What Actually Happened to H100 Spot?
H100 spot listings broke under $1/hr on several marketplaces as Blackwell supply broadened and older Hopper capacity recycled into spot pools. This was not a single provider cutting prices. It was supply normalization plus demand migration upward: flagship inference moved to B200, freeing H100s for everything else.
The headline number moves fast week to week. Treat sub-$1 as a market regime (abundant Hopper spot, volatile by region), not a guaranteed sticker price on every listing. Prices referenced here reflect conditions at time of writing and can fluctuate with GPU availability.
Why Did Prices Collapse? Three Forces
First, Blackwell cannibalization: teams that paid H100 premiums for flagship inference moved to B200 for better $/token, leaving H100 supply chasing fewer premium buyers. Second, fleet amortization: early H100 hosts already paid down hardware and now compete on marginal cost.
Third, software efficiency: better batching, quantization, and prefix caching cut the GPUs needed per token across the board. Demand per token fell even as token volumes grew. That combination (more supply, less GPU per token) is what breaks a price floor.
| Force | Effect | Who feels it |
|---|---|---|
| Blackwell migration | H100 supply freed | H100 renters benefit |
| Amortized fleets | Lower marginal cost | Spot buyers benefit |
| Serving efficiency | Fewer GPUs per token | All inference teams |
Who Wins Most at Sub-$1 H100?
Batch inference and fine-tuning teams win biggest: fault-tolerant, interruptible, and hungry for cheap FLOPS. If your training checkpoints well and your batch jobs preempt cleanly, this regime is a gift. Size the fleet up and shorten experiment cycles.
Latency-sensitive realtime serving wins less than the headline suggests. The cheapest spot H100 in a far region with preemption risk is not the same product as a reserved low-latency endpoint. Do not put p99 SLOs on preemptible capacity without a fallback pool.
Three Traps That Eat the Savings
Trap one: interruption without checkpointing. A job that restarts from scratch on every preemption can cost more in lost progress than the hourly savings. Trap two: cross-region data gravity. Cheap GPUs plus expensive egress and slow datasets is not cheap.
Trap three: over-provisioning because it feels free. We have watched teams waste 35 to 45 percent of bought compute on idle spot they grabbed opportunistically. Cap per-team quotas and kill idle jobs automatically, even at $0.90/hr.
How to Lock the Gains Without Getting Burned
Blend the purchase: reserved H100s for the steady baseline (the floor you always need), spot for burst, with automatic failover between them. Negotiate reserves against the new spot regime instead of last year's prices. Hosts know the market moved.
Put the two numbers that matter in the contract: effective $/M tokens at your batch shape, and the interruption rate over the trailing 30 days. Hourly rate without those is theater.
What to Do This Week
Audit your current H100 bill this week: what share is on-demand versus spot, what your realized $/M token is, and how much idle you carry. Most teams find 20 percent savings before changing providers, just by rebalancing the mix.
Then compare live H100 spot and reserve listings on ClusterBid inventory to see where sub-$1 actually clears today by region. Set a price alert and move the interruptible half of your fleet first.