The Information Machine
Concluded·Day 15·first covered 9 Sep 2026·2 sources

H100 GPU Rental Rates Climb to $3.28/hr in 2026 Despite Chip Age

The gist

H100 rental prices rising through 2026 despite the chip being three years old and a successor generation shipping signals that AI inference demand is outpacing available compute supply. The projected decline toward $1/hr by mid-2027 will affect cost planning for anyone budgeting AI inference infrastructure.

The full picture

H100 one-year contract rental rates rose roughly 40% between October 2025 ($1.70/hr) and July 2026 (~$2.80/hr), with spot prices reaching $3.28/hr in September 2026 per the Ornn index. Marketplace median rates sit at $3.38/hr, with the cheapest tracked rate at $1.49/hr on Vast.ai. The price increase occurred even after Nvidia's Blackwell generation shipped, contrary to standard hardware depreciation patterns. The primary demand driver is AI inference for reasoning and agentic workloads, which consume far more tokens per request than simple chat. Supply constraints include TSMC CoWoS packaging capacity, HBM memory output, and 36-52 week lead times against a roughly 3.6 million unit industry backlog; HBM costs rose around 30% in Q4 2025. AWS raised H200 instance prices 15% in January 2026. Newer chips command higher rates: the H200 runs $3.59-$4.59/hr, the B200 $5.63-$8.64/hr, while AMD MI300X undercuts Nvidia equivalents at as low as $1.85/hr on Vultr. H100 rates are projected to drift toward $1/hr by mid-2027 as B200 absorbs premium demand, following the depreciation curve the A100 followed.

How it developed
9 September 2026

H100 spot price hit $3.28/hr per the Ornn index, up 22% in a single month; marketplace median $3.38/hr

Sources
The daily email

Want this in your inbox?

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

Subscribe free