The $14B Loophole and Meta's AI Gulag

Okay so today we're talking about Scale AI, which is basically the invisible—

Wait, before you do the whole invisible engine of the AI boom thing, we need to talk about the gulag.

The gulag. Yeah, we're definitely getting to the gulag.

I mean, you cannot bury the lede here! Meta conscripted roughly sixty-five hundred elite software engineers last year, these people making half a million dollars a year, and forced them to do mundane data labeling.

Right, basically clicking boxes for AI.

Exactly. And they hated it so much they literally called the division the gulag. One guy even hijacked a company-wide livestream just to insult an executive over it.

Which is wild. But it actually makes perfect sense when you look at the sequence of events. Because that entire meltdown traces back to June 2025, when Mark Zuckerberg decided he had to have Alexandr Wang.

Right, Wang. The MIT freshman who dropped out to build Scale AI with Lucy Guo back in 2016. He was what, 28 last year?

Yeah, 28, already the youngest self-made billionaire. And Scale was pulling in, I think, about 870 million in revenue. Because they solved the dirty secret of AI.

That it is actually just human labor.

Exactly. You can have all the GPUs in the world, but if you don't have thousands of people labeling the data, your AI is totally useless. Scale had this massive army of over a hundred thousand gig workers on platforms like Remotasks, plus PhDs and lawyers for the high-end stuff.

So Zuckerberg looks at this and goes, I need Wang to run my new Meta Superintelligence Labs. But he can't just buy Scale AI outright because the FTC would block it before he even finished signing the check.

Right. So he pulls the ultimate antitrust dodge. The license and release loophole.

It is so cheeky. Fourteen point three billion dollars for a 49 percent non-voting stake. He essentially paid 14 billion just to poach a 28-year-old CEO and avoid a monopoly lawsuit.

And the industry completely lost its mind. OpenAI, Google, xAI, they all used Scale. The moment Meta bought half of it, everyone panicked that Zuckerberg suddenly had a backdoor to see what training data his rivals were buying. One CEO called it the equivalent of an oil pipeline exploding.

Which, honestly, was a valid fear considering what happened next. Remember the Business Insider story in July?

Oh man, the public Google Docs.

Yes! Scale contractors were literally using publicly accessible Google Docs to track highly confidential training tasks for Google and Meta.

I forgot about that. Didn't some of the leaked files have to do with training AI on zombie apocalypses?

Yeah, zombie apocalypses and, like, advanced plumbing scenarios. Just sitting there on the open internet. It was a total mess. And right around then, Scale laid off 14 percent of its staff. ChatGPT logs leaked where the AI predicted Scale would be dead in 24 months.

Right, everyone thought it was a zombie startup without Wang. But here is the crazy part. That didn't happen.

No, the exact opposite happened. Competitors like Turing and Appen got a massive bump initially, but Scale actually pulled it off.

Yeah, under their new CEO, Jason Droege, the guy who founded Uber Eats, they pivoted hard into defense. They bought a tech firm called ICG Solutions, snagged hundreds of millions in DOD contracts, and now they are projecting almost 2 billion in ARR for 2026.

That's completely insane.

Defied all the odds. But meanwhile, back at Meta, Wang is cracking the whip, causing the whole gulag revolt.

Because Wang's philosophy is data first. He doesn't care about the engineers' egos. He knows you need the absolute highest quality human data, and Zuckerberg backed him up. Zuck had that leaked memo where he basically said, look, our engineers are smarter than third-party contractors, so they have to label the data.

Even if it means paying them a fortune to do data entry. But, and this is the really surprising shift, this is why Meta completely abandoned open source.

Right. RIP Llama.

Exactly. They pivoted to Muse Spark. Closed source, multimodal, highly agentic. And people laughed at Muse Spark 1.0 back in April.

Total yawn.

But Muse Spark 1.1 just came out in July and it is crushing the benchmarks. So as chaotic as the gulag was, Wang's brutal strategy might actually be working.

Well, with Meta spending, what was the CapEx projection? Like 125 to 145 billion to buy 1.3 million GPUs this year? It better work.

Seriously. The whole thing just shows that the real bottleneck in AI isn't silicon anymore. It's the human supply chain. Who controls the data, and who's willing to spend 14 billion dollars just to hire the guy who knows how to wrangle it.

Or willing to force their senior developers into a data-labeling sweatshop.

Ha, exactly. Next time, we'll dive into what those DOD contracts actually mean for Scale, but for now, we'll leave it there.