Clara Bennett: Finn, hi — I'm handing you one fact and I want your first instinct before we do anything else with it.
Finn Brooks: Hit me, I'm already stressed about something and this will either help or make it worse.
Clara Bennett: March 2026. Google informs Meta — through Google Cloud — that it cannot fulfill the Gemini compute capacity Meta wanted to purchase. Meta was the most heavily affected enterprise customer. Financial Times broke it in late June. Now — first instinct.
Finn Brooks: My first instinct is that this is like — wait, no, I have the analogy — it's a landlord telling their richest tenant the building is full. Except the landlord is also running a competing business out of the penthouse. That's not a capacity problem, that's a power move dressed as a capacity problem.
Clara Bennett: That analogy is doing real work. The important nuance is that GPU scarcity is genuine — demand for the chips powering AI inference genuinely exceeds supply right now. But given real scarcity, Google still chose who gets the last apartment. Those are two separate decisions.
Finn Brooks: Right, and Reuters said they couldn't verify the FT details, so we don't actually know the choosing part — we just know Meta got cut off and had to quietly deal with it for three months before anyone said anything.
Clara Bennett: Three people familiar with the matter — that's what the Financial Times had. And what those three people described isn't a press release problem. Meta instructed employees to use AI tokens more efficiently. Not 'consider optimizing.' Instructed. That means someone wrote a mandate, sent it internally, and employees running content moderation jobs were checking their token balance before they could proceed.
Finn Brooks: Wait — content moderation specifically?
Clara Bennett: Content moderation, harmful-content detection, scam identification, customer service, advertising tools, coding assistance. That's the list. Those aren't experimental projects — that's Meta's operating infrastructure. When AI tokens get rationed, those pipelines slow. That's compute-gating landing on real work.
Finn Brooks: Okay I — no, I hear that, but here's where I can't fully follow you. Meta is building LLaMA. They're scaling in-house chips. Like, a company that infrastructure-native being this wrecked by one competitor's capacity decision — I mean, doesn't that feel like a temporary gap they just... badly miscalculated?
Clara Bennett: That's exactly the gap. What you're building toward and what you need for inference right now — those aren't the same thing. LLaMA doesn't replace Gemini inference capacity in March 2026. The chips aren't there yet. So the ambition is real, the dependency is also real, and Google's restriction landed precisely in that window.
Finn Brooks: No, I don't buy that they didn't see this coming.
Clara Bennett: They may not have. But here's what I actually want to push on — because three things tightened in the same window, and people are collapsing them into one story, and that's where I think we lose the plot.
Finn Brooks: Wait — you mean the OpenAI thing and the Anthropic thing.
Clara Bennett: Exactly. OpenAI limiting a new model after discussions with the Trump administration over government security concerns — that's not GPU scarcity. That's a security directive. And Anthropic's Fable, which is the consumer version of Mythos, facing U.S. export restrictions that the Trump administration then announced plans to lift — that's regulatory export control. Three distinct mechanisms.
Finn Brooks: Okay but — no, hang on, I'm actually the one saying separate them. That's my point. The a16z structural bottleneck framing, the 'compute is the most scarce resource' language — that applies to Google-Meta. It does not apply to why OpenAI locked a model after a government conversation. Those are not the same pressure.
Clara Bennett: The mechanisms differ — I won't argue that. But from where a developer or enterprise customer sits in that same window, infrastructure access, model access, and export access all narrowed simultaneously. Alphabet is a direct competitor to Meta, OpenAI, and Anthropic. That convergence is— I mean, you can't call it coincidence and you can't call it one story. It's both things at once.
Finn Brooks: Correlation is not causation, though. Like, that's— that's the thing. Three separate actors, three separate decisions. Google Cloud rationing Gemini compute is a supply problem. The Trump administration leaning on OpenAI is a security problem. Fable getting export-blocked is a trade problem. Lumping them is sloppy.
Clara Bennett: The causes aren't identical — that's fair. But the pattern is still real. In practice, the question worth asking is: what does it mean when all three channels tighten at once, regardless of why?
Clara Bennett: The honest landing for me is this: Google restricted Meta's access even though Meta was willing to pay. The constraint was infrastructure, not price. And whoever locked in long-term GPU supply and vertical integration before this moment — they hold a moat rivals genuinely cannot buy their way out of. That's where Alphabet sits right now. On both sides of the equation.
Finn Brooks: Yeah, and — I mean, maybe Meta's answer is just to stop buying from rivals entirely. Like, accelerate the in-house stack, stop the dependency cold.
Clara Bennett: That's a 2028 story. In March 2026, Meta's content moderation ran on rationed tokens.
Finn Brooks: Ugh. Yeah. That's the part that actually sits with me.
Clara Bennett: Good place to stop. Thanks for working through it with me.
