The 8x forecast error

Dario Amodei put a number on his own surprise in early May: "We had planned for a world of 10x growth per year. In the first quarter of this year, if you were to annualize it, we saw 80x growth — in revenue and usage." That is an eightfold miss on a plan that already assumed AI demand would explode. The company whose entire business is forecasting that demand was off by nearly an order of magnitude, in the direction of more.

It raised its rate limits by up to sixteenfold across every tier — and customers still hit the ceiling.

The texture is in the failure mode rather than the headline figure. Anthropic raised the rate limits on its most capable model by ten- to sixteen-fold across every customer tier — and customers kept slamming into the new ceilings anyway. One client reportedly ran up around $500 million a month after never setting a usage cap. Amodei framed the resulting compute shortage not as a product fault but as a forecasting one: they correctly predicted the wave and still could not build ahead of it, and are renting outside data-center capacity to keep up.

What's worth holding onto is that the constraint has moved. For most of the software era the question was whether anyone would want the thing; here, demand is treated as a given and the binding limit is how many chips a company can secure. That reframes the AI race as a contest over physical capacity, not interest — which is also why skeptics press the other half of the ledger: a run-rate that doubles in weeks says nothing about what it costs to serve, and Anthropic's own profitability forecast doesn't arrive until 2027. The bet now is whether the compute can be bought faster than the demand can outrun it.