Zuckerberg’s AI Manifesto Lost it at Compute

Among the many slippery bits of Mark Zuckerberg’s AI manifesto is this: “superintelligence-for-all” empowerment will never happen if Meta and other hyperscalers monopolize compute. Control over compute is a tool of domination as powerful as any of the frontier model secret weights that Zuckerberg slags.

The Atlantic’s Will Oremus observes wryly that in the manifesto, “a man who single-handedly commands a $1.5 trillion tech giant comes out as a critic of corporate power.” Platformer’s Casey Newton notes that while Zuckerberg “celebrates individual voices and democracy,” his company “is working to warp the democratic process through secretive data center deals.” Through those deals, Meta is trying to catch up with the other four hyperscalers (Amazon, Google, Microsoft, and Oracle), who together reportedly own more than 70% of the world’s AI compute. Zuckerberg doesn’t talk about this source of power. Instead, he focuses further up the stack on models and applications. Here, he says, “[r]ather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.”

But what about the physical infrastructure needed to produce that superintelligence? Infrastructure power is technology power. The Commerce Department’s NTIA, in its 2024 report on open weight models (p.30) noted that access to compute is a significant chokepoint for innovating with such models. As David Widder and co-authors put it, “the computational resources needed to build new AI models and use existing ones at scale… are scarce, extremely expensive, and concentrated [and] this significant resource asymmetry undermines any claims to democratization that the availability of ‘open’ AI models might be used to support.” (p.8)

In other words, frothing about AI democracy and liberty is useless without addressing compute. Zuckerberg knows this. That’s why he acknowledges compute as a pinch point in the flow of freedoms he champions. “No matter how intelligent AI becomes, there will always be a finite amount of compute and therefore an opportunity cost for how we use it,” he writes. His response to this scarcity is for the tech giants to build ever more compute so that there’s enough for training “recursive self-improvement” while also “committing the significant majority towards people's individual goals.” Here, as elsewhere in the manifesto, governance choices are the tech companies’ to make. They will decide how compute is allocated. They will own the means through which people self-actualize. Every “exceptionally capable personal agent” and every “powerful tool to build businesses” and all the other compute-driven wonders the manifesto teases will depend on the few companies controlling compute.

For the conveniently libertarian political theory of Zuckerberg’s manifesto, this concentrated private power is unimportant. The only real threat to AI-enabled individual liberty is government tyranny. “There is a risk that an imbalance of power between individuals and government leads to a loss of freedom or totalitarian state,” Zuckerberg writes. “This is a constant tension in democracy.” What’s also a constant tension in democracy is how to constrain overweening private power. Technology policy in the American democracy has never been libertarian. Zuckerberg’s only recognition of private compute power is his assurance that Meta (and other benign titans) will distribute compute widely and focus heavily on alignment. In other words, “trust us.” Given Meta’s history on the trust front, as it appeals a pathbreaking New Mexico court judgment that it is a public nuisance, this tack is laughable. And it has been loudly laughed at here, and here, and here. Obviously, binding and transparent regulation is a more effective check. The manifesto will have none of it ,and Meta has been instrumental in making sure the federal government enacts none.

What might a compute policy look like that took self-empowerment and democracy seriously? A person writing a version of Zuckerberg’s screed with an eye to network studies might have cited Ithiel de Sola Pool’s classic 1984 work, Technologies of Freedom. Pool wrote about how new technologies (like “electronic publishing”) might usher in new freedoms through “distributed intelligence” (p. 251) if only government could abstain from centralizing control. This fits well with Zuckerberg’s paean to decentralized open weight models and preference for policy that amounts to little more than facilitating Meta’s business model. But here’s the thing. Pool understood that freedom at the application layer might require regulation at the infrastructure layer in the form of common carriage and access requirements.

For compute, and more generally for cloud services that mediate access to compute, we can look to the same family of tools long used in US regulation of concentrated networks, platforms, and utilities.

1. Remove chokeholds on access to compute where necessary through some combination of nondiscrimination, interconnection, and common carrier obligations. Jai Vipra and Sarah Myers West of the AI Now Institute have a nice report on this from 2023 when the FTC and the Department of Treasury were beginning to sound the alarm on undue concentration and points of failure in the compute and cloud services markets. Last year, Asad Ramzanali of the Vanderbilt Policy Accelerator published a blueprint for federal legislation for the AI stack, including compute, that he described as pro-competition and pro-national security, and which might actually create the balance of power Zuckerberg’s manifesto praises.
2. Ensure that there is some meaningful supply of “public compute” that is not controlled by any company. The scarcity of compute for scientific research has led to nascent federal efforts to amalgamate commercial and government compute for research in the National AI Research Resource. New York has created Empire AI with new compute resources, again mostly for research. One might imagine a federated network of state compute forming a sovereign compute infrastructure of the kind Europe is trying to build. A new idea would be to regularize public compute set-asides as part of data center builds. What if hyperscalers had to set aside a certain amount of  “civic compute credits”, like a public park, for the autonomous use of nonprofits, local governments, and educational institutions? Cable companies had to set aside PEG channels when building out their networks and the FCC reserved noncommercial broadcast channels for public service.
3. Require transparency. One could talk about transparency at each of the layers in the stack, but let’s limit it here to compute. Here too, the libertarian Pool acknowledged that to have freedom at the top layers – the only ones Zuckerberg really addresses – we need to know how companies are managing the infrastructure choke points. In the case of compute, this might include the amount of compute capacity a firm actually controls, to whom that capacity is allocated (including to the owner itself), and pricing terms.

For ten years at least, Zuckerberg has repeatedly promised connection and democracy on his platforms. Instead, we’ve gotten slop and oligarchic control. The AI manifesto promising superintelligence-for-all is cut from the same cloth. But still, it’s worth taking the aspiration at face value and asking what it would take to actually democratize AI at each layer of the stack, starting with compute. It would require that people who want to run models and applications can do so freely without control or interference from infrastructure providers.