> you call your boy up at Nvidia “yo can I get 1000 B300s and slap some Infiniband on that”
> Nvidia says “that’ll be 30 weeks but since you’re my boy I might be able to sneak you into a shipment because someone (wink wink OpenAI) just bought 200k GPUs”
“bet”
> you call your girl at an OEM “can you give me a quote for some air cooled B300s”
> “yea it would be 10 gajillion dollars”
> you call your boy at Nvidia again “yo the OEM just told me it would be 10 gajillion dollars”
> “nah lemme talk to them see what I can do”
> OEM calls back “hey sorry we’ve revised our quote down to 5 gajillion dollars but do > you have colo”
> “bet lemme go talk to my guy”
> you call your colo guy “yo you still got power in Santa Clara?”
> “nah sorry we just filled that but I could put you in Montreal”
> “ok awesome then we’re set”
> “sg but we just talked to the bank can you do like 50% down on a 3 year contract for like $4.50/gpu/hr and btw it’s only 95% uptime on network and GPUs”
????????????
The biggest bottleneck in the GPU industry at the moment is not chips, CoWoS, or even power.
Yes sure the lead times are increasing, Nvidia announces price increases, OEMs secretly increase prices, wafers are more scarce than ever, mini monopolies scattered all over the vast and complex supply chain, but at the end of the day if you have enough money, none of this is relevant.
They’ll tell you it’s 30 weeks to get an NVLink switch, but buy enough or pay enough and this shrinks to 14! The whole industry is purposely built on relationships and is very “iykyk”, but the biggest problem is actually a lot simpler.
The biggest problem is credit. Specifically, the lack of residual markets.
Residual value is basically the value of the asset at the end of its depreciation. Think of a used car after you lease it and what it’s worth. Clearly at least the price of its parts are valuable, if not the functional car.
But no one knows how to model the residual value of GPUs. For example, banks just model it as 0 after some straight line depreciation because there’s no liquid public market.
Residual value is important because it’s the blocking factor behind securitizing the asset and creating an instrument to hedge risk allowing for lower cost of capital. This sucks for smaller players as there’s no access to cheap financing.
When going rates for these loans are SOFR + 900bps (SOFR is the interest rate used to price U.S. dollar loans, derivatives, and bonds) you need massive downpayments which no small player can afford.
To demonstrate, consider this realistic example of expensive financing.
Let’s say you buy B300s, 128 nodes (8 GPUs) for 5 years, $3.50/hr/gpu, 30% down (below current market price btw).
TCV = 128 * 8 * 3.50 * 24 * 365 * 5 = $156,979,200
which means your downpayment is 156,979,200 * 0.3 = $47,093,760.
Keep in mind, you’re praying the SLAs are good, probably 14 week RFS on some SMCI builds where they sourced parts from some corner of Malaysia that can’t be replaced when it breaks exactly a year later.
You got $47,093,760 lying around as a startup? You sure you can immediately fill those nodes? Oh and it’s build to order so they’ll only order the parts once you sign and you can’t even test them before you do.
Also, 128 nodes of air cooled B300s is also about 2.5 MW including cooling. Anthropic’s revenue is about $50m / MW.
Assuming you’re able to squeeze revenue out of your compute like Anthropic you only get $125m for your 128 nodes of B300s. You’re paying $32m more than this which you can think of as the premium you’re paying on your loan.
Ok enough ranting. Let’s dig into residuals, how the industry deals with it currently, how other industries deal with it, and what the future could look like.
What do current large public deals look like?
How are they handling residual value?
In August 2026 NVIDIA filed a document titled Residual Value Guaranty. It promises a minimum value of up to $105B on the data centers SB Energy is building for OpenAI, about 4.25 GW of them. The chipmaker is guaranteeing residual value, case closed, right? Read the filing.
The guaranteed value is defined as data center, power and transmission related costs, and NVIDIA’s own equipment is carved out into a separate repossession clause outside the guaranteed amount. The guaranty covers the building and the power feeding it, not the chips racked inside it. NVIDIA purposely won’t put a number on the GPUs.
What about Lambda’s $926M senior secured loan? It’s priced at SOFR + 300 and 99.5, carries a Baa2 from Moody’s, is secured by GPU servers and the cash flows they generate, and fully amortizes to a 2030 maturity the company describes as aligned with the contracted cash flows and useful life of the infrastructure.
But what’s doing the real work is the announcement: “dedicated to an investment-grade offtaker.” The customer anchors the credit, not an observable liquidation value.
Then there’s CoreWeave, who has borrowed more against GPUs than anyone else on the public record. Its loans are delayed-draw term loans: a group of lenders commits a pile of money up front, and CoreWeave draws it down as GPUs get delivered against a specific customer contract, so the loan and the contract grow together. The biggest one is $8.5B and got an A3 from Moody’s, which CoreWeave called the first investment-grade rated GPU-backed financing.
If a residual value assumption exists anywhere in GPU lending, it should be in these documents, so I read them.
I ran a full-text search of the newest agreement: 35 instances of GPU, and zero for every term a lender would reach for to value hardware against an outside reference. No appraisal, no residual value, no orderly liquidation, no advance rate, no loan-to-value, no borrowing base, no remarket.
How have other used asset industries handled residual value?
How is this emerging for GPUs?
A used car can be priced against the Manheim Used Vehicle Value Index, a benchmark built by statistical analysis of more than 5 million wholesale transactions a year, and sturdy enough that Kalshi has listed contracts settling on it. The same venue lists DDR5 memory spot contracts. In both cases the market can point to an external, transaction-derived price for the thing itself while compute markets have only half the equation.
For GPUs, we do have Kalshi’s H100 ladders which are two-sided binary markets stacked by strike. These are quotable contracts in a still tiny market. Silicon Data also publishes a daily H100 rental benchmark on a Bloomberg ticker and extends it to a 36-month curve. Ornn supplies Kalshi’s settlement.
CME is set to list cash-settled H100 and B200 rental futures on October 5, pending regulatory review. The compute market finally did what commodity markets do. It built a rate.
A rental future is useful to a neocloud whose revenue moves with GPU-hour prices. It does much less for a lender standing in a data room after a default, trying to decide what a particular rack will fetch. Generation, topology, condition, location, power, networking and the cost of redeployment all sit in the basis between the hour and the box.
The academic work is careful about this gap. Federico Bandi and Yinan Su of Johns Hopkins, working with data from Silicon Data and Ornn, find that compute is non-storable, so the usual no-arbitrage link between spot and futures fails (like power markets) and that synthetic futures built from term rental contracts are likely upper bounds on true futures prices.
The two companies closest to this problem are run by the same person, Carmen Li, and between them they’ve built both ends of the market without the middle.
Silicon Data’s answer to the residual is a DCF. Its GPU Residual Value product, launched August 25, takes forward rental rates off Silicon Data’s own curve, assumes utilization decays as newer chips take the workloads, nets out operating costs, and discounts what’s left over the card’s remaining useful life.
That is a real number and it is not a price. It’s the rent index in a different unit. It tells you what a GPU is worth to whoever keeps running it, which is a different question from what a lender gets when they have to sell it, and the second one is the number a rating agency needs.
Compute Exchange, under the same CEO, opened a used GPU marketplace on July 17. It publishes price ranges per model and condition sourced from aggregated supplier quotes, matches a buyer to verified sellers and makes the introduction.
It does not take custody, hold funds or act as a party to the trade, and its own page says the pricing is for reference only. So the model produces a value and the venue produces quotes, and neither one produces a print.
The prints sit with Compute Credit Index Research (CCIR), a benchmark administrator formed in Delaware in March that runs no marketplace and holds no positions. It has three years of executed used-GPU sales: 3,009 sold listings, 10,911 units, $26.3M in total.
To put this in perspective, Manheim runs more than 5 million transactions a year through its own auctions, and three years of the entire documented used-GPU market comes to about 0.3% of one $8.5B CoreWeave facility.
On August 21 the CFTC opened a request for comment on the listing of compute derivatives contracts. Its preliminary view: “compute markets are fragmented and price formation primarily occurs in opaque bilateral transactions, hindering the availability of current and historical price data.” Dominant participants, it adds, “may wield significant pricing power that may lead to manipulability,” and so “compute may not yet exhibit certain of the characteristics of commodities that typically underlie a commodity derivatives market, including fungibility, standardization, and sufficient liquidity.”
The Commission builds its case partly on Bandi and Su, citing the same Johns Hopkins paper three times. Its two tracks finish in the wrong order. The CME filings sit at Approval Pending under a review period that ends September 25, so unless the Commission acts, the contracts are approved by default and begin trading October 5.
The comment period asking whether compute is ready to underlie derivatives at all stays open until October 20. The specific product clears before the general question closes, which means the first regulated compute future could trade for two weeks while the Commission is still collecting answers about whether the market underneath can support one.
Stepping back, there’s clearly a lot of exciting things going on in the compute markets. But, as Sun Tzu once said: “if there is no liquid market but a need for residuals, you must use forwards not futures.”
I do think one day we’ll have enough liquidity to have GPU futures and ABS loans collateralized by the actual GPUs residual value, but I don’t think it will come soon enough. People want (need, begging, pleading) for GPUs NOW and if we don’t figure out a way to hedge risk asap then we might all just be cooked.
Sources
Everything below was read directly. Market figures were checked between August 29 and September 2, 2026; CCIR aggregates are as stamped August 31; the CME and CFTC status is as of September 2.
Filed documents
NVIDIA, Form 8-K filed 2026-08-17 (Item 1.01): the guaranty for the SB Energy PORTS-Pike data centers with OpenAI as tenant, $105 billion cap, about 4.25 GW of IT load. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000069/nvda-20260817.htm
NVIDIA, Form of Residual Value Guaranty, Ex. 10.1 to the Q2 FY2027 10-Q filed 2026-08-26: guaranteed minimum value “comprised of data center, power, and transmission related costs,” “four and one quarter (4.25) GW of critical IT load,” NVIDIA equipment in a separate repossession clause. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda2027q2ex101.htm
CoreWeave, DDTL 5.5 credit agreement dated 2026-08-07, Ex. 10.1 to the 8-K filed 2026-08-10. This is the agreement behind the full-text search (35 instances of GPU; zero for appraisal, residual value, orderly liquidation, advance rate, loan-to-value, borrowing base, remarket). https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/ex101creditagreement.htm
CoreWeave, press release of 2026-03-31 on the $8.5 billion facility: A3 from Moody’s, A (low) from DBRS, “first investment-grade rated GPU-backed financing,” and the delayed-draw mechanics. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000129/ex991.htm and Moody’s rating news https://ratings.moodys.com/ratings-news/462400
Lambda, “Lambda prices $926 million senior secured term loan B” (2026-08-12) and the closing post (2026-08-27): SOFR + 3.00% at 99.5, Baa2 from Moody’s, secured by the GPU servers and their cash flows, fully amortizing to 2030-12-31, “dedicated to an investment-grade offtaker.” https://lambda.ai/blog/lambda-prices-926-million-senior-secured-term-loan-b-facility and https://lambda.ai/blog/lambda-closes-926-million-senior-secured-term-loan-b-facility
Exchanges and the regulator
Kalshi public API, series and markets: the Manheim Used Vehicle Value Index series (KXMANHEIMUSEDCAR), DDR5 spot series (KXDDR5MS), and the H100 monthly ladders (KXH100MS), with Ornn named as the settlement source. https://api.elections.kalshi.com/trade-api/v2/
CME Group and Silicon Data press release, 2026-08-11: H100 and B200 rental futures to list 2026-10-05, pending regulatory review. https://www.prnewswire.com/news-releases/cme-group-and-silicon-data-to-launch-compute-futures-on-october-5-to-unlock-new-way-to-hedge-ai-risks-302848593.html
CFTC industry filings for the Silicon Data H100 and B200 Rental Index Futures (NYMEX), status “Approval Pending (45),” filed 2026-08-11 with the 45-day review ending 2026-09-25. https://www.cftc.gov/IndustryOversight/IndustryFilings/TradingOrganizationProducts/62544 and NYMEX Submission 26-370 https://www.cftc.gov/filings/ptc/ptc08112615405.pdf
CFTC, Request for Comment on the Listing of Compute Derivatives Contracts, 91 FR 54259, published 2026-08-21, comments due 2026-10-20. Every quoted phrase is verbatim from the release, which cites Bandi and Su at footnotes 5, 42 and 43. https://www.federalregister.gov/documents/full_text/text/2026/08/21/2026-17163.txt
Company pages
Manheim (Cox Automotive), Used Vehicle Value Index methodology: “statistical analysis to its database of more than 5 million used vehicle transactions annually.” https://site.manheim.com/en/services/consulting/used-vehicle-value-index.html
Silicon Data: the H100 rental index SDH100RT https://www.silicondata.com/products/silicon-index/h100 ; the GPU Forward Curve (up to 36 months) https://www.silicondata.com/products/forward-curve ; the GPU Residual Value product and its methodology (rental rate x discount x (1 minus operating cost ratio) over expected useful life, rates drawn from the forward curve) https://www.silicondata.com/products/gpu-residual-value ; launch post of 2026-08-25 https://www.linkedin.com/posts/silicon-data_the-gpu-residual-value-launches-today-straight-line-activity-7498011372006424576-idgc
Compute Exchange: the Hardware Market page (price ranges “sourced from aggregated supplier quotes,” introductions only, no custody or funds, pricing “for reference only”) https://compute.exchange/hardware-market ; the marketplace launch release of 2026-07-17 https://www.einpresswire.com/article/927423264/compute-exchange-launches-secondary-gpu-hardware-marketplace-as-demand-for-refurbished-ai-infrastructure-accelerates ; the About page, Carmen Li as CEO of both Silicon Data and Compute Exchange https://compute.exchange/about
CCIR (Compute Credit Index Research LLC): governance page, Delaware LLC formed 2026-03-11, “operates no marketplace, runs no trading desk, and holds no positions” https://ccir.io/documents/governance ; hardware page, three years of executed sales, 3,009 sold listings, 10,911 units, $26.3M https://ccir.io/hardware
Papers and podcasts
Federico M. Bandi and Yinan Su (Johns Hopkins), “(Early) AI Compute Asset Pricing,” arXiv 2607.12156, July 2026: compute is non-storable so the spot-futures no-arbitrage link fails, and synthetic futures from term rentals are likely upper bounds. https://arxiv.org/abs/2607.12156
Hendrik Bessembinder and Michael Lemmon, “Equilibrium Pricing and Optimal Hedging in Electricity Forward Markets,” Journal of Finance, 2002: the standard result that electricity forwards price off expected spot plus a risk premium because power cannot be stored, the analogy behind “like power markets.”
Dwarkesh Podcast with Dylan Patel, the marginal compute discussion referenced above. [add episode link]
Author’s own numbers, not published data: the SOFR + 900 going rate for startup GPU loans, the roughly $50M per MW Anthropic revenue figure, and the 128-node example.
I work at NVIDIA, this is so accurate. Really good article.
the future times is exciting, and also an interesting one to watch.
excellent write!