Why trade compute

In one paragraph: Trion lists capped options on CMPT, the price of one H100 GPU-hour in USD. A CMPT option is a financial position on the price of compute. You do not rent a GPU or sign a power contract. The index is a trimmed, weighted composite of six named marketplaces and clouds, not one vendor's rate card. Buyers and sellers of a GPU-hour both have a reason to trade its price.

The energized-compute stack

Compute is not one market. It is a stack of three markets. Each is priced in its own unit. Each is the input to the next.

A GPU-hour sits at the top of the stack. Its floor is the power and capacity underneath. When a megawatt-hour or a megawatt of capacity gets scarcer, delivering a GPU-hour costs more. That is why CMPT can move on news that has nothing to do with NVIDIA.

Why the spot GPU-hour price moves

Four supply-side bottlenecks dominate. Each is a number you can watch.

1. The supply queue. New GPU capacity needs a grid connection. The US interconnection queue holds roughly 2,060 GW of projects awaiting connection (LBNL "Queued Up", retrieved 2026-09-21) — years of supply that cannot energize quickly. A long, slow queue means new GPU-hours arrive late and spot stays firm.
2. HBM. High-bandwidth memory is the scarce input to every accelerator. SK hynix's HBM is effectively sold out through 2026, and HBM4 is projected to roughly double the per-gigabyte cost of HBM3e (TrendForce, 2026-08-13). Expensive or short HBM means fewer, pricier GPUs.
3. Permits. Local opposition and moratoriums blocked or delayed roughly $200B of US data-center projects in H1 2026 (Data Center Watch, retrieved 2026-09-21). Every blocked megawatt is a GPU-hour that does not come online.
4. Model efficiency. Open-weight models keep pushing the price of a million tokens down. Token demand still compounds, because agents consume many times the tokens of a single chat. Cheaper tokens raise total demand for GPU-hours even as each token needs fewer of them.

Two forces pull opposite ways. Bottlenecks 1–3 hold supply back and push the price up. Efficiency pushes the cost of each token down while multiplying how many tokens get consumed. CMPT is the number where the two forces meet.

Who hedges, and how a capped CMPT option expresses each view

A capped option pays a bounded amount in USDG when the expiry fixing X settles on one side of the strike. One contract is one GPU-hour (1 CT = 1 H100-hr). The options page states the terms; how to trade walks through buying and writing.

• Call strike $2.55 / cap $3.05: pays max(0, min(X − 2.55, 0.50)) per contract.
• Put strike $2.55 / floor $2.05: pays max(0, min(2.55 − X, 0.50)) per contract.

The width — $0.50 — is the most a contract can ever pay. That is what lets the writer sleeve reserve the full liability. The worked examples use an index of $2.31 and illustrative premiums. Premiums are set by the engine from the mark, time to expiry and 30-day realized volatility (CMPT floor 24%), never promised here.

1. A neocloud hedges idle-capacity risk — buys the put

A neocloud commits to GPUs for 1–3 years at a fixed cost and sells the hours on demand. Its risk is the downside. If the supply queue clears and rates fall, it earns less per hour — or hours go unsold.

Buy the $2.55/$2.05 put. Each contract floors one sold hour at $2.55.

Two honest limits. The floor is $0.50 wide, so it covers the "rates soften" case, not the "rates collapse" case. Below $2.05 the payout stops growing while the spot keeps falling. And it hedges the price leg only. An hour that goes entirely unsold earns zero, put or no put.

2. An AI lab hedges inference-cost risk — buys the call

An AI lab buys GPU-hours. Its inference cost is the GPU-hour price divided by how many tokens each hour produces. If HBM or the queue pushes the GPU-hour price up, its margin shrinks.

Buy the $2.55/$3.05 call. Each contract caps one rented hour's upside at $2.55 (plus premium).

Same limit as before: a $0.50-wide cap is a partial cap. Above $3.05 the hedge stops paying while the spot keeps rising.

3. A miner monetizes conversion optionality — buys the call

A Bitcoin miner already holds an energized grid connection. Converting it to AI/HPC earns roughly 2× mining revenue per MWh (Grayscale Research estimate, not verified by Trion). The conversion decision is itself a call option on compute price. Buying the CMPT call is the liquid, bounded form of that option. You pay a small premium, participate if compute re-rates up, and lose only the premium if it does not.

A miner that has already converted is, economically, a neocloud, and hedges like one (buys the put). A miner that wants to sell the optionality backs the writer sleeve. Trion options are fully collateralized, so there is no naked short. Writers deposit USDG and share the pooled sleeve result (FLAT shares).

What the cap does and does not do

Every example shares one feature: the cap (or floor) bounds the payout. That is what makes the writer sleeve fully reserved. It is also the honest limit of the hedge. Each contract covers the first $0.50 of an adverse move and no more. A wider or deeper contract would cover more. Listing more strikes is a product decision, not a promise made here.

Know the risks

• Buyers can lose the whole premium and gas. Payout is capped.
• CMPT is built from listed and marketplace prices at six venues, not from a single executed tape. It can lag a fast repricing.
• Fixings can be challenged, corrected or delayed. At launch, governance and the council are Trion-held.
• The options launch is CMPT-only. GRID and perpetuals are not part of it.

Source trail: LBNL "Queued Up" (retrieved 2026-09-21); TrendForce HBM4 pricing (2026-08-13, retrieved 2026-09-21); Data Center Watch Q1/Q2 2026 (retrieved 2026-09-21); Grayscale Research, "Investing for the Compute Bottleneck" (Sep 2026 — the 2× mining figure is Grayscale's estimate, not verified by Trion); Capped options; Data methodology.

Repository-owned documentation · September 2026 · Educational material, not investment advice and not an audit.