Babbage Index · Research · 022

A GPU’s Useful Life Can Be Read From Its Rental Curve

Depreciation schedules fix an accelerator’s useful life once a year, by memo. The H100 rental market now quotes the asset’s cash yield across a three-year term structure. Read through a simple valuation model, that curve can imply a life close to what hyperscalers book, under assumptions stated here — and shows how far the answer moves when the curve or those assumptions do.
Babbage Research · 022·Published 2026-11-18

A straight-line schedule over six years writes a server down by 16.67% of its cost every year, whatever happens to the price of what it produces. That is what a schedule is: a mechanism with one parameter, useful life, and no variance. Between October 2025 and March 2026, SemiAnalysis’s H100 one-year contract index went the other way, from $1.70 to $2.35 per GPU-hour, a rise of 38%2. The schedule did not notice, and was not designed to.

That design was right while a server’s output was priced by an internal chargeback rate nobody traded. Useful life could not be observed, so a convention beat a guess. For H100s it can now be estimated: rental prices are published daily, a forward curve quotes committed terms out to three years, and one exchange has filed futures on a rental index and another has announced them. This article turns that curve into an implied useful life, shows the arithmetic, and argues that the instability of the answer, more than its level, is the finding.

The rent curve, in stated units

All prices here are in US dollars per GPU-hour for one H100 GPU; no currency conversion is involved. Babbage’s own series are H100 SXM, quoted per GPU; Silicon Data’s post does not state a form factor. Silicon Data’s neocloud H100 forward curve, read from its portal charts as of July 19, 2026, put spot rent at about $2.72 and the 36-month committed term rate at about $2.381. Silicon Data labels the first a spot price and does not state its billing increment; the term rate is a flat price for three years of committed use. The term rate sits about 12.5% below spot (Silicon Data rounds it to about 13%) — the curve is in backwardation, meaning later delivery is priced lower than delivery now.

That curve is Silicon Data’s most recent published reading, from July 19, 2026. Since then, to November 7, 2026, Babbage’s neocloud on-demand average has moved 3.6% and its marketplace median 28.9%, so the base case below is a July estimate, not a current one.

Read the 36-month rate as a level price for a stream of rent that decays at a constant rate δp a year. If the curve is an unbiased forecast, the term rate equals the average of that decaying stream over three years, and solving for the rate that makes the average 0.875 of spot gives δp = 9.11% a year. Reading the term rate instead as the price at the end of the three years gives 4.45%. The levelized reading fits how a fixed-rate contract is priced, so it is the base case; the gap between the two is the first sign of how much interpretation sits inside one number. The inputs are rounded to the cent, so treat the second decimal of these rates as arithmetic, not information.

Babbage’s own history puts that spot quote in context. On July 19, 2026, the same day, Babbage recorded the Vast.ai marketplace median for an H100 SXM GPU at $1.631 and the average on-demand list price across CoreWeave, Lambda, Nebius and Crusoe at $4.475. Silicon Data’s $2.72 sits between them. None is wrong; each is a different definition of “the” H100 hour, which is the problem 014 set out10.

From rent decay to value decay

An operator does not keep the rent; it keeps rent minus cash operating cost. Assume that cost, c, is $0.85 per GPU-hour, covering power, cooling, space, network, staff and the GPU’s share of its host, and that it stays flat while rent decays. That assumption is illustrative, not measured. The margin at acquisition is m0 = $2.72 − $0.85 = $1.87. Because the cost does not fall, the margin falls faster than the rent, by the operating-leverage factor λ = p0/m0 = 1.455. Near acquisition, the margin decays at δm = λ·δp = 13.25% a year. The model holds that rate constant over the life, a constant-elasticity approximation; with cost literally flat, margin falls faster late in life, and the implied life below is 4.64 years rather than 4.85.

Value a GPU cohort — the GPUs bought together, at one price and one time — as the present value of its remaining margin, discounted at a cost of capital r and running to the end of physical service life T. Let k = δm + r:

V(t) = m0 · e−δmt · [ 1 − e−k(T−t) ] / k
The value at time t of margin that starts at m0 and decays at δm a year, discounted at r, until the hardware is retired at T.

Assume the cohort was bought at exactly that value, so its cost is V(0). On these inputs V(0) is 5.91 dollars per GPU-hour of annual capacity times years; at full occupancy, 8,760 hours a year, that values one GPU at about $51,800. The rate at which it loses value at acquisition, −V′(0)/V(0), is the economic depreciation rate. Straight-line over L years writes off 1/L of cost a year, so the life that matches the economic rate at acquisition is:

1 / L* = δm + k · e−kT / ( 1 − e−kT )
The implied life L* is the straight-line life whose first-year rate matches the economic rate at acquisition.

With an assumed cost of capital of 11% a year and an assumed service life of 6 years, the economic rate is 20.63% a year and the implied life L* is 4.85 years. The booked lives in the filings below run from five to six years. On these inputs, the controllers are roughly right only if cost equals the present value of the curve’s margin. A cohort bought for less loses a larger share of its cost each year, so its implied life is shorter: at an illustrative $40,000 per GPU, the same first-year loss in value is 26.7% of cost, a life of 3.74 years. The purchase-cost assumption moves the answer more than any input in the tables below.

How far the answer moves

The level is not the point. Hold the other inputs and vary the rent decay, and the implied life swings widely:

Rent decay caseδp / yrδm / yrImplied life L*
No decay0.0%0.0%8.50 yr
Point-forward reading of the curve4.5%6.5%6.29 yr
Levelized reading of the curve (base case)9.1%13.3%4.85 yr
Fast decay (illustrative)20.0%29.1%3.02 yr

With no decay at all, L* is 8.50 years — longer than the 6-year physical life, because a discounted stream loses value slowly at first, the same reason a mortgage repays little principal in its early years. At an illustrative 20% a year of rent decay it is 3.02 years. The two readings of the same curve alone give 4.85 and 6.29 years.

The note this article expands said a two-to-three-year life would need rent to decay at about 20% a year. That holds only for three years. The required decay rises steeply as the target life shortens:

Target lifeRequired δp / yrMultiple of curve-implied δp
5 yr8.5%0.93×
4 yr13.1%1.44×
3 yr20.2%2.22×
2.5 yr25.6%2.80×
2 yr33.2%3.64×

A three-year life needs 20.2% a year, 2.2× the curve’s rate; a two-year life needs 33.2%, 3.6×. The curve as quoted does not support the short lives sometimes argued for accelerators, unless one expects rent to fall far faster than the people pricing three-year commitments do.

The other inputs matter too. Each row below moves one input from the base case and holds the rest, including the 9.11% rent decay:

Input movedLowL* at lowHighL* at high
Cash operating cost c$0.605.13 yr$1.104.52 yr
Cost of capital r8%4.65 yr14%5.04 yr
Physical service life T5 yr4.25 yr7 yr5.35 yr
Spot rent p₀$2.1024.52 yr$4.4755.21 yr

A higher operating cost shortens the life, because it raises operating leverage. A higher cost of capital lengthens it, for the mortgage reason above. Service life is the largest lever in the table: the base case’s closeness to booked lives partly reflects the assumed 6-year T, and a shorter 5-year T gives 4.25 years. The spot-rent row applies the same curve ratio and cost to Babbage’s marketplace median on November 7, 2026 ($2.102) and to the July neocloud list average ($4.475); it is illustrative, since neither series has its own term curve. A cheaper hour carries more operating leverage, so the same percentage decay erodes value faster.

Where the model holds and where it breaks

The model is deliberately simple, and each simplification marks a regime. It assumes rent decays at a constant exponential rate; a new generation that resets the price of an H100 hour in one step is outside it. It holds margin decay at its acquisition rate, which with flat cost overstates late-life margin; integrating the flat-cost margin directly gives 4.64 years instead of 4.85. It sets purchase cost equal to V(0). It assumes full occupancy, so it values a GPU-hour sold, not one available; and, as 008 argued, even a sold hour is a unit of occupancy rather than of useful work12. It ignores failures and assumes the hardware runs to T. And it treats the forward curve as an unbiased forecast; a term rate also carries the lessor’s credit view, utilization risk and any risk premium, which would make the curve overstate expected decay.

The inputs are thin. Silicon Data’s figures are approximate, read from charts, neocloud-only and from one July date. The leverage factor blows up as spot rent approaches cash cost: when the margin is small, a modest percentage fall in rent is a large fall in margin. So the conclusion holds in the regime the curve describes — a healthy margin over cash cost and gradual decay — and fails in a price slump or a generation step, which is exactly when a carrying value matters most.

Booked lives are set by hand

Hyperscalers set server lives by periodic assessment, and the direction has not been uniform:

CompanyBooked lifeChange disclosed
Microsoft6 yr3Server and network equipment extended from four to six years, effective fiscal 2023
Alphabet6 yr4Servers and certain network equipment extended to six years, effective 2023
Meta5.5 yr5Certain servers and network assets extended to 5.5 years, effective January 2025
Amazon5 yr (subset; others 6)6A subset of servers and networking equipment shortened from six to five years, effective January 1, 2025

Microsoft, Alphabet and Meta lengthened lives; Amazon shortened one subset. Amazon’s change reduced net income by $298 million in the third quarter of 2025 and $677 million over the first nine months6. These disclosures cover server classes, not GPUs specifically, and later filings may revise them. The point is the mechanism: one parameter moves earnings and residual-value policy together, so it moves rarely, late and in round steps. Impairment testing does not fill the gap. As Meta describes it, the test runs only when events or circumstances indicate the carrying amount may not be recoverable, compares it with undiscounted cash flows, and works at the level of an asset group whose cash flows are largely independent of others5. Such a group can be wide enough to net a sagging H100 cohort against a scarce B200 one.

Marking instead of scheduling

The alternative is to keep depreciation as the allocation mechanism but take its parameter from the market: mark each cohort’s residual value to the curve, report the implied life alongside the booked one, and treat the gap as risk to be managed rather than a disclosure. Between July 19, 2026 and November 7, 2026, Babbage’s neocloud on-demand average rose 3.6% and its marketplace median 28.9%; a mark would have moved, and a schedule did not.

The venue for selling that risk does not yet exist. NYMEX filed Silicon Data H100 and B200 rental futures as Submission 26-3707, but they have not listed: CME’s letter of November 4, 2026 extended the CFTC review to January 8, 20278. ICE and Ornn announced cash-settled GPU compute futures on a transaction-based index in May 2026, subject to regulatory approval and with no launch date9. 020 covers what those contracts need from their indexes11.

Until they list, the substitute is a take-or-pay contract with a renter, often a neocloud. It looks like a hedge but carries wrong-way risk: the renter’s ability to pay depends on the price being hedged, so the protection weakens when it is needed. That is a different risk from the volume risk 013 traced in memory take-or-pay agreements, where the buyer, not the seller, carries the commitment13.

One counter is missing everywhere: realized yield per acquisition cohort. Operationally, for each cohort and month, take billed revenue per GPU-hour available, net of power metered at the rack, and subtract the forward rate quoted at acquisition for that month. That difference is the cohort’s basis against the curve. Without it, a cohort whose margin has fallen to zero reads as an occupancy problem, and gets a scheduler change and a sales push instead of a hedge or a retirement.

What Babbage’s indices can and cannot see

Babbage’s GPU index records on-demand list prices from hyperscalers and neoclouds and a marketplace median, all per GPU-hour. That gives the model’s p0 under several definitions, and its movement over time. It does not give δp. Babbage collects no committed-term prices, so it cannot observe backwardation, and the forward line on its GPU page is a statistical projection of the index’s own recent drift, not a quoted or traded term rate; it should not be used as a decay input. Babbage also sees no operating cost, occupancy or cohort data, so it cannot compute λ or the basis counter above. What it can do is show which spot definition a mark uses and how far that definition sits from the others, as the July comparison above does.

References

  1. Silicon Data (C. Li), “GPU Futures: How Compute Is Becoming a Tradable Commodity,” July 17, 2026; forward-curve table, neocloud only, figures read from portal charts as of July 19, 2026. silicondata.com
  2. SemiAnalysis, “The Great GPU Shortage – Rental Capacity – Launching our H100 1 Year Rental Price Index,” April 2, 2026. semianalysis.com
  3. Microsoft, Form 10-K for fiscal 2022, accounting policies (change in estimate). sec.gov
  4. Alphabet, Form 10-K for 2022, property and equipment (change in estimate). sec.gov
  5. Meta Platforms, Form 10-K for 2024, property and equipment. sec.gov
  6. Amazon, Form 10-Q for the quarter ended September 30, 2025, Note 1. sec.gov
  7. NYMEX Submission 26-370, Silicon Data H100 and B200 GPU rental futures, August 2026. cftc.gov
  8. CME Group, letter to the CFTC extending the review of Submission 26-370 to January 8, 2027, November 4, 2026. cftc.gov
  9. ICE, “ICE and Ornn to Launch GPU Compute Futures Contracts,” May 19, 2026. ir.theice.com
  10. Babbage Research 014, Defining the Compute Behind the Price. babbageindex.com
  11. Babbage Research 020, What Compute Indexes Can Learn from Other Rental Indexes. babbageindex.com
  12. Babbage Research 008, The GPU Hour Is the Wrong Unit of Compute. babbageindex.com
  13. Babbage Research 013, The Memory Discount That Strands the Compute. babbageindex.com
Editor’s note
Expanded from Wei’s weekly architecture note “Residual Vega.” Corrections from the note: the NYMEX contracts have not listed; the base-case sensitivity row is 4.85 yr, not 4.90; the decay required for a short life is given as a range; Meta’s 5.5-year change was an extension.

Provenance note, per house style. Quoted: the Silicon Data spot and 36-month rates (approximate, ref 1); the SemiAnalysis index values (ref 2); booked lives, Amazon’s net income effect and Meta’s impairment policy (refs 3–6); futures status (refs 7–9). From Babbage’s committed GPU history: the July 19, 2026 and November 7, 2026 H100 figures. Assumed and illustrative: cash operating cost $0.85, cost of capital 11%, service life 6 years, purchase cost equal to V(0), full occupancy, the $40,000 purchase cost, the 20% fast-decay case, the sensitivity ranges, and the use of one curve ratio at other spot levels. Computed on this page: the backwardation, both decay readings, operating leverage, margin decay, economic rate, implied lives (closed form and flat cost), the per-GPU value, required decay rates, the July to November moves and all values in the three sensitivity tables.