Monday Momentum
Posts
The Labs Just Wrote Their Own Rules

The Labs Just Wrote Their Own Rules

How three AI companies proposed three different governance frameworks in three weeks, and why each proposal maps precisely onto the weakness its author most needs to hide

Justin Wright
July 20, 2026 • Est. Reading Time: 6 minutes

Happy Monday!

Hassabis published his proposal, titled 'A Framework for Frontier AI and the Dawning of a New Age,' on X on July 14. (source: TechCrunch)

On July 14, Demis Hassabis, the Nobel laureate who runs Google DeepMind, published an essay calling for a new organization to regulate frontier AI. The design he proposed is modeled on FINRA, the industry-funded body that polices Wall Street under government oversight. Labs would voluntarily submit models for a review of up to 30 days before release. Once the system proved itself, passing that review would become mandatory to deploy any model to American users.

Read in isolation, it is a thoughtful proposal from a scientist who takes AI risk seriously. Read in context, it is the third act of a play that has been running for several weeks.

On July 2, Sam Altman proposed giving the US government a 5% equity stake in OpenAI through a sovereign wealth fund, Anthropic countered with a "digital dividend" funded by future AI-sector taxes rather than equity, and now Hassabis has proposed an industry-funded standards body. Three of the most powerful people in AI, three different governance frameworks, three weeks. The convergence is a race to write the rules before someone writes them for you.

In three weeks, the three leading US AI labs each proposed a different framework for governing frontier AI. OpenAI offered the government a 5% equity stake. Anthropic proposed a tax-funded digital dividend. Google DeepMind proposed a FINRA-style standards body the industry would fund and help design. Each proposal maps onto its author's competitive position: Altman needs a regulatory shield, Anthropic wants to avoid dilution, and Hassabis wants a referee that can slow a race Google is losing. The labs have stopped resisting regulation; they are now competing to define it.

TL;DR

Three Proposals in Three Weeks

In June, the US government asserted control over frontier AI faster than anyone expected. It pulled Anthropic's Fable 5 off the market with export controls, then gated OpenAI's GPT-5.6 to roughly 20 vetted partners before a full public release. It did all of this without legislation, using phone calls and Commerce Department letters. The labs learned, in the space of a month, that the government could ground their best models at will.

Then the proposals arrived. On July 2, the Financial Times reported that Altman had pitched a 5% OpenAI stake, worth roughly $42.6 billion, directly to President Trump, Commerce Secretary Lutnick, and Treasury Secretary Bessent. Anthropic's digital dividend, first floated in June, redirects future AI-sector tax revenue to the public rather than handing over equity. And on July 14, Hassabis published "A Framework for Frontier AI and the Dawning of a New Age," laying out the most concrete institutional design any lab has put on paper.

Three companies with three mechanisms but one shared recognition: the era of launching whatever you want, whenever you want, is over, and the only question left is who gets to design what replaces it.

Three Labs, Three Governance Proposals

Lab	Proposal	Mechanism	What It Protects
OpenAI	5% government equity stake	Sovereign wealth fund	Regulatory goodwill, political cover
Anthropic	Digital dividend	Future AI-sector taxes	Equity, no dilution before IPO
Google DeepMind	FINRA-style standards body	Industry-funded review	A referee that can slow the field

Each Proposal Fits Its Author

Here is the part that makes this more than a governance story: each proposal maps onto the exact weakness its authors most need to manage.

OpenAI is losing ground as ChatGPT's market share fell below 50% for the first time. Anthropic's $47 billion revenue run rate nearly doubles OpenAI's. Altman needs political cover and a regulatory relationship that protects OpenAI's position. A 5% equity stake buys exactly that: a government financially incentivized to see OpenAI succeed. The most accommodating proposal comes from the company that most needs accommodation.

Anthropic is winning on revenue and preparing an October IPO at a $965 billion valuation. The last thing it wants is to dilute equity or hand the government an ownership position before going public. A tax-funded digital dividend costs nothing today and only activates after profitability, which Anthropic projects for 2028 or 2029. The company with the most to lose from dilution proposed the one framework that involves no dilution.

Google DeepMind is behind as Gemini 3.5 Pro slipped for months over coding failures, forcing a full pretraining restart. Four senior Gemini researchers left for Anthropic in a single week, part of a talent exodus that erased more than $225 billion in Alphabet market cap. A FINRA-style body that reviews every frontier model for up to 30 days before release does one thing very well: it slows everyone down. When you are behind, a universal speed limit is a gift. The lab losing the race proposed the referee that can pause it.

The Convergence Is the Capture

The remarkable thing about Hassabis's proposal is who praised it. Altman, Microsoft's Satya Nadella, and even Elon Musk offered rare public support. When the fiercest competitors in an industry all endorse the same regulatory structure, that is worth examining closely.

The structure Hassabis proposed is funded by the labs, staffed with input from the labs, and calibrated against tests the labs help design. FINRA works on Wall Street because it operates under hard SEC authority with decades of legal precedent. A frontier AI equivalent would start from scratch, with the incumbents holding real influence over the machinery before any independent version exists. Critics have a name for this: regulatory capture. Rules written to make AI safer can wind up entrenching the biggest companies by raising the cost of entry for everyone else.

The concern is not new; Meta's Yann LeCun warned about it in 2023, naming Altman, Hassabis, and Amodei directly and calling their lobbying "a regulatory capture of the AI industry." Three years later, all three have proposed frameworks that would put the government in partnership with, and partly dependent on, the very companies it is meant to regulate. A body funded by the firms it inspects tilts the same way whether the people running it are cynics or saints. The structure does the work.

To be fair, the alternative is not obviously better. The government's improvised approach in June, grounding models by letter and gating releases by phone call, was arbitrary and opaque. A transparent standards body with published criteria would be an improvement over Commerce Department discretion. And Bernie Sanders has proposed something far more aggressive: a one-time 50% stock tax on large AI companies to fund a public wealth fund potentially worth $7 trillion. Against that backdrop, the labs' proposals look a lot more like negotiation (which may be the point).

What This Means for Practitioners

For founders, the message is that the regulatory perimeter around frontier AI is being drawn right now, and the companies inside it are holding the pen. If you are building toward frontier capability, understand that any of these three frameworks raises the cost and complexity of shipping a model. A 30-day review, an equity contribution, or a tax obligation all favor incumbents with the resources to absorb them. Factor the coming compliance layer into your roadmap.

For enterprise buyers, the practical takeaway is stability. Whatever framework emerges, the major American labs are signaling they intend to operate within government-sanctioned bounds. That reduces the risk of another abrupt Fable-style takedown disrupting a model you depend on. It also means the most capable models will increasingly come with a government seal, for better or worse.

For anyone tracking the industry, watch which proposal gains traction, because it will tell you who won the positioning war. If a FINRA-style body forms, Google bought time. If the equity stake advances, OpenAI bought cover. If the tax model wins, Anthropic protected its cap table straight through its IPO.

The Bottom Line

For two years, the AI industry resisted regulation. In three weeks, its three leading labs each volunteered a framework for it; that reversal is the real story. The labs learned in June that the government would regulate them one way or another, and they decided it was better to write the rules than to receive them.

Each proposal is a mirror: Altman's reflects a company that needs a shield, Anthropic's reflects a company protecting its equity through an IPO, and Hassabis's reflects a company that would benefit from slowing the whole field down. They disagree on the mechanism, but they agree on the direction, and on who should hold the pen. That agreement, from three companies that agree on almost nothing else, is either the beginning of responsible AI governance or the most elegant capture in modern regulatory history. Possibly both.

In motion,
Justin Wright

If the only people with the technical expertise to evaluate frontier AI models are the labs building them, is an industry-funded standards body a necessary compromise or an inevitable capture, and is there any version of AI governance that does not ultimately run through the companies it is meant to constrain?

Food for Thought

Quick Hits

Google delayed Gemini 3.5 Pro for months after coding benchmarks fell short, scrapping the base model and restarting pretraining. The new version targets a 2 million token context window and a Deep Think reasoning layer. (9to5Google)
New York became the first US state to pause new hyperscale data centers, with Governor Hochul signing a one-year moratorium on facilities using 50MW or more. State electricity prices are up nearly 68% since 2019. (CNBC)
Anthropic is in early talks with Samsung to manufacture a custom AI chip on a 2nm process, after hiring Clive Chan, an early engineer on OpenAI's chip program. (TechCrunch)
Xi Jinping launched the World AI Cooperation Organization in Shanghai, with 29 countries signing on, pitching China as the AI partner to the developing world and criticizing US "overstretching" of national security. (Al Jazeera)

If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!

Episodes can be found below - please like, subscribe, and comment!

Spotify: https://creators.spotify.com/pod/profile/mostly-humans-podcast/
Apple: https://podcasts.apple.com/us/podcast/mostly-humans/id1831319729
YouTube: https://www.youtube.com/@Mostly_Humans_Podcast