Nearly every company building AI infrastructure has placed a number in its financial statements that deserves far more scrutiny than it currently receives. That number is six. CoreWeave, Oracle, Google, Microsoft, and most of their peers have assigned a six-year useful life to the GPU hardware forming the physical backbone of their AI infrastructure. Those companies spread the capital cost of those assets across their income statements over that same period. Tens of billions of dollars in debt financing, private credit structures, and equity valuations now depend on that six-year assumption.
The Product Cadence That Is Repricing the Hardware Faster Than the Accounting
Nvidia announced earlier this year that Vera Rubin, its next GPU architecture, will begin shipping in Q3 2026, with volume production ramping through Q4 and into 2027. Rubin Ultra follows in 2027. Feynman is already on the roadmap for 2028. The company has explicitly accelerated its product cadence to one major architecture per year, a pace that CEO Jensen Huang has said is unmatched in the industry. Each successive generation delivers performance improvements that are not incremental: Blackwell Ultra delivers up to 50 times better performance and 35 times lower cost for agentic AI compared with Hopper, according to benchmarks cited in Nvidia’s own Q4 FY26 earnings disclosures. Rubin is projected to deliver up to a 10x reduction in inference token cost compared with Blackwell.
The H100, which defined the 2023 and 2024 AI infrastructure buildout, launched at $8 to $10 per GPU hour in 2023 and now trades at approximately $2 per hour. Major providers have already compressed Blackwell B200 pricing from more than $6 per GPU hour to roughly $2.25, and market competition will likely push prices down another 50 to 70 percent as supply scales through 2026 and into 2027. The industry built its AI infrastructure depreciation schedules for a hardware environment that stopped existing before Hopper shipped. That mismatch creates what may be the most consequential underpriced variable in the AI infrastructure investment thesis, and every company in the sector carries that risk inside its financial statements.
The Accounting Choice That Shapes Every Number Downstream
Depreciation is not a technical accounting footnote. It is a forward-looking statement about how long management believes an asset will generate economic returns, and it shapes every financial metric that investors, lenders, and analysts use to evaluate a business. When CoreWeave increased the useful life of its technology equipment from four years to six years in January 2023, ahead of its eventual IPO, it was making a claim about the commercial viability of GPU hardware across a six-year horizon in an industry where the competitive performance frontier was already moving by 40 to 50 percent annually. That change in accounting policy did not reflect new data about GPU longevity. Nvidia had, by that point, already announced the transition from two-year to one-year architecture cycles. The six-year depreciation schedule was adopted precisely when the data supporting it was becoming weaker, not stronger.
CoreWeave is the most visible case but not the only one. Google and Microsoft have settled on five to six year useful life assumptions for their AI server infrastructure. Oracle has adopted similar schedules. Meta extended its networking and server asset life to five and a half years in 2023, at the peak of the AI infrastructure arms race, with every incentive to maximise the earnings benefit that a longer depreciation schedule provides. Operators and auditors are defending these accounting positions through the value cascade argument: GPU hardware does not become worthless when the next generation ships, but instead cascades down from frontier training to production inference to enterprise AI to batch processing, retaining commercial value at each tier even as its premium market position erodes.
The Short Sellers Targeting the Assumption Directly
Michael Burry has estimated that the five largest AI hyperscalers will understate sector depreciation by a cumulative $176 billion between 2026 and 2028 as a consequence of those schedules, padding reported earnings by approximately 20 percent today at the cost of earnings hits of 15 to 25 percent later through impairment charges. Jim Chanos, who correctly identified Enron as a fraud, has called it a massive financial risk specifically for CoreWeave and Oracle, pointing to the rental rate compression already visible in the H100 market as evidence of exactly the kind of value degradation that six-year depreciation schedules assume will not occur.
The First Crack: Amazon’s Policy Reversal and What It Signalled
The most important data point in the AI infrastructure depreciation debate is not a short seller’s thesis or a financial model. It is a line in Amazon’s Q4 2024 earnings disclosure. In February 2025, Amazon shortened its server useful life estimate from six years to five years, explicitly citing the “increased pace of development in artificial intelligence and machine learning” as the reason for the change. The revision was not costless. Amazon took a $700 million operating income hit and booked $920 million in accelerated depreciation in Q4 2024 alone as a consequence of acknowledging that the hardware it had deployed on a six-year assumption would not sustain commercial relevance on that timeline.
Amazon is not a company without financial flexibility. It has the balance sheet, the operational scale, and the diversified revenue base to absorb a $920 million accelerated depreciation charge without existential difficulty. What the disclosure signalled to the market is something more significant than one company’s accounting adjustment: it was the first public acknowledgement from a hyperscaler that the six-year depreciation schedule was not holding against real-world hardware lifecycle data. A company with every incentive to maintain the most favourable depreciation assumptions its auditors would accept chose instead to take an immediate financial hit rather than defend an assumption that Nvidia’s product cadence had already rendered implausible.
The Accounting Divergence That Revealed the Industry Split
The contrast with Meta is instructive and uncomfortable. In the same quarter that Amazon was shortening its useful life estimate and booking $920 million in accelerated depreciation, Meta was extending its server asset life to 5.5 years and booking a $2.9 billion depreciation reduction. Same technology environment. Same hardware generation. Opposite accounting directions. The $2.9 billion difference in Meta’s reported earnings relative to the more conservative treatment is not a small number even for a company of Meta’s scale. It represents the full financial gap between an accounting policy designed to reflect the hardware’s commercial reality and one designed to produce the most favourable income statement impact that auditor standards permit.
CoreWeave maintains its six-year schedule as of Q1 2026. Microsoft and Alphabet hold at six. Oracle is on six. The industry’s response to Amazon’s admission that six years was too long was, for most operators, to continue booking on six years. That decision defers the accounting reckoning but does not eliminate it. The companies that shorten their depreciation assumptions now take the earnings hit on their timeline and under their control. The companies that wait for the market to force the adjustment will take it under worse conditions, typically when contract renewals are compressing revenue at the same moment that impairment charges are hitting reported earnings.
Why CoreWeave Extended From Four Years to Six
The specific history of CoreWeave’s depreciation policy change is worth understanding in detail because it reveals how the incentive structures around accounting policy can diverge from the underlying economic reality. CoreWeave originally depreciated its GPU fleet over four years, a schedule closer to what Nebius and Lambda Labs continue to use. In January 2023, it extended that schedule to six years. The timing matters. January 2023 was approximately eight months after Nvidia had announced its transition to annual architecture releases, a decision that explicitly compressed the economic life of any given GPU generation relative to the prior two-year cadence.
The financial consequence of the change was substantial. Extending the depreciation period from four years to six years reduces annual depreciation expense by approximately one-third on any given asset. For a company with $13 billion of technology equipment on its balance sheet, as CoreWeave had by mid-2025, the difference between four-year and six-year depreciation amounts to hundreds of millions of dollars in annual reported earnings. Those hundreds of millions of dollars are the difference between reported operating losses that look manageable and ones that look structural. They also shape the financial projections investors used to justify CoreWeave’s $80 billion valuation at IPO.
The Forward-Looking Assumption the Market Is Still Testing
Michael Intrator has publicly argued that the data supports the six-year assumption. He pointed to A100 chips from 2020 remaining fully booked in late 2025, and to customers immediately rebooking a batch of H100 chips from 2022 at 95 percent of their original price when a contract expired. Those data points are accurate. They also describe a market condition that existed before Blackwell supply scaled fully and before NVIDIA confirmed Vera Rubin’s timeline. The forward-looking question is not whether H100s retained value in late 2025. It is whether operators deploying hardware in 2025 and 2026 will retain value at the rates current depreciation schedules assume once Rubin enters the market at scale in 2027.
The Value Cascade Argument and Its Limits
The waterfall thesis is the central intellectual defence of extended AI infrastructure depreciation schedules, and it deserves serious engagement rather than dismissal. The argument holds that GPU hardware does not become worthless when the next generation ships. It retains commercial utility at progressively lower tiers of the workload hierarchy. The H100 that was the frontier training accelerator in 2023 is the production inference accelerator in 2025 and will be the enterprise AI accelerator in 2026. The A100 that shipped in 2020 is, as Intrator correctly noted, still fully booked in 2025. The physical durability of GPU hardware is genuinely long, and the workload pyramid beneath the frontier tier is genuinely large and growing.
Where the waterfall thesis encounters difficulty is in the relationship between sustained utilisation and sustained revenue. Those are different quantities, and six-year depreciation schedules are built on assumptions about revenue, not just utilisation. The H100 that was generating $8 to $10 per GPU hour in 2023 is now generating approximately $2 per GPU hour in early 2026. That is a 75 to 80 percent reduction in revenue per GPU hour over roughly 30 months. The hardware is still deployed, still functional, still generating revenue. The revenue it is generating is a fraction of what the original financial models assumed when the depreciation schedules were set.
This distinction between sustained utilisation and sustained revenue per unit is the specific gap that six-year depreciation schedules obscure. A GPU fleet that remains 95 percent utilised at 25 percent of its original rental rate is generating cash flows that do not support the debt structures financed against the original revenue assumptions. The hardware is working. The financial model is not, and the accounting policy that connects the two has not been updated to reflect the actual market trajectory.
The Performance Cliff That Rental Rate Compression Signals
Rental rate compression is not primarily a supply story, though supply scaling contributes to it. It is primarily a performance story. When Blackwell entered the market, it delivered capability improvements so substantial that customers with frontier AI workloads had strong incentives to migrate regardless of whether H100 capacity was available at lower prices. The B200 delivers roughly 40x the performance of an H100 for AI inference. A customer running a production large language model on H100 hardware is not simply paying more than they need to for equivalent capability. They are running at a material competitive disadvantage relative to any competitor who has migrated to Blackwell, because their inference costs per token are orders of magnitude higher.
This competitive performance dynamic is what drives rental rate compression faster than supply scaling alone would predict. When a customer can access 2 to 4 times the efficiency on newer hardware, older hardware becomes acceptable only at steep discounts, not because it no longer functions but because its performance-per-dollar no longer competes. The cascade down the workload hierarchy is real, but customers and operators only deploy older hardware at price points that reflect its performance disadvantage relative to the current generation, and that disadvantage grows with every new architecture release.
The Rubin Transition That Will Reprice Blackwell
Vera Rubin, shipping in Q3 2026, will deliver up to 10x better inference token cost than Blackwell, according to Nvidia’s own projections. When Rubin reaches market scale in 2027, the Blackwell hardware that customers are deploying at premium prices in 2025 and 2026 will face exactly the H100-to-Blackwell dynamic that has already driven H100 rental rates from $8 to $2 per GPU hour. Operators using six-year straight-line schedules for their Blackwell fleets will be managing assets that have already cascaded two tiers down the workload hierarchy before half of their accounting life passes, generating revenue at rates far below the levels those depreciation schedules originally assumed.
The Private Credit Structures That Are Most Exposed
The depreciation debate has consequences that extend beyond the income statements of hyperscalers large enough to absorb impairment charges without existential difficulty. It is most consequential for the private credit structures that have financed the neocloud buildout on GPU-collateralised debt, where the financial models built on extended depreciation assumptions are load-bearing in ways that hyperscaler balance sheets do not replicate.
CoreWeave has borrowed over $10 billion against its GPU fleet. Fluidstack holds $10 billion in GPU-backed debt. Lambda Labs carries $500 million. Crusoe has $425 million. The total quantum of neocloud debt collateralised against GPU hardware whose economic life is genuinely uncertain runs to over $20 billion across a sector where most operators are not profitable on a GAAP basis. These debt structures work when the collateral holds its value relative to the assumptions embedded in the debt covenants. They face material stress when rental rates fall significantly, when utilisation dips because enterprise customers who over-reserved capacity are renewing contracts at lower volumes, or when a hardware generation transition creates a gap between what borrowers can charge for their fleet and what they need to service their obligations.
The Collateral Reality Behind GPU-Backed Debt
Chanos framed the arithmetic precisely: if chips last three years rather than six, companies must depreciate a third of their spending each year rather than a sixth. For a neocloud with $5 billion of GPU hardware on its balance sheet, switching from six-year to three-year depreciation more than doubles the annual depreciation charge, converting reported operating margins that appear manageable into losses that raise immediate questions about the viability of the debt structures sitting above them.
The secondary market for GPU hardware is the most direct evidence available about whether the depreciation schedules embedded in those debt structures reflect commercial reality. Used and refurbished H100s traded as high as $50,000 per GPU during peak scarcity in mid-2024. By early 2026, the same hardware is clearing the secondary market at steep discounts as supply normalised and Blackwell became available. Silicon Data’s analysis of thousands of H100 secondary market listings between June 2024 and December 2025 found that H100 depreciation accelerates sharply after approximately two years of service, with the inflection point producing secondary market revaluations that treat the hardware as short-cycle capital rather than long-term infrastructure. That is not the depreciation profile that six-year straight-line accounting assumes.
The Secondary Market Signal the Accounting Models Cannot Ignore
The secondary market is the collateral test that private credit structures have not yet fully faced. Debt secured against GPU hardware is only as sound as the market’s willingness to pay for that hardware if it needs to be liquidated. When H100 secondary market values were running above retail price due to scarcity, GPU-backed debt had genuine collateral depth. As secondary market prices compress with each new architecture generation, the collateral supporting the debt compresses with them. The debt service coverage covenants that lenders wrote against CoreWeave’s fleet are secured against hardware whose secondary market clearing prices are moving in a direction that is not consistent with the six-year depreciation schedules they were written around.
CoreWeave disclosed $21.6 billion in total indebtedness as of December 31, 2025, with a further $8.5 billion delayed-draw term loan entered in March 2026 to finance capital expenditures for a customer contract. Its interest expense tripled year on year to $311 million in Q3 2025 alone, with $34 billion in off-balance sheet lease commitments through 2028 adding leverage that does not appear on the face of the balance sheet. These are not small numbers, and they rest on collateral whose value is being tested by a hardware market moving faster than the accounting models assumed.
The Price Compression That Is Stress-Testing the Debt Structures
AWS slashed H100 instance prices by up to 45 percent in mid-2025, directly pressuring neocloud margins on spot capacity and accelerating the timeline on which take-or-pay contract renewal conversations are happening at lower price levels. The operators who built their debt structures on take-or-pay contracts signed in 2023 and 2024 at pricing that AWS has since undercut by nearly half are the ones whose covenant compliance faces the most direct stress as those contracts reach renewal. The secondary market pricing data, the AWS price cuts, and Amazon’s own decision to take a $920 million accelerated depreciation charge in Q4 2024 are all telling the same story. The market is pricing GPU hardware on a shorter commercial timeline than the depreciation schedules embedded in the sector’s debt structures assume.
The Take-or-Pay Contract as the Load-Bearing Assumption
The industry’s primary defence against the depreciation critique is the structure of its customer contracts. The take-or-pay model, under which customers commit to paying for reserved capacity regardless of whether they use it, provides revenue visibility that extends across the depreciation schedule and that lenders use to justify the debt covenants they write. If a neocloud has secured three-year take-or-pay contracts covering 80 percent of its fleet, the near-term revenue risk from rental rate compression is limited to the 20 percent that remains exposed to spot market pricing. The depreciation concern becomes a future-cycle problem rather than an immediate one.
The limitation of this defence is that take-or-pay contracts expire. The first wave of take-or-pay contracts written in 2023 and 2024 against H100 hardware are now entering renewal cycles in 2026 and 2027, exactly as Blackwell supply is scaling and Vera Rubin is entering the market. Enterprises that committed to three-year H100 contracts in 2023 at $6 to $8 per GPU hour are now renewing in a market where Blackwell rates are $2.25 and expected to fall further. The renewal conversations happening in 2026 are the first real test of whether the take-or-pay model sustains the revenue assumptions embedded in six-year depreciation schedules across a full hardware generation cycle.
The operators who secured contracts with the largest, most creditworthy hyperscale customers before the rental rate compression began are the ones whose positions are most defensible. The operators who built their fleets against enterprise and mid-market customers on shorter initial contracts are the ones whose renewal dynamics will surface the depreciation problem first.
The Comparative Evidence the Market Has Available
Not all neocloud operators adopted the six-year depreciation schedule, and the divergence between approaches provides a natural experiment that the market is beginning to evaluate. Nebius, which has a nearly identical business model to CoreWeave, providing GPU-as-a-service on Nvidia hardware with take-or-pay contract structures, depreciates its GPU fleet over four years. Lambda Labs also uses a five-year schedule, more conservative than CoreWeave’s six-year position. The difference in accounting philosophy between these companies is stark given that they are deploying essentially the same hardware for essentially the same workloads.
The financial consequences of that difference are concrete and visible. Nebius’s four-year depreciation schedule produces higher annual depreciation charges, compressing reported margins relative to CoreWeave on equivalent revenue and hardware deployment. What it produces in return is a financial model whose assumptions are more consistent with the observed hardware lifecycle, lower exposure to future impairment charges if rental rates continue compressing, and a balance sheet that more accurately represents the economic reality of deploying hardware in a market where the dominant supplier ships a materially superior architecture every 12 months.
As theCUBE Research observed, CoreWeave is the genuine outlier in adopting an aggressive six-year posture despite an exclusively AI-intensive focus. The operators who chose more conservative depreciation schedules chose to build financial models that can survive a faster hardware obsolescence cycle. As the first cohort of neocloud hardware reaches years three and four of its deployed life, the market will begin accumulating actual data on which set of assumptions was closer to accurate. The contracts expiring, the rental rates available on renewal, and the collateral valuations accessible in secondary GPU markets will all provide evidence that the financial models have not previously had to confront.
What the Roadmap Requires Operators to Confront
The most important analytical insight for understanding AI infrastructure depreciation schedules is not which number any particular company has chosen. It is what Nvidia’s own product strategy implies about the appropriate number and whether the companies whose financial models depend on the six-year assumption have genuinely modelled what happens if it proves wrong.
NVIDIA has explicitly stated that no competitor matches its annual architecture cadence and that it intends to maintain it. Vera Rubin ships in Q3 2026. Rubin Ultra targets 2027. Nvidia has already placed Feynman on its 2028 roadmap. This is not a speculative projection. It is a publicly disclosed product strategy from the company that manufactures nearly all of the hardware underpinning the industry’s six-year depreciation schedules. At the same time that operators are depreciating Nvidia hardware across six years, Nvidia is publishing a roadmap that progressively shortens the competitive durability of any given hardware generation rather than extending it.
The operators and investors who built AI infrastructure positions on six-year depreciation schedules are making a specific and verifiable bet: that the waterfall holds at revenue rates consistent with the original models, that take-or-pay contract renewals sustain the revenue assumptions embedded in the debt structures, and that no hardware generation creates a performance step change large enough to render current fleets economically impaired before their accounting life expires. Blackwell delivering 50x better performance than Hopper is the kind of step change that tests the last of those three assumptions directly. Rubin delivering 10x better inference cost than Blackwell will test it again. The six-year depreciation schedule was an accounting policy choice made in 2023. The hardware roadmap it was written against has already shipped three subsequent generations, and two more are publicly confirmed before the first of those schedules expires.
The Market Signal That Has Not Yet Been Fully Heard
Short sellers seeking to push the market lower and operators seeking to defend their accounting policies have largely framed the AI infrastructure depreciation debate as a conflict of financial incentives rather than a dispute over commercial reality. That framing obscures what the available evidence actually shows. Amazon took a $920 million accelerated depreciation charge in Q4 2024 and explicitly cited AI’s pace of development as the reason. H100 rental rates fell from $8 to $2 per GPU hour in approximately 30 months. Used H100s that were trading above retail at $50,000 per unit during the scarcity of mid-2024 are now clearing secondary markets at steep discounts as supply normalised and Blackwell became available.
Blackwell B200 pricing has compressed from over $6 per GPU hour to $2.25 and is projected to fall a further 50 to 70 percent as Vera Rubin enters production. These are not theoretical projections from bearish analysts. They are market outcomes that have already occurred.
The operators and investors who built AI infrastructure positions on six-year depreciation schedules made a specific bet at a specific moment. Many of them made it knowingly, with a clear-eyed assessment of the value cascade thesis and the take-or-pay contract structures that provide near-term revenue protection. The bet may still prove correct in a scenario where the waterfall sustains revenue at rates consistent with the original models across the full accounting life of the deployed hardware. What it is not is a certainty, and the AI infrastructure market is currently pricing it as if it were.
The Data That Will Resolve the Debate
The next 24 months will provide the data that resolves the debate. The first wave of neocloud take-or-pay contracts written in 2023 are renewing now and through 2026. The secondary GPU market is providing real-time collateral valuations that lenders can compare against the book values in their covenant agreements. Vera Rubin will enter the market in Q3 2026 and deliver the Blackwell-to-inference cascade that will test whether H100 rental rates stabilise or compress further. Rubin Ultra will follow in 2027. Feynman in 2028. Each successive generation provides another data point on whether the waterfall thesis holds at the revenue rates the debt structures require or whether the commercial life of any given GPU generation is genuinely closer to three or four years than to six.
The private credit bet on GPU infrastructure has always rested on assumptions about hardware durability that Nvidia’s own roadmap tests with every new product release. The infrastructure being deployed is genuinely valuable. The workloads that will run on it are real and growing. The question embedded in every six-year depreciation schedule in the sector is a narrower and more specific one: whether the revenue generated by that infrastructure over its accounting life will be sufficient to service the debt written against it, at the rental rates available in a market that Nvidia’s product cadence is repricing every 12 months. Amazon has already answered that question for itself and taken the earnings hit. The rest of the sector is still waiting to find out if it answered correctly.
