AI compute prices could jump 10x as chip supply lags demand

AI compute prices could surge tenfold by 2028 as model intelligence outpaces hardware supply, according to tech analyst Dwarkesh Patel, pricing out consumer AI applications.

"If an AI capable of human-level programming could run on one H100, companies might rationally pay more than $250,000 a year to rent that chip," Patel said in a video analysis, roughly 15 times today's spot price.

The argument starts with a mismatch. Anthropic's revenue has reportedly grown tenfold in a year and could reach $100-150 billion by end of 2026, while industry compute capacity grows only about threefold annually. Spot compute costs have already risen more than 40 percent since February. Google is paying SpaceX $900 million a month for 110,000 GB200 and GB300 GPUs, roughly twice the current spot rate.

The economics favor incumbents. If compute costs 15 times more by 2028, new entrants without paying customers would struggle to bid against established players like Google, Anthropic, and OpenAI. Efficiency becomes a premium: when every GPU hour is costly, customers choose the model that completes tasks with the least compute, allowing the best labs to charge more.

Inference vs. Training — The Tension at the Core

OpenAI now devotes at least half its compute to inference, up from roughly a quarter in 2024, according to Epoch data cited by Patel. But frontier labs want to keep investing in training future models. Spending most of their capacity serving current products leaves less for building successors. Anthropic's inference margins have climbed from 40 percent to above 80 percent, Patel estimates, though he calls that a "total vibe claim."

The tension is structural. For labs pursuing artificial general intelligence, inference exists mainly to prove commercial viability and fund the next training run. Converting too much capacity into cloud services would mean a stall in technical progress. With training demand locked in, price increases become the only release valve for the supply-demand imbalance. This dynamic explains why frontier labs are raising API prices even as model efficiency improves — they need the revenue to fund the next generation of training runs.

Supply Bottlenecks Defy Market Signals

Patel invokes the Simon-Ehrlich bet as a cautionary tale. Unlike commodities where price signals eventually draw out new supply, advanced chip production faces physical constraints. Moore's law contributes about 1.4x annual growth but is approaching its limit. New fabrication plants add roughly 1.2x, constrained by ASML EUV lithography capacity through 2030. Reallocating wafer capacity from smartphones and PCs to AI — the largest contributor at 1.8x — is hitting its ceiling. By end of next year, AI's share of TSMC's most advanced N3 node will jump from 60 percent to 86 percent.

When all frontier wafer capacity is consumed by AI, the industry faces a hard landing. Patel acknowledges markets often overcome apparent resource shortages, but argues advanced-chip supply cannot respond as quickly as ordinary commodity production. Unlike copper or oil, where higher prices incentivize new mines and wells within years, leading-edge chip capacity requires multi-year fab construction timelines and EUV lithography machines that ASML cannot produce fast enough. Compute may eventually become cheap again, but the intervening years could bring higher prices and fewer viable competitors.

What This Means for Investors

Nvidia, TSMC, and cloud infrastructure providers stand to benefit from sustained pricing power. But low-margin AI application companies face a brutal squeeze. The Alchian-Allen effect means only the most efficient models will justify the cost of running on expensive hardware. Cheap consumer AI — "short-form video slop," as Patel calls it — could become uneconomic. The result is an extraordinary concentration of economic and technological power in the hands of the few labs that can afford the hardware arms race. For investors, the trade is clear: long the picks-and-shovels names that own the supply chain, avoid the application layer where margins will compress as compute costs rise.

This article is for informational purposes only and does not constitute investment advice.