Philip Winston has a theory about why Google keeps losing the AI race, and it doesn't involve corporate dysfunction or misaligned incentives. In a March 9 essay on metastable.org, he argues that every complex sociotechnical system operates under a hard physical ceiling on its "learning rate" — the maximum speed at which a human-machine system can iterate, test, and improve. Past a certain resource threshold, adding capital or engineers buys you nothing. You're already at the limit.
The argument reframes the question of frontier AI competition. Google, Microsoft, and Meta have each spent billions, yet neither has consistently topped the leaderboards for model quality. The standard explanations — bureaucracy, talent misallocation, misaligned incentives — assume the problem is fixable. Winston's version is more fundamental: OpenAI and Anthropic reached the resource threshold needed to run at the system-wide speed limit early, and since nobody can exceed that limit, their lead is structurally protected against any competitor who avoids a catastrophic operational mistake. A fair objection here is that Google DeepMind has been closing the gap — Gemini 2 has matched or beaten GPT-4 class models on a range of benchmarks — so the divide isn't as clean as Winston's framing implies. But his core claim doesn't require permanent dominance; it requires only that the ceiling prevents anyone from lapping the current leaders, which the data broadly supports.
The macroeconomics analogy is where the essay gets genuinely interesting. Despite decades of IT revolution, real economic growth has remained effectively pegged to its exponential trend for fifty years — no faster, but no slower. Winston suggests AI may play the same role: not the force that breaks through that ceiling, but the increasingly intensive effort required to stay on the existing curve. It's a sobering read for anyone counting on near-term step changes in productivity growth, and it raises questions that Winston is honest enough to leave unanswered.
For the agent ecosystem, the competitive picture this implies is worth sitting with. If the speed limit thesis holds at the foundation model layer, competing with Anthropic or OpenAI on raw model quality through resource intensity is probably a dead end. The more tractable question is whether the agent layer represents a genuinely different performance curve — one governed by iteration speed on tooling, evaluation, and deployment rather than training compute. Winston doesn't address this directly, and it's where his argument has the least purchase: the same ceiling that constrains model training may well constrain agent development too, or it may not apply at all. That's the structural question the next wave of frontier agent startups will have to answer.