Compute Hedging

Blockhouse has been running trading, investing and hedging infrastructure for institutions across crypto, equities, FX and commodities. Now, compute is the fifth.

We have begun sourcing counterparties for cleared block trades in compute futures, typically under $10 million notional, 30 to 90 day tenors, on H100, H200, B200, B300 or A100, with blended Hopper and Blackwell baskets also available. We can originate either side of the trade.

If any of the situations below is recognizable in your own numbers, we’d like thirty minutes. We promise you’ll get something useful out of it whether or not you ever trade with us.

So what changed?

Computing power is rented by the hour and priced like a commodity. Contract H100 rates went $1.70 → $2.35–2.65/hour between October 2025 and March 2026. That's a 40–56% move in five months with B200 falling 31% in three weeks last summer.

Every other major business input has a way to fix its price in advance. An airline locks in jet fuel. A bakery locks in wheat. Compute had nothing.

That ended this year. CME lists compute futures on 5 October, and privately negotiated trades clear today. However, while the instruments now exist, what doesn’t exist is anyone helping companies work out their exposure, size the trade, and execute it.

Where Blockhouse comes in

You tell us the exposure. We come back with a price and then clear it.

Our trades are cash-settled against a published index, so no delivery, no capacity commitment, and a clearing house stands between the two sides rather than us. Typical trade: under $10 million notional, a 30 to 90 day tenor, on a single chip.

Nothing about your infrastructure changes. You buy GPUs from the same provider, in the same region, on the same terms. The hedge sits alongside the business as a financial position, a price agreement, rather than a supply agreement.

What it looks like. You burn roughly 10,000 B200-hours a day which is 900,000 GPU-hours a quarter. At $6.00/hour you budget $5.4M and tell your board gross margin will be 55%.

A major model launches, spot availability evaporates, the index runs to $7.80. Your bill arrives at $7.02M instead, showing up as $1.6M nobody planned for, and 5.4 points of gross margin, for a reason entirely outside your control.

If this was hedged, the contract pays the difference and you pay what you said you’d pay.

Who this is for

AI companies self-hosting on rented GPUs

“Our gross margin is 52% and the board keeps asking why it isn’t 80%.”

You moved off an API to running open-weight models on rented GPUs, because at volume it’s cheaper. Your cost of goods is now GPU-hours. Inference eats 40–50% of revenue in this category against 15–20% for traditional software.

Neoclouds, GPU operators and data centers

“We’ve sold twelve months of capacity and can’t price month thirteen.”

You own or operate GPUs against debt service that doesn’t move when rental rates do. Your contracted book runs out well before your debt does, and the months after that are the ones your lenders ask about. You can sell forward without committing any capacity. It’s cash-settled, so it turns a forecast into a number you can put in a credit memo.

Inference providers

“We sell tokens at a fixed price and buy GPU-hours at a floating one.”

Your output price is contracted and your input price isn’t. Neither leg is hedged, which is most of why gross margins in your category run near 50% rather than the 70%+ of comparable software businesses.

GPU-backed lender: private credit, infra funds, project finance

“We’re lending against hardware we have no way to mark.”

Rental value moved 40% in five months and there’s no observable price for your collateral. NVIDIA is organizing $500B+ of AI-factory financing while backstopping only about 25% per deal. The rest is independent credit forming its own view on residual value, without a mark or a hedge.

Funds, family offices and treasuries

“We want compute exposure and there’s no way to buy it.”

There is no compute ETF, no index fund, no managed product. Allocators cannot buy this at any price today, and BlackRock’s CEO has said publicly that a new asset class will be buying compute futures.

What a conversation involves

Thirty minutes. If it’s useful, twelve months of your compute cost or revenue history under NDA.

You get back an exposure report on your own book: how your realized cost or revenue actually tracked the published indices, where your basis sits, and what a hedge would have returned over the last year.

That report is yours regardless. It’s useful for budgeting and board reporting even if you never hedge anything. If you do want to trade, we can price a cleared 90-day forward now.

What we offer

Block execution: Live

You tell us the exposure. Chip, direction, size, tenor. We price it, find the other side, and get it cleared.

Bespoke hedges: Live, and where the index work pays

We structure protection shaped to your actual book — your chip, region, interconnect grade and contract terms rather than a market-wide average.

Vaults: Building

Packaged compute strategies inside the managed-account structure we already run, for allocators who want the exposure rather than to shed it.

Why us, in brief

A listed contract settles on a market-wide index, and nobody owns the average. That gap is basis, an exchange structurally can’t cover it, because standardization is what makes an exchange work. Pricing it means reproducing the published indices from their source methodologies, which is the work we’ve done.

Blockhouse already runs trading, investing and hedging infrastructure for institutions across four asset classes. Our next frontier is compute.

Hedging GPU cost is the first product. Compute as an allocation, the power leg, and token-denominated hedges follow. We're more than happy to walk through the roadmap if it’s useful.