Open AI, Closed Power? A Critique of Zuckerberg’s The Future Is for Everyone

I find myself agreeing with much of Mark Zuckerberg’s argument for a more open AI ecosystem. That is precisely why I think his argument deserves closer scrutiny.

Mark Zuckerberg has published a substantial essay setting out Meta’s philosophy for what he calls a positive AI future: The Future Is for Everyone.

I read it with considerable interest — and, perhaps somewhat surprisingly, with considerable agreement.

Zuckerberg’s central proposition is that artificial intelligence, and eventually what he calls personal superintelligence, should not be concentrated within a small number of corporations or institutions. Instead, AI capabilities should be distributed widely so that individuals, entrepreneurs, researchers and smaller organisations can use them to pursue their own goals. (meta.com)

That principle resonates strongly with me.

One of my growing concerns about AI is precisely the concentration of capability within a remarkably small group of organisations: OpenAI, Anthropic, Google DeepMind, xAI and, increasingly, Meta itself.

The Stanford AI Index 2026 provides some context for this concern. More than 90 per cent of notable frontier models produced in 2025 came from industry. The frontier of artificial intelligence is increasingly being determined not in universities, public research institutions or small laboratories, but inside a comparatively small number of extremely well-capitalised corporations. (Stanford HAI)

So when Zuckerberg argues that:

“The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic.”

I think he has identified something important. (meta.com)

But there is also something missing. The essay tends to assume that making AI models more open necessarily distributes power.

I am not convinced those two things are the same.

The strongest part of Zuckerberg’s argument

Meta deserves some credit here.

While much of frontier AI has moved towards proprietary models delivered through controlled APIs, Meta has been an important counterweight through the Llama ecosystem and its broader support for openly available AI.

There is an important technical caveat. The term open source is disputed in this context. The Open Source Initiative argues that Meta’s Llama licensing does not satisfy its definition of genuine open-source AI. It may therefore be more accurate in many cases to speak of open-weight models rather than fully open-source systems. (Open Source Initiative)

That distinction matters.

Nevertheless, the broader contribution is real.

Open models allow researchers to inspect systems, developers to experiment without obtaining permission from a platform owner, companies to deploy models locally, and entrepreneurs to build products without necessarily paying a toll to the creator of a proprietary foundation model.

These characteristics can lower barriers to entry. And lower barriers to entry matter enormously.

The OECD has warned that access to computing power and high-quality data may become important barriers to competition within generative AI markets. It argues that the stakes are sufficiently high that competition deserves active attention as the industry develops. (OECD)

On this point, I think Zuckerberg is broadly right.

A world in which every new AI company needs permission from OpenAI, Anthropic, Google or another frontier provider before it can innovate is not particularly attractive.

Neither is a future in which three or four companies effectively determine which forms of machine intelligence humanity may access.

But this is where Zuckerberg’s argument becomes more complicated.

Open models do not necessarily mean open power

Suppose Meta releases an extraordinary model tomorrow and anyone can download its weights.

Has AI power genuinely been decentralised?

Perhaps.

But ask another set of questions.

• Who owns the chips required to train the next generation of models?

• Who controls the cloud infrastructure required to run them at enormous scale?

• Who can afford billion-dollar data centres?

• Who has access to the largest pools of proprietary user data?

• Who owns the operating systems, app stores, social networks and devices through which AI assistants reach billions of people?

• Who controls the energy infrastructure required to support all this computation?

Suddenly, the picture looks rather less decentralised.

The US Federal Trade Commission has already examined the relationships between major cloud providers and frontier AI developers, noting the importance of access to compute, data, semiconductor development and distribution channels. (Federal Trade Commission)

This is why I think we need to move the conversation beyond open versus closed models.

The deeper question is where power sits across the entire AI value chain.

A model can be open while the infrastructure underneath it remains highly concentrated.

And a company can advocate openness at one layer of the technology stack while exercising considerable control at another.

A simple way to see this is to lay out the stack itself, from the layer everyone argues about down to the layer almost nobody discusses:

1. Models & assistants — open or closed; the layer this whole debate is fought over.
2. Platforms & distribution — the channels through which billions of people actually reach an assistant.
3. Proprietary data — training and usage data at a scale only a handful of companies hold.
4. Cloud compute — hyperscale infrastructure controlled by a small number of providers.
5. Chips & capital — arguably the most concentrated layer of all.
6. Energy & infrastructure — the physical footprint, and the communities who live beside it.

Openness at the top of that stack is real, and worth defending. But it does not, by itself, touch the five layers underneath it.

There is an uncomfortable cognitive bias in the argument

Matthew Berman’s discussion of Zuckerberg’s essay, provocatively titled Mark Zuckerberg just called out Dario (and Anthropic), is useful precisely because it places the debate in the context of the competing philosophies emerging among the frontier AI laboratories. (YouTube)

Anthropic has generally occupied a more cautious position around releasing increasingly capable models. Zuckerberg pushes strongly in the other direction, arguing that restricting access risks concentrating power in the hands of a few institutions.

The tension is genuine.

But there is also a potential self-serving bias embedded in Zuckerberg’s position.

This does not mean that his argument is wrong.

It means we should ask whether the philosophy conveniently aligns with Meta’s competitive position.

Meta competes against companies that have built substantial businesses around proprietary frontier models. Making the model layer cheaper, more abundant and more open potentially weakens that advantage.

Meta, meanwhile, possesses extraordinary strengths elsewhere: billions of users, enormous distribution networks, immense computing infrastructure, deep pools of capital, advertising systems, consumer platforms and increasingly AI-enabled hardware.

Economically, therefore, openness at the model layer may simultaneously be:

• good for innovation;
• good for consumers;
• good for smaller developers; and
• strategically very good for Meta.

All four can be true.

This does not invalidate Zuckerberg’s argument.

But it should make us cautious about confusing Meta’s preferred market architecture with the public interest.

As Mike Masnick observes in Techdirt, there is an uncomfortable historical irony in hearing the founder of one of the most powerful centralised platforms of the social-media era warn us about excessive technological centralisation. (Techdirt)

Again, this does not make Zuckerberg wrong. Perhaps it makes the question more important.

Should principles about the future structure of AI depend upon whether we trust the chief executives advocating them?

Surely they should survive regardless of whether the person making the argument is Mark Zuckerberg, Dario Amodei, Sam Altman, Demis Hassabis or Elon Musk.

A free market is not the same thing as laissez-faire

There is another distinction I think is becoming increasingly important.

I am strongly attracted to the idea of a free and competitive market for intelligence.

Competition creates experimentation.

Experimentation creates innovation.

Different models can embody different design philosophies. Open models can compete with proprietary models. Small companies can challenge incumbents. Universities, social enterprises, governments and individuals can build systems appropriate to their own circumstances.

That pluralism seems preferable to technological monoculture. But a free market should not be confused with a market without rules. Indeed, functioning markets require rules precisely because successful companies frequently have incentives to reduce future competition.

We have seen this repeatedly throughout economic history.

• Network effects create dominant platforms.

• Capital requirements create barriers to entry.

• Vertical integration allows companies to control multiple layers of a market.

• Acquisitions remove emerging competitors.

• Proprietary standards create lock-in.

And extraordinary economies of scale can transform an initially competitive market into one where genuine entry becomes almost impossible.

AI potentially contains all of these characteristics simultaneously. So perhaps the objective should not simply be free AI or open AI.

Perhaps it should be contestable AI.

An ecosystem in which no company — including Meta — can become structurally impossible to challenge.

That could mean supporting open models. But it might also mean interoperability, data portability, open technical standards, public-interest computing infrastructure, university access to compute, effective competition policy and scrutiny of vertical relationships between model companies, clouds and semiconductor providers.

The goal is not to prevent companies becoming successful.

The goal is to prevent success becoming permanent structural dominance.

Zuckerberg also creates a false choice

One aspect of the essay troubles me more than others.

Zuckerberg repeatedly contrasts two futures.

• In one, powerful AI is concentrated among a few institutions.

• In the other, intelligence is distributed among individuals whose competing interests provide a system of checks and balances. (meta.com)

It is an elegant argument, perhaps too elegant.

Reality may not offer us such a binary choice.

We might simultaneously need:

• open models and proprietary models;

• entrepreneurial experimentation and safety regulation;

• distributed intelligence and powerful public institutions;

• individual freedom and restrictions on particularly dangerous capabilities;

• commercial competition and publicly funded alternatives.

Even Zuckerberg implicitly acknowledges this. His essay advocates openness while simultaneously supporting government access to advanced model checkpoints, continued semiconductor export controls and close co-operation between frontier laboratories and the US government. (meta.com)

Which raises another interesting question.

If the future is for everyone, what exactly does everyone mean?

Zuckerberg's essay moves rather quickly between universal human empowerment and American technological leadership.

Those objectives may often overlap, but they are not necessarily identical.

Equal access is not equal power

There is another assumption worth questioning.

One of Zuckerberg’s examples imagines a world where everyone has access to a superintelligent lawyer. Rather than one party possessing an overwhelming advantage, universal access produces a more balanced legal system. (meta.com)

It is a compelling thought experiment.

But would access actually be equal?

• Would the person running an AI lawyer on a consumer device possess the same capabilities as a multinational corporation spending millions of dollars on inference, proprietary databases, specialised agents and human legal teams?
• Would a small business with an open model genuinely possess the same economic power as a company controlling global distribution?
• Would a researcher in Nairobi have comparable access to compute as a laboratory in Silicon Valley?
• Does giving everybody the same software remove inequality — or simply move inequality somewhere else in the stack?

These questions matter because democratising a technology is not the same as democratising the resources required to use that technology effectively.

Whose voices are missing from this debate?

Almost everyone quoted so far in this argument — Zuckerberg, Amodei, Altman, Hassabis, Musk, and me — already holds some form of power. That is worth naming, because the loudest voices in the "open versus closed" debate are overwhelmingly the voices of people who would be fine either way.

The people who are not in the room tend to be the people the outcome actually falls on:

• the content moderators and data labellers whose work trains these systems, often for a fraction of the value they help create;
• the communities who live beside the data centres, mines and power plants that make frontier AI physically possible;
• the small businesses, public institutions and universities priced out of the compute needed to compete, regardless of how "open" the model above it is;
• researchers and entrepreneurs outside the handful of well-capitalised hubs where frontier AI is actually built;
• workers whose jobs are redesigned by AI systems they had no part in specifying; and
• the people who will live with the infrastructure, and the climate bill, long after this generation of models is obsolete.

This is where the idea of giving voice to values matters. Mary Gentile's work on the subject is not about winning the argument over whether Zuckerberg or Anthropic is right. It is about practising, in advance, how to raise the question when you are the person in the room without institutional power — the engineer asked to ship a feature, the policy adviser drafting a brief, the procurement officer signing a compute contract. The people missing from this debate mostly cannot script a rebuttal to a Meta essay. But the people who can, inside these companies and the governments dealing with them, can practise asking the question out loud before the decision is made rather than after.

The climate cost sits in the same blind spot

The same concentration this essay has been describing — chips, cloud, capital, energy — is also where AI's climate footprint lives.

Training and running frontier models consumes enormous amounts of electricity and water, and the companies that decide how open a model will be are largely the same companies that decide where a data centre gets built, how much water it draws, and which grid absorbs the load. Those siting decisions are rarely made with the communities who host them at the table, and the emissions and water stress that follow do not fall evenly. They tend to land hardest on the communities with the least power to negotiate terms — which is precisely the pattern this essay has been describing at every other layer of the stack.

An AI policy conversation that only asks "open or closed" and never asks "at what environmental cost, and who bears it" is missing half the power question.

Where this critique connects to the Sustainable Development Goals

Several of the UN Sustainable Development Goals give this argument a sharper, more accountable shape than "open versus closed" ever could:

• SDG 7 — Affordable and Clean Energy: frontier AI's power demand increasingly competes with communities for energy, water and grid capacity.
• SDG 9 — Industry, Innovation and Infrastructure: compute, chips and cloud infrastructure, not model licences, are the real bottleneck beneath "open" AI.
• SDG 10 — Reduced Inequalities: equal access to software is not equal power; without compute, capital and data, openness can widen gaps as easily as close them.
• SDG 13 — Climate Action: the environmental cost of training and running frontier models falls disproportionately on communities with the least power to contest it.
• SDG 16 — Peace, Justice and Strong Institutions: contestable AI needs accountable public institutions, not just generous licensing decisions from a handful of companies.

None of these goals resolve the argument between Zuckerberg and his critics. But they are a more useful checklist than "open or closed" for deciding whether any particular AI company — Meta included — is actually distributing power, or simply relocating it.

So what kind of AI economy do we actually want?

I do not have a neat answer.

In fact, I think premature certainty may be part of the problem.

Instead, I increasingly find myself asking questions.

• How many genuinely independent frontier AI laboratories would constitute a healthy market: four, ten, fifty?
• Should governments invest in public AI infrastructure so universities and smaller companies can compete without relying entirely on Big Tech?
• What components of advanced AI should genuinely be open source, rather than merely open weight?
• At what level of capability does openness create risks that cannot reasonably be corrected after release?
• Who decides where that threshold lies?
• How do we prevent safety arguments from becoming convenient mechanisms for incumbent companies to exclude competitors?
• Conversely, how do we prevent openness from becoming a convenient commercial argument for companies seeking to commoditise the part of the technology stack where their competitors currently hold an advantage?
• Should interoperability and portability become fundamental principles of AI regulation?
• How do we prevent concentration in chips, cloud infrastructure and energy from undermining competition at the model level?
• Can a company controlling billions of consumer relationships ever genuinely be considered a decentralising force simply because it releases model weights?
• How should we distinguish legitimate returns to innovation from economic rents created by technological lock-in?
• And perhaps most importantly: what would an AI industry look like in which no single company could ever become powerful enough to determine the future direction of intelligence for everybody else?

These strike me as more useful questions than deciding whether Zuckerberg or Anthropic is simply right.

I hope Zuckerberg wins this argument — but not the market

There is much in The Future Is for Everyone that I support.

I do not want humanity’s most consequential emerging technology controlled by two or three corporations.

I want open models.

I want researchers experimenting with them.

I want small businesses building on them.

I want students running sophisticated AI locally.

I want universities participating at the frontier.

I want multiple competing approaches to alignment, safety, architecture and deployment.

And I want entrepreneurs to be able to create the next generation of AI companies without first asking permission from today's incumbents.

In that sense, I hope Zuckerberg’s argument for openness succeeds.

But I do not particularly want Meta to win.

• Nor OpenAI.

• Nor Anthropic.

• Nor Google.

• Nor xAI.

Perhaps that is the point of a genuinely free market.

The best outcome is not choosing which benevolent technology company should become dominant.

It is designing institutions, markets and technologies so that none of them can.

• Openness is part of that answer.

• Competition is part of that answer.

• Regulation may also be part of that answer.

• Public infrastructure might be part of that answer.

And there will almost certainly be tensions between them.

Zuckerberg is right to ask who should possess the power of advanced artificial intelligence.

I am simply not convinced that giving everyone access to a model answers the question.

Because the deeper question is not merely:

Who gets access to intelligence?

It is:

Who gets to accumulate power around it?

And perhaps that is the AI question we should be discussing much more seriously.

References

• Berman, M. (2026), Mark Zuckerberg just called out Dario (and Anthropic), YouTube. Watch the video. (YouTube)
• Federal Trade Commission (2025), Partnerships Between Cloud Service Providers and AI Developers. FTC report. (Federal Trade Commission)
• Masnick, M. (2026), Zuckerberg Is Right About Open, Decentralized AI. He’s Also The Last Person You Should Trust To Deliver It, Techdirt. Read the article. (Techdirt)
• OECD (2024), Artificial Intelligence, Data and Competition. OECD publication. (OECD)
• Open Source Initiative (2025), Meta’s LLaMa licence is still not Open Source. OSI commentary. (Open Source Initiative)
• Stanford Institute for Human-Centered Artificial Intelligence (2026), AI Index Report 2026. Stanford AI Index. (Stanford HAI)
• Zuckerberg, M. (2026), The Future Is for Everyone: The Path to a Positive AI Future, Meta. Read the essay. (meta.com)