Sam Lawhead Offtake

Work sample · November 2026

The Compute Book

A GPU-hour has no single price. It has a term structure, a credit spread and a value that depends on the workload. Any company that buys compute and resells it is running a duration book, whether it models one or not. This is that model.

The supply contracts and demand streams are illustrative, sized to a compute aggregator with about $100M of revenue. They are not any company's actuals. Market inputs are dated public prices. Throughput is SemiAnalysis's open InferenceX benchmark data.

3.6x
Spread from cheapest to dearest public price for one Blackwell Ultra GPU-hour
$19M
Paid over market by a 1,000-GPU H100 contract signed in April 2023
6x
Lower token cost on GB300 than H200 at market rent (DeepSeek R1, FP8, 50 tok/s per user)
2x
Margin on the GB200 book when 90% of idle hours can be resold, against none

What it answers

QuestionHowOutput
Reserved or spot, how long, how muchMonte Carlo of the whole book; a common demand factor moves prices and the book's own demand togetherSourcing frontier by tenor and size
H200, GB200 or GB300Measured throughput per chip at a target speed per user, times market rentCost per million tokens; break-even rent
Where the margin comes fromEvery delivered hour transferred to the product at the market priceProduct margin, procurement alpha, idle, quality and shortage
What today's bets do to the P&L36 months of contracts, renewals at each path's price, delivery slips, prepay credit riskDistribution of gross margin and cash; mark-to-market
What a vintage didFixed-price contracts against the one-year contract index since 2023Gain or loss to lease by signing date

Exhibits

Line chart of the one-year H100 contract rate from 2023 to 2026 with three fixed-price contracts: April 2023 at $3.05, minus $19.0M; April 2024 at $2.35, minus $4.4M; October 2025 at $1.70, plus $5.2M.
Same chip, same term, opposite booksA 1,000-GPU, three-year H100 contract signed at the one-year rate in April 2023 ran $19M above market over its life, about 23% of its value. One signed in October 2025 at $1.70 is $5M ahead after a year and still $1.15 an hour under market. The vintage decided the book.
Dot chart of public prices per GPU-hour, July to November 2026. H100 from $2.30 to $5.97; B300 and GB300 from $2.72 to $9.90.
One chip-hour, a dozen pricesPublic H100 prices ran 2.6x from cheapest to dearest between July and November 2026. Blackwell Ultra ran 3.6x, from Microsoft's five-year offtakes near $2.72 to $2.91 all-in to on-demand lists at $9.50 to $9.90. The spread is term, size and credit, not silicon.
Two bar charts. Left: rent per GPU-hour, H100 $2.97, H200 $3.80, B200 $5.90, GB200 $6.90, GB300 $6.00. Right: cost per million output tokens for DeepSeek R1 at FP8, H100 $4.35, H200 $2.73, B200 $0.99, GB200 $0.69, GB300 $0.45.
Pricier hours, cheaper tokensPriced at market rent, a GB300 hour costs about 1.6x an H200 hour and makes a million tokens for $0.45 against $2.73 (DeepSeek R1, FP8 on every chip, 50 tokens a second per user). For training, the same rents sit near fair value relative to H200. The market prices the hour for training, and the inference gain goes to whoever routes the work.
Line chart of expected three-year gross margin against GB200 GPUs committed for three years, for three resale capabilities. Cannot resell: best 288 GPUs, $17M. Resells 60%: best 432 GPUs, $27M. Resells 90%: best 432 GPUs, $34M.
Liquidity is what makes a big commitment paySourcing a GB200 pool for the illustrative book over 1,000 simulated paths, with demand correlated with prices. A book that cannot resell idle hours should commit 288 GPUs. One that resells 90% should commit 432 or more and earns twice as much. Against buying everything spot, the three-year lock also moves the 5th-percentile outcome from a $10.6M loss to a $5.9M gain.
Waterfall chart of 36-month gross margin: products at market cost plus $50.2M, brokered spread plus $16.2M, bought below market plus $8.6M, idle hours minus $0.6M, paid not useful minus $4.2M, short and bought spot minus $28.9M, own research at market minus $18.5M, gross margin $22.7M.
Split the margin before you trust itTransfer every delivered hour to the product at the market price and the book's 36-month margin separates into what the products earn, what the supply desk bought below market, and what idle hours, failed hours and spot shortfalls cost. In this book, being short is the largest drag.

Method

Market. Each chip's expected price path is calibrated so its average over a tenor equals the quoted term rate, with a step down as the next generation arrives. A shared AI-demand factor moves every price and the book's own demand together, so the book runs short exactly when spot is dear. Volatility comes from the SemiAnalysis H100 and B200 composites.

Supply. Take-or-pay contracts carry price, prepayment, availability, goodput, a delivery slip drawn on each path and the provider's default risk on the prepaid balance. Expiring contracts renew at each path's prevailing price.

Allocation. Each month and GPU pool, contracted customers are served first, then hosted training and inference, then on-demand. Shortfalls are bought at spot plus a premium, with less filled when the market is tight. Research takes its non-deferrable share, the marketplace resells part of what is idle, and elastic research uses what nobody bought.

Margin. Products are charged the market price for every hour they use. The supply desk keeps the difference between that and the contract price, and carries idle hours, failed hours and the spot premium. The components sum to cash gross margin on every path, and the model checks that identity on each run.

Hardware. Output tokens per second per GPU come from SemiAnalysis's InferenceX Data API at a fixed speed per user and a fixed precision. They are priced at this model's market rents, not the benchmark's own cost assumptions.

Limits. Index levels and list prices are not transaction prices. The H200 and B200 reference rents differ widely across indexes. Hopper benchmark coverage for newer models is thin. The supply and demand book is invented to show the mechanism.

Files

The workbook models one GPU pool with every formula live: the monthly book, the margin split, a decision grid and a 1,000-path Monte Carlo run through Excel Data Tables (press F9). It reconciles to the Python engine to the cent.