◆Kinetic Alpha

Practice focus · Energy & Compute

Compute is congealed electricity. Almost none of it is priced that way yet.

The first new major commodity in decades is being financialised in public, and the complex is being assembled out of order — futures listed before benchmarks are governed, indices published before their denominators are measured, credit extended before any of it can be marked. This section follows the stack from the bottom up: what a GPU-hour physically is, where its electricity is priced, who gets to define the benchmark, how paper converts into racks, what the tokens on the other side are worth, and who is financing the whole thing.

Every piece runs on primary data — measured power traces, hourly settlements, contract texts and settlement metadata pulled from the venues themselves — and ships with the tooling to re-run it.

401,379
hourly ISO settlements
ERCOT 4 hubs + PJM 3 hubs + DOM zone, Jul 2023 – Jul 2026
0.1 s
GPU power trace resolution
measured H100 workloads, CC-BY, 5 workload families
1,024
vLLM serving runs joined to power
Llama-3.1-70B on 4×H100 — 11.6M tokens/GPU-hour measured
5
index administrators catalogued
component-level methodology, four on the CFTC track and one outside it

1 · The denominator — what a GPU-hour actually is

Every index, every hedge and every credit structure downstream rests on a measurement nobody publishes: how much energy a GPU-hour consumes, and how many tokens it produces. The work starts here because everything else inherits the error.

A Taxonomy of Compute Cost

The denominator question asked end to end. One kilowatt through one chip costs its owner about $1.07 an hour and reaches the buyer as anything from nine cents to fifty dollars a million tokens — 543 times over the same physical input. A seven-part address (silicon, cloud form, channel, tenancy, workload, deployment, terms) names where in the resale chain any given price sits, twelve levers are ranked by how much of the gap each closes, and three worked ledgers run an enterprise bill, a seat margin and a per-task dealer book. Two of the largest rents in the stack are bounded rather than measured

Also in Tools
Silicon to Tokens — the primer

The physical layer beneath every economic question in AI, for a markets and risk audience. Serving a model is two different jobs — reading the prompt is a math problem, writing the answer is a data-moving problem — and each phase uses only half the chip you paid for, which is why decode sets the margin of the whole business. Plus the software stack that multiplies token output without buying a chip, one GPU's walk through the cascade to its shutdown floor, and the three relationships that bound a forward curve on something that cannot be stored

The denominator problem

Measured H100 workload power against the nameplate-TDP convention: duty factors 0.795–0.880, nameplate overstating device energy 12–21%, facility derates of 73–80%

Data centers are not a flat load

Measured facility profiles joined to 401,379 hourly DA settlements: the load-weighted price runs up to 9.4% above flat-block, and the premium peaks at 40% utilization — the ramp phase, not saturation

The inference spark spread

The heat rate of inference measured at 11.6M tokens per GPU-hour from 1,024 vLLM runs joined to NVML traces — the conversion term the whole complex needs and no administrator publishes

2 · The power leg — where the electricity is actually priced

A data center is a load shape before it is anything else. The energy work exists because the compute work needs a fuel leg that is priced hour by hour, not as a flat annual block.

3 · The index layer — five administrators, no rulebooks

Once compute has a price it needs a benchmark, and a benchmark is a governance object before it is a number. This is where the practice has spent the most time, because it is where the asset class is least finished.

Two Indices, Two Jobs

Kalshi’s reported compute plan puts Ornn’s on-demand index under the perpetual and Compute Desk’s one-year index under the dated future. The case for Ornn under the event markets and Compute Desk under both futures: funding noise, one- and two-day margin, one reference price per venue, and a route for funds to hold compute without the roll. Plus four corrections to the power section of Kalshi’s white paper, where power is shown at 16% of a GPU-hour against a measured 3.5% to 6.5%

Price the Work, Not the Token

A token is how AI is billed, not what a buyer needs done. Six kinds of token index sorted by what each holds fixed; what an index of work would be; and what Ornn’s daily token data shows when the two move apart. In two weeks of October the average Anthropic token got 38% cheaper while the same models cost 3% more. Closes by matching each index to the exposure it fits: a buyer’s own work, a lab’s revenue per token, or the gap between selling tokens and renting chips

The Shape Is the Method

Seven firms now price the same GPUs and three draw forward curves for a market where no forward trades. What on-demand, interruptible and reserved prices each measure; how the same H100 rates read as contango on one comparison and backwardation on another; a test of what hourly index data is worth as a hedge; the five design choices that separate two indices of the same price; and why a one-year reserved benchmark fits the exposure the market carries, on three conditions

The Contract Lost Its Date. The Seller Never Had a Hedge.

The question the delay does not ask. A working paper from HEC Montreal runs a factor model over the rental rates and finds a broad index hedges 99.7% of a flexible buyer’s variance against 28.2% of a locked operator’s — structural, not a defect of any administrator. Per-generation contracts, which is what NYMEX filed, lift the operator to about 0.37; nothing reaches the configuration residual. Two H100 indices from different publishers correlate at 0.17, so which one settles is part of the contract. Closes with nine instruments for the operator’s side and which are buildable before November 9

Also the contract layer
The Kinetic Alpha Compute Index

Every benchmark published so far prices the rent; this prototypes the index that prices the whole hour — rent less power at the meter, less space, plus the change in value. It decomposes by unit and by piece of the money and both cuts reconcile to the headline exactly, which is what makes each sub-index a leg someone is long and someone is short. Interactive: move any input and watch both cuts still add up

Five Indices, One Price

What a benchmark has to do before a contract can reference it — the six models of administration, the four series nobody publishes, and why the harder deficiency is independence rather than data. Including August's two exceptions: a regulated third-party administrator over Compute Desk's benchmarks, and an open-sourced construction anyone can rerun. With the insurance industry's contributor mutual as the structural precedent, antitrust case included

Open Source Is Not an Audit Trail

Transparency is not reproducibility. NATIVX published its whole index engine under Apache-2.0 — the most inspectable compute index in existence — so we cloned it and recomputed from the constituents published beside each value. Ten of thirteen equalled one vendor's list price to the cent, two matched no observation at all, and the checksums hashed nothing. NATIVX shipped a fix six days later: 64 of 64 values now reconcile exactly, and the auditable history that created falsifies its own forward curve by 6.3×

The Compute Crack Spread

The futures settle on rent, which barely moves; tokens reprice in hours and have no index. C = T·p − r, built from a live endpoint book — with the CFTC's August 19 request for comment read as a specification the token leg mostly passes and the rent leg mostly fails

The Contract Has a Date. It Needs a Dealer.

Who actually holds a floating GPU-hour exposure — read off the filings, the calls and the lender decks. Every balance sheet with a real exposure prices it fixed, no hedger is named anywhere, and the CME filing never says which Silicon Data tier settles it. Compute is running the oil sequence backwards

The Long End Has Two Sellers Now

The Treasury and the AI build-out are selling duration to the same buyers — three crisis paths, four transmission channels, and a 100 MW project run through each, which shows the merchant case dying of rental deflation rather than of rates

The Other Side of the Spark Spread

The merchant power desk mapped function by function onto an inference lab — a three-leg chain from power to GPU-hour to token with two heat rates in it, only one of which is physics, and the finding that at 0.8% of token revenue the lab's trader hedges absolute dollars and the tail rather than margin

The compute risk premium got its first print

The academy's first brief on compute futures, reviewed: cost-of-carry is dead on arrival, term rentals are contaminated proxies through the physical access wedge, and hedging pressure sets the premium — plus four corrections where the market has moved past the model

The contract got a date

The first US compute futures list on NYMEX on October 5 — and the dispersion ladder they must navigate, from 1.4× same-chip variance to a 2.6× tier spread that sits outside the settlement sample

The index was the wedge. The product is the risk desk.

An eleven-product audit of the administrator whose numbers the futures settle against — seven exposures mapped to five participant classes, an honest routing table, and the utilization benchmark nobody publishes

How the indices actually calculate price

Component-level catalogue of every administrator positioned to bear settlement weight — and the finding that not one publishes a rulebook, IOSCO statement, audit or oversight committee

Verifiable is not replicable

A fifth administrator, outside the CFTC perimeter, disclosing more of its construction than any of the four inside it — and still not reproducible. Publication answers IOSCO Principle 9, not Principle 11

Implied compute forward curves

Kalshi's ladders settle on Ornn OCPI, so a strip of digitals is a discretised distribution — the first continuously published compute term structure, with a $1.23/GPU-hour convenience yield measured against reserved-tier term sheets

Four curves, one commodity

Kalshi, Architect, CME × Silicon Data and ICE × Ornn — the race to price the GPU term structure, and how much of it settles on one index family

The compute complex

The six-level hierarchy the rest of the work indexes against: benchmark, grade, region, firmness, tenor, venue

4 · Delivery — where paper becomes racks

A cash-settled contract is a bet on an index. A contract that converts into capacity is forward procurement, and the difference decides whether hedgers show up at all.

Every Seat but the Dispatcher

An inventory of 64 firms and standards in six seats — brokers, desks that take the other side, order books and marketplaces, routers and token venues, prices and contracts, plumbing — mapped by who closes the deal and how long it runs. Software closes deals by the request and by the hour; people close term deals. The seat nobody holds is the dispatcher, which decides where a workload runs and under what contract. Eight frictions ranked, five conditions for a market with few frictions, what a dispatcher does to chip-loan tranches, and who gains

The Venues and the Tranches

Eleven venues and instruments for compute mapped on two axes, tenor and settlement form: SF Compute’s minute-resolution order book, Computable’s calendar weeks, Hyperbolic’s principal book, Compute Exchange’s reserved and hardware markets, Andromeda’s certified contracts, Kalshi on Ornn, the pending futures, rated neocloud loans, onchain loans, Nvidia’s rent-back and residual value cover. What each solves for a cluster owner and the two exposures none covers — listing its own cluster, and a price at the one-to-three-month tenor — because no public block standard exists. The four-part combination that would close the gap. And the financing finding: seven rated chip loans already sort into three tiers by the customer behind the contract, from SOFR plus 2.25 to nine percent, so compute will tranche by counterparty, tenor and basis before order of loss

When the Building Outlives the Chip

A data centre is three assets on three clocks — the building at fifteen to forty years, the chip’s accounting life at four to six, its frontier life at one to two — and only the slowest is financed like infrastructure. The last cycle bridged the gap with the neocloud balance sheet; the next is being bridged by sovereign programs, frontier labs and enterprises on open-weight models, each owning a cluster it did not size to rent out. What that does to market structure: the traded instrument becomes a one-to-three-month block, basis becomes the pricing problem, the lender’s collateral becomes a chip in someone else’s building, and someone has to hold the block. Six services sequenced against the owner’s need, and a readiness map showing the market has built the top and bottom of the stack and skipped the middle

The Reserved Contract, Unbundled

Compute Desk's Nodal future and the ComputeConnect exchange-for-physical, from the seller's chair: a neocloud with 1,024 uncontracted B200s walked through the reserved contract, the cleared strip and the EFP, ten classes of buyer, and Nodal Clear's margin and waterfall as the thing that replaces a $104 billion bilateral backlog. With an interactive ledger — the strip matches the reserved contract at about 71% of hours sold at market, and a 30% rally is a $12 million cash call

Also the credit layer
Where the Risk Goes

The market-maker's question rather than the product question: after every announced instrument does its job, what exposure is left and who holds it. Eight risks with natural sellers, ten structures ordered by what exists, and five residuals no instrument absorbs — the tenor tail, basis, the jump wing, embedded utilization and correlated credit. Includes the SA-CCR calibration nobody has raised, and the seven observable spreads that price an inventory which cannot be hedged outright

A Signal Is Not a Curve

Compute has prices worth watching, not prices you can lean on, size into and exit from. A five-part test for tradability — fungibility, two-way size, transferability, settlement integrity, credit intermediation — run against the exhibits in Liquid Compute's primer, and the finding that the one genuinely executable trade in it is executable because it needs racks and a sales team rather than financial infrastructure. Plus the $5.15 tenor discount decomposed into five strips, only the last of which, about $0.35, any instrument exists to isolate

Paper becomes racks — the first compute EFP

Architect × Compute Desk's exchange-for-physical network read through the crude, gold, gas and metals precedents, with basis tables and convergence mechanics

Two venues, one index

Kalshi and Polymarket settle on the same Ornn URL — the first live compute spread with the index basis verified at zero, and still unreadable. Disclosure is necessary and visibly not sufficient

Also in Predictive Markets

5 · The product leg — the tokens compute actually produces

GPU-hours are the input. Tokens are the output, and they have their own price, their own volatility and — so far — no settlement-grade index. This is the newest thread and the one with the most open ground.

Give Us This Day Our Daily FLOPs

Where the Flop Network actually sits, read from a v0.1 teaser and a manifesto rather than from the token. Every benchmark and contract announced this year settles on the rented hour; Flop prices the delivered floating-point operation instead, which puts it one node to the right of all of them and makes its competitors the inference routers rather than the index providers. A work market is not a capacity market: all the utilisation risk lands on the miner. Against it — the hardware guidance attracts the depreciated tail, a FLOP is not a FLOP once decode is memory-bound, and a subsidised price in a floating numeraire is not a signal

The Other Side of the Crack Spread

The demand-side companion. Every announced compute derivative settles on rent, but the buyer whose bill justifies the market pays for tokens, not GPU-hours — so it has no r to hedge. Its spend is S = p·h·N against the producer's C = T·p − r, which inverts the sign on price and introduces a heat rate set by prompt design rather than physics. Price is the small term and falling; quantity is the risk and rising, correlated across every buyer on the day a model generation ships

The token leg has a price screen. Nobody has built the index.

A designed-but-unsited token futures contract tested against the venue that could settle it: 56 live quotes, a 7× spread on identical open weights, and the four places the index breaks

Hedging the inference book

Fixed subscriptions downstream, floating token costs upstream — a commodity retailer's book. Token forwards versus GPU futures, with the efficiency curve as the unhedgeable basis between the units

6 · The credit layer — who finances the racks

Compute became an asset class the moment lenders started underwriting it — and in August 2026 the largest alternative-asset managers in the world agreed it was underwritable collateral. This is the applied end of everything above: the index work becomes a validation program, and the measurement work becomes a collateral curve.

The Refiner and the Merchant

Two ways to pool compute, and the listed chip-hour futures fit only one. The refiner sells tokens against a crack spread whose efficiency wedge the contracts cannot span — about 3% explanatory power on the Federal Reserve paper’s token-hedge result. The merchant buys blocks on commitment and resells by the hour, breaking even near 74% fill at $2.50 in and $3.40 out, and is the natural seller of contracts designed to collapse its own spread. Includes the regulatory paradox: the opacity behind the CFTC’s September 21 delay to November 9 is what makes the merchant’s margin, and the 2021 electricity retailers as the precedent

Also the credit layer
The Customer Is the Collateral

Where compute credit actually stands: $48B of chip loans, $155B of landlord vehicles and $61B of building-backed bonds, with a rated ledger showing that the same borrower on the same hardware pays 2.25 points over the benchmark behind a Meta contract and 5.50 when the loan outlasts its contracts. The counterparty opens the grade and the structure sets the notch. Plus the Federal Reserve paper on the multi-borrower pool nobody has issued, which finds the senior slice is the piece most understated when recovery falls in the same conditions that cause default — and the four things missing before a pool is buildable: no public default, no recovery print, no hedge requirement, no loan-to-value test against an index

The contest can't be hedged. The financing can.

The BIS models the build-out as a contest rather than a mania — so the racers won't hedge, the demand migrates to whoever financed them, and it lands in a market roughly 120× too small

Compute became collateral. NVIDIA sold the floor.

Six MOUs to mobilize over $500B into vehicles issuing debt secured by GPU compute — and the only quantified NVIDIA exposure in the announcement is a written put on the depreciation of its own hardware

Hedging the GPU balance sheet

Settlement mechanism, standardization tier and index validation for specialty compute lenders — five checks that require no administrator cooperation, with the power leg grounded in ERCOT and Dominion

Volta absorbs, Trillium distributes

Two structures for the unrated side of the market, running the same pipe in opposite directions — exposure ladders on both sides and the hedge program each would need

7 · Synthesis — driving the whole complex

The workbench that connects every layer above: chips to megawatts to basis, with the options, margin and take-or-pay economics attached.

Where this connects

The threads that run out of this one.