The Second Rental

The idea

This note began with a report that Vista's research desk sent me: Katherine Bourzac's "Hyperscale data centers in the AI era", written for MIT Technology Review and published in February 2026. Its discussion of how quickly AI hardware goes out of date prompted a plain question: when does replacing a GPU actually pay? The desk also sent me a second MIT Technology Review report, Edd Gent's "Understanding World Models", published in August 2026 and built on interviews from June. It explains AI systems that model how an environment behaves, and their uses in generating synthetic training data, testing the policies that drive robots, and planning actions. It added a possibility that complicated the answer: could offline simulation of that kind give older GPUs useful work?

Bourzac's report carried a question that Mihir Kshirsagar of Princeton's Center for Information Technology Policy had posed on October 15, 2025, as "Lifespan of AI Chips: The $300 Billion Question", after the sum the largest builders were expected to spend on AI infrastructure in 2025 alone. His answer was blunt: the chips, he wrote, have "a useful lifespan of one to three years", yet companies depreciate them over five or six. And in September CoreWeave, which depreciates its GPUs over six years, showed investors an A100 contract, on a chip launched in 2020, that "Extends into 2029", and a three-year H200 renewal at a "Premium to original contract."

So I set out to model the decision, drafting with the help of an AI assistant. On the evening of October 4 I sent Vista's Desk the manuscript, called "Conditional replacement thresholds for AI compute", with its model, its inputs and three suites of tests, and asked the Desk to find what was wrong with it. Over three rounds it found real faults, among them that my first version let an operator walk away from a customer's protected contract for nothing, and I fixed them. The Desk was wrong several times too, and conceded each time it was shown why. By the third round it found nothing it called material.

The argument has two loud sides. Michael Burry wrote in November 2025 that the largest builders would understate depreciation by $176 billion from 2026 to 2028 by stretching the lives of their chips; CoreWeave answers that its old chips keep renting at nearly their first prices. Much of the ground between them has been walked already, a good deal of it in the last few weeks, and the people who walked it are named where their work comes up. What we have not found elsewhere is my draft's ceiling on a swap, worked per megawatt in closed form, with the contracts, the dark months, the retrofit and the resale kept separate, and set against CoreWeave's own equipment cost per megawatt; and a break-even life for CoreWeave's equipment that counts the cost of borrowing and the data center's own equipment, set beside the company's renewals and the market's term prices.

The superintendent's question

For more than twenty years Harold Zurcher kept the buses of Madison, Wisconsin, on the road. He was superintendent of maintenance at the city's Metropolitan Bus Company, and month after month, bus by bus, he made the same small decision: put in a new engine now, or let the old one run another month. A new engine cost a good deal of money. An old one cost more to maintain with every mile, and the longer it ran the likelier it was to fail on the road. In 1987 the economist John Rust published a paper with the superintendent's name in its title, "Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold Zurcher." From ten years of monthly records for 104 of the company's buses, Rust worked out what Zurcher's choices revealed about the costs he was weighing, and the paper became a classic, still taught to students of economics.

A GPU fleet puts Zurcher's question with two changes. The first is that nothing wears out the way an engine does. A chip does the same work in its sixth year as in its first; what changes is what it can earn. The second matters more. A new engine displaced nothing but the old one. A new chip in a full data center displaces whatever else could have used its power, and in 2026 power is the thing a data center cannot simply buy more of. So the question an operator asks of an old chip is never quite whether it is worn out. It is whether the megawatt it stands on could be earning more.

The shadow price of a megawatt

Start with what is scarce. In 2026 a data center can order chips, at a price and with a wait, but it cannot order electricity. A new connection to the grid takes years, so the power a site already has is the thing every chip in it competes for. Economists call the value of one more unit of a binding constraint its shadow price, and my draft is, at bottom, an attempt to put a number on the shadow price of a megawatt.

It begins by counting what fits. A server of eight A100s, the chip NVIDIA launched in 2020, draws at most 6.5 kilowatts; eight H100s draw 10.2 and eight B200s 14.3. Allow a tenth more for networking and storage and a fifth more for cooling, and one megawatt from the utility holds about 930 A100s, 590 H100s or 420 B200s.

GPUs per megawatt = 1,000 × n / [P × (1 + e) × PUE]

Here n is the eight GPUs in a server, P the server's maximum draw in kilowatts, e the allowance for networking and storage, and PUE the power the whole building uses for each unit the computers use.

Then it prices the megawatt. On October 4, CoreWeave's posted rate for an eight-GPU A100 machine came to $2.70 a GPU an hour, for H100s $6.16 and for B200s $8.60. Rented every hour of a year at those rates, a megawatt of A100s would bring in about $22 million, and a megawatt of either newer chip about $32 million. The newer chips earn some 45% more from the same power, though their machines also carry eight times the local storage. Lambda's price list tells a different story. It asks about what CoreWeave asks for an A100 but only about $4 for an H100, so at Lambda a megawatt of H100s brings in less than a megawatt of A100s. A price list is a seller's opening line, not a market, and these two disagree.

So should an operator pull out its A100s and put in H100s? The new chips earn more, but they cost money, the hall goes dark while the swap is made, and the wiring and cooling may need work. My draft asks the question the way an engineer would. Hold everything else fixed, and ask how much the new equipment can cost before the swap stops paying. Call that ceiling K*:

K* = PV(new) − PV(old) − R − X + S + ΔV

PV(new) and PV(old) are the present values of each path's operating cash over the horizon, after power, service and the site's fixed costs; R is the retrofit; X anything paid to end the old customers' contracts; S what the old chips fetch if sold; and ΔV the difference in what the two fleets are worth at the end, discounted.

The answer on CoreWeave's prices, over three years, with 70% of hours paid, 60% of the posted price realized, three months dark for the swap and $2 million of retrofit, is about $6 million a megawatt, if the old A100s fetch nothing when they come out and neither fleet is worth anything after three years. That is the most the H100s for a megawatt could cost and still pay for themselves. Sold at Silicon Data's September estimate of an A100's value, about $5,000, the old chips would add some $4.6 million a megawatt and lift the ceiling to about $10.6 million. CoreWeave's own books carry about $23 million of technology equipment for each megawatt it has running: $33.8 billion at cost against about 1.5 gigawatts of active power at the end of June. Either way the swap loses money, and the A100s stay.

My draft's base case, about $13.5 million a megawatt a year for H100s at 70% of hours paid and 60% of the posted price, sits between CoreWeave's own revenue per megawatt of active power in the second quarter, about $7 million a year, and the roughly $40 million a megawatt it says short contracts signed in the third quarter fetch.

From my draft's model, run from its own code and inputs. Prices posted October 4, 2026, divided by eight; 70% of hours paid, 60% of the posted price realized, power at 8 cents a kilowatt-hour, three months dark and $2 million of retrofit for a swap, 12% discount rate. A paid share above 100% means the swap cannot pay at that cost.

The old chips are also cheap to keep. Under the same assumptions, a megawatt of A100s covers its power and service when about one hour in ten is paid, and covers the site's fixed costs too at about one in five. Nothing on the income statement forces them out. The only thing that can is a better use for their megawatt, and the draft's countercase shows what that takes: an old business already weak (four hours in ten paid, prices falling 30% a year, no contract left to protect it) and a new one strong (nine in ten paid on a three-year contract). Then the ceiling rises to about $42 million, and the swap pays even at $20 million. What retires a chip in a full building is its customers and its rivals.

Others got here before my draft. When NVIDIA argued in March 2025 that, megawatt for megawatt, its new chips left the old ones with little to do, Glenn Lockwood answered that the argument said nothing about the price of the new equipment. Mihir Kshirsagar, in a second post in December 2025, set an owner comparing the revenue per kilowatt of old and new chips in a building short of power; Square Tower wrote in January that old chips earn only above the opportunity cost of new ones; and American Compute's model of whether to refresh 256 H100s, published in April, found that replacing them with B200s wins at 80% utilization, because it holds the number of chips fixed rather than the power, and the B200s rent for about twice as much. In September Jared Watkins separated three lives for a GPU: the years until its rent no longer covers the cash it costs to run, about twenty-five by his extrapolation; the year or so it takes to pay for itself; and the years until a newer chip wins its slot, 5.6 for an H100 and 9.9 for an A100, depending on whether the building has the power and cooling the newcomer needs. He put a liquid-cooling retrofit at about $2 million a megawatt, the figure my draft assumed. What the draft adds is the swap written out as a ceiling per megawatt, with the contracts, the dark months, the retrofit, the resale and any payment to old customers each kept separate, so an operator can put in its own numbers and see which one decides.

The one-hoss shay

In September 1858 The Atlantic Monthly printed a poem by Oliver Wendell Holmes about a deacon who noticed that carriages never simply wear out. They break down, and they break down at the weakest spot. So he built one with no weakest spot, every part exactly as strong as every other, and it ran for a hundred years to the day and then went to pieces "all at once, and nothing first," Holmes wrote, "just as bubbles do when they burst." Economists who measure capital borrowed the deacon's carriage, the one-hoss shay, as the name for a machine that gives full service every year of its life and then none.

A GPU is very nearly a one-hoss shay. An A100 does the same arithmetic in its sixth year that it did in its first. That is why the argument about how long these chips last keeps going in circles: one side points at chips still running and the other at prices still falling, and both are right. What a chip can do and what it is worth are different things. Its worth is what is left of its rentals.

Harold Hotelling put this in a single line in 1925. The value of a machine at any moment is the present value of the net rentals it has left to earn, and depreciation is how much that value falls in a year:

V(t) = Σ_{s > t} c(s) / (1 + r)^{s − t}    D(t) = V(t) − V(t + 1)

where c(s) is the cash the chip earns in year s after power and service, and r the cost of the money tied up in it. A shay that earns the same rent every year loses value slowly at first and quickly at the end, like a mortgage being paid down. A chip whose rent falls each year loses value fastest at the start. Charles Hulten and Frank Wykoff, studying the prices of used machines and buildings in 1981, found that the second pattern is the usual one: for most of the assets they studied, value fell by a roughly constant share each year.

For GPUs the curve has been drawn already. Chris Zeoli worked through an H100 bought in late 2023 for about $40,000 and found that half of what it netted over six years came in its first two; six years, he concluded, is a defensible length with the wrong shape. Silicon Data publishes a monthly estimate of what used GPUs are worth from discounted forward rents, and Benedict Amissah-Ocran and Mark Tettey Ayumu set out the same method for lenders in August. The measure Hulten and Wykoff relied on, the price of used machines, is starting to appear for GPUs, and it points to long lives. Kai Golden, Grant Benson and Adam Phillips studied Hydra Host's own 67 sales of used GPUs, $136 million of them between 2023 and early 2026, and found that chips two and three years old fetched about 58% of their launch list price. Luxor's index of asking prices put refurbished eight-GPU H100 machines at about 76% of new in August. Asking prices are not sales, and Hydra raises money for GPU funds, but neither is anywhere near the fifth of the new price a short life would imply. What follows applies Hotelling's line to CoreWeave's own numbers.

The builders have each made their own guess, and the guesses have moved, mostly toward longer lives. Here is what they tell their shareholders about servers:

Each company's annual or quarterly report for the year of the change (Nebius in its Form 20-F and Form 6-K reports). Amazon's 2025 change covered a subset of servers and networking equipment.

A GPU on a take-or-pay contract is a shay twice over. It earns full rent every month until the contract's last day, and then whatever the market will pay. Put CoreWeave's numbers into Hotelling's line and the shape of the bet appears. In 2025 its equipment earned about 23 cents of EBITDA a year for each dollar it cost. A five-year contract at that rate brings back $1.15 for each dollar, before interest; discounted at 10%, it is worth about 87 cents. On the fleet's average yield, then, the first contract does not quite pay for the chip. CoreWeave sees it differently: "we have built a business whose economics do not rely on recontracting after initial customer term," its finance chief said in August, and the illustrative contract it showed analysts in 2025, as Kerrisdale Capital reconstructed it, earned far more than the fleet's average. Either way, the second rental is where any surplus lies. If the chip keeps working for eight years and keeps nearly all of its first cash margin, the dollar is worth about $1.21; at half that margin, about $1.05; if nobody rents it again, 87 cents.

The six-year life in the accounts sets its own bar. At the end of a five-year contract, straight-line depreciation leaves a sixth of the chip's cost on the books. To earn that back over three more years, the chip must keep about 30% of its first cash margin; over a single year, about 80%. The cost of running a chip does not shrink with its rent, so in price terms the bar is higher: re-renting at roughly half the first price for three more years, 47% to 58% depending on how much of the rent goes to power, staff and buildings, or at 85% to 88% for one. CoreWeave's own evidence clears it. On its November 2025 call, the chief executive, Michael Intrator, said the company's first expiring contract for more than 10,000 H100s had been renewed two quarters early "at a price within 5% of the original agreement", after a first contract that, given when H100s began shipping, can have run only two or three years. In September it showed investors an A100 contract, on a chip launched in 2020, that "Extends into 2029", priced, it said, in line with its usual one-year terms. The market is less sure. Five-year rentals listed by Ornn in September imply that a chip rented for five years today would re-rent, when the term ends, at about 54% of its contract price for an H100, 46% for Blackwell, 32% for the H200 and 79% for the A100: about at the bar for the newest chips, below it for the H200, and comfortably above it only for the chip whose fall is mostly behind it.

The rent record

My draft assumed that once a chip's price protection runs out, its rent falls 15% a year. That number deserves a test, because the whole second rental hangs on it, and there are two records to test it against: the prices the clouds post, and the prices the market pays.

The posted prices first, and they have not fallen at all. CoreWeave's on-demand rates are the same in every archived copy of its price page since each first appeared: $4.76 an hour for an H100 since April 2023 and $2.21 for an A100 since August 2022, counting the GPU alone. Lambda's have risen, its eight-GPU H100 machines from $2.59 a GPU-hour in 2023 to $3.99 today. CoreWeave also says its average H100 price at the end of 2025 was within 10% of where it began the year, that its A100 price rose in 2025, that prices for every chip from the A100 to the newest rose in the first quarter of 2026, and that in July it raised them about 25% across the board. The posted rates did not move in July, so that rise was in prices it does not post. On its own word, old chips have been getting dearer.

The market's record is steeper, and stranger. SemiAnalysis's index of what H100s rent for fell about 32% in 2024 and 29% in 2025, then rose 17% in the first nine months of 2026. Its A100 index stands where it began in June 2023, and market prices for the H200 and B200 have risen. The same firm's survey of one-year H100 contracts, the nearest public thing to a renewal, shows the price a year later down by a median of about 15%, but the range runs from a fall of 30% to a rise of 60%.

The market's forward prices tell a different story again. Ornn, which publishes an index of GPU rents, lists rentals of different lengths, and a five-year rental of an A100 costs 80% of a one-month rental; for an H100 it is 60%, for Blackwell 54% and for the H200 44%. Read as the market's average expectation, those prices imply rents falling about 9% a year for the A100, 20% for the H100, 24% for Blackwell and 33% for the H200. The oldest chip is expected to hold its rent best, perhaps because most of its fall is behind it. Ornn's own line is that a new generation "does not, by itself, make its predecessors economically obsolete." CCIR, comparing posted rents for five generations on September 28, finds rent about 21% lower for each year of a chip's age, but adds that "rent tracks capability": once that is taken out, the old chips carry no extra discount for being old. Per kilowatt of chip rating, an A100 now rents for about 13% more than an H100.

The SemiAnalysis index changes run from each year's last reading to the next. Ornn's ratios are five-year rental prices as a share of one-month prices, read as an average with no discount or risk premium, so the implied declines are a rough guide. CCIR's figure compares chips of different ages on one day, so it mixes age with capability, and its rents per kilowatt use each chip's rated power.

So my 15% is a fair average for the H100 across a slump and a rebound, and too harsh for the A100, whose market rent has gone nowhere in three years; no record shows a steady decline, and CoreWeave's own renewal prices, which the second rental actually needs, are not public. A falling rent and a long life go together comfortably. A chip whose rent halves still earns for years, because the cash it costs to run is small beside what it rents for; Chris Zeoli puts an A100's rent at two and a half to four times the roughly 40 cents an hour it costs to run.

Prices of this kind are where the cobweb lives. Supply answers price with a lag, because a data center takes years to build, so capacity tends to arrive just as the scarcity that called for it passes. The chips going in now, at rents CoreWeave says are setting new highs, will reach their second rentals in the early 2030s, and nobody knows what the power market will look like then.

What the price assumes

As I wrote in my book, The Scholars' Treasure (chapter 7, page 86), business school at the University of Chicago gave me "a front row seat to the rise and fall of the dot-com era." What mattered then was raising money: "Would your idea work? What did it matter? Who cared?" A GPU fleet is the opposite kind of business. Whether its idea works is measured every hour, in dollars, and the measurement that matters most is what a chip earns in its sixth year.

CoreWeave rents GPUs by the hour, mostly on contracts of about five years that customers must pay whether they use the chips or not. It is the purest listed bet on how long a GPU keeps earning, because GPUs and the servers around them are most of what it owns. At the end of June its books carried $33.8 billion of technology equipment at cost, about $23 million for each megawatt of the roughly 1.5 gigawatts it had running, and depreciated it over six years, a life it raised from five in 2023.

Look at the second quarter of 2026 through that lens. Revenue was about $2.6 billion. Adjusted EBITDA, the company's own measure of earnings before interest, taxes, depreciation and some other charges, was $1.51 billion. Depreciation and amortization took $1.39 billion of it. That left a cushion of about 8% between what the equipment earned and what the accounts said it used up, and the company reported an operating loss of $49 million. Had every dollar of that depreciation been on a five-year life, the loss would have been about $330 million; on four years, about $750 million.

Warren Buffett has no patience for earnings that leave out the cost of the machines. "References to EBITDA make us shudder," he wrote in his letter for 2000, and asked whether management thinks "the tooth fairy pays for capital expenditures." His alternative, which he called owner earnings, subtracts what a business must spend to keep its equipment whole. For CoreWeave that means replacing its fleet once in each life, and the data center's own equipment in its turn. Jim Chanos ran this arithmetic in October 2025, when seven years of depreciation on CoreWeave's capital already came to more than its EBITDA, and Insider Monkey ran it again on October 2. At the second quarter's rate, on the company's adjusted EBITDA, owner earnings before interest and tax come to about $0.4 billion a year if the equipment lasts six years, $1.2 billion if seven and $1.8 billion if eight. Adjusted EBITDA adds back $165 million a quarter of stock paid to employees. Count that, and the data center's own equipment, and owner earnings fall below zero at six and seven years and come to about $0.5 billion at eight.

Owner earnings = EBITDA − (equipment at cost ÷ life)

At $89.62 on October 2, CoreWeave's shares were worth $49.4 billion. Add $35.6 billion of debt and take away $5.5 billion of cash, both as of June 30, and the business was valued at $79.5 billion, about 13 times the second quarter's adjusted EBITDA at an annual rate, which sounds ordinary, and about 170 times its fuller owner earnings even on an eight-year life, which does not.

Second-quarter 2026 adjusted EBITDA times four, less technology equipment at cost on June 30 divided by the life; the second column also takes off stock pay ($165 million a quarter) and replaces data center equipment over twelve years and software over four and a half. Before interest and tax. The multiple is the enterprise value over those owner earnings; the growth is what they would need each year for ten years, taxed at 21% and then growing 3% a year, for their value at 10% to equal the enterprise value. The growth is counted as if it cost nothing to build, which flatters the price.

A fast-growing company is caught in the middle of a build, and some of the June equipment was still being fitted out, so one quarter's earnings understate what the fleet will earn. A steadier measure is the yield: EBITDA a year for each dollar of equipment at work. In 2025, measured against the equipment in service through the year, it was about 21 to 23 cents, on the company's adjusted EBITDA, which adds back stock pay (about a ninth of it in the second quarter of 2026). From that yield follows the break-even life, the years a dollar of equipment must keep earning to pay back itself, its share of the data center's own equipment, and the cost of the money.

8.3% is CoreWeave's weighted average interest rate in the second quarter of 2026. Data center equipment and software, about 20 cents for each dollar of technology equipment, are charged over twelve years; software lasts less, so this flatters the result slightly. Pretax, on adjusted EBITDA.

The market values CoreWeave at 1.7 times the net book value of its property and equipment. A company is worth more than its assets only if its investments earn more than they cost, so the price is saying that CoreWeave's chips land on the right side of this table. On the yields in its 2025 accounts, that means earning for about six and a half to nine years, longer than the six in its accounts. Or it means yields rising toward a quarter of the equipment's cost a year, and in that case six years is about enough. The company says it is seeing both: prices up about 25% across its chips in July, five to ten points more expected contribution margin on contracts signed in the second quarter, and old chips rented again at nearly their first price.

Others have read the price this way, and got there first. Cape Fear Advisors, in September, valued the contracted part of CoreWeave at about 59% of its enterprise value and called the rest a call option on used GPUs, riding on "a price nobody publishes, six years out, where no market trades." Kerrisdale Capital, betting against the stock in 2025, reconstructed an illustrative contract CoreWeave had shown analysts, in which years five and six were rented again at 75% of the original price, and worked out that, on one industry cost model for a rack of the newest chips, an operating margin of about 20% on a six-year life falls to nothing on four. Sunil Dutta and Panos Patatoukas of Berkeley found that the lives implied by analysts' forecasts for the five largest builders rose from three years in 2015 to ten in 2026, and that their share prices imply fourteen. Our addition is a break-even life that counts the cost of borrowing and the data center's own equipment, worked from CoreWeave's own yield; it says that for this company the yield, more than the life, is what the price is waiting on.

There is one more thing in the filings that settles who carries the risk. CoreWeave's contract-backed loans, it told investors in September, typically amortize "such that the debt is fully repaid by the time the customer contract expires." Those loans are $13.6 billion of its $35.6 billion of debt. The first contract pays the lenders who lent against it; the noteholders, including the buyers of the $4.2 billion convertible it sold in September, wait on the second rental, and the shareholders stand last. Jamin Ball called it the second life of a GPU in May, and Chris Zeoli showed in September how GPU loans are built to be repaid before the chips must be sold. The second rental is the equity.

Where the model points in listed markets

None of these is a recommendation; each is a claim the market will test, with what would prove it wrong.

CoreWeave: the books hold for now. Straight-line depreciation over six years asks a chip coming off a five-year contract to re-rent at roughly half its first price for three more years. CoreWeave says its first large H100 renewal came in within 5% of the original price, and it is renting A100s into 2029; the market's term prices put the newest chips at about the bar. Neither points to a write-down before the end of 2027. Falsified if CoreWeave shortens the useful life of its technology equipment, or records a material impairment of it, in any filing through its report for the second quarter of 2027.

CoreWeave: the price needs more than the books. At the yield in its 2025 accounts, a dollar of CoreWeave's equipment repays itself and the cost of borrowing only if it earns for about six and a half to nine years; counting stock pay and the data center's own equipment, owner earnings at the second quarter's rate are below zero unless the equipment lasts about eight years. We expect the yield to stay below a quarter of the equipment's cost, because most of its revenue comes from multi-year take-or-pay contracts whose prices are fixed when they are signed. Falsified if adjusted EBITDA for the four quarters to June 30, 2027, divided by the average technology equipment at cost over those quarters, reaches 25 cents or more.

CoreWeave: old chips keep their rent while power is scarce. My draft's model says a chip is retired by a better use for its megawatt, and that at CoreWeave's own equipment cost of about $23 million a megawatt, swapping new chips into an old hall rarely pays. So old chips should stay in service and keep a price. Ornn reports that A100 occupancy rose from 74% to 90% between March and September 2026 while its spot index rose 20%, and CCIR finds that per kilowatt of chip rating an A100 now rents for more than an H100. Falsified if, on September 30, 2027, CoreWeave's posted on-demand price for an eight-GPU A100 or H100 machine is more than 25% below its October 4, 2026 price, or the A100 is no longer offered.

The cloud companies: long lives stay. Microsoft, Alphabet, Meta and Oracle have each lengthened the lives of their servers since 2022, and only Amazon has stepped back, for some servers, in 2025. If chips earn as long as the shadow price of power suggests, nobody else will need to step back soon. Falsified if Microsoft, Alphabet, Meta, Oracle or Nebius shortens the useful life of its servers or GPUs in any filing through September 30, 2027.

The data that we're missing is...

• Second rentals, contract by contract. What expiring A100 and H100 contracts actually re-rent for, as a share of their first price, how quickly, and for how long. CoreWeave gives examples; nobody gives a schedule.
• What customers pay, as against what the price lists ask. Committed-use discounts, credits and the share of a fleet sold on demand at all.
• Paid hours on old chips. How many hours a year an A100 outside a long contract is actually rented, and at what price.
• What a megawatt of new equipment costs, installed. By generation, including networking, storage and any new cooling.
• Used prices for whole servers. What a three-year-old eight-GPU H100 server actually sells for, and to whom. The public records so far are one seller's own sales (about 58% of launch list for GPUs two or three years old) and an index of asking prices (about 76% of new); closed prices for whole servers are scattered.
• Failure rates by age. How many GPUs in a large fleet fail each year, in their first, third and fifth year of service.

Primary research

Likely, we can get some of it from primary research, which is what Vista is all about. The people who priced, financed, audited and resold this equipment know now what the filings will take years to show.

For a client engagement, Vista combines this report with that primary research.

The checkpoint, set down now

We will score this note on September 30, 2027, and publish the result on the scorecard whether it flatters us or not. Each test sides with the price:

• The yield rises. CoreWeave's adjusted EBITDA for the four quarters to June 30, 2027 reaches 25 cents or more for each dollar of technology equipment at cost, averaged over those quarters.
• Second rentals hold, as a rule. By September 30, 2027, CoreWeave reports, in a filing or on a call, that capacity coming off contract was renewed or rented again at 90% or more of its original prices on average across that capacity. A single example does not count.
• The market comes round. In its figures for September 2027, Ornn's five-year rental price for an H100 is 75% or more of its one-month price (60% in September 2026). If Ornn no longer publishes it, the test is void.
• The profit appears. CoreWeave reports operating income above zero, after depreciation on its six-year life, for the four quarters to June 30, 2027.

Three or four passes and the price's reading gains ground; one or none and ours holds; two is too early to say. Four yes-or-no tests cannot settle a question about size; they can show which way the evidence is moving.

John Rust's achievement was to read a man's costs from his choices. Nobody could see what Harold Zurcher was weighing when he pulled an engine, but anyone could count the miles at which he did it, and from the counts Rust recovered the trade-off. A GPU fleet will be read the same way. The filings will not say what a three-year-old chip is worth; the renewals will. Each time a five-year contract runs out, somebody decides whether a chip from 2023 is worth renting again, and at what price. The deacon's shay ran its hundred years and went to pieces all at once. A first contract ends just as punctually. What a chip earns the morning after is the second rental.

Questions this note answers

How long do AI chips like NVIDIA's GPUs last?

A GPU does the same work in its sixth year as in its first, so its working life is set by what it can earn. In a data center that has run out of power, a chip is retired when its megawatt can earn more with something else. CoreWeave has rented out chips of a kind launched in 2020 on a contract that runs into 2029.

What useful life does CoreWeave use to depreciate its GPUs?

Six years, straight line, for all its technology equipment. CoreWeave raised the life from five years on January 1, 2023, and has recorded no material impairment of it.

Is CoreWeave's six-year depreciation too aggressive?

Not yet, on the evidence so far. Six years asks a chip coming off a five-year contract to re-rent at roughly half its first price for three more years, because the cost of running it does not fall with the rent. CoreWeave says its first large H100 renewal came within 5% of the original price; the market's five-year rental prices imply about half for H100s and Blackwell and about a third for the H200.

When does it pay to replace older GPUs with newer ones?

Only when the new chips' extra earnings from the same power pay for them. In my draft's model, a megawatt of H100s brings in about 45% more than a megawatt of A100s at CoreWeave's posted prices, but over three years the swap pays only if the new equipment for that megawatt costs less than about $6 million, or about $11 million if the old chips are sold. CoreWeave's books carry about $23 million of equipment for each megawatt it runs.

What does CoreWeave's stock price assume about how long its GPUs earn?

At $89.62 on October 2, 2026, CoreWeave was valued at $79.5 billion, 1.7 times the book value of its property and equipment. At the yield in its 2025 accounts, about 21 to 23 cents of adjusted EBITDA a year for each dollar of equipment, a dollar of GPUs repays itself and the cost of borrowing only if it keeps earning for about six and a half to nine years. The price needs longer lives, richer rents, or both.