THE PAPERWORK. $14.3 billion, 49% of a company, and the founder leaving with the stake. Then Meta closed the model it used to give away.
JUNE 12, 2025. META ANNOUNCED A $14.3 BILLION INVESTMENT IN SCALE AI. The round valued the company at over $29 billion. Meta was buying 49% of it.
And the founder was leaving. Alexandr Wang — 28 years old — stepped down as chief executive to join Meta.
Which is the whole transaction in one sentence: Meta paid fourteen billion dollars, took less than half of a company, and got the man instead.
WHY 49% AND NOT CONTROL? BECAUSE THE POINT WASN'T CONTROL.
Scale AI's business was producing and labelling the data that AI labs train on. Wang started it in 2016 as a 19-year-old MIT student — labelling street photographs for self-driving cars. Then he dropped out, went through Y Combinator, and by 2025 he was supplying the same labs Meta was losing to.
That's the asset. Not the servers. The training pipeline that the rival labs had come to depend on.
So a 49% stake is a very specific instrument. It doesn't give Meta control. It makes Scale a compromised vendor for everybody else. If your data labelling contractor is nearly half-owned by your biggest competitor — and the man who built it now works there — you have a problem you can't fix with a contract.
Meta didn't buy a supplier. It made the supplier unusable to its rivals and hired the only person who knew how to run it.
NOW GO BACK A YEAR AND SEE WHY THEY PAID.
July 2024: Zuckerberg published a post saying that starting the next year, future Llama models were expected to become "the most advanced in the industry."
Llama was the open-weight strategy — championed by chief AI scientist Yann LeCun, a Turing Award winner. Give the model away, let developers build on it, and nobody gets locked into a rival's closed model. Meta spent to make it happen: about $70 billion in capital expenditure for 2025, fourth-highest in big tech behind Amazon, Google and Microsoft.
Then Llama 4 shipped in April 2025 — and it was neither the most advanced model in the industry nor the best open model. Chinese open releases, DeepSeek among them, had passed it.
$70 billion of capex and the strategy was losing on its own terms.
JUNE 30, 2025: META PULLED ITS AI TEAMS INTO ONE UNIT — META SUPERINTELLIGENCE LABS — AND MADE WANG THE COMPANY'S FIRST-EVER CHIEF AI OFFICER.
That was widely read as the end of LeCun's influence, and the read was right. LeCun left Meta after twelve years, and by January 2026 he was executive chairman of his own company, AMI Labs, in Paris. In March it closed a $1.03 billion seed round at a $3.5 billion pre-money valuation.
The man who defined Meta's open-model era left and raised a billion dollars to do the opposite of what Meta was now doing.
AND THE PRICES THEY WERE PAYING FOR RESEARCHERS GOT ABSURD.
Sam Altman publicly claimed Meta was offering OpenAI researchers signing bonuses of up to $100 million. The Wall Street Journal reported one researcher turned down an offer worth about $1.5 billion over at least six years. Meta called that characterisation "inaccurate and ridiculous."
Whichever version is right, the direction is unambiguous: this was a company paying scarcity prices because it had already spent the money and had nothing to show for it.
JULY 30, 2025: THE LETTER THAT ANNOUNCED THE REVERSAL.
Zuckerberg published a note declaring that developing superintelligence was now in sight — and warning that Meta would have to be careful about what it chose to open source.
That sentence was the product plan.
Inside Meta, Wang's team was building a model codenamed Avocado. On April 8, 2026, it shipped as Muse Spark — the first model out of Meta Superintelligence Labs, built over about nine months, and described by the company as the first production of a ground-up overhaul of its AI efforts.
And you couldn't download it. Artificial Analysis, which benchmarked it independently, called it the first frontier-class model from Meta since Llama 4 Maverick — and noted the thing that matters more than any score: it is the first Meta model not released as open weights.
Meta said it hopes to open-source future versions. But the company that made open weights the centre of its AI strategy shipped its best model closed.
THEN IT STARTED SELLING THE THING THE OPEN MODEL WAS SUPPOSED TO PREVENT.
July 2026 earnings call: Meta had rebuilt its assistant on Muse Spark. About six weeks later came the app.
SEPTEMBER 8, 2026: MUSE — a personal agent, not a chatbot. It has its own computer and tools in the cloud — Meta calls it the Muse secure virtual machine — so it can act on your behalf: open a browser, fill out a form, list a couch on Marketplace, send the email, book the thing.
It reads your email, calendar, payments, health, shopping and smart home, and it's reported to be modelled on the open-source agent OpenClaw. There's a free tier, and paid tiers at about $20 and $100 a month depending on usage. A confidential version, where the whole machine is encrypted with a key only the user holds, is promised for later.
So the arc from June 2025 to September 2026 is: buy the data pipeline and the man, form the lab, close the model, then put it in an app with a subscription.
SPLASH'S READ: The usually-told version of this is a redemption story — aging giant hires wunderkind, giant recovers. That's not what the documents show.
What they show is a company that changed its licence the moment it had something worth keeping.
Llama was open because Meta was losing and openness was distribution. Give it away, get the ecosystem, make the closed labs compete with free. Nobody can call that hypocrisy — it's a rational strategy and it worked for years.
But the second Meta believed it had a frontier model, the licence changed. Not a little. Completely: from download-anything to closed, with a subscription behind it.
Which is the same sentence this site keeps arriving at, one more time: New Balance kept the factories. Vans sold the machine. Yeezy never owned a mold. Suge Knight's catalog went for $18 million. Ownership isn't a position you hold. It's a decision you make at the moment the asset becomes valuable — and the person holding the paperwork makes it alone.
Meta paid $14.3 billion for 49% of a company and got the founder. Then it spent a year and a half closing everything he built at the open-model company he joined.
That's not a pivot. That's what happens when the paperwork finally matters.