A team had an 8x H100 cluster running at 12% utilization. Bought for a project that shrank in scope halfway through.
Someone on the team submitted a spec to GPUaaS.com. Tier, count, the overnight hours nobody was touching, the region the cluster sat in. Four days later a buyer had signed.
The buyer had already been vetted before the spec reached them. Real workload. Real payment history. A compliance posture that matched the deal.
Key takeaways
A well-prepared spec, real interconnect numbers, a named-failure-type SLA, a benchmark under real load, matched a vetted buyer within four days
Buyers are vetted for real workload, payment history, and compliance posture before a listing ever reaches them
Windowed availability lets a cluster owner keep using hardware during business hours while a buyer's job runs overnight, on the owner's own schedule
Full-time listings without an early-exit clause can lock a cluster owner out of reclaiming hardware for months if internal needs change
A cluster at 10% utilization has 90% of its capacity producing no return while operating costs run the same either way
◆ WHAT THE SPEC INCLUDED
Details that closed the deal in four days
The spec included the interconnect bandwidth, the real number, not just the GPU model name. An SLA naming specific failure types instead of one uptime percentage. A benchmark run under sustained load a week earlier for an unrelated reason.
◆ COMPLIANCE, BEFORE RATE
SOC 2 came up in the first conversation
Compliance came up in the first conversation, before rate. The team had a SOC 2 baseline from an earlier audit cycle.
The cluster kept running the team's own workloads during business hours. The buyer's job ran overnight, in the window the team specified. Nothing else changed.
◆ SIX WEEKS IN
The operating cost never changed. What sat against it did
Six weeks in, the arrangement was still just the overnight hours. The operating cost of running the cluster hadn't changed. A buyer's payment now sat against those hours instead of nothing.
◆ A SECOND TEAM, A DIFFERENT STRUCTURE
Full-time, no exit clause
A second team listed 16 GPUs full-time instead of windowed hours. Twelve-month term. Flat monthly rate. In month four an internal project needed the cluster back. The listing terms hadn't included an early-exit clause. The team waited out the remaining eight months before reclaiming the hardware.
◆ A THIRD TEAM, A SPLIT FLEET
12 listed, 28 kept internal
A third team ran 40 GPUs across five racks. Listed 12 full-time, kept 28 for internal work. The 12 listed GPUs covered close to a third of the fleet's total operating cost by the end of the first quarter.
90%
of a cluster's capacity produces no return at 10% utilization, while operating costs run the same regardless
GPU utilization audit patterns across enterprise fleets, 2026
GPU utilization audits across enterprise fleets turn up numbers in the same range as the first team's 12%. A cluster at 10% utilization has 90% of its capacity producing no return while the operating cost runs the same either way.
The early-exit clause the second team skipped is a single line in the listing terms, set once at submission. Windowed availability, like the first team used, keeps that flexibility by default since the hours are already fixed to a schedule the owner controls.
See what demand looks like for your idle hours.
Submit availability, get matched with vetted buyers. For single-GPU or month-to-month demand, packet.ai handles self-serve access with 24/7 human support.
No. Availability windows and any early-exit terms are set by the cluster owner at submission. A team can list only overnight hours, only weekends, or specific months, and keep the rest for internal use, so the hardware remains available for its own workloads on the schedule the owner sets.
Buyers are checked for a real workload spec, payment history, and compliance posture before ever being matched to a listing. A cluster owner isn't exposed to unknown counterparties, and the matching process is designed to surface serious, qualified demand rather than speculative inquiries.
It depends on whether the hardware is genuinely no longer needed internally. Windowed availability keeps flexibility by default since only specific hours are committed. Full-time listings can generate more revenue but should include an early-exit clause if there's any chance the hardware will be needed back before the term ends.
Real interconnect bandwidth numbers, not just the GPU model name. An SLA naming specific failure types rather than a single uptime percentage. A benchmark run under sustained load if one is available. Listings prepared this way tend to match with serious buyers faster than a bare spec sheet.
Yes. Teams with larger fleets commonly list a portion full-time or windowed while keeping the remainder for internal workloads, rather than choosing one approach for the entire fleet.
