Blog ▸ Rent Out Your H100 Cluster: Committed Terms vs On-Demand for Providers
GPU Infrastructure
Spot and committed H100 both say $2.50. One pays for 720 hours a month, the other for whatever got booked. The provider math on which to build the fleet around.
Rent Out Your H100 Cluster: Committed Terms vs On-Demand for Providers
GPUaaS.com Team
GPU Infrastructure
October 7, 2026
No items found.
A provider renting out an H100 cluster on committed terms in Q4 2026 earns $2.50 to $3.00 per GPU-hour on every contracted hour. The same cluster on a spot marketplace earns around $2.50 at the headline rate but only on the hours it is actually booked, which for most providers is well under half of them.
This is the provider-side view of the two-price market. Buyers see two H100 prices, spot and committed, and wonder why both are right. Providers see the same two prices from the other side and have to decide which one to build a business on. The decision is not about the rate. It is about what fraction of the hours in a month the rate is paid for, and what the cluster costs during the hours it is not.
Key takeaways
Committed H100 SXM clusters quote at $2.50 to $3.00 per GPU-hour through the GPUaaS.com network this month, paid on every contracted hour
H100 spot has sat near $2.50 for most of 2026, but spot revenue is paid only on booked hours and preemption churn keeps utilisation low
An 8-GPU H100 node costs roughly the same to run at 0% and 100% utilisation: power, cooling, colocation and depreciation do not stop
At 40% spot utilisation, $2.50 spot earns the equivalent of $1.00 per GPU-hour against a committed floor of $2.50
One-year H100 contract pricing rose from about $1.70 to $2.35 per GPU-hour between October 2025 and March 2026, per SemiAnalysis, and the committed band has held above that since
H100 requests still rank fourth in the network, behind Blackwell parts and H200, and they are price-sensitive. Documentation wins them
◆ THE COST SIDE
An idle H100 costs about the same as a busy one
The arithmetic that decides this question starts with what the cluster costs when nobody is using it. An 8-GPU HGX H100 node draws around 10 kW at load and most of that at idle, because the fans, the CPUs, the NICs and the HBM refresh do not switch off. Colocation power is contracted, not metered by GPU utilisation. Depreciation runs on the calendar. Support staff are paid monthly. The supply anchor post puts the all-in idle cost of an H100 at roughly $30 a day, and the busy cost is not much more.
That means revenue per GPU-hour is the wrong number to optimise. The right number is revenue per GPU-hour multiplied by the fraction of hours that earn it, minus a cost that does not move. A rate that looks good on a marketplace listing and is paid for a third of the month can lose to a lower rate paid for all of it.
The depreciation clock post works through the same point from the buyer side, for enterprises with owned clusters sitting idle. The provider version is the same clock with a listing attached.
◆ SPOT VS COMMITTED, ONE 8-GPU H100 NODE, ONE MONTH
Channel
Rate per GPU-hour
Hours paid
Effective rate
Node revenue, 30 days
Spot marketplace, 30% utilisation
$2.50
216 of 720
$0.75
$4,320
Spot marketplace, 50% utilisation
$2.50
360 of 720
$1.25
$7,200
Spot marketplace, 70% utilisation
$2.50
504 of 720
$1.75
$10,080
Committed, bottom of band
$2.50
720 of 720
$2.50
$14,400
Committed, top of band
$3.00
720 of 720
$3.00
$17,280
Illustrative. Spot rate from published H100 spot indices near $2.50 in 2026. Committed band from the GPUaaS.com October 2026 benchmark. Utilisation figures are assumptions; marketplace fees and preemption gaps are excluded and would lower the spot rows further.
3.3x
monthly revenue on the same H100 node at the bottom of the committed band against a spot listing at 30% utilisation, at the same headline rate
Illustrative, from the table above
◆ WHY SPOT UTILISATION IS LOW
Spot fills racks. It does not pay for them.
A spot marketplace is designed for the buyer who wants an H100 for six hours and does not mind being preempted. That buyer is real and there are a lot of them, which is why spot listings show activity quickly. But the activity is bursty, the jobs are short, and the gaps between them are the provider's problem. A node that is booked for a four-hour job, idle for three hours, booked for two, idle overnight and booked again at 9am US-East has had a busy day on the dashboard and a 40% day on the invoice.
Price competition makes it worse. The spot rate is set by whoever is desperate enough to undercut, so the $2.50 headline is a ceiling a provider rarely holds. The two-price market covers why spot clears where it clears, and the provider lesson is that spot is a channel for soaking up genuinely spare hours, not a plan for the fleet.
Committed demand works the other way. The buyer is trading flexibility for a rate inside the band, and in exchange the provider gets every hour of the term paid whether the buyer's job runs or not. That is why the committed band can sit at the same number as spot and still be worth two to three times as much.
◆ WHERE H100 DEMAND IS
Fourth in the ranking, price-sensitive, and winnable
H100 is no longer the default ask. The October demand snapshot ranks it fourth, behind B300, B200 and H200, and the buyers who still specify it are either running a pipeline that works and do not want to requalify, or running inference where they have done the cost-per-token arithmetic and H100 won. Both groups are price-sensitive in a way Blackwell buyers are not.
That cuts both ways for a provider. The request volume is lower than it was a year ago, but the requests that arrive are committed-term asks from buyers who know what they want, and they tend toward 8 to 32 GPUs on six to twelve month terms. A provider with H100 racks is competing in a thinner market with a buyer who will sign for a year if the quote lands inside the band and the questions are answered.
The questions are the usual ones. Inter-node fabric and speed, since multi-node training is the most common H100 ask. Support hours and replacement policy. Uptime commitment. Where the data sits. A quote that arrives with those written down wins at $2.80. One that arrives without them loses at $2.50 to a provider who had them ready.
◆ PRICING INSIDE THE BAND
Term length first, then size, then fabric
The committed band is a range because term length, cluster size, interconnect and region each move the final number. A provider pricing a request should start with term: a twelve-month commitment is worth the bottom of the band because it pays for 8,760 hours with no gaps, and a three-month commitment belongs nearer the top because the provider carries the re-fill risk sooner. Size comes next. Contiguous, InfiniBand-connected capacity at scale prices above the same GPU count scattered across nodes on Ethernet, and a request that can tolerate non-contiguous allocation should get the lower number.
Region moved very little this month. The October benchmark found similar bands holding across most sourcing regions, so a European or APAC provider should not discount for location. The regional difference is in how fast the request fills, not in the rate.
One more thing worth knowing: a provider does not have to choose. H100 capacity between committed contracts can be offered as inference through packet.ai, where it earns on tokens served rather than on term, and that fills the gaps a committed-only fleet would otherwise leave idle.
List your H100 cluster for committed demand.
One node is enough to start. No cost to apply or list. Committed requests arrive qualified: spec, term, region, interconnect.
On committed terms through GPUaaS.com, H100 SXM clusters quote at $2.50 to $3.00 per GPU-hour this month, paid on every contracted hour. An 8-GPU node at the bottom of that band earns about $14,400 a month. On a spot marketplace the headline rate is similar but only booked hours pay, so effective revenue is typically a third to two-thirds of that.
For a provider with fixed costs, yes. An H100 node costs roughly the same to run idle as busy, so revenue depends on paid hours rather than headline rate. Committed terms pay every hour of the contract. Spot pays only booked hours, with preemption gaps and price competition on top. Spot is useful for soaking up spare hours between contracts, not as the plan for the fleet.
Yes, though it ranks fourth behind B300, B200 and H200 in requests through the GPUaaS.com network. The buyers who specify H100 have working pipelines or have done the cost-per-token arithmetic, ask for 8 to 32 GPUs on six to twelve month terms, and are price-sensitive. Documented interconnect, support and uptime win those requests.
Inside the published band, $2.50 to $3.00 this month. Twelve-month terms at the bottom, three-month terms toward the top. Contiguous InfiniBand capacity above non-contiguous Ethernet. Region barely moves the number this month. Price certainty rather than silicon and the quote lands where buyers expect it.
Roughly $30 a day per GPU all-in, covering contracted power, cooling, colocation, support staff and depreciation, and almost none of that falls when the GPU is idle. An 8-GPU node therefore costs around $7,000 a month whether it earns $0 or $17,000, which is the whole reason paid hours matter more than headline rate.
Yes. The same hardware can be listed for committed cluster demand through GPUaaS.com and offered as inference capacity through packet.ai, which earns on tokens served rather than on term. That fills the gaps between committed contracts without putting the fleet on a spot marketplace.
