Depreciation is the biggest risk in a GPU fleet, and now it is tradable.

A GPU system loses most of its market value while it is still running production workloads. That is the uncomfortable arithmetic at the center of every AI infrastructure balance sheet right now. The hardware works fine. The accounting says it has years of useful life left. The market disagrees, and the market is the one that sets the number you can actually recover when you sell, refinance, or write down.

For a long time depreciation on a compute fleet was a passive line item, a straight-line assumption stamped into a model and revisited once a year. It is not passive anymore. The value erosion of an installed base is now a measurable, market-driven schedule, and as of this year it is something a CFO can hedge financially rather than simply absorb. This piece lays out the size of the risk, why the standard accounting treatment understates it, and how an index-referenced position lets an operator put a floor under the residual value of a fleet that is already racked and running.

The size of the risk

Start with the magnitude, because it is easy to underweight. The used market for the prior accelerator generation tells the story plainly. Refurbished H100 units traded as high as roughly $50,000 in the middle of 2024, then fell sharply as supply loosened. By 2025, second-hand H100s under a year old were changing hands in the $18,000 to $25,000 range, and units at two years and beyond were down to roughly $7,000 to $12,000. One large cloud reportedly cut its on-demand H100 pricing by about 44% in a single move in mid-2025. None of that is gradual. It is a repricing event that lands on the asset whether or not the depreciation schedule has caught up.

Put that next to the cost basis of a current system and the exposure becomes obvious. A complete HGX B200 NVL8 server, something like a Supermicro SYS-A22GA-NBRT, a Gigabyte G894-AD1-AAX5, or a Dell PowerEdge XE9680L, lands in the low-to-mid six figures fully built with eight Blackwell GPUs, dual Intel Xeon 6980P or AMD EPYC Turin head nodes, ConnectX-7 fabric, and the supporting NVMe and memory. A fleet of a few hundred of those systems is a nine-figure asset. If the market value of that asset erodes by a third or more inside one generation cycle, the dollar amount at risk is larger than almost any other controllable line on the infrastructure P&L. Power is a meaningful cost. Depreciation, measured at market rather than on a schedule, is frequently the larger number.

That last figure is worth sitting with. In late 2025, Michael Burry argued that the five largest AI hyperscalers would collectively understate depreciation by roughly $176 billion between 2026 and 2028, on the logic that hardware built for a 12 to 18 month obsolescence cycle was being depreciated over four to six years, flattering sector earnings by around 20%. You do not have to accept the headline number to accept the mechanism. The gap between the schedule and the market is real, it is large, and it accrues to whoever holds the asset.

Why straight-line depreciation misleads

Straight-line depreciation is an accounting convention, not a law of physics and not a GAAP mandate. It spreads cost evenly across an assumed useful life because that is simple and auditable, not because economic value actually drains evenly. Over the last two years, Amazon, Alphabet, and Microsoft each extended the official useful life of their servers to six years, up from a prior three to four year standard. That change reduces annual depreciation expense and improves reported margin. It does not change what a buyer will pay for the box.

The problem is that the real obsolescence schedule is event-driven. Value does not decay smoothly. It steps down when a successor ships and the market reprices substitution. NVIDIA announced the Blackwell platform on March 18, 2024, with B200-class GPUs carrying 208 billion transistors on a custom TSMC 4NP process and the GB200 NVL72 cited at up to 30x LLM inference performance versus the same count of H100s. Then, at GTC 2025, NVIDIA introduced the B300, or Blackwell Ultra, which began shipping in the second half of 2025. The B300 carries 288GB of HBM3e, roughly 50% more memory than the B200's 180 to 192GB, and around 55% higher dense FP4 throughput, at 1,400W per GPU versus the B200's 1,000W. The day a B300 system is generally available is the day the market starts discounting the B200 forward, regardless of where the straight-line curve sits.

That is the obsolescence delta, and we have written about how it surfaces in a forward curve in The forward curve, the spot price, and the obsolescence delta. H100 value over time is the cleanest illustration that this is repricing rather than scheduled decline: at roughly two years post-launch, used H100s traded at a median near 61% of the contemporaneous new price, then moved to around 69% by 2025 at roughly three years post-launch. A smooth schedule does not move back up. A market does, because the denominator (the new price) is falling too. If your model treats depreciation as a fixed annual percentage of original cost, it is describing an accounting artifact, not the recoverable value of your fleet.

The fleet-as-aircraft analogy

The discipline that already solved this problem is aviation finance. An airline does not think about a 737 as a depreciating block of metal on a straight line to zero. It thinks in residual value, the price the aircraft will fetch in the secondary market at each point in its life, and it manages the fleet against that curve. Lease rates, sale-leaseback structures, part-out economics, and fleet renewal decisions all key off residual value, and there is an entire market of appraisers and lessors whose job is to keep that number honest.

Compute asset teams at the largest operators are moving toward the same posture. The question stops being how many years until the books say zero and becomes what this fleet is worth in the secondary market at each quarter from here, and what the exposure is if that curve drops faster than planned. Once you frame it that way, two things follow. First, residual value, not utilization, becomes the variable you manage. A fully utilized B200 fleet that is about to be repriced by B300 substitution is still losing balance-sheet value every week. Second, residual value, like an aircraft's, can be hedged once there is a liquid, observable reference price for the asset. Aviation has its appraisal benchmarks. Compute now has the Rillor Compute Index.

How to hedge it financially

Here is the mechanism, stated plainly. You own an installed base of B200 systems. Its forward value is falling as B300 supply ramps. You want to offset that erosion without selling the hardware, because the hardware is busy earning revenue. So you take a short position referenced to the B200 forward price at a licensed trading venue. If the B200 forward value falls, the short gains, and that gain offsets the markdown on your physical fleet. If demand surprises to the upside and B200 holds its value, the short costs you, but your fleet is worth more, so the position has done its job either way. That is a hedge, not a bet. It converts an open-ended depreciation risk into a known, bounded cost.

The reference price that makes this possible is the Rillor Compute Index, a 30-day rolling-blend forward price per SKU computed from active Rillor forward contracts. Because the index is built from real bilateral contracts between verified buyers and sellers rather than from quotes or surveys, it is a defensible settlement reference, which is exactly what a cash-settled derivative needs. Third-party exchanges and funds license the index as a settlement feed and API and build cash-settled instruments against it: a B200 forward, a B300 perpetual, a basket. Your treasury desk transacts those instruments at the licensed venue. The hedge for an entire fleet can be expressed in a handful of index-referenced positions sized to the SKU mix you hold.

This is also why the hedge belongs to a fund or a desk and not to a procurement team. A fund can express a view on compute prices without owning a single GPU, transacting only the cash-settled instruments that licensed venues build against the index. The same plumbing serves an operator who wants to neutralize the depreciation on a fleet it does own. The fleet stays racked and earning, while the price risk it carries moves to a desk that can hold it.

Why this stays off Rillor's physical book

A point that matters for any CFO evaluating counterparty and regulatory risk: the depreciation hedge does not live on Rillor's book, and it cannot. Rillor's contracts are bilateral OTC forwards on complete OEM GPU systems with the intent of physical delivery, always. A 10% deposit is posted at execution, the balance at delivery, the seller posts a performance bond, an independent escrow agent holds funds, and the end customer of record is captured for NVIDIA channel compliance. The CFTC distinguishes a forward contract, where a commercial buyer and seller agree to deliver a specified quality and quantity at a future date, from cash settlement, which pays the cash value of an underlying or an amount based on the level of an index. Rillor sits squarely on the delivery side of that line and never cash-settles its own contracts.

The hedge is, by definition, a cash-settled product. It pays the change in an index level, not a server on a pallet. So it lives downstream, at the licensed venues that build instruments against the Rillor Compute Index. That separation is deliberate and it is the load-bearing design choice. Rillor owns and controls the index and licenses it as a feed, which is the moat, and we keep the physical forward book clean of speculative cash-settled positions. You get the hedge you need at a regulated venue, and Rillor stays a delivery market. The two pillars reinforce each other precisely because they do not mix.

SKUs decay differently, so size the hedge accordingly

Not every accelerator is on the same curve, and a hedge that ignores the difference will be mis-sized. The single most important input is whether a successor is shipping.

The B200 is the textbook steep case. The B300 is in market, it is a clear capability step (more HBM3e, higher FP4 throughput), and substitution is live, so the B200 forward decays on the obsolescence curve and the short should be larger and nearer-dated. The GB200 NVL72 follows the same logic at rack scale, with the GB300 NVL72 already shipping, so the exposure is best hedged at the rack rather than the node.

The AMD Instinct MI355X is the instructive counterexample. Built on 4th-generation CDNA with 288GB of HBM3e, 8TB/s of bandwidth, 256 compute units, around 20.1 PFLOPS of FP4/FP6 and 10.1 PFLOPS of FP8 at a 1.4kW TDP, it reached general availability in October 2025 with no publicly shipping successor on the board. With no announced next part pulling the forward down, MI355X value is governed by demand rather than by a substitution event. Its decay is flatter and its timing is less certain, which argues for a smaller, longer-dated hedge. Holding the two side by side is the whole point: a one-size depreciation assumption applied across a mixed fleet of B200, GB200, MI355X, and H200 will be wrong on every line. You can see the live per-SKU forward prices on the marketplace and the full catalog under SKUs.

Translating the hedge into a residual-value floor

The deliverable a CFO actually needs is a number that goes into a model, so end here. The hedge converts an uncertain depreciation curve into a residual-value floor you can defend in a board deck.

Mechanically, you do three things. First, replace the straight-line assumption in your fleet model with the index forward curve for each SKU you hold, so the model reflects market value rather than schedule value. Second, size index-referenced shorts at a licensed venue against the SKUs with the steepest, nearest decay, weighting the B200 and GB200 exposure heavily and the MI355X exposure lightly. Third, the strike of those positions becomes the floor: below it, the hedge gains offset the fleet markdown, so the modeled recoverable value of the fleet does not fall past that line. The all-in cost of carrying that floor is the premium on the positions, a known number you can amortize, rather than an open-ended impairment risk you cannot.

The result is a balance sheet where depreciation on the largest asset class is bounded rather than discovered after the fact. That is the shift. The risk did not get smaller. It got tradable, and a risk you can trade is a risk you can put a price on and stop fearing.