Meta's Deal Creates a New AI Chokepoint

Google Splits With Scale AI

One of the basic rules of a high-stakes, zero-sum competition is that you probably shouldn't let your biggest rival look over your shoulder while you're working on your secret plans. If you are, say, Coca-Cola, you don't hire a team of PepsiCo executives to taste-test your new formula. It’s just good sense.

And yet, in the hyper-competitive race to build artificial intelligence, something very much like this was happening. All the big players—Google, Microsoft, OpenAI, Meta—were, to varying degrees, relying on the same crucial third-party supplier, a startup called Scale AI. Scale’s business isn’t making chips or servers; it’s providing the one resource that has become just as scarce and valuable: high-quality, human-annotated training data. It hires armies of experts—scientists, historians, coders—to teach the AI models how to reason, how to be more accurate, how to be less weird. This is the secret sauce. To get this service, the AI labs hand over their prototype models and proprietary data to Scale.

This was a workable, if slightly tense, arrangement. Until last week. That’s when Meta announced it was taking a 49% stake in Scale AI.

This is, you know, a problem. The reaction from Scale’s other customers was swift and predictable: They are all running for the exits. Here is Reuters on the fallout:

Companies that compete with Meta in developing cutting-edge AI models are concerned that doing business with Scale could expose their research priorities and road map to a rival, five sources said. By contracting with Scale AI, customers often share proprietary data as well as prototype products for which Scale’s workers are providing data-labeling services. With Meta now taking a 49% stake, AI companies are concerned that one of their chief rivals could gain knowledge about their business strategy and technical blueprints.

Of course they are. Continuing to use Scale would now be like letting Meta’s engineers sit in on your most sensitive AI strategy meetings. You’d be paying your top competitor for the privilege of seeing your homework. Which is awkward.

What this really exposes is that the frantic, capital-intensive AI arms race has bottlenecks everywhere. We talk a lot about the hardware chokepoints: the scramble for Nvidia’s GPUs, the reliance on TSMC’s fabs, the desperate search for enough electricity to power it all. But the Meta/Scale deal reveals a critical “soft” chokepoint in the supply chain. It turns out that building superhuman intelligence requires a huge amount of very specialized human intelligence, and the market for providing that service just got violently bifurcated.

The result is a land grab. The remaining independent data-labeling firms are seeing their demand triple overnight. And the big AI labs are scrambling to bring this work in-house, creating their own secure, internal teams of expert trainers. The AI supply chain, once a web of uneasy frenemies, is rapidly Balkanizing into walled gardens. The idea of a neutral Switzerland for AI data may be dead. The battle for silicon was just the start; the war for the human teachers has now begun in earnest.

Your New Coworker Is a Bot

For a while, the polite fiction in corporate America was that AI was a “copilot,” a helpful assistant that would make human workers more productive, not replace them. It was a comforting story, but the quiet part is getting louder. The latest executive to say it out loud is Allison Kirkby, CEO of BT Group, who noted that AI will likely lead to even deeper job cuts than her company had already planned.

This isn’t a one-off. It’s part of a trend. It seems we are rapidly moving past the “augmentation” phase and into the “automation” phase. The reason isn’t just that the models are getting smarter; it’s that the fundamental job of the AI is changing. We’re moving from chatbots to what the industry calls “agents.”

The difference is subtle but profound. A chatbot is a tool you operate; you give it a prompt, it gives you a response. An agent is a worker you manage; you give it a goal, and it autonomously figures out the steps to achieve it. It's the difference between asking an AI to draft an email for you and asking it to manage your entire marketing campaign. The first is a productivity tool. The second is a potential replacement for the person who used to run the marketing campaign.

Nvidia CEO Jensen Huang, who sells the shovels for this particular gold rush, put it this way:

There's a trillion dollar opportunity out there to revolutionize the way companies are built and the products that companies make. Not just the way they run, but the products that they make. Every part of it is going to be revolutionized by agentic AI.

The reason for all this excitement is, as usual, money. An internal study from accounting firm RSM found that while copilots offered a nice productivity boost of 5% to 25%, AI agents that automated entire workflows delivered efficiency gains of up to 80%. A 25% productivity boost is great. An 80% boost is… a different conversation entirely, one that often ends with someone getting a severance package. This is why RSM is planning to invest $1 billion in AI agents, and why Microsoft is laying off software engineers while also boasting that its AI now writes a huge chunk of its own code.

Of course, there are problems. Giving a piece of software your corporate credit card and telling it to book travel is a big leap of faith. A 2024 KPMG report found that public trust in AI is actually decreasing, and the technical hurdles to giving agents secure access to sensitive systems are enormous. And the compute power required for millions of autonomous agents to run complex tasks will be astronomical, which helps explain the trillions of dollars being poured into new data centers.

Still, the incentives are clear. You can have a debate about whether AI will ultimately create more jobs than it destroys, as Mark Cuban argues, or whether we are automating human labor into obsolescence, as people like Anthropic’s Dario Amodei have warned. For now, though, the people signing the checks seem more focused on that 80% number.

The Scoreboard

• AI: Why Is Meta Fueling the Frenzy for AI Talent and Acqui-Hires (ARPU)
• Semiconductor: Chinese AI Companies Dodge U.S. Chip Curbs by Flying Suitcases of Hard Drives Abroad (WSJ)
• Semiconductor: Taiwan adds Huawei, SMIC to trade blacklist amid escalating US-China tech rivalry (SCMP)
• Cloud: Oracle Cloud Infrastructure Sales Projected to Grow by 70% (Fast Company)