All posts
EconomicsH100PricingCapacity

H100 spot pricing recap: Q1 2026

The H100 spot market did something unusual this quarter: prices fell during the same window training demand spiked. Here is what we saw and what we think drove it.

Riya SaitoCapacity & Pricing· April 1, 20267 min read

The expected pattern for H100 spot pricing in Q1 was straightforward: continued elevated demand from a handful of large training programs, modest supply growth as new buildouts came online, and net upward pressure on hourly rates. That is not what happened.

Across our pool, the median H100 SXM 80GB hourly rate went from $2.84 in early January to $2.31 by the end of March. That is an 18.6% decline during a quarter when several customers told us their training spend went up.

Three things drove it

  • Large reserved-capacity contracts shifted from 12-month to 18-24 month terms, freeing operators to list more spot inventory without worrying about contract coverage
  • MI300X started cannibalizing the H100 mid-tier — workloads that previously needed H100 for memory bandwidth were testing well on MI300X at 70% of the price
  • Two major operators added inference scheduling that lets them mix training and inference on the same fleet, which absorbs idle hours that previously showed up as unutilized capacity

What it does not mean

It does not mean H100 demand is collapsing. It means H100 supply is finally catching up to H100 demand for the workloads where H100 is the right answer. Frontier training programs are still capacity-constrained — those contracts are signed long, signed early, and signed at premium rates that do not show up in the spot market.

It also does not mean prices keep falling. We expect the H100 spot floor to stabilize in Q2 around $2.10-$2.20, then start rising again in Q3 as inference demand absorbs the slack. If you have a workload that wants to lock in 12 months of capacity, this quarter is a reasonable time to do it.

What we are watching

The MI300X cannibalization story is the one to watch. If AMD's serving stack continues closing the gap on time-to-first-token and structured output reliability, we expect another 10-15% of H100 spot demand to migrate by year-end. That would put further pressure on H100 pricing without changing the underlying demand for compute.