There’s a structural problem in AI that most companies don’t want to say out loud: they can’t reliably serve the models they’re shipping. I used Google’s Gemini 3.1 Pro preview as the concrete example because the failure mode was obvious within hours—“no capacity available” errors, paying customers getting locked out, and then bans. And this is Google. Arguably the most infrastructure-rich company on Earth. If they can’t keep up with demand on a major launch, it tells you where the whole industry is heading. I break down why this is bigger than “rate limits.” First, compute has shifted from training to inference—serving the model is now the bottleneck. Second, agents create an agentic multiplier: one autonomous coding session can burn hundreds of thousands to millions of tokens, and research shows token usage can vary by ~10x across today’s shipping coding agents. Third, supply is not catching up: data center build-outs are slowed by permitting, power, and local opposition, and memory is a real constraint (“RAM megadon”) with new fabs not meaningfully helping until 2028. The takeaway for 2026 is simple: we’ll see more quotas, credits, and lockouts, plus more specialized inference hardware. The real competition is shifting from “who has the best model” to “who can serve a great model at scale.” A 95-score model is worthless if users can’t access it.