The building lasts decades. The chip inside it is competitive for two years and useful for five or six. The last cycle solved that mismatch by letting a handful of neoclouds borrow against the chips and sell long contracts to a handful of labs. The next cycle will not look like that, because the buyers are changing.
Three new kinds of owner are buying chips and renting the space around them: governments that want compute inside their borders, frontier labs that want out from under their cloud landlords, and enterprises that can now run open-weight models on hardware they control. Each of them owns a cluster for its own use, not to rent out. Each of them will, at times, have more compute than it needs, and at other times less.
That is the opening. A market with many owners, many configurations and no standing off-take contracts needs a place to sell a month of reserved capacity, a way to check the price, and someone who will hold the block while it sells. Those three things, in that order, are where the value is. Cash-settled futures are the last step, not the first.
The case against is real, and section 04 gives it its due. The five largest buyers still hold well over half the installed chips. Frontier training stays with whoever owns the newest hardware at scale. Most enterprises have never run a cluster. The thesis is a shift in share and in who needs an intermediary, not a replacement of the hyperscale model. Our verdict: plausible, already visible in the data, and under-served by the products being built. The exchanges are building hedges for a market that still lacks a secondary market to hedge.
A data centre is three assets on three clocks, and only the slowest one is financed like infrastructure.
The building is the long asset. Its value depends on power, cooling and location, none of which care which chip is inside. The chip runs on two shorter clocks. Its accounting life is four to six years: CoreWeave uses six, Lambda five, Nebius four, and the three largest clouds moved from three or four years to five or six between 2020 and 2024. Its frontier life, the window in which it is the chip you would choose for training, is one to two years. After that it cascades: training in years one and two, real-time inference in years three and four, batch work in years five and six.
The gap matters because the two assets are financed differently. The building is financed against a lease, at infrastructure rates, by investors who understand real estate. The chip is financed against a customer contract and a guess about resale value, at equipment rates, by lenders who are still learning the asset.
The last cycle bridged the gap with the neocloud balance sheet. CoreWeave carried $35 billion of debt at June 30, 2026, on facilities analysts put at 60 to 70 percent loan-to-value, maturing between 2026 and 2028. Its $8.5 billion loan in March 2026 was secured by chips and a Meta contract worth at least $19 billion. On July 1, 2026 Nvidia began renting back unused capacity from its neocloud partners at a fixed rate, so that lenders would keep lending.
That financing dictated the market’s shape. Lenders want contracted off-take, so capacity is sold long, to a few large counterparties. Demand does not have that shape: training runs, evaluation cycles and inference bursts want capacity in weeks, with the option to stop. Hyperbolic’s October paper puts it plainly: the constraint on short-duration supply is not that operators have overlooked it, it is that their creditors will not let them serve it. Change who owns the chip and you change what the lender wants, and with it the shape of the market.
Owned chips in rented buildings. Three buyers who are not sellers by trade.
The next wave of chips will be bought by their users, not by landlords, and housed in buildings those users rent.
The mechanism is simple. A building built in 2024 will see four or five generations of chip pass through it. Whoever owns the building needs tenants for thirty years. Whoever owns the chip needs to earn back its cost in four. The neocloud tried to be both, and borrowed to do it. The next owner will be one or the other. Three kinds of buyer are stepping up to own the chip.
Governments. Sovereign compute is now a procurement category. The point is control: a cluster inside the border, on a national network, running models the state can inspect. These buyers do not rent out capacity. They allocate it, often below cost, to domestic firms and universities.
Frontier labs. The labs that rented everything in 2024 are building their own capacity because the landlord’s margin is now the largest line in their cost of revenue. Patience Capital’s reading of Anthropic’s S-1 is the cleanest statement of the arithmetic: rented compute at about $10 per watt against depreciation of $3.50 and operating cost of $2 per watt when owned, with hyperscalers’ share of the lab’s capacity falling from all of it before 2026 to under a fifth in 2028. A lab that insources stops paying an 80 percent margin to a landlord that also competes with it for customers.
Enterprises on open-weight models. This is the buyer the thesis turns on. A closed model can only be reached through its vendor’s API, so the enterprise rents by the token and never touches a chip. An open-weight model can be downloaded, fine-tuned and run on hardware the enterprise owns. Once the model is free to move, the enterprise can buy the chip, rent rack space in a colocation building and pay for power. The hardware decision becomes a capital decision like any other.
What these three buyers share is the thing that matters for market structure. None of them is a seller of compute by trade. Each sizes its cluster to a baseline it expects to run, and each will be wrong about that baseline in both directions. A government cluster sits idle between research programs. A lab finishes a training run and has a month of spare capacity before the next one. An enterprise pilots an agent, discovers it needs three times the throughput, and needs it next week.
The result is a market with many more owners, many more configurations, far less contracted off-take and a recurring need to sell or buy a few weeks of reserved capacity. That is the secondary market for reserved compute, and today it barely exists.
Each of the three buyers is already visible in public data. So is the resale behaviour the thesis predicts.
| Signal | What it shows | Source |
|---|---|---|
| 184 sovereign AI projects in 67 countries by mid-2026; 80% of new projects in the first half of 2026 were infrastructure, up from 62% | Governments are buying compute, not models | CNAS, Aug 20, 2026 |
| Most sovereign model projects are fine-tunes of foreign open-weight models; three-fifths of those bases are American, Llama the most common | Sovereign clusters run open weights: the buyer owns the hardware and borrows the model | CNAS, Aug 20, 2026 |
| Canada’s Sovereign Compute Infrastructure Program, call opened April 15, 2026: approximately C$890 million over seven fiscal years from 2026–27 for a Canadian-owned supercomputer | A state buying and owning its own cluster | Informed Clearly, corroborated by the program page at Innovation, Science and Economic Development Canada |
| Anthropic’s capacity mix from its S-1 and press reports: hyperscalers from 100% before 2026 to under 20% in 2028; internal capacity from zero to over 60% | A frontier lab insourcing its chips | Patience Capital, Nov 1, 2026 |
| Rented compute at about $10 per watt against $3.50 of depreciation and $2 of operating cost per watt when owned | The margin that drives the lab to own | Patience Capital, Nov 1, 2026 |
| Open-weight models processed 29% of tokens through Vercel’s gateway in June 2026, up from about one-ninth two months earlier, for under 4% of spend | Enterprise traffic moving to models that can be self-hosted | Zentera, Jul 28, 2026, citing Vercel |
| Nvidia and more than 120 companies formed the Open Secure AI Alliance in July 2026 to promote open models run on enterprise infrastructure | The chip vendor backing the on-premises buyer | Zentera, Jul 28, 2026 |
| Compute Exchange opened a used and refurbished hardware marketplace on July 17, 2026; requests from hundreds to tens of thousands of chips; used units 40 to 60% below list | A secondary market for the chip itself, at institutional size | Blockspace via Yahoo Finance |
| Hyperbolic sources supply from neoclouds, enterprise idle capacity and customer resale; $80 million run-rate by September 2026 from near zero in June 2025 | Owners already resell what they do not use, through an aggregator | Hyperbolic, Oct 26, 2026 |
| Nvidia’s July 1, 2026 backstop: it rents back unused capacity from neocloud partners at a fixed rate so that lenders will lend | Even the neocloud model needs a buyer of last resort for idle capacity | Capacity, Aug 12, 2026 |
Two things stand out in the table. First, the owners are already reselling: Hyperbolic’s supply comes partly from enterprise idle capacity and customer resale, and Nvidia had to become the buyer of last resort for its own partners. Second, the used-chip price is wide and inconsistent across sources, from $6,000 to $22,000 for an H100 in one survey to $25,000 to $35,000 in another from the same publisher five months later. A spread that wide is not a market. It is the absence of one, and it is the clearest sign that the price-discovery layer has not been built.
Five ways the thesis fails. The first two are strong.
Concentration does not move. The five largest buyers hold roughly 58 to 60 percent of installed AI chips on Epoch AI’s count, and their capital budgets dwarf everyone else’s. Nvidia’s $500 billion financing platform with six institutional partners, with up to $125 billion backstopped by Nvidia itself, is aimed at hyperscalers and labs, not at a thousand enterprises buying eight-chip servers. Share can shift at the margin while the centre of gravity stays where it is.
Frontier work stays with the newest chip at scale. Training a leading model needs tens of thousands of the current generation, networked tightly, in one place. No enterprise and few governments will own that. The thesis only ever applies to the cascade, the chip’s years three to six, and to inference and fine-tuning. That is a large market, but it is the cheaper end of it: Compute Desk’s one-year reserved index has the B200 at a September 2026 average of $5.02 an hour, and a used H100 system sells for 40 to 60 percent below its list price.
Owning is harder than renting. A cluster needs a power envelope above 20 kilowatts a rack, liquid cooling, a scheduler, a team that can diagnose a failing memory module, and a warranty path when one of the eight cards dies. Colocation space with that envelope is scarce; industry reporting through 2026 describes enterprises reserving space 18 to 24 months ahead. The lead time on power, not chips, decides who can own. Many enterprises will look at that list and keep renting.
The chip is cheap only while it is busy. Self-hosting an open model beats the hosted price only above a sustained volume, roughly 500 to 800 million tokens a month in published estimates. Below that, idle chips are the most expensive way to buy tokens. The enterprise that over-buys will want to resell, which supports the secondary-market half of the thesis, but it will also learn the lesson and buy less next time.
The hyperscaler can match the economics. If a lab or enterprise can run a cluster at depreciation plus operating cost, so can the cloud, and the cloud has the lower purchase price and the diversified workload. Patience Capital’s own conclusion is that hyperscaler returns will fall and rental rates will normalise as greenfield capacity arrives. A cloud that cuts its margin from 80 percent to 50 percent removes much of the reason to own.
Narrower than the headline, still large. Sovereign buyers own for reasons that are not economic, so price competition does not move them. Labs own because their scale now justifies it, and that logic extends to any model company past a few gigawatts. Enterprises own the inference baseline for workloads that cannot leave the building, the regulated and export-controlled share of enterprise AI, and rent the rest. In every one of those cases the owner has a cluster it did not size to rent out, and will at some point have too much or too little of it.
Price discovery moves from the long end to the short end. The intermediary’s job moves from financing the chip to placing a block of time.
In the last cycle the price was set years out, between a neocloud and a lab, and nobody outside the deal saw it. The indices that exist today were built around that market, and they measure different things. Compute Desk marks a one-year reserved contract in the United States, blending completed transactions with neo-cloud list prices and excluding hyperscalers. Ornn prints executed on-demand trades. Silicon Data normalises posted and transacted prices across more than thirty platforms. Hyperbolic’s comparison of the three on the B200 shows how far apart the methods land: Compute Desk averaged $3.80 an hour from mid-November to the end of 2025 and $5.02 in September 2026, with a standard deviation of $0.42 over the sample, against $1.19 for the on-demand series. The spread between them is not noise. It is the term structure and the provider mix, and nobody publishes the tenor in the middle.
Five things change when the chip has many owners.
- The traded instrument is a block of reserved time, one to three months long. That is the tenor between on-demand and the one-year contract, and it is the one a cluster owner actually has to sell or buy. No index tracks it today.
- Basis becomes the pricing problem. Chip generation, condition, network topology, location, power contract and the provider’s tier all move the price. A grid that prices any configuration against a reference is worth more than the reference itself.
- The lender’s collateral changes. It stops being a customer contract and becomes a chip in somebody else’s building. That lender needs two things it does not have: a resale curve for the chip by generation and condition, and a record of how hard the chip has been used. The used-hardware price brackets the rental price for older chips from both sides, which makes the secondary hardware market a pricing input for the rental market, not a separate business.
- The short-duration premium is the arbitrage. Hyperbolic’s reference trade buys 256 B300 chips on a 36-month contract and sells them in weeks, at a modelled 2.1 times return before hedging. The premium exists because lenders force neoclouds to sell long. Owners who do not have lenders do not have that constraint, and a secondary market for their surplus compresses the premium.
- Someone has to hold the block. An owner with a month of spare capacity wants cash now, not a listing. The buyer of that month wants a week. The party in between carries the inventory, the credit and the basis. That is the merchant seat, and it is the one the comment file at the Commodity Futures Trading Commission asks to be kept separate from whoever publishes the benchmark.
The net effect is a market that looks more like physical power than like a futures-first commodity: many nodes, each settling on its own hardware class and location, with an aggregator doing the work of the grid. In power, the benchmarks that won were the ones attached to the deepest physical delivery point. Compute will go the same way.
Futures, secondary markets or tranches? The question is sequence, not choice.
Each service solves a different party’s problem, and the owner’s problem comes first.
Start with who has what problem. The new owner has a quantity problem: too much capacity this month, too little next quarter. It does not have a price problem, because it is not selling compute for a living; it bought the chip to run a model. The lender behind that owner has a residual-value problem: what is the chip worth in three years, in somebody else’s building, with no customer contract attached. The merchant who holds a block between seller and buyer has a price problem: the block can lose value while it sells. Only the third party wants a futures contract.
| Service | Whose problem it solves | What must exist first | Who occupies the seat today |
|---|---|---|---|
| Resale venue for reserved blocks, with delivery | The owner’s surplus and shortfall | A standard block (chip, count, network, term, location), a control plane that can hand over a specific cluster on a date, verified counterparties | Compute Exchange (reserved capacity and forward requests, no custody), Hyperbolic (physical settlement), Corgi (resale in development) |
| Price check against a basis grid | Both sides of every resale | Completed-deal data across configurations, grade and condition, a reference per chip | Three headline indices, none at the one-to-three-month tenor; grading from the comment file and the hardware marketplaces |
| Block holder | The seller who wants cash now | Capital, credit lines, a hedge, and distance from whoever publishes the benchmark | Hyperbolic says it will; neoclouds cannot, because their lenders require off-take |
| Loans against an owned chip in a rented building | The owner’s financing and the lender’s collateral | A resale curve by generation and condition, a usage record and a remarketing path, which means the first two rows | Only at neocloud scale: CoreWeave’s $8.5 billion facility, Nvidia’s rent-back. Nothing for an enterprise with 64 chips |
| Fungible units priced into tranches | Investors who want a yield on compute without running it | All of the above, plus an independent administrator and a grade system (frontier, mid, economy) | Index designs and 26 registered futures-based funds, one live. No pooled physical product yet |
| Cash-settled futures | The block holder and the lender | A benchmark that survives the Commission’s review, and a physical layer for it to reference | CME (H100 and B200 on Silicon Data; review extended to November 9, 2026), ICE (on NativX and on Ornn), Nodal (on Compute Desk’s one-year reserved index); Kalshi’s fully collateralised contracts on Ornn, trading since early 2026, with an amendment stayed by Commission staff until November 19 |
- The resale venue and the price check come first. They need no capital and no ruling, and they are what the owner actually asks for. Compute Exchange and Hyperbolic are building this layer now, and both say the same thing: a venue that takes positions is a competitor, not a marketplace.
- The block holder comes second, and needs the first two. It cannot hold a block it cannot price, and it cannot sell a block without a venue. The premium is real: Hyperbolic’s modelled 36-month B300 contract returns 2.1 times unhedged and 1.7 times after buying a put on the final year. The seat is also the one the comment file wants kept away from the index.
- Finance for the owned chip comes third. Lenders will lend against a chip in a rented building when they can see a resale curve and a usage record. Until then the only chip-backed loans are at neocloud scale, with the vendor standing behind them. This is the largest prize, because it is what lets the enterprise and the sovereign buy the chip at all without a landlord.
- Tranches come last. A pooled, graded, yield-bearing compute asset is the end state, and every input to it is a row above. Building it first produces a product nobody can value.
- Futures ride alongside the block holder, not ahead of it. They hedge the merchant and the lender. They do not hedge the owner, whose exposure is quantity, and whose configuration sits a long way in basis from any headline index.
So: not futures, not tranches, and not a secondary market alone. A secondary market with a price check and a holder, which is the merchant model from our earlier work, with the holder kept at arm’s length from the benchmark.
The market has built the top and bottom of the stack and skipped the middle.
Built. The used-hardware market is real and institutional: Compute Exchange’s hardware venue takes requests up to tens of thousands of chips. Three headline indices print daily, and they disagree in ways that are informative once you know their methods. Aggregators run control planes that can deliver a cluster on a date; Hyperbolic’s is sourcing from idle enterprise capacity already.
Pending. Three exchanges have filed futures. CME’s H100 and B200 contracts on Silicon Data were due October 5, 2026; the Commission extended its review to November 9 and CME says the effective date will be announced. ICE is waiting on the Commission’s request for comment, which closed October 20, with contracts on the NativX energy-normalised index and on Ornn’s transaction index. Nodal has a contract on Compute Desk’s one-year reserved index. Kalshi’s fully collateralised contracts settling on Ornn have traded since early 2026; the Commission treated them differently because there is no leverage, though staff stayed a Kalshi amendment on August 21 until November 19 and the firm has since asked to list compute derivatives afresh. When the futures list, they will hedge the merchant and the lender. They will not change the owner’s day.
Missing. Four things, and they are the four the new owner needs.
- A resale venue for reserved blocks at the one-to-three-month tenor, with a standard block definition and physical delivery. Compute Exchange handles reserved capacity by request; Corgi lists resale as coming. Neither publishes a traded price for the block.
- A price for that tenor, and a grid that adjusts it for configuration, condition and location. The comment file’s loudest demand, per-fixing disclosure of contributor count and input mix, is the precondition for anyone trusting such a grid.
- A usage record. Lenders, buyers and insurers all need to know how hard a chip has run. Nobody publishes it, and the owner has no incentive to volunteer it until a loan or a resale price depends on it.
- A lender for the owner. Chip-backed loans exist only at the scale of CoreWeave’s $8.5 billion facility, with a Meta contract and an investment-grade rating behind it. An enterprise buying 64 chips for a colocation cage finances them like office furniture, if at all.
The pattern is the one our enterprise-hedge work found: the obstacle is market structure, not index design. The indices can be built in weeks. The resale venue, the usage record and the owner’s lender are the slow parts, and they are what makes the index worth settling on.
Seven signals. Three confirm the thesis, three kill it, and one is the tell.
Every page opened on November 3, 2026.
Pieces this article responds to
- GPU Rental Indices Challenges: Let Me Count the Ways, Blu Putnam, Hyperbolic, November 2, 2026. The three-index comparison, the Compute Desk B200 averages and standard deviation, and the distinction between futures and prediction markets.
- GPUs are financed for years. AI workloads run for weeks, Hyperbolic, October 26, 2026. The short-duration premium, the reference trade and the power-market analogy.
- Anthropic and the AI compute profit pool, Jimmy Su, Patience Capital, November 1, 2026. The lab-insourcing case and the per-watt economics.
Public sources
- CNAS Insights: Sovereign AI’s Second Wave Is Coming Into View, August 20, 2026
- CoreWeave’s debt hits $35bn, Capacity, August 12, 2026
- CoreWeave raises $8.5 billion GPU loan backed by Meta deal, Bloomberg Law, March 2026
- Resetting GPU Depreciation, theCUBE Research, November 22, 2025
- Evaluating AI’s economic viability starts with GPU lifespan, a summary of a Goldman Sachs report
- Nvidia’s chief executive on a seven-to-eight-year chip life, October 2026
- Compute Exchange launches secondary GPU marketplace, Blockspace via Yahoo Finance, July 17, 2026
- The Weights Are Coming Home, Zentera, July 28, 2026
- Nvidia H100 price in 2026, CloudZero, October 2026
- Understanding the GPU lifecycle, WhiteFiber, 2026
- Private AI infrastructure costs 2026, ecn-apps, 2026
- Buy, Lease, or Colocate, VDF, 2026
- ICE announcements of compute futures on Ornn (May 19, 2026) and on NativX (July 1, 2026)
- CME Compute Futures, Ironbeam, October 2026
- GPU Compute Futures and the CFTC, DeSilva Law, October 2026, on the Commission’s request for comment on compute derivatives and the comments filed to it
- Compute Desk index methodology, as quoted by Hyperbolic
Assumptions and limits
- Useful-life ranges in Figure 1 combine accounting schedules, design life and vendor claims. They are not one measure.
- The 58 to 60 percent hyperscaler share of installed chips is Epoch AI’s estimate as carried in our earlier work, not re-verified here.
- Compute Desk levels are the averages Hyperbolic computed from the published index; we have not reproduced them from Compute Desk’s own page. Compute Desk indexes H100, H200, B200 and B300 only, so no reserved print exists for the A100.
- Used-chip prices vary by a factor of three across published sources. We treat the dispersion as a finding, not as noise, and none of the figures should be used as a mark.
- Compute Exchange and Hyperbolic are parties with a position in this market. Their published claims about their own volumes are not independently confirmed.
- The positions on the readiness map are our judgement. They are meant to be argued with.
- We have not read the Anthropic S-1 directly. The capacity-mix figures are Patience Capital’s estimates from the filing and press reports.