CoreWeave depreciates its GPUs over six years. Nebius depreciates the same chips, bought from the same vendor, deployed in the same business, over four.
Both companies are audited. Both file with regulators. Neither is accused of fraud. And the difference between those two numbers, on identical hardware, is the entire neocloud business model reduced to a single disagreement, because a depreciation schedule is a claim about how long an asset earns, and the whole enterprise is a bet on that claim.
The word amortization is doing double duty here, and the doubling is the point. Accounting amortization is the schedule over which the cost of a GPU gets recognized against revenue, and it determines reported profit. Debt amortization is the schedule over which the money borrowed to buy that GPU has to be repaid, and it determines survival. Those two schedules are set by different parties for different reasons, they do not have to agree, and the space between them is where a neocloud lives or does not.
What a neocloud actually is
The category emerged around 2023 and describes a specific animal: a cloud provider whose entire business is renting accelerated compute, without the general-purpose infrastructure, enterprise software, or diversified revenue of a hyperscaler.
CoreWeave is the reference case, having gone public on NASDAQ in March 2025 at forty dollars a share. Nebius, Lambda, and Crusoe are the other names that matter, along with a long tail of smaller operators. What they share is a structure: buy GPUs at scale, secure power and space to run them, sign multi-year contracts, and rent capacity at a rate that has to cover depreciation, interest, power, and operations while leaving something over.
The reason the category exists at all is that hyperscalers were capacity-constrained faster than they could build, and a specialist willing to take balance-sheet risk could stand up accelerated compute quicker than a general-purpose cloud could reconfigure for it. Early access to new NVIDIA platforms has been part of the arrangement, with neoclouds frequently first to deploy a generation, which is a commercial advantage and also a dependency worth naming. Allocation of scarce new-generation hardware is a decision the vendor makes, and a business whose competitive position rests partly on being favoured in that allocation is exposed to a relationship rather than to a market. The same dynamic governs any industry where one supplier controls the critical input, and it rarely favours the buyer over a full cycle.
The growth figures are genuine and worth stating with their dates attached. CoreWeave went from $229 million in revenue in 2023 to $1.92 billion in 2024 to $5.13 billion in 2025, with adjusted EBITDA of roughly $3.09 billion on that last figure, which is about a sixty percent margin. Nebius reported $399 million in Q1 2026, up 684 percent year over year. Contracted backlog at CoreWeave reached $66.8 billion at the end of 2025 and $99.4 billion as of the end of March 2026.
The figures that sit next to those are equally real. CoreWeave’s GAAP net loss for 2025 was $1.167 billion, wider than 2024’s. Capital expenditure was $10.31 billion against $5.13 billion of revenue, a ratio of two to one, and the 2026 capex plan runs to thirty or thirty-five billion against revenue guidance of twelve to thirteen. Total debt reached roughly $21.4 billion with $1.2 billion of interest expense in 2025.
So: sixty percent adjusted EBITDA margins, a billion-dollar net loss, and capex at twice revenue. Those three facts are not in tension. They are what it looks like when a company is converting borrowed money into depreciating hardware faster than the hardware can pay for itself, on the expectation that it eventually will.
The hardware itself is worth a note because it sets the capital intensity. A single GB200 NVL72 rack runs roughly three million dollars, closer to four all-in with networking and storage, and it draws one hundred and twenty to one hundred and forty kilowatts, which means the facility housing it is as much of the investment as the silicon, and the thermal and power infrastructure required to run it cannot be redeployed to a different chip generation as easily as the chip can be swapped. The magnets, minor metals, and specialty materials inside that rack are their own supply question, and the semiconductor supply chain that produced the accelerators is the constraint on how fast anybody can deploy them regardless of financing.
The two amortizations
Set the schedules side by side and the structure becomes legible.
An eight-GPU H100 server from an OEM runs roughly $250,000 to $400,000 as of early 2026. Take $300,000 as a working figure. On a six-year straight-line schedule that server carries about $50,000 a year in depreciation. On a four-year schedule it carries about $75,000. Across a ten-thousand-GPU fleet the difference is on the order of thirty million dollars a year in reported expense, on hardware that is physically identical.
Now the other schedule. GPU-backed asset loans in this market typically run sixty to seventy percent loan-to-value with terms of twelve to thirty-six months, frequently through special purpose vehicles that isolate the collateral. CoreWeave’s GPU-backed borrowing has priced at roughly eight and a half percentage points above the benchmark rate, which puts the coupon in the nine to ten percent range.
Notice the mismatch. The accounting says the asset earns for six years. The debt says repay me in one to three. That gap has to be bridged by refinancing, by cash flow, or by new capital, and every bridge assumes the collateral is still worth something when you cross it.
The reason the depreciation schedule matters so much more than an accounting choice normally would is that it is simultaneously a profitability claim, a collateral valuation, and a covenant input. Lengthen it and reported profit improves, asset values on the balance sheet hold up, and coverage ratios look better. None of that changes the cash. What it changes is every number a lender is looking at.
The comparison to other asset-backed industries is instructive because they solved this problem institutionally rather than by assumption. Commercial aircraft carry published residual value curves, standardized appraisal by accredited firms, active lease and sale markets, and lenders with decades of loss data. Shipping has the same apparatus. Both took a long time to build, and both permit five-year secured paper because the residual is a market fact rather than a management estimate. The absence of that apparatus is the single clearest difference between GPU lending and every other large asset-backed market, and it is a gap that institutions rather than spreadsheets have to close. GPU lending has none of that scaffolding and is writing similar paper anyway.
Why the depreciation argument is not a technicality
The disagreement runs across the entire industry and it has been moving in both directions, which is the tell that nobody knows.
In 2023 the major hyperscalers extended server useful life from three or four years to six. Across roughly three hundred billion dollars of combined capital expenditure, that change reduced reported depreciation expense by something on the order of eighteen billion dollars annually. CoreWeave made the same change in January 2023, from four years to six, before going public.
Then it reversed in places. Amazon shortened its server estimate from six years to five in February 2025, citing the increased pace of development in artificial intelligence and machine learning, taking a $700 million operating income hit and booking $920 million of accelerated depreciation in the fourth quarter of 2024 alone. Meta went the other way, extending to five and a half years and booking a $2.9 billion depreciation reduction. Microsoft and Alphabet held at six. CoreWeave held at six. Nebius sits at four, Lambda at five, and the CEO of Scaleway has said publicly that his company assumes a three-year horizon on the basis of an eighteen-month generation cycle.
When the companies with the most information about an asset class disagree by a factor of two about how long it earns, the disagreement is the finding. Someone is wrong, and being wrong in the generous direction means profits are overstated now and write-downs arrive later.
Michael Burry’s version of the claim is that hyperscalers will cumulatively understate depreciation by roughly $176 billion between 2026 and 2028. Treat that as what it is, which is a position taken by an investor who benefits if it is believed, and the same treatment applies to every number in this subject including the optimistic ones. The forecasting problem is structural rather than a matter of anybody’s honesty, since the useful life of a class of asset that has existed in its current form for about four years cannot be established by observation yet.
The complicating fact arrived from the vendor. NVIDIA announced in 2025 a move from a roughly two-year product cadence to an annual one. A shorter generation cycle means the installed base becomes previous-generation faster, which is an argument for shorter useful lives, delivered by the company with the strongest interest in customers buying new hardware sooner.
The value cascade, and the bull case stated properly
The six-year defenders are not simply being aggressive, and their argument deserves its strongest form because it might be right.
The claim is the value cascade. A GPU’s economic life does not end when it stops being the best chip for frontier training. It cascades downward: from training the largest models, to fine-tuning, to high-value inference, to batch inference, to cheaper regional capacity. Each tier tolerates older hardware and pays less for it, and the asset keeps earning at a declining rate rather than falling off a cliff.
The load-bearing part of that argument is inference. Training is periodic and brutally sensitive to being one generation behind, because a competitor with newer silicon trains faster and cheaper. Inference runs continuously, at global scale, and is far less sensitive to generational position, since what matters is cost per token delivered rather than time to complete a training run. If inference comes to dominate GPU utilization, older hardware retains commercial value considerably longer than a training-only view implies.
The analysis arguing for the cascade and the six-year schedule makes the case that AI factories bend rather than break useful-life assumptions, and the observed rental market has partially supported it, which is discussed below.
The honest problem with the cascade is that it requires a market that does not yet exist in mature form. Cascading value assumes a redeployment channel: someone to sell or lease four-year-old capacity to, at a price, with a functioning secondary market to establish what that price is. Secondary markets for aircraft and ships took decades to develop and have standardized appraisal, established residual curves, and lenders who understand them. The scrap and secondary markets that eventually developed for industrial metals offer a rough template, and they required decades plus standardized grading before anybody would lend against the residual. Nobody knows with confidence what a used GPU cluster is worth, and a depreciation policy that assumes a redeployment market is a policy that assumes an institution. The recycling and recovery economics that work only when a waste stream is concentrated and arrives at a known address are the closest available analogue, and they took decades to develop for materials with far simpler valuation problems.
What the rental price actually did
The observable price is the closest thing to a market verdict, and it has been violent.
One-year H100 rental rates ran around eight dollars per GPU-hour in early 2024. By October 2025 the same metric had fallen to roughly $1.70. Then it rose about forty percent to $2.35 by March 2026 on a wave of inference demand that had not been priced in, before softening again by May. Reported effective pricing tells a similar story from the contract side, with market estimates putting CoreWeave’s realized rate around $2.80 to $3.20 per GPU-hour against on-demand list rates several times that, and Nebius nearer $2.20 to $2.50 given heavier discounting to AI startups. List price and realized price are different numbers in this business, and the gap is where the long contracts sit.
Read that sequence carefully, because both camps get something from it. The collapse from eight to $1.70 is exactly what the short-useful-life argument predicts: new generations arrive, the previous generation reprices, and the asset earns far less than the original underwriting assumed. The recovery to $2.35 is exactly what the cascade argument predicts: inference demand absorbed the older fleet and restored pricing power.
What both readings share is the volatility, and the volatility is the actual problem. An asset whose rental rate falls seventy-nine percent and then rises thirty-eight percent inside eighteen months is not the kind of collateral that supports multi-year secured lending under any conventional underwriting standard. No aircraft lender or shipping lender would write five-year paper against an asset with that price behaviour without a hedging mechanism, and in most cases would not be permitted to.
That is the structural observation worth carrying: the neocloud sector has built an asset-backed lending market around collateral with no residual value curve, no standardized appraisal, and no hedge. Commodity markets solved that problem with futures and options, which is why an aluminium smelter or a copper miner can lock a forward price and a GPU lessor cannot.
Asset-backed lending against depreciating collateral has a well-understood failure mode and it is worth spelling out because it is mechanical rather than speculative.
Secured GPU debt is collateralized by the hardware. If the market value of that hardware falls faster than the loan amortizes, coverage ratios degrade. Degraded coverage lets lenders demand additional collateral, accelerate repayment, or decline to extend new draws. Each of those tightens the borrower’s position, which forces asset sales or distressed refinancing, which pushes market prices down further, which degrades coverage again.
The payment calendar is the near-term constraint. Roughly $4.2 billion in scheduled debt amortization and vendor financing payments came due in 2026 for CoreWeave, and the sector as a whole carries more than twenty billion dollars in GPU-backed debt.
The mitigations are real and worth crediting. Non-recourse structures and special purpose vehicles limit contagion from one asset pool to the parent. Take-or-pay contracts lock revenue ahead of depreciation. A large contracted backlog is genuinely de-risking, because it converts a demand question into a counterparty-credit question.
The short-seller case argues that the arithmetic does not clear even so, modelling GB200 economics at roughly twenty percent EBIT margins under the company’s own six-year depreciation and eighty-seven percent utilization assumptions, falling toward zero or negative under four to five year lives, while noting that borrowing at nine to ten percent to serve A-rated counterparties is a thin structural position. That is a document produced by a party with a disclosed short interest, which does not make it wrong and does mean it should be read the way the bull research should be read. Both documents are arguing about the same unobservable number, and both would be reporting different figures if they held opposite positions. Reading interested documents against each other rather than deferring to either is the only available method when the underlying quantity is unobservable, and it is the same discipline required of every demand forecast in a capital-intensive materials industry.
The balance-sheet arbitrage
Here is the part of the model that has nothing to do with GPUs and everything to do with where capital intensity is allowed to sit.
When a hyperscaler builds its own capacity, the spending is capital expenditure. It lands on the balance sheet, it depreciates, and it pressures free cash flow. When the same hyperscaler rents capacity from a neocloud under a multi-year contract, the spending is operating expense recognized over the contract term. The capacity arrives either way. The accounting treatment does not.
The magnitudes are not marginal. Meta has entered neocloud agreements reported at up to $62.2 billion, running through 2031 and 2032, which spreads to something under ten billion a year in operating expense rather than appearing as capex. Microsoft has roughly $60 billion in comparable agreements against 2026 capex guidance near $190 billion and operating cash flow forecast around $200 billion, which would put capex at ninety-five percent of operating cash flow before the neocloud spending is counted.
So the neocloud takes capital intensity off the balance sheet of a company with an AA credit profile and puts it onto the balance sheet of a company borrowing at nine to ten percent. That is a transfer of financing cost from a cheap borrower to an expensive one, and it happens because the expensive borrower is willing to accept the asset risk that the cheap borrower would rather not carry.
Whether that is a service or a hazard depends on where you are standing. From the hyperscaler’s position it is capacity without balance sheet consequences. From the neocloud’s position it is contracted revenue that makes the debt raisable. From a system position it is leverage migrating toward the entity least able to absorb a repricing. That migration pattern is not new, and the history of infrastructure financed by parties other than the ultimate beneficiary suggests losses tend to land where the equity is thinnest rather than where the demand was created.
The power arbitrage, which is the real one
Underneath the financial engineering sits a physical asset that is scarcer than GPUs, and the operators know it.
CoreWeave and Nebius have each secured roughly 3.5 gigawatts of contracted power capacity. Very little of it is energized. CoreWeave has targeted around 1.7 gigawatts of active power by the end of 2026 against 43 sites and something over 850 megawatts active, with 3.1 gigawatts contracted.
That gap between contracted and active power is the operational centre of the business. Contracted power is an option. Active power is revenue. Converting one into the other requires substations, transformers, interconnection agreements, and utility timelines that do not compress for anybody, which is why grid interconnection queues rather than chip supply have become the binding constraint on capacity growth. The transformers, switchgear, and high-voltage equipment required to energize a site carry lead times of their own, measured in years rather than quarters, and they are competing against every other electrification project for the same manufacturing capacity.
Which reframes what a neocloud actually sells. The GPUs are procurable with money. Power in a specific location at a specific date is not, and an operator who secured an interconnection position in 2023 holds something that cannot be bought at any price in 2026 because the queue is the queue. Valuing these companies as cloud providers on revenue multiples misses that a meaningful share of the enterprise value sits in power rights, which is closer to how a pipeline or a transmission asset gets valued. The long history of infrastructure assets whose value sat in a right-of-way rather than in the equipment is the relevant precedent, and the accounting conventions for that kind of asset look nothing like a cloud provider’s.
The energy procurement decisions that follow from that are the reason these companies end up negotiating with utilities, reactor operators, and county commissions rather than only with chip vendors.
Concentration, and what take-or-pay does not cover
The contract structure is the de-risking mechanism and it has specific edges.
Microsoft accounted for approximately sixty-seven percent of CoreWeave’s 2025 revenue, under a master agreement dating to February 2023 with pre-IPO pricing. Meta and OpenAI commitments signed in late 2025, worth $14.2 billion and $6.5 billion of incremental backlog respectively, do not flow through the income statement until 2026 through 2028 capacity comes online. Nebius signed a five-year Meta arrangement reported at $27 billion, structured as roughly $12 billion fixed and $15 billion optional.
Take-or-pay means the customer pays whether or not it consumes the capacity, which converts utilization risk into counterparty risk. That is a genuine improvement and it protects only through the initial term. Counterparty quality is doing a great deal of work here: a take-or-pay contract with an AA-rated hyperscaler is a different instrument from the same contract with a venture-funded startup, and part of what the neoclouds bought with thin margins was the right to serve borrowers whose credit made the debt raisable at all. The concentration risk that follows from depending on a small number of large buyers is the standing hazard of that trade.
The mismatch is the term structure. Weighted average contract duration runs around five years. A six-year depreciation schedule therefore requires that contracts either renew at acceptable rates or be replaced with equivalents at a point when the hardware is four or five generations old. The take-or-pay covers the first period. The residual value assumption covers the second, and the residual value assumption is the thing nobody can price.
Concentration compounds it. A customer that is two-thirds of revenue is also, functionally, the entity that determines whether the depreciation schedule was correct, because its renewal decision at year five is the market test. That is a genuinely unusual position for a supplier to be in, and it is the reason customer concentration in this sector is not merely a revenue risk but a valuation risk, in the way a single dominant buyer reshapes any supply relationship: the anchor customer’s renewal is simultaneously the cash flow and the appraisal.
The circular financing question, audited
The relationships in this sector are unusually tangled and the tangle deserves precise treatment rather than either dismissal or scandal framing.
NVIDIA holds equity positions in neoclouds, has entered capacity arrangements with them, and in some structures has provided backstops. Neoclouds use that capital and those commitments to buy NVIDIA hardware. Hyperscalers contract with neoclouds for capacity built on that hardware, and some hyperscalers are also NVIDIA investors and customers.
What is legitimate: vendor financing is ordinary in capital equipment industries, strategic investment in customers is ordinary, and a chip vendor with an interest in its ecosystem having capacity to deploy silicon is not by itself improper.
What warrants scrutiny: revenue that returns to the vendor as demand for the vendor’s product, equity marks that depend on the customer’s ability to raise further debt, and a chain in which the same dollar can appear as revenue at multiple points. None of that is fraud and all of it makes the aggregate demand signal harder to read, because a share of reported demand originates inside the same set of balance sheets. The comparison worth holding is to vendor-financed telecom buildouts, where equipment makers lent to carriers to buy equipment, demand looked robust while the lending continued, and the writedowns arrived on both sides simultaneously. That precedent does not determine this outcome and it does establish that the structure has a known failure mode.
The audit standard that holds is the one the whole subject requires: when a forecast is produced by a party with a position, the forecast is a document with an author who wanted something. That applies to NVIDIA’s total addressable market estimates, to neocloud backlog disclosures, to hyperscaler capex guidance, and to short reports, symmetrically.
The claims that do not hold up
An audit, since this sector generates confident assertions in both directions.
Neoclouds are just resellers with no moat understates the power position, the deployment expertise, and the contracted backlog, and misses that being early in an interconnection queue is a durable advantage even if nothing else is.
Neoclouds are the picks and shovels of AI so they win regardless is the mirror error. A picks-and-shovels business that borrows at nine to ten percent against collateral with no residual curve is not insulated from the outcome. It is levered to it. The suppliers of scarce inputs to a boom generally do better than the boom’s operators, and a neocloud is an operator rather than a supplier.
Sixty percent EBITDA margins mean the business is highly profitable conflates adjusted EBITDA with cash generation. Adding depreciation back to earnings is defensible for a durable asset and question-begging for an asset whose durability is the contested variable, which is why the same company shows a sixty percent adjusted margin and a billion-dollar GAAP loss. The two figures are not measuring different things badly. They are two positions on one assumption.
The backlog guarantees the revenue overstates take-or-pay. Backlog is contracted revenue subject to counterparty performance, and its conversion depends on capacity coming online on schedule.
GPUs will be worthless in three years is not supported by the rental data, which showed a recovery rather than a collapse to zero, and which suggests the cascade has some reality to it.
GPUs will earn for six years is not supported either, because the assumption requires a redeployment market that has not been demonstrated.
Depreciation is just an accounting choice ignores that it drives collateral values, covenant compliance, and the reported profitability that makes further borrowing possible.
What the neocloud model is actually telling us
Strip the sector down and what remains is a well-defined bet with a knowable structure and an unknowable input.
The structure: borrow at nine to ten percent, buy hardware, secure power, sign five-year take-or-pay contracts with investment-grade counterparties, and earn the spread between contracted revenue and the sum of interest, power, operations, and true economic depreciation. Every term in that expression is observable except the last one.
The input nobody has: how long a GPU earns. Six years and the model produces roughly twenty percent EBIT margins and a viable business. Three to four years and the margins compress toward zero, the collateral is worth less than the loans it secures, and the equity is a call option on inference demand growing fast enough to bail out the schedule.
That is not a criticism of the operators, who are transparent about the assumption and whose disclosures state it plainly. It is a description of what is being purchased when anybody buys the equity, lends against the hardware, or signs a long contract, and it is the same structural pattern that shows up wherever long-lived financing meets short-lived physical assets: five-year silicon on debt priced for a longer horizon, in a facility financed on paper longer still. The building has a thirty-year life. The power contract runs fifteen or twenty. The customer contract runs five. The silicon inside is arguable at four. Four different clocks on one asset, and the material and energy commitments underneath them run on schedules of their own.
The thing to watch is not the revenue growth, which is real and will continue for as long as the contracts convert. It is the residual value question resolving in public. That happens the first time a large fleet of four-year-old accelerators has to be sold or re-leased at an observable price, because until that transaction occurs there is no market evidence about the assumption underwriting the entire sector, only two audited companies disagreeing about it by a factor of one and a half.
There is a second thing to watch that gets less attention, which is the conversion rate from contracted power to active power. Backlog converts to revenue only when capacity energizes, and the schedule for energizing capacity is controlled by utilities, equipment manufacturers, and permitting processes at the county level rather than by the operator, including the water withdrawal approvals that cooling requires before anything energizes. A slipped interconnection date does not reduce the debt service, and the gap between a 3.5 gigawatt contracted position and a 1.7 gigawatt active target is where that risk lives.
The ten-lecture briefing on how AI data centers work runs the physics, the money, and the politics of that arrangement in sequence, and the money lens has a specific job: to insist that every figure state its unit, its stage, and its date. A sixty percent margin, a billion-dollar loss, and a hundred-billion-dollar backlog are all true about the same company in the same year, and which one you lead with is a decision about what you want the reader to conclude.
Which is also why the sector resists the bubble framing in both directions. A bubble implies the assets are worthless and the demand imaginary, and neither is established: the contracts are signed with solvent counterparties, the capacity is being consumed, and the rental market recovered when inference demand arrived. What is established is that the equity value depends on an assumption about asset life the operators themselves cannot agree on, financed with debt that matures before the assumption resolves. That is a specific fragility rather than a general delusion, and it has a test date attached.
CoreWeave says six years. Nebius says four. Same chips, same customers, same business. One of them is going to have been right, and the difference will be settled by an auction nobody has held yet.
Leave a Reply