Do you guys see what’s happening in the market right now?
Deals are clearing above $24/gpu/hr for some B300s with pricing for short term (<1 year) compute hovering above $7. Actually by the time you finish reading this it will have gone up again. I’m not even going to entertain the ridiculous VR pricing that’s starting to float around, let alone everyone’s promises of Q1 deployments.
Maybe one day I will share my thoughts on the current state of the public and private credit markets (no one in SF somehow cares that the 10y just passed 5%) and potential oversupply, but today I want to talk about trading compute
There’s an emerging market of compute derivative products and I believe this market could fundamentally change how neoclouds and anyone adjacent can grow as well as protect themselves. A 5 year deal asks me to take a huge bet on demand to secure an affordable rate. I'd like to pay for protection against expensive GPU-hours while leaving myself room to change how many hours I rent. A supplier or trader willing to carry the price risk could earn that premium.
For an inference cloud, this is a pretty urgent problem. If I sell a customer fixed-price service while my GPU bill floats, I've taken a position on compute prices whether I meant to or not.
Buying before the customer signs
Say I'm quoting a fixed price to a video gen customer with a distribution deal pending. They'll decide whether to launch in 12 months. If they do, I need 2,048 B300s for their first 3 months, months 13-15. That's 4,485,120 GPU-hours at 730 hours per month.
A vendor offers me those 2,048 B300s starting today at $4.15/gpu/hr if I sign for 5 years:
2,048×4.15×24×365×5=$372,264,9602{,}048 \times 4.15 \times 24 \times 365 \times 5 = \$372{,}264{,}960
That buys 60 months of compute starting now. The customer has outlined a three-month launch window a year away. Ay, there's the rub.
If I make a firm reservation for their launch now, I’m committed even if the deal falls apart. If I wait, the customer might launch just as everyone else wants the same GPUs. I'd like their success to be good news for my margins too. In fact, I'd pay a bit more on average if it meant I could resize the fleet more often and take some of the sting out of an expensive renewal. The question is how much that protection costs.
A call option would help here. I pick a rental benchmark and a price I'm worried about paying above called the strike. For this launch, suppose I buy a call on the average rental index over months 13-15 with a $4.50 strike and settlement at month 15. If that average is $8, the call pays $3.50 per covered hour. If it's $2, the call pays nothing and I get the cheaper rent.
If my average rental price S matches the index, the math becomes:
S−max(S−4.50,0)=min(S,4.50)S-\max(S-4.50,0)=\min(S,4.50)
That's a $4.50 ceiling on the matched quarter's rent before the option premium and financing. The rental purchase stays separate so I can choose whether to make it after the customer decides.
That still leaves me with a customer expecting 2,048 GPUs on launch day. I'd have to secure them through a separate capacity agreement or accept the risk of sourcing them later. Reserving capacity has its own cost and supplier risk. For the example below, I’m assuming I can source the GPUs and cover the rental invoices until the call pays cash at month 15 (this is a massive assumption).
The (pseudo) forward curve
Before I can decide whether this is worth buying, I need a price for the call. The hedge covers a specific future quarter so I need a reference price for those future hours.
A forward curve puts today's prices for future delivery periods on a timeline. For this call, I want the price I could agree today for the customer's months 13-15 rental index. That will be the starting input for the option calculation. Forward prices also reflect what the market charges for taking risk, so I wouldn't read the curve as a pure forecast of where GPU rents will end up.
Ideally, I'd ask a dealer for that quarter's forward quote. For now I've got the vendor's package menu for the same 2,048 GPUs: 1 year at $6.25/gpu/hr, 3 at $4.85, 5 at $4.15. Same nodes, start date and service, paid monthly with nothing down.
The three-year price includes the first year which I can also price separately so we can subtract it to see what the remaining two years contribute:
3×4.85−1×6.253−1=4.15\frac{3\times4.85-1\times6.25}{3-1}=4.15
Doing the same math between the 5 and 3 year packages gives $3.10 for years 4-5. My initial block prices are therefore $6.25, $4.15 and $3.10.
This is a pseudo forward curve using the package-unbundling construction Ornn describes. The $4.15 is inferred from two whole contracts. The vendor hasn't offered to sell me just years 2-3 at that price. The packages also pay for availability, credit and commitment terms which Bandi and Su examine. A dealer's index-forward quote could differ.
Payment dates matter too. At an assumed 10% continuous annual discount rate, I subtract the discounted package bills and divide by the discounted GPU-hours in the extra months. With month-end payments, the blocks become $6.25, $4.04 and $2.80, the solid line below.
read image description
ALT
Figure 1. Hypothetical whole-term quotes and implied blocks. Flat monthly prices are assumed. Solid: 10% continuous discounting with month-end payments. Dashed: no discounting. Blocks cannot be bought separately.
The customer's launch window falls inside the second block and I'll assume its monthly price is flat and use $4.0377 as my starting proxy for the quarter.
The price of the future
The curve is only the starting point. For the same forward level, higher volatility in price gives the call more value because its upside can grow while its payout stays at zero below the strike.
Let's price it in a simple model. I'll assume I can buy and sell matching index forwards today and at month 12, for all the hours I need. I can borrow and lend at a fixed 5% continuously compounded rate, and I'll leave out fees, default and margin constraints. With those assumptions, I can use forwards and cash to replicate the option's payout, which gives us a price.
Say the $4.0377 forward goes up or down 50% at month 12 then does the same again before the quarter average settles at month 15. At each step I give the up and down outcomes a 50% pricing weight each which keeps the weighted next value equal to the current forward. These are weights for pricing the hedge, not a forecast of GPU rents. The two steps cover different lengths of time so these moves don't amount to a constant annual volatility either.
read image description
ALT
Figure 2. Hypothetical two-step model. Forward values today and at month 12 lead to the months 13-15 average index, settled at month 15. Each step has 50/50 pricing weights. Only the two-up outcome pays above the $4.50 strike, with 25% pricing weight.
The quarter average ends up around $9.08, $3.03 or $1.01. Only two up moves get it above the $4.50 strike. That's the outcome that pays.
At the top price, the call pays $4.584896 per covered hour. That outcome gets a 25% pricing weight (50% × 50%). Discount the weighted payout back 15 months:
C0=e−0.05×1.25[0.25×4.584896]=1.07678C_0=e^{-0.05\times1.25}\left[0.25\times4.584896\right]=1.07678
The call covers the customer's 4,485,120 GPU-hours during months 13-15. At $1.07678 per covered hour, that's $4.83M paid today to protect the price during that three-month launch window. The GPU rental invoices are still paid separately.
Optionception
Buying this call now commits $4.83M of cash before the customer has signed. In this model I could buy it and sell it later if plans change. It’d be nice to also have a quote for keeping more cash in the business now while fixing what it would cost to buy the call in a year.
A compound option (option to buy option) lets me pay today for the choice of buying that same call at month 12. I fix that later purchase price now at $1 per covered hour. That later dollar buys the option, not the GPU rentals.
Start at month 12 and work backwards. If the forward went up to $6.0566, the call I can buy is worth:
C12=e−0.05×0.25[0.5×4.584896]=2.26397C_{12}=e^{-0.05\times0.25}\left[0.5\times4.584896\right]=2.26397
I'd pay $1 for something worth $2.264. If the forward went down, the call is worth zero and I leave it alone. And if the customer cancels while the right is still valuable I can still exercise or sell it. To get today's price I weight those outcomes and discount them:
V0=e−0.05[0.5×max(2.2639708−1,0)]=0.60116\begin{aligned}V_0&=e^{-0.05}\left[0.5\times\max(2.2639708-1,0)\right]\\&=0.60116\end{aligned}
For the same launch window that's $2.70M now plus $4.49M at month 12 if I exercise. The possible later payment has a pricing-weighted present value of $2.13M, exactly the difference between the two upfront prices.
Say I've set aside $4.83M for price protection on this launch. Buying the call now uses it all today. The compound leaves $2.13M available. That could help fund this customer's development work while I wait on the launch. The compound gives me a different cash schedule to choose from. In the frictionless pricing model I could also borrow to buy the call now.
read image description
ALT
Figure 3. Hypothetical payments for 4,485,120 covered GPU-hours. The compound leaves $2.13M available today and requires another $4.49M at month 12 if exercised. The call settles at month 15 whichever way I buy it.
What would this change for the fleet?
Suppose the launch goes ahead and this video gen customer becomes an ongoing workload. I'll start a fresh five-year simulation at that point, with 2,048 GPUs of demand and a new hypothetical rental menu. Reserving everything exposes me to unused hours. Renewing annually gives me more chances to resize, but also more chances to encounter expensive prices. I want to see whether calls make that second choice easier to live with.
I ran a simulation with 4,000 made-up five-year paths for rental prices and customer demand. Annual renewal averaged $4.72 per hour across those histories, and its worst 5% averaged $10.91. Adding calls raised the first number to $4.88 and brought the second down to $7.20, including the option premiums.
I tried five purchasing policies on the same paths, including what I'd pay to cover shortfalls, what I'd recover from spare capacity and what the options cost. Here's what I compared:
- A: fix all 2,048 GPUs for five years.
- B: renew annually, sizing the purchase to demand observed at renewal.
- C: fix 60% of the original fleet and renew the remaining need annually.
- B + calls / C + calls: add $4.50 monthly index calls for operating months 13-60, on fixed 2,048/819.2-GPU notionals respectively.
I made price and demand tend to move together, so the business can find itself needing more GPUs just as they get expensive. It can resell only 70% of spare capacity at 80% of the index and pays 125% for shortfalls. I let it buy enough to serve all demand here, so I'm comparing the bills.
For prices I used 50% annual volatility. The demand driver uses 35%, with 0.6 correlation between its shocks and price shocks. I clip and rescale demand to keep its expected level at 2,048 GPUs. This five-year operating case uses a fuller hypothetical rental menu and 10% discounting, separate from the three-month launch hedge above. I price the calls using the same process that generates the paths, so their expected discounted profit is zero.
read image description
ALT
Figure 4. Hypothetical simulation, including premiums. The separate bars show equal-weight mean path cost and the mean of each policy's own worst 200 of 4,000 paths. Each path divides discounted cost by discounted requested GPU-hours.
That makes annual renewal much more interesting to me. I still get to resize the purchase each year. Adding calls trades a slightly higher average cost per hour for less expensive bad outcomes. If those are the outcomes that would wipe out my margin or force me to scramble for cash, I'd want that quote.
What this means for the future
I think these early forward trades are the beginning of a much bigger market. Once there are credible prices for future GPU-hours (and residuals), buyers can start asking for protection around their own plans. A customer might want a fixed bill. A supplier might want some income locked in. Someone else might be happy to take the price risk for a premium. Those are different needs and a five-year rental contract is a pretty blunt way to handle all of them.
For a smaller operator this could change which customers are even worth pursuing. I might be perfectly capable of serving this video gen customer and still be unable to carry years of rental commitments while they sort out distribution. A reasonably priced hedge could let me protect the launch's rental-price exposure while I wait. I'd still reserve against demand I trust. I'd have another way to buy for the demand that might arrive.
And yes, I think the operators who understand the financial side will have an advantage. Being very good at inference won't stop the rental bill from eating the margin. I need to know which risk I'm getting paid to carry and what it would cost to pass it to someone else.
That's why I want to see a proper market develop here with quotes in useful sizes and counterparties that can pay when prices move. I'd like to be able to take this hypothetical purchase to a dealer and get an actual choice: here's the cost of reserving, here's the cost of protecting the price, and here's the cash each would tie up. For a startup trying to grow, having that choice before committing billions of dollars could matter a lot.
Sources
Trades and regulation
- FalconX: H100-index OTC swap announcement.
- Wintermute: H100 forward announcement.
- CFTC: Sep 21review-extension letter. Extends the NYMEX compute-futures review through Nov 9.
Data and methodology
- Ornn: Forward curves and package unbundling, public preview and OCPI methodology. The curve construction and on-demand rental benchmark referenced here. OCPI excludes reserved contracts. I used these as references and reproduced no Ornn data.
- US Treasury: Daily par yield curve, September 2026. The 10-year yield was 5.17% on Sep 25.
Research papers
- Sergey V. Chernenko, Krista B. Schwarz and Jonathan H. Wright: The Information Content of Forward and Futures Prices (2004). How forward prices reflect both expectations and the price of risk.
- Federico M. Bandi and Yinan Su: (Early) AI Compute Asset Pricing, v3. Compute pricing, non-storability and the distinction between term rentals and forward contracts.
- John C. Cox, Stephen A. Ross and Mark Rubinstein: Option Pricing: A Simplified Approach (1979). The binomial replication logic behind the option-pricing example.
- Mark Davis, Walter Schachermayer and Robert Tompkins: The Evaluation of Venture Capital As an Instalment Option (2003). Staged investment and the option to make a later payment.
Hypothetical prices and option terms were used for the worked examples and simulation.
