Blog ▸ Presell Your GPU Capacity: How to Book Revenue Before the Hardware Lands
GPU Infrastructure
Blackwell nodes take 16 to 26 weeks to arrive and the deposit left in October. List the capacity the day the order is confirmed and the contract is waiting when the truck is.
Presell Your GPU Capacity: How to Book Revenue Before the Hardware Lands
GPUaaS.com Team
GPU Infrastructure
October 13, 2026
No items found.
GPU capacity that has not been delivered yet can be listed on GPUaaS.com against its delivery date, and buyers planning a quarter ahead will commit to it. With Blackwell node lead times running 16 to 26 weeks, a provider who waits for the hardware to arrive before listing it gives up four to six months of booked revenue on a rack that is already paid for.
This post covers how presell works on the network, which buyers commit to future capacity and why, what a provider has to be able to say about undelivered hardware for a buyer to sign, and the two risks that have to be managed: slipped delivery and spec drift.
Key takeaways
Capacity on order can be listed with a future available date. It enters the same request queue as live capacity, flagged by date
Blackwell node lead times from order to delivery run 16 to 26 weeks. That is the window a presale fills
Buyers who commit to future capacity are planning a Q1 start, locking a committed band before it moves, or reserving against a known allocation shortage
A presale needs four things in writing: the order confirmation, the delivery date, the exact configuration, and the facility it will land in
Slipped delivery is handled by a stated buffer and a start-date clause in the three-party contract, not by hoping
B300 and GB300 NVL72 are the parts most often presold this month. A handful of buyers are already asking about Vera Rubin
◆ THE WINDOW
Sixteen to twenty-six weeks of paid-for hardware earning nothing
A B300 node ordered in October lands between February and April. The deposit went out in October. Depreciation on the balance sheet starts when the asset is capitalised, which for most providers is delivery, but the financing cost starts when the money leaves. For a provider carrying a fleet on debt, that is four to six months of interest on $500,000 to $795,000 a node before the first GPU-hour is billed.
The conventional sequence is to wait for the hardware, rack it, burn it in, and then go looking for a buyer, which adds another four to eight weeks before the first contract starts. The alternative is to list the capacity the day the order is confirmed, let buyers commit against the delivery date, and have the contract start the week the nodes pass burn-in. The supply anchor post covers the general case; this post is the detail.
The B300 node price breakdown puts owned cost at $3.06 to $8.05 per GPU-hour over three years depending on purchase price and facility. Every month the node sits unbooked, the effective figure rises, because the same capital is amortised over fewer earning hours.
16 to 26
weeks from order to delivery on a Blackwell node in Q4 2026. Capacity listed on day one has a contract waiting when the truck arrives
OEM lead times as compiled in the GPUaaS.com B300 node price post, October 2026
◆ WHO COMMITS TO CAPACITY THAT DOES NOT EXIST YET
Three buyer reasons, and all of them are about the calendar
The first is a planned start date. An enterprise AI team with a project funded for Q1, or an AI product company with a model launch in February, knows now that it needs 32 B300s in March. Committing in October against a stated March delivery gives them certainty, and it gives them the current band. A minority of requests through the network this month are for a Q1 2027 start, per the October demand snapshot.
The second is the band itself. B300 committed capacity quotes at $5.00 to $6.50 per GPU-hour this month, below B200, and buyers who read the benchmark know that is unusual. Locking a twelve-month term at this month's band against a Q1 delivery is a rational hedge for a buyer who expects the band to firm up once demand catches up with allocation.
The third is allocation. On GB300 NVL72, and in Europe on anything Blackwell with residency documentation, buyers know supply is short and will reserve against a provider's confirmed allocation rather than wait and compete for it in March. For a provider holding a confirmed European B300 allocation, that is the strongest position on the platform right now.
◆ WHAT A PRESALE LISTING NEEDS, AND WHY
Item
What to provide
Why the buyer needs it
Order confirmation
OEM or distributor confirmation with quantity and SKU
Proves the allocation exists and is yours
Delivery date
OEM-stated ship date plus your rack and burn-in buffer
Sets the contract start date and the slip clause
Exact configuration
GPU, form factor, CPU, memory, NICs, fabric, cooling
Prevents spec drift between what was sold and what lands
Facility
Named site, country, power and cooling confirmed for the density
Residency, and proof the hardware can actually be sited
Support and SLA
Same written policies as a live listing
A presale is a normal contract with a later start
GPUaaS.com provider operations guidance for presold capacity, October 2026.
◆ THE TWO RISKS
Slipped delivery and spec drift, and how the contract handles both
Delivery slips. Supermicro build capacity was named as the B300 bottleneck in October and OEM dates move. The way to handle it is not to promise the OEM date; it is to list the OEM date plus a buffer, four weeks is typical, and to write a start-date clause into the three-party contract that says what happens if the buffer is exceeded. Usually that is a short grace period, then an option for the buyer to walk or for the network to source interim capacity. A provider who states the buffer up front and hits it is trusted on the next presale. One who promises the OEM date and misses it is not.
Spec drift is the quieter risk. A node ordered as HGX B300 with ConnectX-8 and liquid cooling can arrive, after an OEM substitution, with a different NIC or a different CPU, and a buyer who committed against the original spec has a legitimate complaint. The fix is to list the exact configuration from the order confirmation and to notify the buyer of any change before delivery, not after. The DGX vs HGX post covers how much the same eight GPUs can vary around them.
Both risks sit with the provider, and both are smaller than the risk of an unbooked rack. The contract is between the provider, GPUaaS.com and the buyer, so the terms are agreed once and the provider is never negotiating slip clauses with an unknown counterparty.
◆ WHAT IS BEING PRESOLD NOW
B300, GB300 NVL72, and the first Rubin questions
B300 is the most presold part this month, because it is the most requested part and its allocation is still arriving. GB300 NVL72 is second, because rack-scale buyers expect to wait and prefer to reserve. H200 and H100 are rarely presold, since live supply is available. A handful of buyers have already asked whether Vera Rubin can be presold against a 2027 allocation; it cannot be quoted yet, but a provider with a confirmed Rubin allocation date should say so on the listing, because the question will be routine by Q1.
Presell also works for a provider who is expanding rather than entering. A current provider adding a second rack can list it on order and have the term start on delivery, without any gap between the capex and the revenue. The one node post covers the listing itself; a presale is the same five fields with a date in the future.
For capacity that will be sold as inference rather than as a committed cluster, packet.ai can onboard a provider ahead of delivery too, so that token demand is routed to the nodes the week they pass burn-in.
Hardware on order? List it today.
Five fields and a delivery date. No cost to apply or list. Buyers planning a Q1 start are already asking.
Yes. Capacity on order can be listed on GPUaaS.com against its delivery date, with the order confirmation, exact configuration and facility documented. It enters the same request queue as live capacity, and buyers planning a quarter ahead commit to it on the current band with the contract starting on delivery.
The listing carries the OEM date plus a stated buffer, typically four weeks, and the three-party contract has a start-date clause covering what happens if the buffer is exceeded: usually a short grace period, then the buyer may walk or the network sources interim capacity. State the buffer honestly and hit it.
Three reasons: a planned Q1 start date that needs certainty now, a wish to lock the current committed band before it moves, or a known allocation shortage on parts such as GB300 NVL72 or European Blackwell with residency documentation, where reserving beats competing for it later.
The OEM or distributor order confirmation with quantity and SKU, the delivery date with your buffer, the exact configuration including fabric and cooling, the named facility with power and cooling confirmed for the density, and the same written support and SLA policies as a live listing.
B300 most often, then GB300 NVL72. H200 and H100 are rarely presold because live supply exists. Some buyers are already asking about Vera Rubin against 2027 allocation; it cannot be quoted yet but a confirmed allocation date is worth stating on a listing.
No. It is the same three-party contract between the provider, GPUaaS.com and the buyer, with a later start date and a slip clause. The provider is not negotiating delivery terms with an unknown counterparty, and the buyer is not paying before the capacity is live.
