On March 12, the New York Times reported that Meta had quietly delayed the rollout of its next-generation foundational AI model. Code-named internally — though the company won’t confirm it — something like “Avocado,” the model had been positioned by Mark Zuckerberg last July as one that would “push the frontier.” Instead, it fell short of rivals from Google, OpenAI, and Anthropic on internal benchmarks for reasoning, coding, and writing. Meta declined to comment on the record. The delay was framed as a setback.

Then the discourse did what the discourse does. Within weeks, Substack essays and Hacker News threads were declaring that AI itself was hitting a wall — that the scaling laws were dead, that the age of exponential progress was over, that we’d all been sold a bill of goods. The tone was equal parts vindication and despair, depending on what the writer had invested in the narrative.

Both reactions miss what’s actually happening. The Meta delay isn’t evidence of a ceiling on AI capability. It’s evidence of something more interesting: the end of the era where simply throwing more compute at a problem counted as a strategy.

The Benchmarks Are Now the Product

For the past three years, the AI industry has operated on what you might call the “announcement economy.” A lab would train a larger model, post a chart showing it outperformed the competition on a handful of benchmarks, and the press would dutifully report that the future had arrived. The model itself could be buggy, hallucinate freely, and refuse to answer basic questions — none of that mattered, because the chart was the product.

Meta’s delay signals that this game is winding down. Internal testing, according to the Times, showed the model underperforming not on some esoteric measure of “intelligence” but on the categories customers actually pay for: reasoning, coding, writing. These are the benchmarks that matter when you’re selling API access, not conference talks.

One engineer at a mid-sized AI startup described the dynamic to me over Slack this week: “We used to evaluate models by whether their release blog post had a cool chart. Now we’re running them on our actual eval suite — on our actual customer data — and the differences are way flatter than anyone wants to admit.”

That’s the shift nobody’s headline quite captures. The slowdown isn’t in what models can do. It’s in how much more they can do than the model that came out six months ago.

The Spending Hasn’t Slowed — It’s Just Moved

If you’re looking for evidence that the industry is panicking about a plateau, the capital expenditure numbers tell a very different story. Microsoft alone reported $19.7 billion in capital expenditures in its most recent quarter, overwhelmingly driven by AI infrastructure. Google’s CapEx was $17.2 billion. Amazon’s was higher still. This is not the spending profile of an industry that thinks the returns are drying up.

What’s changed is where the spending is going. The marginal dollar is no longer chasing a bigger parameter count. It’s going into inference infrastructure, into tool-use capabilities, into the unglamorous plumbing that makes models useful in production rather than impressive in demos.

This is terrible news for the labs that built their entire valuation story on being six months ahead of everyone else. If the next model isn’t dramatically better, the switching costs between providers collapse. Enterprise customers start negotiating on price. The moat shrinks.

The Real Losers Aren’t Who You Think

The conventional wisdom — already hardening into a Substack genre — is that the AI slowdown means the bubble is popping, the hype was overblown, and the skeptics were right all along. This is a comfortable story if you’ve spent two years insisting that none of this was real.

But the actual losers are more specific. The labs whose advantage was purely architectural — who had the biggest cluster and the newest training technique but no meaningful distribution — are now competing on a flattening curve. The startups that promised AGI by 2027 and raised at valuations that priced in that timeline now have to explain to their boards why the thing they were building is starting to look like a commodity.

Meanwhile, the companies with distribution and integration advantages — Microsoft embedding Copilot in every enterprise license, Apple wiring intelligence into the OS layer — are better positioned, not worse. A plateau in raw model capability favors the integrators over the researchers.

In a hotel bar at a machine-learning conference last month, a researcher from one of the major labs put it bluntly: “Everyone’s terrified that the thing they’re racing toward isn’t going to exist. But the thing they’re racing toward was never the real product. The real product was always going to be boring infrastructure, and nobody wanted to admit that.”

He was half right. The real product was always going to be infrastructure. The difference is that boring infrastructure, built well, is worth trillions. The AI industry is just now confronting the uncomfortable reality that it has to actually build it.

Sources