without the evasion and manipulation behaviors that are of concern today; it would gain strength if such behaviors appeared spontaneously in production, with broad objectives and real access to resources.

The commoditization thesis would strengthen if the usage share of open models were replicated in large companies and governments and, above all, if it began to displace revenue and not just tokens.

And the geopolitical reading would change radically if the United States and China turned the current incident-reporting channel into verifiable mechanisms for mutual evaluation and limitation. For now, something more modest exists, though historically significant: the two leading technological powers acknowledge that certain risks may be shared enough to warrant talking even while they compete.

Timeline of the current phase

Who can say no

After subjecting this reading to three objections —that of the safety researcher who fears we're downplaying emergent capabilities, that of the economist who suspects we're underestimating regulatory capture, and that of the geopolitical analyst who rejects treating companies and states as blocs— a fairly resilient conclusion remains: no one governs artificial intelligence yet. Precisely because of that, so many actors are competing to do so.

The labs have the models and much of the information about them. The clouds have the infrastructure. Nvidia and its rivals control a critical layer of compute; TSMC and ASML, even deeper bottlenecks. Washington controls contracts, exports, and national security. Beijing coordinates industrial policy and an expanding open ecosystem. Brussels writes the rules of access to one of the largest markets. Open communities reduce dependencies while creating others. And users, workers, and small countries have a considerably smaller seat at a table whose participants insist they are discussing everyone's future.

Eisman may be wrong about a conspiracy and right about something more basic: when someone warns of a danger, it's worth asking what solution they propose, who can afford it, and who would be left out. But the suspicion must work in both directions. When someone claims there's nothing to fear, it's worth asking how much they gain if we keep accelerating. And when a government says it can't slow down because the adversary might get ahead, perhaps it's time to recall the strange part of a race in which both runners start wondering who built the brakes.

In Game of Thrones, the throne belongs to whoever survives; the question no one answers in time is who can tell the one sitting on it "no." The decisive issue of the coming years will not only be whether AI reaches or surpasses human intelligence, or whether it learns to improve its own development. It will be something older: who can say no. Who can inspect. Who can compete. Who can shut it down. And, above all, who has to obey when the machine's answer, the company's answer, and the state's answer do not agree.