Can you resell GPU compute you already bought?

If you are about to sign a multi-year reservation for GPU capacity, the question worth asking before the question about price is simpler: if your plans change, can you get out? Most buyers assume the answer is no, that a reservation is a sunk cost the moment the contract is signed. That is not how the market actually works in 2026. Reserved capacity, offtake, and prepaid compute can be resold or assigned, and whether you can do that, on what terms, at what cost, is itself something you negotiate before you sign, not something you discover after.

Reserved, on-demand, and spot are three different products

A GPU price is not one number. It depends on what you are buying.

• On-demand is pay-by-the-second, no commitment. Modal’s public pricing page, checked September 2, 2026, lists on-demand serverless rates of $7.10/hr for B300 and $3.95/hr for H100 SXM5. Lambda’s public pricing page, also checked September 2, 2026, lists on-demand H100 SXM at $3.99 to $4.29/hr depending on cluster size.
• Reserved is a forward commitment, usually one to three years, priced lower per hour in exchange for the commitment. Compute Exchange’s public reserved-GPU pricing page, published April 10, 2026, states that a one-year reserved H100 SXM contract runs near $1.89/hr and a three-year reserved H100 PCIe contract near $1.84/hr, against an on-demand H100 SXM rate of about $2.99/hr quoted on the same page.
• Spot is unreserved, interruptible capacity priced off whatever is momentarily idle, and it can be cheaper than either, or briefly more expensive, depending on demand at that moment.

A reservation is functionally a forward contract: you are agreeing today to a price and a quantity of compute delivered over a future period, in exchange for a discount off the on-demand rate. Comparing $1.89/hr reserved (Compute Exchange, April 10, 2026) to $3.95/hr on-demand (Modal, September 2, 2026) makes the reservation look like a clear win, roughly 2.1 times cheaper. But a forward contract only pays off if you actually use, or can otherwise monetize, what you locked in. That second part is the part buyers skip.

The secondary and offtake market already exists

The market has an answer to “what if I overcommitted,” and it is not new. Two examples, described from what each company states about itself publicly.

SF Compute’s public homepage, checked September 2, 2026, describes itself as built around exactly this: “Reserve the compute you need and resell what you don’t, so idle GPUs become revenue instead of sunk cost.” It states that its scheduler “allows tenants to sell back their idle compute to others,” and separately describes itself as “the only cloud designed to let you sublease your reserved capacity when plans change.” That is a public company stating, on its own site, that resale of reserved capacity is a supported product feature, not an edge case.

Compute Exchange operates a public marketplace for reserved and secondary GPU hardware. Its blog post on reserved GPU marketplaces, published March 4, 2026, describes rising secondary-market activity as expiring A100 and H100 contracts enter availability channels, and puts refurbished 2-year-old H100 units at roughly 30 to 40 percent of their historical peak price. The same post notes that this resale activity had not, as of its March 2026 publish date, produced the price relief a buyer might expect: it describes H100 pricing as having risen a further 15 to 20 percent over the period it covers, with several neocloud providers announcing additional increases.

Two things follow from those two sources. First, resale and reassignment of reserved capacity is a real, named, publicly marketed product category, not a theoretical workaround. Second, having a secondary market does not guarantee it will bail you out at a good price. A market where sellers currently outnumber willing buyers, or where the exiting party is desperate, prices accordingly.

Exitability is a negotiated term, not a fallback

Because a secondary market exists does not mean your specific contract lets you use it. Whether you can resell, assign, or downsize your reservation, and whether the counterparty can block or tax that transfer, is written into the contract you sign, and it is far easier to negotiate before you sign than to discover after.

Concretely, consider what having no exit clause costs on a generic, round-number commitment. Take a buyer 18 months into a 3-year reserved commitment on a 1,000-GPU allocation, with no right to resell or assign, whose workload has since gone away.

The remaining calendar exposure on that commitment is 1,000 GPUs times 8,760 hours per year times 1.5 years:

1,000 x 8,760 x 1.5 = 13,140,000 GPU-hours still owed.

At the low end of Compute Exchange’s April 10, 2026 reserved-H100 range ($1.07/hr), that is 13,140,000 x $1.07 = about $14.1M still contractually owed. At the one-year reserved rate on the same page ($1.89/hr), it is 13,140,000 x $1.89 = about $24.8M. That is the size of the number an exit clause is protecting you against, on a 1,000-GPU commitment, and it is why “can I get out of this” belongs in the term sheet, not in a follow-up email eighteen months later.

The commodity analogy: you are long compute, and that has basis risk

Signing a multi-year reservation makes you long compute at a fixed price, the same way a company that locks in a forward price for oil or wheat is long that commodity. The risk is not that the underlying commodity stops being useful. The risk is that the market price of the thing you are long moves against you while you are stuck holding it, unable to exit at a price close to what the market has moved to. That gap between what you are locked into and what the market is currently pricing is basis risk, and it is exactly what a resale or assignment right is meant to let you close.

This has already happened once in GPU pricing, not hypothetically. Introl’s analysis of the H100 rental market, published January 12, 2026, describes H100 prices falling from an earlier peak of about $8 per hour to a more recent range of $2.85 to $3.50 per hour, a decline it puts at roughly 64 percent. Introl attributes that earlier peak to late 2024. Anyone who locked in a multi-year rate near the peak Introl describes, with no resale right, would have been left holding a contract priced near $8 an hour while the market it describes repriced to less than half of that.

Run the same arithmetic as the exit-clause example above, on the same 1,000-GPU, 18-months-remaining commitment: a gap of $8.00/hr minus about $3.18/hr (the midpoint of the $2.85 to $3.50 range Introl cites) is $4.82/hr. 13,140,000 x $4.82 = about $63.3M of value between what a stuck buyer would be paying and what the same capacity costs at the more recent range Introl describes. Whether that gap is real cash lost depends entirely on whether the contract allows an exit into the lower-priced market, a resale to someone else, or a renegotiation. Without one of those three, it is a paper loss you cannot realize, and a bill you still have to pay.

What to ask before you sign

None of this requires being a GPU engineer. It requires asking the questions that turn an assumed sunk cost back into something you have leverage over:

1. Can this reservation be resold or assigned to another party, and does the counterparty’s consent get unreasonably withheld, or does it come with a fee or a haircut on the price?
2. Is there a public or private secondary market for this specific type of contract right now, and what is it currently pricing similar capacity at?
3. What happens to the remaining term if your usage drops to zero? Is there a minimum-commitment floor, a downsizing right, or a hard obligation to pay for the full remaining calendar hours regardless of use?
4. If the market price for this class of GPU falls after you sign, as it has before, does your contract let you benefit from that in any way, or are you locked at the signing-day rate for the full term?
5. Who is on the other side of the resale market if you need to use it, a broker’s aggregated marketplace, or a bilateral deal you would have to originate yourself under time pressure?

A reservation priced well on day one and a reservation you can actually exit are two different things, and only one of them shows up in the headline rate. Ask about the second one before you sign the first.