Locking in compute: why long-tenor, prepaid GPU contracts work, and when they stop

The short version

A long-tenor GPU contract with a customer prepayment is a trade, not a concession. The customer buys reserved capacity and a locked price. The neocloud and its capital partner buy contracted cash flow and less capital at risk. In 2026 that trade is working because both sides fear the same thing: scarcity.

It stops working when expectations diverge. If customers believe rates will fall, or hardware will be obsolete before the contract ends, they stop committing. If providers believe rates will rise, they stop offering long terms at fixed prices. The skill is knowing which side of that line the market is on, deal by deal.

The 2026 backdrop: a steep tenor curve and rising capex

B300 shows how steeply the market prices commitment. The same GPU rents for roughly $7.90 an hour on demand and about $4 to $5 an hour on a multi-year contract. That gap is the price of certainty, and scarcity sets it.

Meanwhile the cost of building new capacity is rising. Memory is the driver, and it shows up directly in server quotes.

Exhibit 1. The 2026 backdrop in nine signals

Signal	Data point	Source
B300 on-demand	Median ~$7.89/GPU-hr across 13 providers, flat over the past 90 days (September 2026)	GetDeploying B300
B300 reserved	Median $5.71/GPU-hr reserved vs $7.87 on-demand (September 14, 2026)	AIMultiple GPU Index
B300 36-month take-or-pay	2,304 B300s, ~$21M revenue per quarter; implies ~$4.16/GPU-hr (our arithmetic)	Axe Compute 8-K
HGX B300 server (8-GPU)	Early-allocation pricing $640K to $690K per 8-GPU server incl. networking, 20+ week lead times (July 2026)	Rack Rate
Server DRAM	Contract prices up ~90% quarter-on-quarter in Q1 2026	TrendForce via Barrack.ai
AI server pricing	Builders told Nvidia customers of >15% increases for early-2027 shipments	Implicator.ai
Prepay norms (large)	Nscale weighted-average prepay 23% of contracted TCV	Nscale S-1
Prepay norms (large)	Nebius expects >$9B of customer prepayments in 2026; in its Q2 deals that included prepayments, they covered 50 to 60% of related capex	App Economy Insights
Prepay norms (extreme)	Some 1-year rentals require prepays up to 100% of contract value	SemiAnalysis

Source: As linked in each row, September 2026 unless dated.

Source: GetDeploying (September 2026); AIMultiple GPU index (September 14, 2026); Axe Compute Form 8-K, implied rate is GPU Capital arithmetic.

Show data

A multi-year B300 commitment prices at roughly 50 to 60% of on-demand

B300 rate	$ per GPU-hour
On-demand median, 13 providers	$7.89
Reserved median	$5.71
36-month take-or-pay, implied	About $4.16

The takeaway: a multi-year B300 commitment prices at roughly 50 to 60% of the on-demand rate. Owners discount that deeply because certainty is worth that much to them, and rising build costs mean the next cluster will not be cheaper.

Why it makes sense for customers

For an AI company running production inference or training, compute is cost of goods sold. Locking it in is the same logic an airline uses to hedge fuel.

• Capacity is the scarce asset, not price. When on-demand capacity sells out, the cost of not having GPUs is lost revenue. A dedicated, contracted cluster removes that risk.
• A lower hourly rate for longer commitments. Providers price tenor. Each additional year of commitment typically buys a lower locked rate, because the provider recovers its capital with more certainty.
• Prepayment is cheap financing, if your capital is cheaper. Every contract rate embeds the provider’s cost of capital. A well-funded customer that prepays is replacing the provider’s expensive capital with its own cheaper cash, and should be paid for it in rate.
• Budget certainty matched to revenue. Inference businesses increasingly sign their own multi-year customer contracts. A fixed compute cost lets them underwrite their own gross margin.
• Protection from build-cost inflation. With memory and server prices rising, capacity contracted today at today’s capex is a hedge against what the next cluster will cost.

Why it makes sense for neoclouds and capital providers

For the owner of the hardware, the biggest risks are an empty cluster and a falling rate. A long, prepaid contract addresses both at once.

• Less capital at risk from day one. A prepayment funds part of the purchase. The owner’s exposure is the net amount, not the sticker price of the servers.
• Utilization is contracted, not forecast. Take-or-pay terms replace a utilization assumption with a signed obligation for the full term.
• Rate decay is deferred. A fixed rate holds for the whole contract. Market-wide price declines only hit the owner when the contract rolls off and the capacity is re-rented.
• Financeability. Lenders size debt against contracted cash flows. A signed, prepaid contract is what turns a GPU purchase into a bankable asset. Nscale describes its take-or-pay model as the way it secures long-term cash-flow visibility for capital-intensive buildouts.
• Skin in the game. A customer that has prepaid a meaningful share of the contract is far less likely to walk away, and an unamortized deposit protects the owner if it does.
• Faster capital recycling. Nebius has said its Q2 2026 contracts are expected to pay back capex and operating costs in about 22 months, helped by customer prepayments (shareholder letter, August 12, 2026). Capital that comes back sooner can be redeployed into the next deal.

The tenor-for-rate trade

Longer tenor lowers the hourly rate but raises the total contract value, and a prepayment set as a percentage of the full contract grows with it. That is the lever most customers underestimate.

An illustration using our current B300 rate ladder: one 8-GPU HGX B300 node costing $640,000 including networking (the low end of current allocation pricing), 25% of total contract value prepaid, 100% contracted utilization.

Exhibit 3. The B300 tenor ladder, per node

Term	Locked rate ($/GPU-hr)	Contract value per node ($)	Prepayment per node ($)	Prepay as % of node cost	Owner capital at risk ($)
2 years	5.00	700,800	175,200	27%	464,800
3 years	4.80	1,009,152	252,288	39%	387,712
4 years	4.40	1,233,408	308,352	48%	331,648
5 years	4.10	1,436,640	359,160	56%	280,840

Contract value is 8 GPUs × the locked rate × 8,760 hours a year × the term. The prepayment is 25% of contract value. Node cost $640,000.

Source: GPU Capital calculation; node cost is the low end of Rack Rate’s HGX B300 allocation range (July 2026).

• Owner capital at risk
• Covered by the customer prepayment

Contract value is 8 GPUs x the locked rate x 8,760 hours a year x the term; the prepayment is 25% of contract value. Moving from two years to five cuts the customer’s rate by 18%.

Source: GPU Capital calculation; node cost is the low end of Rack Rate’s HGX B300 allocation range (July 2026).

Show data

Longer prepaid terms cut the owner’s capital at risk from 73% to 44% of the node

Term and locked rate	Owner capital at risk	Covered by the customer prepayment
2 years at $5.00	73%	27%
3 years at $4.80	61%	39%
4 years at $4.40	52%	48%
5 years at $4.10	44%	56%

Moving from two years to five cuts the customer’s rate by 18%. The owner’s capital at risk falls from 73% to 44% of the node cost, even though the hourly rate is lower. Both sides can be better off, which is why this structure has become standard for large neoclouds.

The flip side is also visible in the table: the customer’s total commitment roughly doubles. Whether that is sensible depends entirely on where the market goes next.

When it stops making sense

Long tenor works when both sides fear the market moving against them. It breaks when one side becomes confident the market will move in its favor. These are the signals we watch.

Exhibit 5. The signals that end the trade

Signal	Who walks away	What to watch
Supply catches up	Customer: why lock today’s rate if next year’s is lower?	Gap between 1-year contract and on-demand rates narrowing; memory and HBM supply easing (TrendForce projects HBM bit shipments up 50 to 60% in 2027)
Hardware generation outruns the term	Customer: paying for a chip superseded twice before the contract ends	Rubin, B300’s successor, began production shipments in August 2026; compare contract length to Nvidia’s release cadence
Customer capital becomes expensive	Customer: prepaying ties up scarce cash	Funding conditions for AI startups; customers asking for lower deposits or shorter terms
Counterparty risk on the prepay	Customer: a prepayment is an unsecured loan to the provider	Provider leverage, vendor-financing dependence, requests for escrow or step-down deposits
Rates rising fast	Provider: a fixed rate gives away upside	Contract-rate index momentum; renewals at or above original rates
Demand shock	Both: take-or-pay becomes stranded cost for the customer and credit risk for the owner	Inference efficiency gains, model consolidation, customer revenue concentration

Source: GPU Capital. Rubin shipments: NVIDIA Q2 fiscal 2027 results, August 26, 2026.

The customer-capital signal is already visible at the smaller end of the market. In our experience, proposals with 25% prepayments and five-year terms take far longer to close than shorter terms with smaller deposits. Large, well-funded buyers can prepay heavily; many growth-stage inference companies cannot, or will not.

The counterparty point deserves emphasis. A customer that prepays is lending to its provider. As vendor-financed and backstopped neoclouds proliferate, sophisticated customers will ask who holds the hardware, who holds the debt, and what happens to their deposit if the provider fails.

How GPU Capital structures around this

We treat tenor and prepayment as a single pricing decision, and we underwrite what happens after the contract ends, not just during it.

• Price the ladder, let the customer choose. We offer a rate for each tenor so the customer can trade term for price on its own terms, rather than negotiating one number.
• Deposits amortize, they don’t sit idle. The prepayment is credited back against monthly billing over the term, so the customer sees it working and both sides’ exposure declines together.
• Rates locked flat, decay underwritten at roll-off. We assume an annual market decline that applies when capacity is re-rented, and lower utilization after the contract ends.
• Match term to useful life. We underwrite a six-year useful life, with lower rental rates assumed at re-lease; actual life may be shorter. We avoid terms that outlast the hardware’s economic relevance for the customer; that is where defaults come from.
• Judge structures on capital at risk. A structure that earns similar dollars on less net capital is better, because the freed capital funds the next cluster.
• Fill the gap programmatic lenders leave. We finance $1M to $5M deployments, up to the full capital stack, that are often too small and too contract-specific for large-scale lenders.

If you are a neocloud with a customer ready to commit, or a customer weighing term against price, we would like to hear from you.