Sm0ke

The @AnthropicAI S-1 is out and most of the takes I’ve seen are just copy/paste of metrics without much analysis. For me, the most interesting stat is revenue per dollar of compute. In Q1'25, Anthropic brought in $0.40 of revenue for every dollar it spent on compute. By Q2'26 that was $2.08. Quarterly revenue hit $11.6B, up 14.7x Y/Y, while compute and infrastructure costs grew 4.0x to $5.6B. Put another way, compute went from 177% of revenue to 48% in a year! This efficiency helps explain the huge bet Anthropic seems to be making: the company has signed roughly $830B in compute commitments running out to 2036. Meanwhile RPO is $10.8B, with about 75% due within 12 months, and nearly half of revenue comes through cloud marketplaces. The commitments stretch out a decade, but most of the revenue Anthropic can see is due within a year. They’re betting revenue keeps outrunning compute costs, and so far it is. It also shows where the margin in AI infrastructure is headed. The hyperscalers turn megawatts into GPU hours and sell them on long-term contracts. That locks in revenue, but it also caps it. The labs take on the commitments and keep the value they create turning those hours into high-value tokens. As the cost of a megawatt evens out across providers, the money will be in how much value you can squeeze out of each one. The same logic should apply to inference providers like @baseten and @fal. And the next step is bigger - moving up the stack. If frontier labs, inference providers and other neoclouds go from tokens to business output, abstracting tokens from customers while mixing/matching the best tokens to accomplish a task, they can capture even more revenue. This will happen while their costs plateau, giving them even more leverage. That’s the part I’m watching most closely.