Dwarkesh Patel, the podcaster and blogger known for long-form interviews with AI researchers, published a back of the envelope model arguing that computing power could become far more expensive over the next few years. His premise starts with an observed figure: Anthropic’s revenue has grown roughly tenfold year over year, putting it on pace to end 2026 with an estimated $100 billion to $150 billion in revenue. If that growth rate held, Patel calculates, Anthropic would need to reach $1 trillion in revenue by the end of 2027, a hypothetical he uses to test what has to be true about the market underneath it.

Total compute available to leading labs grows only about threefold each year, Patel notes, citing research from Epoch AI. For revenue to outpace compute tenfold while compute merely triples, one of three things must happen: lab margins increase, the price of compute itself rises, or labs redirect a larger share of their compute from training new models toward running existing ones for paying customers.

Patel argues all three are already underway, based on figures he treats as current estimates rather than confirmed disclosures. He puts Anthropic’s inference margins at roughly 40 percent in 2025, rising to more than 80 percent this year, a number he ties to SemiAnalysis reporting that industry-wide blended inference margins now exceed 70 percent. Spot prices for AI chips have risen more than 40 percent since bottoming out in February. OpenAI’s share of compute spending devoted to inference rather than training, about a quarter in 2024 according to Epoch AI, is likely close to half today, by Patel’s estimate.

The clearest evidence of tightening supply, Patel writes, sits in the tranche of compute large labs actually need: dedicated capacity with enough security and scale to serve customers, not spot instances. Google reportedly pays SpaceX roughly $900 million a month to lease 110,000 Nvidia chips, a mix of GB200 and GB300 models, about double the spot rate for that hardware, according to a Yahoo Finance report Patel cites.

Patel’s model suggests margin expansion alone cannot explain a jump to trillion-dollar revenue. For margins to do all the work, they would need to reach the mid-90th percentile by the end of 2027, a level he calls implausible even for a market leader. That leaves rising compute prices as the more likely driver, assuming the revenue trend holds at all, which Patel repeatedly flags as uncertain rather than settled.

To illustrate how far prices could move, Patel runs a thought experiment rather than citing an existing price. If a chip roughly as capable as an Nvidia H100 could run software engineering work at a human professional’s level, current market salaries for engineers imply that chip should command more than $250,000 a year in rent, near fifteen times today’s spot price. He is explicit that this is speculative math, not an observed rate, and that it assumes a flood of AI labor does not depress the market value of software engineering the way large influxes of comparable human labor eventually do.

The strongest challenge to this forecast is one Patel does not fully resolve: inference cost per unit of AI capability has fallen sharply and consistently as labs ship smaller, more efficient models that match the output of costlier ones from a year or two earlier. A tenfold rise in the price of a chip hour does not translate into a tenfold rise in the cost of finishing a given task if the software running on that chip keeps getting more efficient at the same pace or faster. Patel’s own figures hint at this tension: Anthropic’s inference margins are expanding even as compute prices climb, which is only possible if efficiency gains, utilization, or pricing power are absorbing part of the difference.

If Patel’s forecast holds even partially, the practical effect favors efficiency over raw scale. Cheaper models tuned tightly to a narrow task could out-compete frontier models on cost per completed job, pushing more production workloads toward smaller open-weight or distilled systems and reserving frontier models for work that genuinely requires them. Teams planning a 2027 AI budget should model both scenarios: what a task costs if compute prices triple, and what it costs if the efficiency curve keeps bending the way it has for the past few years. That gap, not the raw price of a GPU, is what should decide the next build versus buy call.

Dwarkesh Patel published this analysis on his blog on July 29, 2026.