The Compute Desk — Novation White Paper 01

Large buyers of electricity, gas, or jet fuel employ someone whose job is deciding when to buy it and from whom, and runs a function whose job is the contracts: what was committed, what was delivered, and what was paid. Buyers of AI compute, now one of the largest lines in a corporate technology budget, mostly have neither. They sign a contract, run their work where they have always run it, hold the commitment in a footnote, and find out what it cost when the invoice arrives.

This paper is about the two jobs that are missing: what a compute desk does and what the record underneath it holds, why nobody has either, what they are worth, and where they go once a buyer has them. It starts with a single purchase, because the theory is only worth reading if the purchase is worth improving.

00 · One purchase, as it happens today

A company needs 9,450 chip-hours of H100 capacity in an eight-way contiguous topology, within fourteen days, for a training campaign it can checkpoint every thirty minutes. It has one supplier, on a contract signed last year. The engineer who owns the run asks the incumbent for a quote, gets a rate card, and books it. Nobody asks anyone else. Nobody writes down what the alternative would have cost. Nobody checks whether the hours should have retired the commitment signed last year, or whether they ran above it. The invoice arrives, and it is what it is.

Here is the same purchase with somebody whose job it is.

Three things in that picture are the whole argument. The cheapest rate card lost the piece that mattered: Supplier A's interruptible $1.94 delivers at $2.40 once transfer, setup, expected reruns, and the wait for a window are priced, and the reruns carry an eight-way training run past the deadline, so it is rejected for training and used for the staging and evaluation pieces that can take interruption. The dearer card won: Supplier B's $2.24 delivers at $2.40 and holds the deadline. And the two savings figures are kept apart on purpose, with the second measured against a route that works. Against the incumbent, the campaign was leaving $6,143 on the table. Against the best feasible route the buyer could have reached alone — Supplier A's reserved tier at $2.50 delivered, which also holds the deadline, had anyone asked — the desk added $945. A route that misses the deadline is a rejected option, never a baseline.

The second number is the one a fee has to be judged against, and it is small. On this campaign, the value of a desk is mostly in a question the buyer never asked, and only a little in the answer. The ratio is this campaign's, not the market's. Whether real purchases look like this one is the question this paper cannot answer and Novation's first year is designed to.

01 · What a desk is, what a record is, and why compute has neither

A desk is a person, or a small group, whose job is a recurring buying decision: how much, from whom, for how long, in what shape, and at what risk. Load-serving utilities run one for electricity. Airlines run one for jet fuel. Every large industrial runs one for gas. They exist because the spend became large enough that somebody had to own the decision.

A record is the function beside the desk whose job is the contracts: every commitment with its floor, its term, and its expiry; every purchase authorized before it is made; every invoice matched to the contract it was billed under; and the position — what is committed, what is forecast, and what has been consumed — stated as one number, built for a controller to close the month from. The large corporate buyers of volatile inputs we have studied — electricity, gas, jet fuel — each run one, and it grew in a particular order: the bill and the contracts first, then measurement of what was delivered, then advice on the next purchase, and only after years of all three, execution in the company's name.

Compute has neither, and the reason is not that nobody noticed. The spend arrived faster than the function. Ninety-eight percent of FinOps practices now manage AI spend; two years ago it was 31%. The discipline names AI cost management as its single largest skills gap. Cost-management teams stay at eight to ten practitioners plus contractors even above $100 million of cloud spend. And the role is being created in front of us: Anthropic staffs a compute finance team that owns the cost-allocation methodology beside an engineering head of compute; OpenAI posts a Director, Compute and Infrastructure FP&A, and a senior manager of accounts payable for compute whose job is to translate contract terms into invoice validation and payment rules; CVS Health and a national mortgage lender have posted AI FinOps leads with budgets, chargeback, and commitment management in the description. At most buyers the job is one line in a head of infrastructure's job description, next to the scheduler, and the contracts are a tab in a spreadsheet.

The contracts are already large enough to need an owner. In their most recent filings, non-lab public companies disclosed multi-year cloud commitments of about $16.7 billion (Salesforce, contractual commitments including infrastructure providers), $7.7 billion through fiscal 2032 (Palo Alto Networks, cloud), $6.8 billion (Adobe), a $6.0 billion cumulative minimum through March 2031 (Snowflake, amended April 2026), at least $5.6 billion through February 2036 (Palantir, amended March 2026), $5.3 billion (DoorDash, from $1.07 billion), about $4.8 billion on cloud services through 2030 (ServiceNow, from $805 million a year earlier), $4.0 billion through May 2031 (Pinterest, signed May 2026), and $3.7 billion (Atlassian, doubled in a year). Roughly fifteen such companies disclose $1 billion or more and about forty $100 million or more. Where the mechanics are stated, the contract is a dollar minimum with a shortfall true-up — Snowflake is “required to pay the difference,” MongoDB must “pay for such capacity irrespective of actual usage” — and Tempus AI carries a $39.5 million liability for an Azure commitment “in excess of forecasted usage,” assumed with its Paige acquisition. Snowflake reports third-party cloud costs, “including those related to AI inference and graphics processing units,” at about 75% of its cost of product revenue in the quarter to 31 July 2026, up from 70% a year earlier. These are finance objects, and finance holds them in a footnote.

The decision is also a committee's, and that is why it falls between chairs. Infrastructure owns the scheduler. Finance owns the budget, the contracts, and the baseline. Procurement owns the terms. Security approves the suppliers and the data access. A product owner needs the work done by Thursday. The purchase is made the way it was made last year, because that is the one option nobody has to approve, and the commitment is renewed at whatever the incumbent quotes, because nobody holds the drawdown.

The energy analogy explains the two jobs. It does not establish that compute must grow the same market institutions as electricity, or that an independent company captures the value. A refinery needs procurement expertise and may supply it internally, buy it as advice, or lean on an intermediary. The desk this paper describes is the second of those. It keeps the order in which every other desk grew and changes only the cadence: the record is the bill and the contracts, held continuously; advice and measurement arrive every day rather than once a quarter; and execution is earned by that record rather than sold ahead of it.

Where it goes is further than a purchase. A company that rents compute at scale will run it the way a company that burns jet fuel already does: term contracts for the base load, financial contracts for the price risk, and a desk for everything that cannot wait. One system connects the teams who create the demand, the finance function that pays for it and administers the contracts, and the risk function that watches the exposure. It is to a company's compute what an ERP system is to its money, and what an energy trading and risk management system is to a utility's power book. Section 09 describes it in layers. Section 05 describes the two layers built today: the record, which enters through what finance already holds, and the desk underneath it.

02 · Seven jobs nobody at the buyer holds all at once

See what they actually pay. Almost no buyer can state its blended rate per GPU-hour without a week of work, because the number is spread across committed discounts, on-demand overflow, interruptible capacity, and an enterprise agreement that bundles compute with storage, networking, and support. Three professional index families now publish compute prices and disagree with each other, because they measure different chips, regions, configurations, and contract shapes under the same name. One of them publishes the size of the problem: on 8 September 2026 Compute Desk's Hopper index printed $2.66 blended against $2.97 listed, and Blackwell $4.37 against $5.30 — advertised premiums of 11.7% and 21.3%, from a benchmark with no stake in this argument.

Hold every commitment in one place. The commitment is signed once a year against a forecast nobody owns, and then it lives in a spreadsheet, a footnote, and the inboxes of the people who signed it, with its floor, its ramp, its expiry, its prepayment, its credits, and its side letters in different places. Too little, and the overflow lands at on-demand rates. Too much, and the unused portion is a write-off at a hundred cents: AWS's Reserved Instance Marketplace excludes Convertible RIs and Savings Plans and caps lifetime resale at $50,000, and all four major providers forfeit unused term-committed inference capacity. Nobody sees the drawdown pace until the term is nearly over. The sell side has organized itself around getting that signature early and large: CoreWeave described “a typical five year contract,” reported about $104 billion of backlog at 30 June 2026, and says its committed contracts are take-or-pay; Nebius said roughly 70% of the deals it closed in the same quarter carried customer prepayments. And a commitment is a floor, not a ceiling. Google Cloud's chief executive was reported in July 2026 as saying its customers run about 50% above their commitments, and that Google rents capacity from CoreWeave and Nebius to bridge them. When a deadline-bound job arrives above the floor — an unplanned run, an evaluation before a launch — the only channel open inside a week is the incumbent's on-demand rate, unless somebody qualified the alternatives months ago.

Authorize compute before it is spent. Project budgets are set once a year and exhausted in a quarter. In a February 2026 survey of 500 finance leaders at companies with a thousand or more employees, 79% reported AI cost overruns in the prior year, and the shops with the most mature cost management overran most; in a mid-2026 survey of 700 engineering leaders at companies of the same size, 72% had an unexpected AI cost spike, 52% said there was no clear cost owner, and 56% said AI budgeting was guesswork. Both are vendor surveys. What they describe is a budget that nothing reserves when the work is approved and nothing debits when it runs, so the first notice is the invoice.

Compare two suppliers on like terms. The advertised rate omits moving the data, standing the environment up, work rerun after an interruption, and what the hardware delivers. On the H100 PCIe sample in one benchmark study, inter-provider FP16 throughput ranged 34.5%; on the SXM sample, 10.4%; and device throughput predicted application throughput weakly. The study (Slynko et al., “Did You Win the GPU Cloud Lottery?”, GPGPU '26: eleven providers, 3,534 GPU instances over four months, by authors at Silicon Data and William & Mary) argues for one thing: measure the buyer's own work on the actual allocation before committing to it. It does not license a fleet-wide estimate, and this paper does not make one.

Look anywhere but the incumbent. Most companies renting AI compute rent it from one vendor on a commitment signed before anyone knew what the year's work would look like. Qualifying a second supplier means a security review, a contract, an acceptance test, and scheduler integration — weeks of specialist work nobody is free to spend. So the buyer's whole price history is one seller's list, and in 2026 one supplier can mean no supplier: one Azure customer waited ninety days across five regions for any GPU quota at all. On 4 September 2026 the first alternative to a hyperscaler was posted between 1.12× and 1.72× below it.

Say which of their work could move, and act at the speed prices move. Deadline tolerance is not in the job logs, and neither is interruption tolerance. Both live in engineers' heads. In one survey of 103 practitioners, 96% had submitted work they described as non-urgent and only 3% required zero delay; the same survey found roughly 80% of users mispredict their own job completion time by 10% or more. The flexibility is real and the self-report is unreliable, so somebody has to ask — and count by spend, not by job, because Section 03 shows the two answers differ by an order of magnitude. Meanwhile the prices move inside a year: AWS raised Capacity Block reservation prices three times in 2026, about 15% in January, 20% in July, and 15% in October, and CoreWeave raised prices about 25% in July. Requoting five suppliers is a week of somebody's month. The mandate to manage AI spend arrived; the capacity did not.

Reconcile the invoice, and prove any decision was right. Invoices arrive from different suppliers in different units, and the largest GPU cloud's billing export carries usage quantities and no cost fields; OpenAI is hiring an accounts-payable manager for compute to translate contract terms into invoice validation, purchase-order, approval, and payment rules. And nobody wrote the baseline down before the decision, so there is nothing to reconcile the invoice against, and cost conversations in this sector collapse into assertion. Flexera's 2026 estimate of wasted cloud spend is 29%, up for the first time in five years — a respondent self-estimate, and Flexera's figure rather than the FinOps Foundation's.

Three constraints, and a fourth that has arrived

Every purchase is decided against three things the buyer cannot trade away: timing (deadlines, interruption tolerance, how long work can wait), risk (a supplier that may not survive the quarter, data residency, security review, concentration), and the market (what is purchasable today, from whom, in what form). Getting one right is easy. Getting all three right, for every purchase, across several suppliers, is a job.

The third constraint has tightened. Posted prices fell through 2024 and 2025 and turned up from late 2025. Compute Exchange, a request-for-quote marketplace across a hundred providers, says it cannot fill about 60% of its request volume for want of inventory. Bandi and Su's 2026 study of compute futures finds positive risk premia — the sellers are the ones hedging. A desk that finds cheaper capacity is worth most when capacity is loose. In a tight market, whether a qualified supplier will quote firm capacity in the buyer's window is the first question, and an unfilled request is an outcome, not a footnote.

03 · What can actually move

A desk is only worth having if some of the work is genuinely portable. It is, along four seams that are measurable rather than rhetorical. It is also less portable than the systems literature implies, and the narrower boundary is the more useful one.

First, clear one number out of the way. The claim that GPUs run at about 15% utilization appears in nearly every pitch in this sector, including earlier versions of ours. It traces to a sponsored vendor post whose own wording is “based on experience with POCs and new installations”: no sample size, no methodology, no measurement window, and a sample self-selected for already having a utilization problem. Worse, three different quantities travel under the name “utilization,” and conflating them is how this literature gets miscited.

What survives verification is narrower and more useful. Allocation is high, efficiency is not, and demand is granular while allocation is lumpy. Alibaba's production trace finds a median instance request of 0.5 GPUs against median usage of 0.042, roughly a twelvefold gap between the GPU bought and the GPU used. Google measured its own indivisibility cost directly: fixed-size machine types would need 30–50% more resources than requested shapes in the median case. That is a market-design problem rather than a diligence problem. The Alibaba trace is from mid-2020, on a pre-LLM recommender and ads fleet, so it is evidence about lumpiness rather than about a current frontier cluster.

By phase. Large-language-model inference splits into prefill and decode, and the two have genuinely different hardware appetites. Prefill is compute-bound; decode is memory-bandwidth-bound. On Llama-2-7B the arithmetic intensity of the projection layers differs between them by a factor of 1,024. Splitwise, running the two phases on separate hardware pools, reports 1.4× more throughput at 20% lower cost, or 2.35× at the same cost and power. Those cluster-scale figures come from a simulator validated to under 3% error rather than from a deployed cluster; only the cache transfer was measured on hardware.

By hardware. Accelerator choice is not a single ranking. On the Gavel throughput matrix, moving from a K80 to a V100 buys ResNet-50 a 9.6× speedup and a deep-RL workload only 2.2×, a 4.4× spread in how much different models value the same upgrade. A Transformer gains nothing at all from P100 to V100 while ResNet-50 gains 2.6×. Matching work to silicon is worth real money, and it is exactly the calculation a merit order performs in power: cheapest first, subject to what each unit can actually do.

By deadline. Users report tolerance directly, in the survey figures above. One workload taxonomy splits urgent, priority, and normal work 5 : 35 : 60.

By interruptibility. Every measured interruptible discount in this literature is 70% or better against on-demand, and systems built for preemption demonstrate that training survives it at measured cost. SkyPilot's brokered plan across regions and clouds beat single-region interruptible capacity on both axes at once: $18.40 against $21.80, and 21 hours against 34.

And the constraint that sets the ceiling

Between 2020 and 2023 the share of GPU-time held by large all-or-nothing jobs, the kind that need every GPU at once or none, rose from roughly 60% to 94–96%, while the single-GPU population's share fell from 68% to under 2%. In one 2023 trace, pretraining is 3.2% of jobs and 94% of GPU-time, while the easily divisible evaluation work is 92.9% of jobs and 0.8% of GPU-time. The divisible work is numerous and small. The GPU-time is concentrated and indivisible.

Three further constraints are load-bearing and each is measured. All-or-nothing scheduling is fundamental rather than a preference, and in the Alibaba trace 85% of instances carry that requirement, accounting for 79% of total GPU demand. Keeping a job's GPUs close together is worth 22–35% of throughput: the same four GPUs deliver 103.6 images per second on one server and 80.4 split across two. And moving a running job costs tens to hundreds of seconds rather than milliseconds.

Two conclusions follow, and they point in opposite directions on purpose. A market that requires most work to be finely divisible is betting against a clear trend. A desk that prices the indivisible block correctly, and only moves what can move, is not. That is why we name the movable share before we quote a saving, and why a desk that reports a small movable share has done its job rather than failed a sale.

04 · What the market already tells you

The strongest evidence that timing has value in compute is not a forecast. The industry already sells the product, at an administered price, with nowhere to trade it.

Nine competing firms independently arriving at exactly −50% for a 24-hour deferral is an administered price, in the way a regulated tariff is, and not a market clearing. Two limits travel with it. These are token prices on managed inference, not GPU rental rates, so they establish that buyers value timing flexibility and nothing about the discount a GPU buyer can obtain for deferring a training job; that discount is measured per workload, in a quote round. And identical posted discounts do not prove that competition could not have produced them. Latency is priced in both directions too: on identical Claude Opus 5 weights, input runs $2.50 per million tokens on the batch tier, $5.00 standard, and $10.00 in fast mode, a 4× spread on the same weights determined entirely by when you are willing to receive the answer. Whether buyers value deferral is settled. What the price would be if anyone ran an auction is not.

Prices are far apart, and power does not explain it

On a single day in one marketplace's live order book, 39 offers for H100 SXM on-demand capacity ranged from $1.34 to $6.18 per GPU-hour, a 4.62× spread. Across providers the range runs from $1.73 to $12.29.

It is tempting to attribute locational price differences in compute to electricity, the way a power trader would. The arithmetic does not support it. The full 4.49× spread in lower-48 industrial power rates moves a GPU-hour by $0.226, about 6.8% of the $3.31 median. It cannot explain a 4.6× spread inside one marketplace, a 3× hyperscaler premium, or a 1.68× spread between two AWS regions on an identical SKU. Locational pricing in compute is real, but its drivers are interconnect, availability, contract structure, and hardware quality. Claiming that power explains it is the fastest way to lose a technical reader.

You cannot get out of a position

AWS's Reserved Instance Marketplace excludes Convertible RIs and Savings Plans entirely and caps lifetime resale at $50,000. Azure's own documentation warns that reservations do not guarantee capacity. All four major providers sell term-committed inference capacity in which unused capacity is simply forfeited. A buyer who over-commits to a hyperscaler has no exit, and a buyer who over-reserves GPU capacity has only a thin secondary market — a few marketplaces where a reservation can be resold, at whatever the market will pay that week — which makes commitment a one-way decision at exactly the moment the technology is changing fastest. This is the most under-priced risk in compute procurement and almost nobody names it, and it is why the record starts with the commitments rather than with the purchases.

Meanwhile the mismatch sits on the supply side's balance sheet, in public filings. CoreWeave's Q2 2026 Form 10-Q shows a weighted-average remaining lease term of 12 years against a revenue book of which 41% is recognized within 24 months and 80% within 48, with 98% of revenue from committed contracts and 72% from three customers. That is a textbook hedging need: a long asset, a short book, concentrated counterparties, no instrument. It is also hedging pressure from sellers, which is what the first academic asset-pricing work on compute finds. Bandi and Su construct a compute futures return panel from term rental contracts and report preliminary evidence of a positive risk premium, implying the pressure comes from providers rather than buyers.

The instruments arrived before the cash market

Six exchanges or platforms have announced listed compute contracts. One is live, and no listed, cleared compute futures contract has traded on any designated contract market.

The only trading actually happening in listed compute price risk is on a binary event exchange, in monthly strike prices, at a few hundred thousand dollars of open interest. Everything else awaits regulatory approval. The CME contracts did not list on 5 October 2026 as planned: the CFTC extended its review to 9 November, citing “novel or complex issues,” having written in its August request for comment that it “preliminarily believes that compute may not yet exhibit certain of the characteristics of commodities” that underlie a derivatives market — “fungibility, standardization, and sufficient liquidity.” That request closes on 20 October. The accounting treatment of a compute hedge — whether a cash-settled contract on a rental index qualifies as a hedge of a physical commitment — is unresolved, and a finance organization cannot book what its auditors have not ruled on.

The three index families behind those contracts are routinely described as interchangeable, and they are not. Compute Desk has two exchange relationships in three months — ComputeConnect with Architect Financial Technologies on 8 July 2026, the first US exchange-for-physical network for GPU capacity, and financially settled futures with Nodal Exchange announced 3 September — and it operates the physical settlement layer underneath them. Nodal's release says Compute Desk's backers include “the CFOs of hyperscaler companies,” a claim this paper attributes to that release and has not verified. Silicon Data has the CME listing, now under extended review. Ornn has Bloomberg distribution, one announced ICE relationship with no launch date, and the live Kalshi dependency. They are competing on distribution rather than on methodology, which is a different race from the one the sector describes.

And that sequencing is itself the risk. Brorsen and Fofana's study of agricultural futures found that an active cash market is necessary for contract success, a single variable that perfectly predicted which commodities sustained futures markets. Compute has built the derivatives layer first, on three competing indexes, and neither Ornn nor Silicon Data publishes a full methodology. The benchmark layer is the least verified part of the entire stack, and it is the part everything else settles against.

05 · The record, and the desk watching

What follows is the record, and then the engine stage by stage, with the part of the buyer's problem each answers. It enters a company read-only, and it enters through finance. Given read access to the bill, the contracts, and the budgets — the things finance already holds — the record prints a first position within two weeks: every commitment it has been given, every project budget, every invoice it can match, with its coverage stated as a share of spend and what it could not yet match listed beside it. Given, when the buyer is ready, read access to the scheduler's job logs, an hour with the engineers, and an approved list of suppliers, the desk underneath it watches the work that runs and prints, every day, what it would have done and what that would have cost, next to what the buyer actually paid. It has no write access to anything. Where the buyer requires it, the record runs in the buyer's own tenancy: control plane only, and the data never leaves. Two stages, bidding and dispatch, are recorded rather than placed while it watches. They write down what would have been bought and what would have moved, and touch nothing. A ninth stage, settlement, is not offered at all.

5.0 · The record, from the second week

The order matters, and it is the order in Section 01. The record comes before the desk because the contracts are where the money already is, because finance can grant access to its own bill and its own order forms without asking infrastructure for anything, and because a record that holds the commitments is useful on a day the desk finds nothing to move. It holds six things.

It is built for a controller to close the month from, for a treasurer to see the prepayments, which are unsecured credit from the buyer to a supplier, and for a project owner to see what is left. The budget mechanism is the one borrowed most directly from how a finance organization already works: a request is authorized against a named budget before the work runs, the amount is reserved at authorization and debited at the invoice, and a request that would breach the budget stops and waits for a named person. None of it needs a job log. All of it is on the screen in the second week, and every number says where it came from and what changed.

The seventh object in the record is the desk's, and it arrives with the job logs: for every piece of work that ran, what the buyer paid, what the desk would have bid and where, what that would have cost delivered, and the difference, weighted by spend and summed since the watch began. The stages that produce it follow.

5.1 · Characterize: which work can move, by spend

Somebody asks the engineers, job family by job family, what can wait and what can be interrupted, and weights the answer by GPU-time rather than by job count. Section 03 explains why the two weightings differ by an order of magnitude and why the spend-weighted answer is the only one that matters. The job logs say what ran; only the engineers know what could have run elsewhere, and the desk writes their answer down once and then holds it against every job it sees. A small movable share is not a failed sale. It tells the buyer to negotiate the commitment rather than shop the overflow, and the desk says so.

5.2 · Slice: the requirement, written for every piece before any price

A training campaign is not one purchase. It is staging, then training, then evaluation, and each needs a different machine for a different window with a different tolerance for interruption. Section 03 found four seams along which work separates: phase, hardware, deadline, and interruptibility. The desk cuts every job family along them into pieces that can be priced apart, and for each piece writes the requirement first: chip class, quantity, topology, region, start window, deadline, checkpoint tolerance, where the data sits, and who can say yes. It is written before a single price is looked at, so the comparison cannot be refitted to whatever comes back. This is the cheapest stage and the one that most changes the answer, because it converts "get me some H100s" into a product that can be quoted.

5.3 · Price: every qualified supplier at delivered cost

The advertised rate is an input to the price, not the price. The delivered cost of an hour at a given supplier is the rate, plus data transfer at the buyer's own negotiated egress rate, plus setup, plus expected rerun after interruption, plus the cost of waiting for a window, adjusted for what the hardware delivers against its nameplate. What it is compared against is not the incumbent's rate card either but the buyer's marginal cash cost of staying put: what the hour actually adds to the bill under the contracts as they stand, given the commitment it draws down or fails to, the credits and tiers the agreement applies to uncovered usage — at AWS, uncovered usage is priced at on-demand rates after the organization's own discounts, not at list — the other services consuming the same minimum, and the term remaining. A recommendation that lowers the GPU rate and raises the total bill is the failure this comparison exists to catch. What is being priced is a chip class, in a region, for a delivery window: an H100 in US-East on Tuesday is a different product from an H100 in US-West on Friday, for the same reason a megawatt-hour at two points on a grid is two products — moving between them costs real money. Compute has no physical network model to take a shadow price on, so what a desk produces is a set of independent local prices rather than one decomposable price, and this paper claims nothing more.

Comparable offers require a comparable request. The same specification goes to every supplier the buyer would accept, with seven terms the supplier can see before it decides to answer: a named buyer with a confirmed window and the authority to solicit a quote; one standardized request; a bounded response window and a quote expiry, with no use of the quote to pressure another seller; an agreed test on the actual allocation, in a reserved window, with a named payer; an objective acceptance condition with a remediation period and a stated consequence; confidentiality on the negotiated price and the test result; and a responsibility matrix for provisioning, outages, cancellation, invoicing, and escalation. Suppliers cooperate with that because it reduces their sales effort on a funded buyer. They do not cooperate with a fishing expedition, and they are right not to. Where a marketplace is the route, its terms govern; SF Compute's forbid sharing account access with a third party without written authorization, and a desk assumes no integration right a platform has not granted in writing.

While the desk is watching, most of its prices are asks. Three things go into the delivered-cost map: posted rates across the qualified panel, the buyer's own purchase history, and any standing quote a supplier will give against a named buyer's demand profile. Every number carries its source. A supplier will not quote firm capacity into a request nobody will fund, so the first live quote round runs on the first purchase the buyer makes during the watch, whoever the buyer ends up buying from. From that day the record carries real offers beside the asks, and the gap between the two is itself a finding. Every offer is marked for what it is: a firm offer carries a quantity and a hold time and can be chosen; an indicative price is information, and in a tight market a supplier that quotes indicatively without holding inventory is the case the fill-rate test in Section 08 counts separately.

5.4 · Forecast: this week's price, and how much to commit

Nobody publishes a forward price for a specific buyer's fleet. Two index administrators publish a forward curve for a chip class, and the two disagree. The desk builds the buyer's own forward view from that buyer's purchase history and the quotes and asks just collected, on two horizons. The short one is this week and next, for each piece's class of compute, and it decides where the piece should run. The long one is next quarter, and it serves the decision that matters most and is made least often: how much to commit for how long. Every point of the bill committed and unused on a $60 million fleet is $600,000; every point of unplanned overflow is bought at on-demand rates. The same forecast, run against the floors, the overflow, and the project budgets in the record, produces the budget-at-risk: the probability of breaching the enterprise budget or any project's, and what drives it. The forecast does not need to be right about the market. It needs to be right about the buyer's own demand, and that is a question the buyer's engineers can answer if somebody asks.

5.5 · Schedule: where and when each piece should run

For every piece, the desk answers one question against three constraints: where and when should this run, given the deadline, the risk limit, and the project budget it was authorized against. Retiring the commitment first is the usual answer, because a dollar of committed capacity left unused costs a hundred cents — but it is a heuristic the solver checks against the contract, not a rule it applies: whether this hour is eligible against that commitment, whether another commitment overlaps it, what credits and tiers apply, what other services are drawing on the same minimum, and how much term remains all change the answer, and the marginal cash cost in the record is where they are priced. What remains is placed where it costs least delivered, inside the buyer's concentration cap per supplier, inside the regions its data may enter, and in an order that respects the work: staging before training, training before evaluation, with the overlaps the engineers allow. The limits — the budget, the approval threshold, the concentration cap, the approved panel, the regions — are not constraints the desk trades against. They are the box finance draws around it. During the watch this is a recommendation, printed against what actually happened. When the desk acts, it is the plan.

5.6 · Bid and dispatch, recorded rather than placed

Bidding puts each piece to the qualified suppliers and takes the best firm offer that meets the constraints. Dispatch sends the work through the buyer's own scheduler or router — SkyPilot, Kubernetes, Slurm, Ray — and confirms it ran; the desk tells the router where, and never replaces it. Where a supplier's terms bar orders placed through a third party, the desk prepares the order and the buyer submits it under its own identity: AWS's service terms, as updated on 1 October 2026, bar submitting Capacity Block requests “through any third party (e.g., 'proxy purchasing'),” and an API is not a permission. During the watch, neither stage places anything. Against every job that ran, the record shows what the desk would have bid, to whom, and when, and what would have moved. On the days it would have left the work where it was, it says so. On the days no qualified supplier would have quoted inside the window, it says that too, and counts it, because in a tight market an unfilled request is the finding. Nothing changes until the buyer approves it, and Section 09 says what has to be in writing before it does.

5.7 · Verify: test the allocation, then reconcile every invoice

Before the full spend is committed, the chosen allocation runs the buyer's own workload, or a representative slice of it, under an acceptance condition agreed with the supplier in advance. Silicon Data sells a benchmark suite for this; where it fits the workload, the desk uses it rather than building its own, because the value is in the independent selection and interpretation of the test, not in owning the tool. A failed test has an agreed consequence — repair, reallocation, cancellation on agreed terms, or a buyer-approved exception — because a supplier sells hardware and uptime, not a warranty on the buyer's training throughput, and no report can convert a shortfall into a refund the supplier did not sign.

After every purchase, the invoice is reconciled twice: to the contract it was billed under, in the record, and to the baseline written at stage 5.3. The FinOps Foundation's FOCUS 1.4 specification, published in June 2026, now carries invoice-detail and billing-period datasets built for exactly this; FOCUS 1.5, scheduled for ratification on 3 December 2026, adds model and token columns and a SKU price dataset, and no GPU-specific fields, so the GPU columns remain the record's own. The reconciliation is becoming a data import. What it reconciles against — a contract held in one place, and a baseline written before the decision by a party paid only by the buyer — is not.

5.8 · Settle: last, and not offered

Paying suppliers on measured delivery is the ninth stage, and the one this paper describes without offering. It is the only stage that would put the desk in the payment path, and Section 09 says what has to exist before it does: the desk has to have acted for real, and the credit and collateral machinery has to be built first. Until then the buyer pays its suppliers, and the record reconciles.

5.9 · Qualify: the four questions underneath every price

Every price above comes from a supplier that has answered four questions.

Identity: is this who they say they are. Technical: can they deliver this topology, in this region, in this window. Security: does the buyer's own review pass. Financial: will they exist at the end of the contract. The four are assessed separately because a supplier can pass three and fail one, and because private suppliers will not disclose what the fourth needs. Where disclosure is limited, the record states a confidence level and caps the exposure; it does not print a reassuring score. CoreWeave's 10-Q for the quarter ended 30 June 2026 shows 72% of revenue from three customers against a twelve-year weighted lease term, total indebtedness of $35.6 billion by its own statement, $1.18 billion of interest expense for the half-year, $5.5 billion of cash, and $103.7 billion of remaining performance obligations with 41% due within 24 months. That is a supplier a buyer can carry, with the exposure known — and a prepayment of 15–25% of contract value, CoreWeave’s own disclosed average across its active contracts, is credit to the supplier, unsecured unless the contract says otherwise, which the record shows the treasurer before the placement, not after.

A fifth question sits underneath the four, and it moves. Export controls on advanced chips now reach toward remote access: the Remote Access Security Act passed the House 369–22 in January 2026 and sits in the Senate, not yet law, while the Bureau of Industry and Security enforces the existing rules — a $252 million settlement in early 2026. The qualification file screens the counterparty and its owners against the restricted-party lists, records where the hardware sits, and asks the supplier what it does about its other tenants; the record tracks the rule as it moves and re-opens the screen when the lists or the law change. It is a tracked obligation, stated at its size, not a catastrophe and not an afterthought.

5.10 · What the record prints every day

The deliverable is the record, kept continuously. From the second week, with its coverage stated as a share of spend and its unresolved exceptions listed first: the position — committed, forecast, and consumed, with what is above the floor and what it cost; every commitment's drawdown pace and what expires when; every project budget, authorized, reserved, consumed, and remaining; every invoice reconciled, with the exceptions first; the budget-at-risk; and the supplier panel, with who is qualified, what they quote, and measured throughput where a test has run. From the first job-log print: for every piece that ran, what the buyer paid; what the desk would have bid, where, and when; what that would have cost delivered; and the difference, weighted by spend and summed since the watch began. The days it would have stayed put. The days nobody would have quoted. The forward view on both horizons. And for every purchase the buyer actually made: the requirement as written, the offers received and refused, each at delivered cost, the test result, the decision and the alternatives it rejected, and the invoice reconciled to the contract and to the baseline.

It belongs to the buyer and it exports. A sales deck deletes the rejected routes. This record keeps them, with the numbers they were judged on, so any decision can be audited later against what actually happened. An engineer who keeps a spreadsheet beside it is saying the record has not earned trust yet, so every exception is shown, every number says what changed and why, and every figure carries its lineage. That is most of what a buyer pays for; none of the seller-paid alternatives in Section 07 produce it; and every later layer in Section 09 is built on it.

5.11 · Two answers, kept apart

Every day the desk returns two answers, and the discipline is in keeping them separate. The view is whether there is money here: the price gap, the movable share, the delivered-cost map. It moves with the market. The readiness is whether the buyer can act on it: the supplier not yet through security review, the engineers who have not yet confirmed the work can move, the project budget not yet named, the delegated authority not yet signed. It moves with the organization, which is much slower, and it is the one that usually says not yet. A crude oil desk holds a constructive view and still declines the trade. Publishing the separation is a commitment not to convert every constructive view into a purchase, and a desk that sometimes says stay put is the only proof of neutrality a buyer can see.

5.12 · The part that decides is a solver

Most software that buys things puts a language model at the decision point. This desk puts a solver there: the schedule in 5.5 and the bid in 5.6 are the output of an optimization against the three constraints, with the buyer's limits as hard bounds, and the same inputs produce the same schedule every time. The language model has two bounded jobs: it reads unstructured quote replies and contract clauses into typed fields, each with a confidence score and the page it came from, and it drafts the explanation of a decision already made. It never decides, and a parsed quote is treated as untrusted input and bounds-checked before it can enter the optimizer. That is necessary for a saving to be reconciled to an invoice, though not sufficient on its own: the versioned record of inputs, approvals, and outcomes, and the baseline written before the decision, are what make the reconciliation auditable. Reproducible is not the same as right: the answer is as good as the contract reading, the workload characterization, the forecast, and the quote underneath it, and each carries its own source and date so that a wrong answer can be traced to the input that produced it. It is what makes this desk the compute equivalent of the planning models that tell a refinery which crude to buy.

06 · What a decision is worth

6.1 · Three baselines, never one

Every saving is measured against three baselines, and the desk's value is the difference between the last two.

1. Stay put. What the purchase costs at the incumbent, as it stands.
2. Best feasible alternative without a desk. What a competent internal team with one negotiated second supplier and a router would achieve. Agreed with the buyer before the work starts.
3. The plan with a desk.

The watching desk makes this observable every day rather than argued once. The buyer's actual purchases are the first baseline. The second is agreed before the watch starts. The desk's own record is the third.

Novation's published model returns the first minus the third. On a representative fleet — $60 million a year of AI compute, 30% of the work able to move, 15% of that held back to retire an existing commitment, 60% of the rest actually moved, and a 1.45× gap against the cheapest qualified supplier — it returns $2.53 million a year after deducting rerun waste, extra checkpointing, slower interconnect, data transfer at the buyer's own rate, supplier onboarding, the buyer's cost of adopting the desk, and the desk's subscription. That is 4.2% of the bill. It is the purchasing opportunity, and the model that produces it is public, so anyone can run it on their own inputs. The waterfall is five lines, never one: a gross benefit of $3.74 million ($2.85 million from the price gap on the spend moved, $0.89 million from the interruptible discount on the share that can take interruption); less $0.71 million of execution cost; less $0.14 million of first-year implementation; less the $0.36 million subscription the proposed tiers charge a fleet this size; $2.53 million net. The subscription is a flat line, not a share of anything, and it is meant to be read against the gross.

The desk's incremental value is a different number: the second baseline minus the third. It is computed per buyer, and this paper claims no figure for it because none has been measured. Section 00 shows why the distinction matters: $6,143 against the incumbent, $945 against the best route the buyer could have reached alone. Whether real purchases look like that campaign is the question the first year answers.

6.2 · Four value lines, reported separately

The lines overlap and are never added. Delivered performance is measured on the buyer's work, not estimated for the fleet. And every line is a running total: a desk's value is a stream of small numbers, no single day proves anything, so each line is reported cumulatively against the baseline, dated, and never as a per-event win.

A record is not a saving. Holding the contracts in one place prevents the liability and the write-off, gives a controller the accruals, the prepaid drawdown, and the true-up liability the month-end needs, and shows a treasurer the prepayments. None of that prints a saving, and this paper counts none for it. The one exception is the second line, counted only on reconciled actuals: a commitment that was sized to measured demand, against the shortfall or the overflow it replaced. Everything else the record does for finance is stated, not priced, and it is the reason a buyer keeps the subscription in a quarter when the desk finds little to move.

6.3 · What the model does not price

Availability. A cheap quoted GPU is not an available, contiguous allocation at the required time. The record carries firm quantity, start window, topology, quote expiry, and cancellation rights, and an unfilled request stays an outcome. Deadline risk. Expected rerun cost is not the cost of a missed launch; the buyer sets the acceptable miss probability and the plan is built to it. Commitment cash flows, except through the marginal cash cost. A spend commitment, a discounted-usage instrument, and a capacity reservation are different contracts; "retire the floor first" is a heuristic rather than a calculation; and uncovered usage at a hyperscaler is priced at on-demand rates after the agreement's own discounts, not at list. The record holds all three contracts as what they are and prices every hour at what it adds to the bill. The back-cast. A past purchase re-run as the decision stood at the time is a diagnostic. Historical alternatives are unobserved, and the saving it prints is not a saving.

About a third of the plausible input range in the published model falls below its own 12% screening rule on the movable pool. That is a rule for deciding whether to talk, not a probability of success, and the grid is published whole so a buyer can find their own cell.

Everything in the left column of that figure works over the phone with two suppliers. None of it waits on an exchange. The right column is real and larger per dollar, and it is not Novation's to build; Section 09 says what it would take.

07 · Who else is on the field

Every function in Section 05 exists somewhere, and one company performs most of the desk's. Nobody who holds the buyer's contracts, budgets, and purchases in one record is paid only by the buyer: the advisers who are paid only by the buyer — NPI Financial says plainly that vendors do not compensate it — hold none of those, and the platforms that hold some of them are paid by sellers, by a share of savings, or hold only the hyperscalers. That is the whole of the distinction. It is real, it is smaller than "nobody does this," and it is not a moat on its own; what can become one is the combination — the buyer's contract economics, the buyer's work measured on each allocation, alternatives that can actually be executed, and a decision record the buyer owns — and only through repeated use. So this paper makes the smaller claim.

Every index administrator in this market now has a commercial stake in a place where compute changes hands: Compute Desk runs the settlement layer; Silicon Data's chief executive also runs Compute Exchange, a marketplace; Ornn operates its own marketplace, Ornn Compute, beside its index. This desk administers no index, trades nothing for its own account, and holds no capacity. The follow-up question is always why not simply resell an index, and the answer is what an index does not do. It prices an hour. It does not report what that hour delivered on the buyer's own work, because an index has to aggregate to one printable number. The desk runs the buyer's own work on the buyer's own allocation, before the work starts and again at settlement, and that record is what no index subscriber holds.

The spend-management question, asked plainly. Finance already has tools, and the record does not replace them. Ramp governs token spend, for free, and does it well. Flexera, Apptio, CloudZero, and Vantage see the hyperscaler bill and automate hyperscaler commitments, and Vantage now meters GPU usage on CoreWeave and Nebius clusters at rates the customer supplies. Tropic tracks AI vendor contracts and benchmarks the rates. What none of them does is hold a GPU cloud's contract and invoice beside a hyperscaler's, authorize a project against a budget before it spends, measure what an allocation delivered, or decide where work should run on price and deadline. Where they are right, the record says so and leaves them in place; it ingests what they hold. The question a finance organization should ask is not which of these to buy, but which of them holds the commitment that is about to be renewed, and the answer today is a spreadsheet.

The managed-platform question, asked plainly. A buyer can hand the whole problem to Andromeda or Modal and get one invoice. That is often the right answer, and a decision record says so when it is. It is the wrong answer when the work cannot leave the buyer's own contracts, when the spread is the buyer's margin, or when finance needs to know what was paid and why. The revealed preference at the best-funded AI companies is for one counterparty. A desk asks the buyer to hold several, and its fee has to beat that convenience on outcome.

The FinOps Foundation's framework already includes workload placement, with baselines, business cases, and outcome validation. What its members report lacking is not the capability on paper. It is the standing capacity to exercise it against AI spend, a record that holds the contracts across suppliers, and a party that holds that record and is paid only by the buyer. Its 2026 survey also says where the capacity would sit: 78% of FinOps practices report to the CTO or CIO and 8% to the CFO, so a record sold to finance alone is a record without the job logs, and the platform team is in the first meeting. That is the gap, stated at its true size.

08 · What has to be true

A desk is a set of hypotheses. These are the ones that matter, each with a test that settles it inside fifteen months. The starred four are make-or-break.

Two things are not on that list, and were on earlier versions of it. A supplier set below four is not a failure: two credible suppliers create a choice, and twenty unqualified ones do not. A price gap below 1.35× is not a failure: the gap and the share moved trade off against each other, so the condition is a curve, and a commitment or assurance engagement can pay at a smaller gap. Neither survives contact with the three baselines in Section 06.

One thing is on it that was not. The first starred belief is now about finance rather than infrastructure, because the record enters through the bill and the contracts, which finance can grant on its own authority, and the job logs follow. If the fourth belief fails — if finance grants the record and infrastructure never grants the logs — the product is a contract-administration tool for AI compute, which is a real product and a smaller one, and it is priced as one.

If the starred beliefs fail, the desk narrows to the purchase decision in Section 09, sold as a project service in the segment where it paid, or sells its method to a platform that needs one, or stops. That is written down before the first watch, not after the third.

09 · Where the desk goes

Section 05 describes the record and the desk as they enter: read-only, one buyer's contracts and one buyer's fleet. This section describes where they go. A buyer adds the layers below in the order it earns them, each built on the record the one before it produced, and on the way the desk passes through two gates: first it begins to act, and much later it begins to settle. Every design problem in these layers was argued over and instrumented in electricity between 1982 and 2007, and where an established market has solved one, the subsections say how and what the compute version looks like. None of it is required for the value in Section 06. All of it is the destination. The table is in the order a buyer earns the layers, which is not their number.

The order is the argument. The corporate buyers of volatile inputs we have studied built the function in one order: the bill and the contracts first, then measurement of what was delivered, then advice on the next purchase, and only after years of all three, execution in the company's name. This paper keeps that order and changes its cadence. The record is the bill and the contracts, held continuously. The watching desk is measurement and advice delivered every day instead of once a quarter. Execution is earned by the record, not sold ahead of it. Settlement comes after execution, and only with the credit machinery built first.

The first gate is acting. The desk executes only when four things are in writing: a delegated authority with limits — a threshold below which it acts alone, a concentration cap per supplier, the qualified panel only, the regions its data may enter, and the project budget it may draw on; security evidence the buyer has accepted for a write path to the scheduler; liability terms, including whose loss it is if a supplier fails mid-run (the buyer's, and the buyer is told so before the placement, not after); and a watching record long enough that the buyer has seen the desk call the market right, and seen it say stay put. The buyer keeps the contracts. The desk signs nothing.

The second gate is settling, and it is the last thing the desk does. Paying suppliers on measured delivery is the one stage that would put the desk in the payment path. It is not offered until the desk has acted for real, and not before the credit and collateral machinery exists, for the reason 9.5 gives: every organized market that built the matching engine first paid for it later. Whether the desk runs settlement itself or routes it through a third-party settlement layer is decided then. Until then, the buyer pays its suppliers and the desk reconciles.

The foundation is a single result. In 1984, Bohn, Caramanis, and Schweppe showed that on a network the efficient price at a point is a distinct price, because delivery between points consumes real resources and is capacity-constrained. Prices differ by place and by time as a consequence of the optimization, not as a market imperfection. Their architecture — a physical settlement underneath, with contracts as financial overlays on top — is the most transferable idea available to compute, with one exception stated in Section 05: compute has no physical network to take a shadow price on, so its local prices do not decompose.

9.1 · Grading

A comparable GPU-hour is not a rate card entry. It is a rate plus fabric class, the size of the high-bandwidth domain, whether the machines are physically adjacent, memory per device, storage tier, driver and collective-library versions, region, data transfer at the buyer's own rate, expected interruption rate, the supplier's financial standing, and a measured collective bandwidth rather than an assumed one. Grain got graded elevator receipts in 1859 and became tradeable. Compute has not been graded, and the inter-provider throughput ranges in Section 02 are what ungraded looks like.

An index cannot grade a particular allocation, because an index aggregates to one printable number and that is what makes it an index. An index administrator can sell a test tool, and one does. What neither can do is run the buyer's own work on the buyer's own allocation under an acceptance condition the supplier has agreed to. That is a desk function, and it produces a record that no index subscriber holds: what was asked, what was offered, what was accepted, and what was delivered, fifty purchases deep.

9.2 · Buying, and the auction it becomes

Today, almost all of it happens by asking. The buyer states what it needs, several qualified suppliers answer with a price and a quantity, and the buyer takes the best offer that clears its constraints. This is unglamorous and it is also how most of the world's oil, gas, and industrial chemicals are still bought. Bilateral trading is not a transitional stage on the way to an exchange. In oil it has coexisted with exchange trading for four decades.

A buyer's tool that only works once a liquid exchange exists is a bet on market structure. This one is not. The desk prices every route open to the buyer and compares across them in the same units, and the auction below is what it does when a marketplace offers one.

The auction version is worth describing because it is where bidding goes if a marketplace ever clears one, and because it changes the buyer's economics rather than the buyer's question. Sellers submit price-and-quantity curves rather than single prices. ERCOT caps these at ten steps, with a one-megawatt minimum: ten is enough to express a real cost structure and few enough to keep the clearing legible. Offers sort ascending, bids descending, and the market clears where they cross.

Everyone transacts at the marginal price rather than at their own bid. Uniform pricing is the standard in every major power market because the alternative, pay-as-bid, makes bidders forecast the clearing price instead of revealing their costs, and in doing so disadvantages smaller and less sophisticated participants. For a mid-sized compute buyer that second point is the whole argument. The case was made definitively for California in 2001 and never seriously overturned, though the formal literature finds the ranking on productive efficiency ambiguous.

Two design features matter more than they look, and both bear directly on compute.

Start-up costs make the arithmetic lumpy. Because units have minimum run times and costs to start, dispatch is an integer problem, and under those conditions there is in general no single uniform price that both clears the market and leaves every committed participant willing to follow the efficient schedule. A unit can be in the optimal solution and still lose money at the clearing price. Power markets fixed this with top-up payments, plus a substantial literature on choosing the price that minimizes them. GPU clusters have exactly this structure: provisioning, container pull, weight loading, fabric initialization, and minimum billable blocks are all real setup costs. A compute market that prices only a per-hour rate, with no start-up component, is under-designed for the same mathematical reason a power market would be.

A menu of reliability tiers does the allocation. In 1987 Chao and Wilson showed that a supplier facing customers with different tolerances for interruption should offer a menu of contracts, each with a different probability of being served under scarcity, at a different price. Customers sort themselves on their own private valuations, and scarce supply reaches those who value it most without the supplier knowing anyone's willingness to pay. That is precisely the interruptible / standard / reserved menu every cloud already sells. What almost none of them does is price it as a designed sorting mechanism rather than as three separate product lines. The theory says the price differences between the tiers, not their levels, do the allocative work. This is the most immediately actionable idea in this paper for anyone who sells compute.

9.3 · Forecasting

A power desk forecasts three ways and cross-checks them. It models the supply stack, building up from the units that exist and running them against forecast demand. It models the history statistically. And it reads the forward curve straight off a liquid cleared market. Compute can do the first, can do the second badly, and cannot do the third at all.

The fundamentals are more knowable than the sector assumes. Compute supply is unusually well disclosed, because most of it is financed. Datacenter build schedules, power interconnection queues, lease terms, chip shipment guidance, and multi-year capacity commitments all appear in filings, earnings calls, and press releases. A desk that assembles those into a build-out model has a genuinely forward view of when capacity of a given class arrives in a given region. This is closer to how a gas trader models liquefaction trains coming online than to anything in software procurement.

The statistics are thin and are moving in one direction more than they cycle. This is the strongest current objection to compute as a traded commodity, and it deserves its own paragraph rather than a footnote. Craig Pirrong's point is that a futures market needs volatility around a trend, not the trend itself, and compute pricing has been trend-dominated. Over 50 trading days in spring 2026 the H100 hyperscaler index posted a coefficient of variation of 0.47% and was unchanged on 78% of daily transitions. He notes the irony that a proponent whose firm produces one of the compute indexes had dismissed DRAM futures for exactly this reason. And those index figures come from a vendor's own series, one that vendor sells and that underpins the CME contract, which is the same objection this paper makes about the benchmark layer generally.

The counter is in the same window. B200 rental pricing moved 24% over those same 50 days, with a coefficient of variation more than twenty times higher. Volatility in compute lives in the newest silicon, where supply is genuinely scarce, not in the mature generation where the index is calm. That is an argument about which contract to list first rather than about whether the market exists.

The forward curve is missing, and its absence is the reason a desk forecasts per buyer. With no cleared curve to read, the desk builds one for a single fleet from that buyer's own completed purchases, live supplier quotes, and the capacity known to be arriving. It produces bands rather than points, and it is worse than what a power desk carries. It is also enormously better than the alternative, which is renewing at whatever the incumbent quotes because nobody has a view.

Where compute price forecasting sits today is roughly where power forecasting sat in the early 1990s: the physical fundamentals are modelable, the price history is short and unstable, and the market-implied curve has not arrived. That took power about a decade to fix, and the fix was not better mathematics. It was a cash market in which anyone could verify what had actually been bought, and at what price.

9.4 · Settlement

Settlement is where compute is harder than electricity, not easier, and it is the part the sector has thought about least. A kilowatt-hour is a kilowatt-hour. A GPU-hour is not a GPU-hour. Whether a cluster delivered eight healthy GPUs at full interconnect bandwidth for the hours it billed is a question that surfaces days after the invoice, and it cannot be answered from the invoice.

So a desk measures at three moments. At provisioning, it runs an acceptance test across the actual allocation before any work starts, and rejects the allocation if it fails. During the run, it records interruptions, restarts, and delivered throughput against what was promised. After delivery, it matches each invoice line to the decision that produced it, including the quote source, the timestamp, the constraint that bound, the policy that applied, the person who approved it, and the fill. That last record is what turns a claim into a number, and it has to be written at decision time, because it cannot be reconstructed afterward.

Power markets learned to expect a long settlement tail. ERCOT settles the same operating day up to four times: an initial statement at day 5, a final at day 55, a true-up at day 180, and a hard backstop at 365, because metering disputes surface late. Compute will need the same shape. A market that settles once at delivery will relitigate every large trade by email.

Financial settlement is a different thing and is worth separating cleanly. In a financially settled contract no chips change hands; the two parties settle the difference between the contract price and a published reference price in cash. That is how most commodity price risk is actually transferred, and it requires exactly one thing compute does not yet have: a reference price both parties trust. Three indexes that disagree with each other cannot serve. Which is why the register of what was actually bought, at what price, and what was actually delivered is the precondition for the instrument layer, rather than a by-product of it.

9.5 · Risk

Four risks sit on a compute buyer today, and three of them are unmanaged almost everywhere.

Counterparty risk. Will the supplier still be there on Friday? Neocloud credit stress is live: a 2026–2028 refinancing wall, a visible down round in August 2026, and an NVIDIA backstop introduced in July 2026 renting back unused partner capacity specifically so lenders would extend credit, which is a signal about utilization risk rather than about confidence. A buyer whose campaign sits at a supplier that fails mid-run bears that loss. Screening a counterparty's financial standing alongside its contract terms, security posture, and hardware is ordinary practice on any commodity desk. It is not free: it costs a subscription, an analyst, and a stated confidence level where a private supplier will not disclose, and Section 05 prices it that way.

Volume risk. You committed to capacity you no longer need, and there is almost no way out: hyperscaler resale is capped or excluded, unused committed capacity is forfeited, and the marketplaces where a GPU reservation can be resold are thin. In every mature commodity market this is the risk a secondary market exists to solve. Compute barely has one, so the management available is on the way in, through a commitment sized to measured demand, contract length, and staggered expiries rather than one large renewal — and on the way through, by holding the drawdown pace in the record so that a shortfall is seen a quarter out rather than at the true-up. The record states volume and price risk together, against the budget, as budget-at-risk: the probability of breaching the enterprise budget or a project's, and what drives it.

Price risk. You fixed and the price fell, or you floated and it rose. Today this is managed with contract structure, and the commitment sized to measured demand is the buyer's hedge. When a listed contract actually trades, and when an accounting treatment exists to hold it, the exposure can be managed with an instrument, and the desk's job at that point is to decide how much of it to fix, which is a position decision with a stated tolerance rather than a technology decision. It is always an overlay on a physical purchase, and it is never the headline.

Performance risk. You got the chips and they delivered less work than the rate card implied. That is the grading and acceptance-test problem from 9.1, and it is the one risk here that a buyer can fully retire on its own.

And the discipline that makes the rest credible: the desk never signs as principal, holds no inventory, and takes no position of its own. Whose loss is it if a supplier fails mid-run? The buyer's. The buyer has to be told that before the placement, in writing, not after.

The credit machinery is not optional and it is not glamorous. In the GreenHat default in the PJM market, $179m was socialized against $559,447 of posted collateral, a ratio of 320 to 1. In ERCOT's February 2021 storm, a pricing error ran for 32 hours and produced $16bn of over-priced energy, about $5.1bn net of hedges, which regulators declined to correct. In California's summer of 2000, market power accounted for 59% of the increase in wholesale expenditure. Every one of those is a design failure rather than an argument against design. They are also the reason a compute market's credit and collateral machinery has to be built before its matching engine.

9.6 · What an organized market was worth in electricity

Electricity restructuring is the only large natural experiment in turning a bilateral, vertically integrated commodity into an organized market, and the gains were real, repeatedly measured, and smaller than the advocates promised.

The result that maps most directly onto a buyer's desk is Mansur and White's study of a region moving from bilateral trading into an organized market. Bilateral trading captured only about 45% of the achievable gains from trade; the organized market delivered $293–316m a year against $135m, for a one-time implementation cost of roughly $40m. That is the measured case for an exchange over a broker, and it deserves a straight answer: the organized market changes where you buy. It does not decide what to buy, when, or whether. Wholesale power prices have been public for twenty-five years, and every load-serving utility in every one of those markets still runs a dispatch desk, because a public price does not know your load, your contracts, or your tolerance for risk.

Two qualifications belong with it. That study measures a transition into an organized market from a bilateral one that already had a defined product, published settlement, metered delivery, and an established set of counterparties; compute has none of those yet. And organized markets are neither free nor automatically good.

ERCOT's move from zonal to nodal pricing cost $544.4m, plus roughly $175m borne by market participants, against an original projection of $108–157m; depending on the benefit estimate accepted, payback is somewhere between 1.7 and 9 years. Borenstein and Bushnell, examining 1997–2012, found no statistically discernible effect of restructuring on retail rates at all, and the identification problem governs everything: the states with the worst prices selected into it. If an organized compute market forms, a desk gets better price discovery, an exit from unwanted positions, and a hedge, and gets easier to run. If one never forms, everything the desk does has to be done by hand. A buyer needs the desk in both futures. Only one of them needs the market.

Appendix B · What past markets teach

Every intermediary in a commodity market that failed left a record, and the record rarely says what the folklore says. Five cases shape the design in this paper.

Deutsche Börse Cloud Exchange, 2013–2016. A neutral marketplace for infrastructure-as-a-service, backed by an exchange operator, with five providers at launch. The shareholders resolved to wind it up on 5 September 2016 and gave no reason; the parent’s 2016 annual report records €38,000 of sales revenue for the exchange that year, and no volume or customer count was ever disclosed. The trade press consensus was that the technology worked and the buyers did not come, because a neutral venue for a non-scarce good solved no pain the buyer felt. The record never mentions standardization. The lesson it does carry: a credible neutral sponsor could not manufacture demand for neutrality itself. Buyer-paid independence is a feature of this desk, and a feature is not a reason to buy. This paper does not sell it as one.

Enron bandwidth, 1999–2001. Enron's market died before Enron did, from a fiber glut, carriers who refused to connect to Enron-controlled pooling points, and a liquidated- damages provision the industry would not sign; most of its recorded trades were with other energy traders. Schwartz's retrospective describes an OC-12 circuit at $12,000 a month plus roughly $8,000 to reach the delivery point — $20,000 delivered, before any comparison with what the incumbent circuit cost, which he does not give. The access cost is the durable lesson and the desk prices it at the buyer's own rate. The other durable lesson is Arbinet's, from the bandwidth exchange that survived, still in its 10-K a decade later: sellers withdraw from any intermediary whose price transparency erodes their margin. That is a first-order risk for a pricing desk even with no marketplace, and it is why Section 05 asks suppliers for a bounded quote round with a confidential price rather than a public one.

GreenHat in PJM, 2018. A financial-rights portfolio of 889 million megawatt-hours, the largest in the market, held on $559,447 of collateral, because the credit model valued positions on a three-year historical average rather than live marks. Members lost about $179 million. A market can have standardized contracts and functioning auctions while its credit controls fail. A buyer's desk carries no clearing waterfall, but it does carry the buyer's prepaid cash and deadline at a supplier that may fail. So Section 05 assesses financial resilience separately, caps exposure, and prefers staged payment, and a saving reported before the invoice settles is treated as a mark on a stale price.

Utilitywise, 2012–2019. The largest UK energy broker went from a £200 million market capitalization to administration in five years. It recognized commission on estimated consumption and was forced to refund £7.6 million when customers used less than the volumes it had been paid on. A fee charged as a share of savings, on reported savings, before reconciliation, is that structure. This desk charges a flat subscription on the spend the record administers — the bill and the commitments, not the movable share — and a fixed fee per purchase decision, and takes no share of savings until they are reconciled actuals, and then only at the buyer's request.

Amazon EC2 Spot, 2009–2017. The spot price was generated from a hidden reserve rather than discovered from bids, and when the bidding model became inconvenient AWS replaced it in 2017 with administered prices that adjust "gradually." Supplier-published spot prices are quotes to verify, not market data. Section 04's price evidence is used for the size and shape of the gaps between suppliers, and for nothing about how any one of them will clear.

And one about markets rather than intermediaries. Brorsen and Fofana found that an active cash market perfectly predicted which agricultural futures contracts survived. Compute has built its derivatives layer first, on three competing indexes, none of which publishes a full methodology, the first listed contracts did not list on their 5 October 2026 target because the CFTC extended its review to 9 November, and its request for comment on compute derivatives closes on 20 October. The desk in this paper needs none of that to arrive. If it arrives, the desk prices the contract as an overlay on a physical purchase, because a cash-settled future cannot finish a training run.

Method

Every claim in this paper was checked against a primary source. Several widely repeated figures did not survive the check and do not appear above:

• “GPUs run at 15% utilization” is unsourceable. It traces to a sponsored vendor post with no methodology.
• Sub-second live migration does not exist in the literature. The most-cited system reports 28–228 seconds and calls it “tens of seconds.”
• “H100 fell from $8 to $2” compares a hyperscaler rate to a marketplace rate. The hyperscaler series rose from $7.76 to $9.34 before falling to $6.26.
• No NVIDIA training/inference revenue split exists. Eight quarters of filings and CFO commentary contain none.
• No published figure exists for the share of inference traffic that is batch or deferrable. Five independent datasets were checked. Any number you have seen was invented.
• Differences in electricity prices cannot explain differences in GPU prices. Power accounts for 6.8% of the median rate.
• The Compute Desk index figures above carry their own caveat. Its published term-structure module was labeled “Last 60 days” while showing data four months old, and one headline token price does not reconcile with the table beneath it. The Blackwell blended and listed lines diverged over the same thirty days on a merchant base of roughly thirty-two chips across six sellers. We use the blended-versus-listed gap because it is directional, large, and published by a third party with no interest in our argument — not because the benchmark is beyond question. Reference all three index families plurally and depend on none.
• Two figures that are routinely misattributed. The 29% wasted-cloud-spend figure is Flexera's State of the Cloud 2026, not the FinOps Foundation's survey, and both are respondent self-estimates rather than measurements. And the claim that energy brokerage compresses to roughly 3% of spend at large industrial accounts, which we had used as a fee benchmark, could not be sourced independently: brokers publish a 3–10% range of annual bill value, and the trade conventionally quotes in mils per kilowatt-hour rather than as a percentage of spend.
• Four claims are carried at the strength of their source and no stronger. Andromeda is described from its own site and press, and no source found confirms that it acts as principal or earns a spread. Compute Desk's backers are as Nodal Exchange's release states them, not independently verified. Google Cloud customers running “about 50% above their commitments” is a paraphrase of its chief executive carried by every outlet found, and is cited as reported. The 79% overrun figure is a vendor survey.
• Five claims did not survive a second check in October 2026, against two independent reviews, and were replaced. The campaign's second baseline had been a route that could not hold the deadline; it is now Supplier A's reserved tier, which can, and the interruptible tier is shown as priced and rejected for the training piece. Overflow priced at list does not hold for AWS, whose documentation prices uncovered usage at on-demand rates after the organization's discounts, so the paper prices every hour at its marginal cash cost. No FinOps platform connecting to a GPU cloud stopped being true when Vantage began metering GPU usage on CoreWeave and Nebius clusters in February 2026; the precise gap is the contract, the invoice, the authorization, and the delivered performance. No buyer-side party paid only by the buyer does not hold for NPI Financial, which states that vendors do not compensate it; the claim is now about who holds the record. And GPU fields scoped for FOCUS 1.5 does not hold: the 1.5 release adds model and token columns and a SKU price dataset, and no GPU-specific fields.
• A third check, by three external models against the filings, changed three figures and confirmed seven that it had disputed. Palantir's commitment is $5.6 billion to February 2036 after a March 2026 amendment, not $1.95 billion to 2033; AWS raised Capacity Block prices three times in 2026, not twice; and Together AI's $1.99 is its preemptible tier, now labeled. The Salesforce, ServiceNow, DoorDash, Atlassian, Palo Alto Networks, Tempus, and Andromeda figures were challenged and held on re-reading the sources.

Where two peer-reviewed sources conflict, both appear in the text. Where a figure is a working-paper result that the published version states more cautiously, we say so. Where a company's status is self-reported, it is labeled. The full evidence pack, with all corrections, conflicts, and unverifiable claims, is available on request, and the objections that survive verification are collected in the counter-evidence paper.

Selected bibliography

Bohn, R., Caramanis, M. & Schweppe, F. (1984). Optimal Pricing in Electrical Networks over Space and Time. RAND Journal of Economics 15(3), 360–376. · Schweppe, F., Caramanis, M., Tabors, R. & Bohn, R. (1988). Spot Pricing of Electricity. Kluwer. · Hogan, W. (1992). Contract Networks for Electric Power Transmission. Journal of Regulatory Economics 4(3), 211–242. · Wilson, R. (2002). Architecture of Power Markets. Econometrica 70(4), 1299–1340. · Chao, H. & Wilson, R. (1987). Priority Service. American Economic Review 77(5), 899–916. · Scarf, H. (1994). The Allocation of Resources in the Presence of Indivisibilities. JEP 8(4), 111–128. · O'Neill, R. et al. (2005). Efficient market-clearing prices in markets with nonconvexities. EJOR 164(1), 269–285. · Gribik, P., Hogan, W. & Pope, S. (2007). Market-Clearing Electricity Prices and Energy Uplift. Harvard WP. · Borenstein, S. (2005). The Long-Run Efficiency of Real-Time Electricity Pricing. The Energy Journal 26(3). · Borenstein, S., Bushnell, J. & Wolak, F. (2002). Measuring Market Inefficiencies in California's Restructured Wholesale Electricity Market. AER 92(5), 1376–1405. · Fabrizio, K., Rose, N. & Wolfram, C. (2007). Do Markets Reduce Costs? AER 97(4), 1250–1277. · Davis, L. & Wolfram, C. (2012). Deregulation, Consolidation, and Efficiency. AEJ: Applied 4(4), 194–225. · Cicala, S. (2015). When Does Regulation Distort Costs? AER 105(1), 411–444 — contested by a Comment (Han, Houde, van Benthem & Abito, 2021) and Reply. · Cicala, S. (2022). Imperfect Markets versus Imperfect Regulation in US Electricity Generation. AER 112(2), 409–441. · Craig, J. D. & Savage, S. J. (2013). Market Restructuring, Competition and the Efficiency of Electricity Generation. The Energy Journal 34(1). · Wolak, F. (2011). Measuring the Benefits of Greater Spatial Granularity in Short-Term Pricing. AER P&P 101(3). · Borenstein, S. & Bushnell, J. (2015). The US Electricity Industry After 20 Years of Restructuring. NBER WP 21113. · ERCOT (2026). Nodal Protocols §§4, 6, 9. · Potomac Economics (2021). 2021 State of the Market Report for ERCOT; IMM letter, PUC Project 51812, 4 March 2021. · Mansur, E. & White, M. (2008). Market Organization and Efficiency in Electricity Markets. Discussion draft; publication status unconfirmed. · Triolo, R. & Wolak, F. (2022). Quantifying the benefits of nodal market design. Energy Economics 112. · Brehm, P. & Zhang, Y. (2021). The efficiency and environmental impacts of market organization. Energy Economics 101. · Jha, A. & Wolak, F. (2019). Can Financial Participants Improve Price Discovery and Efficiency in Multi-Settlement Markets with Trading Costs? NBER WP 25851; journal version unverified. · Joskow, P. (2008). Lessons Learned from Electricity Market Liberalization. The Energy Journal 29(SI2), 9–42. · Kahn, A., Cramton, P., Porter, R. & Tabors, R. (2001). Uniform Pricing or Pay-as-Bid Pricing. The Electricity Journal, July. · Fabra, N., von der Fehr, N. & Harbord, D. (2006). Designing Electricity Auctions. RAND 37(1), 23–46. · Cramton, P., Shoham, Y. & Steinberg, R. (eds., 2006). Combinatorial Auctions. MIT Press. · Jeon, M. et al. (2019). Analysis of Large-Scale Multi-Tenant GPU Clusters for DNN Training Workloads. USENIX ATC. · Weng, Q. et al. (2022). MLaaS in the Wild. NSDI. · Hu, Q. et al. (2021). Characterization and Prediction of Deep Learning Workloads in Large-Scale GPU Datacenters (Helios). SC. · Hu, Q. et al. (2024). Characterization of Large Language Model Development in the Datacenter (Acme). NSDI. · Xiao, W. et al. (2018). Gandiva. OSDI. · Narayanan, D. et al. (2020). Heterogeneity-Aware Cluster Scheduling Policies for Deep Learning Workloads (Gavel). OSDI. · Patel, P. et al. (2024). Splitwise. ISCA. · Pope, R. et al. (2023). Efficiently Scaling Transformer Inference. MLSys. · Zhong, Y. et al. (2024). DistServe. OSDI. · Shukla, D. et al. (2022). Singularity. arXiv:2202.07848. · Sukprasert, T. et al. (2024). On the Limitations of Carbon-Aware Temporal and Spatial Workload Shifting in the Cloud. EuroSys. · Strati, F., Ma, X. & Klimovic, A. (2024). Orion. EuroSys. · Schwartz, A. (2002/03). Enron's Missed Opportunity. BRIE WP 152, UC Berkeley. · Brorsen, B. & Fofana, N. (2001). Success and Failure of Agricultural Futures Contracts. · Cronon, W. (1991). Nature's Metropolis. W.W. Norton, ch. 3. · Bandi, F. & Su, Y. (2026). (Early) AI Compute Asset Pricing. arXiv:2607.12156. · Pirrong, C. (2025–26). “Going Vertical” and “Will Compute Futures Compute?” Streetwise Professor. · FinOps Foundation (2026). State of FinOps 2026. Linux Foundation, 19 February 2026; n = 1,192, $83bn+ annual cloud spend. · Flexera (2026). State of the Cloud 2026. 18 March 2026; n = 753. Source of the 29% wasted-spend figure. · Ehrig, S., Frechet, M. & Ignjatovic, V. (2026). Agentic AI in Energy Trading: From Fragmented Market Data to Explainable Trading Intelligence. Capco. · Compute Desk (The Compute Index, Inc.) (2026). Regulated blended and listed-only indexes, free tier, retrieved 8 September 2026; Nodal Exchange announcement, 3 September 2026; Architect Financial Technologies ComputeConnect announcement, 8 July 2026. · CFTC (2026). Request for Comment on the Listing of Compute Derivatives Contracts, 21 August 2026; comment window closes 20 October 2026. · Synergy Research Group (2026). Q2 2026 cloud infrastructure services review, 30 July 2026; and Neocloud market forecast, 2 April 2026. · Gartner (2026). Worldwide IT Spending Forecast, 27 July 2026. · IDC (2026). Worldwide AI Infrastructure Tracker, 21 July 2026. · Epoch AI (2026). Frontier labs don't use most AI compute. 20 May 2026; modeled estimates. · CoreWeave, Inc. (2026). Form 10-Q, quarter ended 30 June 2026. · Microsoft, Amazon, Alphabet and Meta (2026). Quarterly results for the period ended 30 June 2026. · Salesforce, Palo Alto Networks, Adobe, ServiceNow, DoorDash, Atlassian, Snowflake, Pinterest, MongoDB, Tempus AI (2025–26). Forms 10-K and 10-Q, purchase-commitment and cost-of-revenue disclosures. · CoreWeave and Nebius (2026). Second-quarter 2026 results and earnings calls: contract length, backlog, take-or-pay terms, customer prepayments. · DoiT International / Sapio Research (2026). Survey of 500 finance leaders at companies of 1,000 or more employees, fielded February 2026; source of the 79% overrun figure. · Harness (2026). Survey of 700 engineering leaders, mid-2026; source of the 72%, 52%, and 56% figures. · CFTC (2026). Extension of review of the CME Group compute futures submissions to 9 November 2026. · Nodal Exchange (2026). Announcement of financially settled compute futures with Compute Desk, 3 September 2026. · Stripe (2026). Announcement of the OpenRouter acquisition, 19 August 2026.

Novation · White Paper 01 · September 2026, revised 12 October 2026. Figures cited to primary sources. Caveats travel with the numbers. The charts on this page are drawn in the site's own line grammar; the numbers and their caveats are unchanged.