TL;DR — Key Takeaways
Zuckerberg is warning against concentrated AI power, but Meta’s position on openness appears to shift with its competitive standing.** Meta promoted open models when it was behind**, yet its new Muse Spark model is closed and sold commercially.** Every major AI company supports openness where it weakens rivals**, while protecting the technology, infrastructure or distribution layer it controls.** Self-interest does not make the argument wrong:Wider access to advanced AI may still produce more public value than concentrating it inside a few companies. The real test comes if Meta takes the lead.**Zuckerberg’s commitment to openness will be clearer when releasing a frontier model could threaten Meta’s own advantage.
Mark Zuckerberg does not give many interviews. This week he gave four.
An opinion column in The Wall Street Journal. A sit-down with the Journal’s Meghan Bobrowsky. A conversation with the Financial Times). And an interview with The New York Times, which he grants to traditional news organizations about as often as Halley’s Comet comes around. Four outlets, one week, one message. You think he was making a point?
The message is that superintelligence will be the most profound technological advance we see in our lifetimes, provided its power is not restricted to a few institutions. The discourse coming out of the other labs building this technology is filled with doom. The data supports optimism. Distributing the technology creates more value than harm by a considerable margin.
It may surprise people who expect me to take the other side, but I agree with most of it.
Concentrating the most consequential technology of the century within three or four companies is a genuinely bad outcome, and the companies arguing hardest for restrictions happen to be the ones that would benefit most from it. Zuckerberg told the FT there is a question of regulatory capture here, and he is right that there is. A mandatory review period of 30 or 60 days does not sound like much until you notice how fast the field moves and who currently holds the lead. Anthropic and OpenAI are asking Washington for rules that would freeze the board in a position that happens to favor Anthropic and OpenAI. That deserves scrutiny no matter who raises it.
But I want to talk about who is raising it, and when.
The Weapon of Whoever is Behind
Meta open-sourced Llama in 2023. The company did not do that because Mark Zuckerberg woke up one morning holding a copy of the GNU Manifesto. Meta did it because Meta was behind. Google had a lead. OpenAI had the product everyone was talking about. Meta had capable researchers and no realistic path to catching up by charging for API access nobody was clamoring for yet.
So Meta gave the models away. Free models commoditize the layer where your competitors are trying to build a business. Free models buy developer mindshare you cannot purchase any other way. Free models turn a deficit into distribution. It was a smart move, and at the time most of us read it correctly as a competitive maneuver rather than a philosophical awakening.
Openness was the weapon of the company that was losing.
Then the position changed. After Llama 4 underdelivered, Meta put $14.3 billion into Scale AI, hired founder Alexandr Wang to rebuild the AI organization, renamed it Meta Superintelligence Labs, and assembled an expensive roster of researchers around him. That team built a model. The model is called Muse Spark. It is not open. This month Meta began selling access to it for the first time.
Same company. Same chief executive. Same week, he is telling four news organizations that centralizing AI power would amount to abandoning our values.
I am not calling that a lie. I am calling it a position.
Kettle, Meet Pot
I wrote last week about the Open Weights and American AI Leadership letter, the one Jensen Huang used his first-ever post on X to promote, and about my grandmother’s observation that it is easy to talk when it is someone else’s tuchas on the line. Every signatory on that letter wants openness at the layer directly above wherever it keeps its own scarcity. NVIDIA wants abundant models because abundant models consume accelerated computing. Microsoft wants interchangeable models because Azure stays sticky either way. Application vendors want models cheap because models are an input cost.
Meta signed that letter. Meta also sells a closed model.
The pattern does not stop at the American border. Days before Zuckerberg’s media tour, Xi Jinping stood up at the World Artificial Intelligence Conference in Shanghai and called on the world to encourage open source, openness, collaboration and sharing. Moonshot AI released Kimi K3 at the same conference. Chinese labs have been shipping capable open-weight models at prices American providers have struggled to match, and they have been doing it while export controls limit their access to the best silicon.
China is not championing openness out of conviction any more than Meta was in 2023. China is behind on compute and ahead on distribution, so China argues for the terms that favor distribution. Same weapon, different hand, aimed at a different moat.
Notice what is absent from Xi’s enthusiasm for openness. An open internet. An uncensored public square. Openness turns out to be a policy about models rather than a principle about information.
Both Things are True at Once
None of this makes the argument wrong. As I said, I agree with most of it.
Self-interest and public benefit are not mutually exclusive. The internet was not built by companies with spotless motives. Mistral’s commitment to open weights is probably sincere, and it is also the only strategy available to a company that cannot outspend OpenAI at the frontier. Playing the hand you were dealt is not a character flaw. It is what operators do.
So Zuckerberg can be entirely correct that concentrated AI power is dangerous, and entirely motivated by the fact that Meta does not currently hold the concentrated power in question. Both statements survive contact with each other.
What I would ask readers to do is stop treating these arguments as testimony and start treating them as positions. Before you weigh what a company says about how the AI market should work, look at where that company sits in the AI market. The correlation is close to perfect, and it runs in one direction. Openness for the layer I do not own. Control for the layer I do. That is the Indispensability Trap operating exactly as designed.
The Test That is Coming
Meta has projected up to $145 billion in capital spending this year. It reports earnings this week, and analysts expect it to go free cash flow negative. The company laid off 8,000 people in May and attributed the cuts to the cost of its AI ambitions. That is an enormous amount of money riding on the proposition that Meta eventually builds something worth protecting.
If it does, we will find out what Zuckerberg actually believes. The interesting question is not whether he is right today. It is what he says on the morning Meta finally has a frontier model that leads the field, and somebody asks him to release the weights. I have a guess. So do you.