How compute could become 10x+ costlier as AI capabilities and monetization outpace supply, and a look at the implications if Anthropic hits $1T in 2027 revenue
If a human-level software engineer that could run on an H100 equivalent, at current market rates for software engineers, that H100 should rent for over $250k a year.
Dwarkesh PodcastDwarkesh Patel
Context & Ripple Effects
The analysis extends a growing mismatch between AI revenue ambitions and the capital required to secure compute. Bain previously estimated that projected demand could require $2T in combined annual AI revenue by 2030, with an $800B shortfall likely under its assumptions; this story frames what happens when capable systems make that scarce capacity economically defensible at far higher prices.
Anthropic is a useful test case because its economics are already tied tightly to external infrastructure: it projected lower gross margin amid higher inference costs, while its prospective Meta capacity-rental arrangement illustrates how major model developers may need to source compute beyond their primary cloud partners.
First-order effects
If software-agent output can be sold at labor-like prices, scarce H100-equivalent capacity could command much higher rental rates, raising operating costs for AI developers and their customers.
For Anthropic, a path toward very large revenue would make long-term compute access and inference efficiency central constraints on margins, not merely infrastructure procurement; its earlier margin pressure from higher inference costs underscores that exposure.
Second-order effects
Cloud providers and data-center operators would gain leverage over model builders, encouraging more capacity-reservation, rental, and partnership arrangements like the reported Meta-Anthropic talks.
AI vendors would face stronger pressure to price by useful work delivered, limit costly workloads, or shift demand to more efficient models as compute becomes a larger share of cost of goods sold.
Third-order effects
If capability gains consistently outrun supply additions, AI could evolve less like conventional software and more like a capacity-constrained utility market, where access to compute shapes which firms can scale.
The key uncertainty is whether efficiency improvements and new capacity offset demand. The broader funding gap outlined in Bain's projected compute-revenue shortfall suggests that monetization, not technical demand alone, will determine how durable higher compute prices are.
The trend: AI is moving toward agentic unit economics in which the market value of completed work increasingly sets the ceiling for compute spending, while infrastructure supply sets the floor on margins.
I think this is the most interesting prediction about what will happen in this world of 10x compute prices: If you can train the best, most efficient model, then you'll be able to charge MUCH higher margins than you can today. [image]
I haven't the slightest idea whether Dwarkesh is going to be right or wrong on this, but I bought a fuck ton of $ORCL calls over the past few days just in case he is right...
Interesting piece. If compute gets much more expensive, maybe the only economically useful application of AI will be in hardening cybersecurity against AI.
Interesting thoughts from Dwarkesh. tldr; The key economic metric for the intelligence age is: > Intelligence per unit of compute A metaphor: miles per gallon of a combustion engine. But intelligence is a much more subtle and nuanced resource than thermal energy. The distinction …
If $$ per flop grows faster than hardware improves, old GPUs stop depreciating. An H100 that rents for 15x more in 2028 is an appreciating asset. Yet public clouds carry them on five to six-year straight-line depreciation schedules
New blog post on what would be true about the world if trendline continues and leading lab hits $1T in revenue by the end of next year. In other words, why compute might get 10x+ more expensive in coming years https://www.dwarkesh.com/... [image]
Interesting blog post here, although it is funny to me that it still downplays the take. Anthropic run-rate revenue grew 10x last year, in the first four months of this year it grew at an annualized rate of ~87x! I still feel like this simple stat is criminally underrated.
“Yeah bro we're getting slammed out here and need a pump ASAP, go tell Dwarkesh to publish some bullshit about how OpenAI sees $1tn of revs by FY27, he doesn't give a shit”
I think you can believe that OpenAI and Anth are seeing revenue going up and to the left and still be bearish. All that revenue's gotta be coming from someone.