The pricing beat CoreWeave by 20 percent. Then your training job stalled for six days because the GPUs were never actually in your provider's data center.
You signed a contract for 64 H100s. The pricing beat CoreWeave by 20 percent. Then your training job stalled for six days because the GPUs were never actually in your provider's data center, they were subleased from a company two layers up the chain, and your support ticket had to travel through three separate helpdesks before anyone with actual rack access saw it.
This is the scenario more buyers are running into as the neocloud market crowds with new logos. Gartner now projects the AI cloud market will reach $267 billion by 2030, with neocloud providers capturing 20 percent of that spend, up from a near-zero share a few years ago. That growth has pulled in a wave of resellers, brokers, and marketplace operators wearing neocloud branding without owning the hardware behind it. Before you sign anything, there is one question that separates a real infrastructure provider from a middleman: do you own the GPU?
Sunny Smith, founder and CTO of Massed Compute, put it plainly on the DataStorage.com Podcast: the first thing any buyer should ask a prospective GPU provider is whether they own the hardware. Most brands calling themselves neoclouds do not. They are brokers, reselling capacity they lease from someone else, who may have leased it from someone else again.
This is not a technicality. Ownership determines three things that show up the moment something goes wrong: support quality, pricing transparency, and reliability. When a support ticket has to be copy-pasted between three companies before it reaches someone who can actually touch the rack, resolution stalls. Smith described this exact failure mode: intermediaries relay tickets instead of resolving them, and the buyer absorbs the delay.
Capital intensity is why so few players actually own fleets. A single B300 server runs approximately $750,000. Building and operating a fleet at scale requires balance sheet commitments most companies cannot or will not make, which is why financial engineering, leasing, subleasing, and reselling, dominates the middle of the market. Owning GPUs is a capital bet. Reselling them is a marketing decision.
Asking the ownership question directly is the fastest filter, but it is a proxy for a longer list of operational realities buyers need answered before signing:
If your provider does not own the rack, they cannot walk over and diagnose a failed NVLink connection or a thermal throttling issue. They are waiting on someone else's field engineer, on someone else's SLA.
Multi-layer brokering adds margin at every hop. The rate you are quoted may already include two or three markups you cannot see, which limits how much negotiating room actually exists.
Brokers renew or lose their underlying leases based on decisions made by companies you have never heard of and have no contract with. Owners control their own capacity roadmap.
Some GPU operators, including fleet owners, only handle hardware provisioning and leave workload tuning entirely to the customer. Others provide deeper operational support. Neither is wrong, but you need to know which one you are buying before the invoice arrives, not after.
Two forces are converging to make provider vetting more urgent right now.
First, the neocloud category has scaled fast enough to attract capital and brand building that has outpaced actual fleet ownership. Gartner's Enrique Castera has described the shift as enterprises needing to diversify beyond hyperscalers while simultaneously tightening technical controls and risk management around the newer providers they add. That second half of the guidance, tighter controls, is the part buyers skip when a pricing sheet looks attractive.
Second, the supply picture is shifting underneath every contract signed this year. CoreWeave expanded its Meta agreement to $21 billion in April 2026, and Nebius holds Meta commitments reported as high as $27 billion, concentrating enormous capacity behind a small number of owner-operators. At the other end of the spectrum, Meta Compute and SpaceX have both begun selling surplus GPU capacity, a signal that the multi-year shortage era is starting to loosen for buyers willing to shop multiple providers instead of locking into one. In a market where real owners are consolidating hyperscaler-scale contracts and surplus capacity is starting to trade hands, buyers who cannot tell who actually owns their GPUs are negotiating blind.
There is also a scarcity layer beneath the GPUs themselves that most procurement teams do not budget for. NVMe storage prices have roughly tripled amid supply constraints expected to persist into 2027, according to Smith's account on the podcast. Storage adjacent to the GPU, with high east-west bandwidth, is no longer a line item you can treat as an afterthought. A provider that owns its compute but bolts on generic storage from a third party carries the same fragmented-accountability risk as a broker reselling GPUs. This is closely related to how CoreWeave's recent revenue miss signaled a turning point for AI infrastructure spending more broadly.
Run any prospective provider through these checks before signing a contract, not after the first outage:
| Question to Ask | What It Reveals |
|---|---|
| Do you own this hardware, or lease/resell it? | Hardware access |
| How many hops between contract and rack? | Support latency |
| What happens if the underlying lease isn't renewed? | Renewal risk |
| Does support include workload tuning or just provisioning? | Support depth |
| Where does storage sit relative to compute, and who owns it? | Data gravity risk |
Generic containers tuned for one GPU generation can leave 30 to 50 percent of performance on the table when deployed on different hardware, a gap that in-place optimization can recover but generic support will not touch. Data gravity, being landlocked by where your data lives, is the AI infrastructure constraint nobody puts in the RFP but everyone eventually pays for. Backblaze and Wasabi have built entire businesses around eliminating the egress penalty that turns a routine provider switch into a five-figure surprise bill.
If you're already running workloads on a provider and are not certain whether they own their fleet, this is worth resolving now rather than at renewal. Ask your account team the ownership question directly. A provider confident in its answer will usually volunteer specifics, data center locations, fleet size, ownership structure, without hesitation. Evasiveness on this specific question is itself a signal.
If the answer reveals you're several hops removed from the physical hardware, that does not necessarily mean you need to migrate immediately. It means you should price out the egress cost of moving your data alongside any GPU migration, budget extra time into any support-dependent timeline, and treat your current contract length as a negotiating lever rather than a locked-in commitment, particularly with surplus capacity from Meta Compute and SpaceX starting to soften the market.
Ask directly in writing, and cross-check their answer against public data center announcements, fleet size disclosures, or capital expenditure reporting if the company is public. A provider that owns its fleet will typically be specific about locations and scale.
Not always. Resellers can offer flexibility and access to capacity an owner-operator doesn't have available. The risk is in not knowing which one you're buying, since support responsiveness and pricing stability differ significantly between the two models.
Storage that isn't physically adjacent to the compute, with sufficient east-west bandwidth, becomes a bottleneck regardless of how fast the GPUs are. A provider that treats storage as an afterthought often has the same fragmented-accountability problem as a GPU broker.
There's no universal figure since it depends on how many hops exist in the chain, but each layer of leasing or subleasing typically adds its own margin. The practical impact is less about a fixed percentage and more about reduced room to negotiate, since part of the rate you're quoted may already reflect markups from parties you never see.
No, though the risk profile differs. AWS, Azure, and Google Cloud own their own data centers, but capacity allocation and internal prioritization can still create availability gaps during periods of high demand. The ownership question matters most with newer, smaller-branded neoclouds where reseller arrangements are more common and less disclosed.
Ownership is the cheapest due diligence question in the entire contract. Ask it before the pricing, not after the outage.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
