Nadella posted on September 13. Altman on the 14th. Zuckerberg on the 15th.

Their companies compete for the same AI budget, yours included, and by Tuesday all of them had named the same thing as the danger they worry about most. AI power ending up in too few hands.

They all agree on the danger. They split, hard, on the fix. And the fix each one prefers happens to be the product his company sells. That’s the part a buyer should care about, because when you pick an AI vendor you’re also signing up for its bet, and the bet decides how locked in you are, who sets your price, and whether your access survives the next strategy change.

Who said what

All three were answering Dario Amodei’s September 12 essay asking labs to slow how fast frontier capabilities improve. I covered the essay and Anthropic’s $517 billion compute bill last week, so I’ll skip it here.

Altman’s post, as Fox Business reported it, lists two ways this goes very badly. Losing control of the future to AI is the first. The second: “we could end up in a world with too much concentration of power,” with one person or company using extraordinarily powerful AI to impose a worldview on everyone. He wants a narrow middle path, and OpenAI has signed on to embedded outside evaluators.

Zuckerberg got there first, in an August 10 essay called “The Future is for Everyone.” If superintelligence is held by a small number of individuals, businesses, governments, or AI itself, he wrote, outcomes will be “less favorable for everyone else” (Fox Business). His fix is distribution. On September 15 he added that “any policy that slows American model releases — even by a month — could add significant risk to American leadership while letting foreign models race ahead” (Business Today).

Nadella was the one talking to buyers. “Broad access and choice at every layer of the AI stack; enterprise control of learning loops and models,” plus a line that no organization should depend on any one model provider, per Unite.AI.

Three bets, priced out

The bet	Lock-in	Who sets the price	Continuity of access
Closed models, paced (Anthropic, OpenAI)	High. Your prompts and evals are tuned to a model you don’t hold	The vendor	Whatever the contract says
Open weights (Meta)	Low for the version you downloaded	You, through hardware and staff	Permanent for that version, uncertain for the next
Choice inside a platform (Microsoft’s pitch)	Moves from the model to the platform	The platform	As long as you stay on it

That table is my reading, not anyone’s pricing sheet. Two rows deserve more than a cell.

The closed row has a live example. On August 28 OpenAI told Cursor it would end its model-access contract after SpaceX bought the company. Cursor survived it because those models were about 5% of its traffic. Most companies have never measured their own number.

The open row has a catch in Meta’s own history. Muse Glimmer is real: 30 billion parameters, Apache 2.0, on Hugging Face, and quantized it fits on one 24GB consumer GPU, VentureBeat reported. But Meta moved away from open weights in April when it launched Muse Spark as a closed model, then reversed in August and promised Spark 1.2’s weights “soon.” Two turns in four months. The file you downloaded stays yours. The next version is someone else’s call.

And no, Washington won’t referee this. On September 14 Donald Trump, on speakerphone with Jensen Huang, called the backlash against AI “a hoax,” TechCrunch reported.

My bias, and what I’d do anyway

At IQ Source we build almost everything on Claude. That puts us on Amodei’s side of this, and Anthropic also declined to sign Jensen Huang’s open-weights letter in August. Read the rest knowing that.

I still wouldn’t pick a team. None of these three will call you before changing their mind, and one of them already changed it twice this year.

You might think the safe move is going all in on open weights. No, not that either. A model that fits on a gaming card doesn’t do the same work as one running in a data center, and running your own model means people on payroll who know how.

What you can control is where your AI’s working parts live. Prompts, test cases, evaluation data. If those exist only inside a vendor’s console, you’ve handed over the only key to your own house.

So, a few things I’d want answered this month. Which of your tasks would run fine on a 30B model in your own rack? What does one week without your main provider cost you? Do your tests run against two models, or one? Who on your team actually knows?

We can build that map with you, step by step, showing which model each part of your operation leans on and what breaks if that model changes price, license or owner. Better to have it before one of these three changes bets.

Map which models your operation depends on