The AI data center buildout is the largest capital expenditure cycle in the history of technology. A genuine, serious debate has emerged over whether it is overbuilt, and that debate turns on a single question that happens to sit squarely in the world of IT asset disposition: how long does a GPU actually stay useful, and what happens to it when it does not? Here is a balanced look at the argument, and the disposition angle almost no one is covering.
TL;DR
The AI infrastructure debate is no longer about whether AI demand is real. By 2026 it had shifted to the financial mechanics beneath the buildout, and one of them is fundamentally an asset-disposition question:
- The buildout is staggering. Estimates put 2026 hyperscaler AI capex around $725 billion, part of a projected multi-trillion-dollar cumulative spend through the end of the decade.
- The debate centers on GPU useful life. Hyperscalers depreciate AI hardware over roughly 5 to 6 years, while critics (most prominently Michael Burry) argue the real economic life is closer to 2 to 3 years given NVIDIA’s roughly annual architecture cadence.
- The gap is enormous. Burry estimated the industry may understate depreciation by roughly $176 billion across 2026 to 2028 if the shorter useful life is correct.
- The bull case is the “value cascade.” GPUs do frontier training for a year or two, then cascade to inference and lower-tier work, extending their useful life, which is how operators justify the longer schedules.
- The stranded-asset scenario is the bear case. If the buildout outpaces monetization, or if hardware becomes uneconomic faster than expected, the result is stranded GPU and data center assets, distressed liquidation, and early retirement waves, potentially concentrated around 2027 to 2028.
The disposition angle that almost no coverage addresses: whether the buildout continues at full speed or cools, enormous volumes of AI hardware are heading toward retirement, and the entire debate is, at its core, an argument about depreciation and disposition. What retired and stranded AI hardware is actually worth, and how efficiently it can be recovered, is not a footnote to this debate. It is close to the center of it.
Note: This article discusses a contested financial and economic debate for educational purposes. It is not investment advice, and ROC Telecom is not a financial advisor. It presents multiple perspectives on an unresolved question and does not predict market outcomes.
The Scale of What Has Been Built
To understand the debate, start with the numbers, because they are genuinely without precedent.
The four largest hyperscalers spent roughly $410 billion on infrastructure in 2025 and have guided toward approximately $725 billion for 2026, a jump of more than 70 percent. Goldman Sachs has modeled trillions of dollars of cumulative AI capex through the end of the decade, with baseline estimates implying roughly $765 billion in annual AI capex in 2026 growing toward $1.6 trillion by 2031. By some measures, technology equipment and software investment reached roughly 4.4 percent of US GDP in 2025, near the dot-com peak, and AI-related capital expenditure has become a primary support for US economic growth.
This is not a sector allocating capital in the ordinary sense. It is a sector at war over it, deliberately engineering a year of negative free cash flow (for the first time in decades for some of these firms) on the conviction that AI’s revenue potential justifies the sacrifice. Whatever one concludes about the bubble question, the scale is real, the demand for chips is backed by purchase orders sold out well into the future, and the physical buildout is happening.
The question is not whether the buildout is real. It is whether the economics underneath it hold.
The Core of the Debate: How Long Does a GPU Last?
Strip away the noise and the entire AI-bubble debate reduces, to a remarkable degree, to one variable: the useful life of a GPU. This is where the disposition world and the financial world meet.
Why Useful Life Is an Earnings Lever
Depreciation spreads an asset’s cost across its estimated useful life. A longer useful life means a smaller annual expense and a larger reported profit today. The math is direct: a six-year schedule recognizes about 16.7 percent of a chip’s cost per year, while a three-year schedule recognizes 33.3 percent. Stretching the schedule from three years to six roughly halves the annual depreciation hit, which flows straight to reported earnings.
Over recent years, major hyperscalers progressively extended their assumed useful life for this hardware, moving from three or four years toward six. At the buildout’s current scale, that shift is seismic for reported earnings.
The Bear Argument
The most prominent skeptic is Michael Burry, known for anticipating the 2008 housing crisis. His argument is simple and hard to dismiss: NVIDIA now ships a new architecture on roughly an annual cadence (A100 in 2020, H100 in 2022, Blackwell in 2024, Blackwell Ultra in 2025, with Vera Rubin generations projected for 2026 and 2027), each generation dramatically more efficient per watt, which renders prior-generation silicon uneconomic for frontier training almost immediately. Depreciating a chip over six years, he argues, implicitly assumes five or six chip generations coexist productively, which strains credibility. His estimate: the industry may understate depreciation by roughly $176 billion across 2026 to 2028. The precise figure matters less than the mechanism, and the inputs are public, audited, and sitting in the filings.
The Bull Rebuttal: The Value Cascade
The counterargument is the “value cascade.” A GPU does frontier training for a year or two, then, rather than becoming worthless, cascades down to high-value real-time inference, then to lower-tier inference and other work, remaining productive for years. Microsoft’s leadership has publicly argued that newer AI models can run on older GPU fleets, and that the company optimizes across all of it. In this view, the longer depreciation schedules are justified because the hardware genuinely keeps producing value well past its frontier-training window.
Notably, even bulls concede the tension. Some hyperscaler leaders have described a “speed-of-light execution” approach precisely to avoid being stuck with several years of depreciation on a single generation, and at least one has cited the rapid pace of NVIDIA migrations as a reason to pull back certain builds. The disagreement is real and unresolved, and it turns on exactly the question ITAD professionals think about every day: what is aging hardware actually still good for, and what is it worth?
The Stranded-Asset Scenario
The bear case does not require believing AI is fake. As one AI builder put it, a bubble is not always about inherent value, sometimes it is about a structural misalignment between the pace of innovation and the accounting used to justify the spending. So what does the downside scenario actually look like, and where does the hardware go?
What “Stranded” Means Here
An asset becomes stranded when it can no longer generate the economic value its cost assumed. For a GPU, that happens if it becomes operationally obsolete or uneconomic to run before its depreciation schedule expires, because a new generation delivers dramatically better performance per dollar. The operator is then carrying the cost of an asset that no longer earns its keep. Scale that across a buildout measured in hundreds of billions of dollars per year, and the stranded-asset question becomes macro-economically significant. Some analysts flag a first wave of stress around 2027 to 2028, as initial data center lease terms come up for renewal and early financing assumptions meet reality.
The Fiber-Glut Parallel, and Its Limit
The most common historical analogy is the telecom and fiber-optic bust of 2000 to 2001, when overbuilding was estimated at around 85 percent and vacancy exceeded 20 percent. The bull version of this analogy is reassuring: the dark fiber stranded in 2001 carried the traffic of the 2010s, because glass in the ground has a thirty-year life and amortized against three decades of later demand.
But the analogy has a crucial limit that cuts the other way, and it is a disposition point. Fiber stranded gracefully because it lasted. GPUs do not. Where dark fiber sat cheaply in the ground waiting for demand to catch up, AI accelerators carry high ongoing costs (power, cooling, maintenance) regardless of utilization, and they age out on a 3-to-4-year cycle. That means overcapacity in AI hardware does not wait patiently to be used, it rapidly converts into obsolescence. A bubble in short-lived assets deflates faster and more completely than one in long-lived infrastructure, precisely because the assets depreciate so quickly. The margin for error is tighter than the dot-com era, not wider.
Where the Hardware Would Go
Here is the part the financial coverage rarely follows through on. If a meaningful slice of the buildout is stranded, that hardware does not vanish. It enters the secondary market, all at once, under distress:
- Distressed liquidation. Operators unwinding stranded capacity would liquidate GPU fleets, potentially in volume and under time pressure, which is a very different disposition problem than an orderly refresh.
- Early retirement waves. Hardware retired ahead of its planned schedule floods the market with equipment that still has real, if diminished, value.
- Recovery under pressure. In a distressed scenario, the difference between an informed asset recovery and a panicked fire sale is enormous, and it determines how much of the stranded value is actually recovered rather than lost.
Even some policy thinkers have begun contemplating this: a widely discussed 2026 analysis floated the idea of turning stranded data center assets into a public cloud resource, a striking sign of how seriously the scenario is being taken in some quarters.
The Case That It Is Not a Bubble
Intellectual honesty requires giving the bull case its due, because it is substantive and does not rest on hype.
First, the demand is real and contracted. Chip inventory has been reported sold out 18 to 24 months forward, representing hardware demand backed by purchase orders rather than projections. Analysts have found little evidence of irrational, speculative spending in the way the dot-com era saw.
Second, the mechanics favor a delayed payoff. Capex in this buildout is a step function: compute, data centers, and power are sunk before a dollar of inference revenue arrives. Monetization is an S-curve that begins only once capacity is running. The two curves are offset in time by construction, which means peak capex intensity is the trough of the benefit-to-cost ratio, not its steady state. The ratio improves mechanically the moment spending shifts from building to running, even if technology progress stopped entirely.
Third, transformative general-purpose technologies historically show a lag. Electricity and the internet delivered little measured productivity gain for years or decades before the benefits materialized. If AI follows that pattern, the absence of immediate GDP impact is early-stage behavior, not failure.
Fourth, capacity remains constrained today. Colocation vacancy fell to record lows by the end of 2025, and one credit analysis noted that AI cloud infrastructure was expected to remain capacity-constrained in 2026 regardless of whether investors believe a bubble exists. It is hard to be simultaneously overbuilt and sold out, which is part of why the debate is genuinely unresolved.
The honest synthesis: the thesis is conditional. Benefits outrun capex if and only if the useful life of the hardware holds up better than the fastest bears fear. That is one variable, and it is, once again, a disposition question.
Why This Is Fundamentally a Disposition Question
Step back and notice what the entire debate is actually about. It is not really about whether AI is useful. It is about how quickly the hardware loses value, what it is worth as it ages, and what happens to it at end of life. Those are the exact questions IT asset disposition exists to answer.
Consider how disposition sits at the center of every branch of the argument:
- The depreciation debate is a useful-life debate, and useful life is determined by what aging hardware can still do and what it is still worth, which is assessed every day in the secondary market.
- The value-cascade bull case depends on secondary utility, the ability of older GPUs to keep earning in inference and lower-tier roles. That cascade is, in effect, a managed depreciation curve, and the secondary market is where its real slope gets revealed.
- The stranded-asset bear case is a disposition scenario, distressed liquidation and early retirement at scale, where recovery expertise determines how much value survives.
- The residual-value question underpins the collateral question. Some analyses warn that GPU-collateralized debt assumes collateral values that may prove illusory in default, because the hardware depreciates like consumer electronics, not like buildings. What that collateral is actually worth is a secondary-market question.
In other words, whether the optimists or the pessimists are right, the resolution runs through the disposition of the hardware. If the bulls are right, it is because the value cascade holds and older GPUs retain useful economic life. If the bears are right, it is because the hardware strands faster than the accounting assumed. Either way, what AI hardware is worth as it ages, and how efficiently it can be recovered or retired, is not peripheral to the AI-bubble debate. It is close to its core.
What This Means for Operators, in Either Scenario
The useful takeaway does not require picking a side in the bubble debate, because the disposition implications point the same direction whether the buildout continues at full speed or cools.
If the buildout continues: the compressed refresh cycles that fuel the depreciation debate mean AI hardware is retiring faster than ever, generating a steady, growing stream of high-value GPUs into the secondary market. Capturing that value through informed recovery is a live opportunity right now, not a future one.
If the buildout cools: stranded and early-retired hardware would enter the market under distress, where the gap between expert recovery and a panicked fire sale is at its widest. In that scenario, disposition expertise is not just about capturing value, it is about limiting loss.
In both cases, the operators who benefit are the ones who treat AI hardware disposition as a strategic function rather than an afterthought. That means understanding real secondary-market values rather than relying on depreciation-schedule assumptions, moving quickly to capture value before generational decay erodes it, and having recovery relationships in place before the hardware needs to move, not scrambling after. The single most useful posture, in a debate this unresolved, is to be ready for either outcome. Well-run disposition is exactly that readiness.
Frequently Asked Questions
The following is general educational information about a contested economic debate, not investment or financial advice.
Is there an AI bubble?
This is genuinely unresolved, and credible analysts disagree. By 2026 the debate had moved past whether AI demand is real (it is, backed by chip orders sold out 18 to 24 months forward) to the financial mechanics beneath the buildout. Skeptics point to aggressive depreciation assumptions, circular financing among key players, negative free cash flow, and the risk that inference revenue cannot grow fast enough at falling prices to service the debt and depreciation. Optimists point to constrained capacity, contracted demand, the natural time lag between building and monetizing infrastructure, and the historical pattern of transformative technologies paying off on a delay. The honest answer is that the question turns on a small number of contested variables, chiefly how long the hardware stays economically useful.
What are stranded assets in the AI context?
A stranded asset is one that can no longer generate the economic value its cost assumed. For AI infrastructure, a GPU becomes stranded if it turns operationally obsolete or uneconomic to run before its depreciation schedule expires, typically because a newer generation delivers dramatically better performance per dollar. Data centers and power infrastructure built for a particular demand level can also strand if that demand does not materialize. The concern is that if the AI buildout outpaces monetization, or if hardware ages out faster than assumed, large volumes of GPU and data center assets could be stranded, with some analysts flagging a potential first wave of stress around 2027 to 2028.
Why does GPU depreciation matter so much to the AI bubble debate?
Because it is the single variable the whole debate turns on. Depreciation spreads an asset’s cost over its useful life, so a longer assumed life means lower annual expense and higher reported profit. Hyperscalers depreciate AI hardware over roughly 5 to 6 years, while critics argue the real economic life is closer to 2 to 3 years given NVIDIA’s roughly annual architecture cadence. If the shorter life is correct, reported earnings across the industry may be overstated substantially (one prominent estimate put it at roughly $176 billion across 2026 to 2028). Whether the longer schedules are justified depends on whether older GPUs retain real economic value, which is fundamentally a secondary-market and disposition question.
What is the “value cascade” for GPUs?
The value cascade is the leading argument for why longer depreciation schedules can be justified. The idea is that a GPU does frontier AI training for its first year or two, then cascades down to high-value real-time inference, then to lower-tier inference and other workloads, remaining productive for several years rather than becoming worthless when the next generation arrives. If the cascade holds, the hardware genuinely keeps earning across a longer useful life. The bear counterargument is that each new NVIDIA generation is so much more efficient per watt that older silicon becomes uneconomic to run for frontier work almost immediately, compressing the real useful life well below the accounting assumption.
How does the AI buildout compare to the dot-com and fiber bubble?
The telecom and fiber-optic bust of 2000 to 2001 is the most common analogy, with overbuilding then estimated at around 85 percent. The reassuring version notes that stranded dark fiber eventually carried the traffic of the 2010s, because fiber has a roughly 30-year life and amortized against decades of later demand. But there is a critical difference that cuts against complacency: fiber stranded gracefully because it lasted and cost almost nothing to maintain, while GPUs carry high ongoing power and cooling costs and age out on a 3-to-4-year cycle. This means AI overcapacity converts rapidly into obsolescence rather than waiting to be used, so an AI bubble, if it exists, could deflate faster and more completely than the fiber glut did.
What happens to GPUs if the AI buildout slows down?
They enter the secondary market, potentially in volume and under distress. A slowdown would likely produce distressed liquidation of GPU fleets, early retirement of hardware ahead of its planned schedule, and pressure to recover value quickly. In that scenario, the difference between informed asset recovery and a panicked fire sale becomes very large, and it determines how much of the stranded value is actually recovered. This is why the stranded-asset question is fundamentally a disposition question: the hardware does not disappear, it needs to be retired, recovered, and remarketed, and how well that is done determines how much value survives.
Is AI hardware a good collateral for debt?
This is a growing concern in credit analysis. A significant amount of AI infrastructure has been financed with debt, some of it collateralized by the GPUs themselves. The worry is that GPU collateral depreciates like consumer electronics rather than like buildings, so collateral values assumed in underwriting may prove optimistic if the hardware ages out faster than expected, particularly at lease renewals around 2027 to 2028. Whether that collateral holds its assumed value is, again, a secondary-market question: it depends on what used GPUs actually fetch as newer generations arrive, which is precisely what the disposition market reveals.
Does ROC Telecom think AI is a bubble?
ROC Telecom does not take a position on whether AI infrastructure is a bubble, and this article is not investment advice. What ROC observes from the disposition side is narrower and more concrete: the entire debate turns on how quickly AI hardware loses value and what happens to it at end of life, which are the questions ITAD exists to answer. Whether the buildout continues at full speed (producing a steady stream of high-value retirements) or cools (producing distressed liquidation), large volumes of AI hardware are heading toward disposition, and how efficiently that hardware is recovered determines how much value is preserved. The disposition implications point the same direction regardless of who wins the bubble argument.
How should operators prepare for either outcome?
By treating AI hardware disposition as a strategic function rather than an afterthought, which hedges against both scenarios. That means understanding real secondary-market values rather than relying on depreciation-schedule assumptions, moving quickly to capture value before generational decay erodes it, maintaining recovery and buyer relationships before the hardware needs to move, and ensuring certified data destruction and chain-of-custody are in place regardless of market conditions. If the buildout continues, this posture captures the value in a steady retirement stream. If it cools, it limits loss in a distressed one. In an unresolved debate, readiness for either outcome is the sound position.
The Bottom Line
The AI-bubble debate is one of the most consequential and genuinely unresolved questions in the economy, and it hinges to a remarkable degree on a single variable that sits in the world of IT asset disposition: how long a GPU stays economically useful, and what it is worth as it ages. The bulls argue the value cascade holds, older GPUs keep earning, and the longer depreciation schedules are justified. The bears argue the hardware strands faster than the accounting assumes, with roughly $176 billion of potentially overstated earnings and a possible stranded-asset reckoning around 2027 to 2028. Both cases are serious, and the resolution runs directly through the disposition of the hardware.
For infrastructure operators, the practical takeaway does not require picking a side. Whether the buildout continues at full speed or cools, large volumes of AI hardware are heading toward retirement, and the difference between capturing that value and losing it comes down to disposition expertise. If the buildout continues, compressed refresh cycles generate a steady stream of high-value retirements to recover. If it cools, distressed liquidation makes expert recovery the difference between preserving value and a fire sale. In an unresolved debate, the operators who prepare for either outcome (understanding real secondary-market values, moving quickly, and having recovery relationships in place) are the ones positioned to come out ahead no matter which way the argument resolves.
How ROC Telecom Helps
ROC Telecom is an R2v3, RIOS, NIST 800-88, and ITAR-compliant ITAD specialist built for exactly the disposition questions this debate raises:
- Real secondary-market valuation of AI and GPU hardware, grounded in an active buyer network rather than depreciation-schedule assumptions
- Rapid asset recovery with speed-to-remarketing that protects value against the generational decay at the heart of the depreciation debate
- Distressed and volume liquidation capability for operators unwinding or consolidating AI infrastructure, where informed recovery beats a fire sale
- Direct buyer relationships across enterprises, neoclouds, and research institutions that convert retired GPUs into recovered capital
- NIST 800-88 data destruction with per-asset serialized Certificates of Destruction and full chain-of-custody, regardless of market conditions
- R2v3 Appendix E materials recovery for hardware genuinely at end of life, with direct-to-refiner processing
- Pre-shipment payment on buybacks, giving operators certainty in either a steady or a distressed market
15+ years of ITAD experience, $25M+ in client capital recovered, 45M+ pounds diverted from landfill.
ROC Telecom does not provide investment or financial advice and takes no position on the AI-bubble debate. We help operators recover value from AI hardware disposition in any market condition.
Request a Free AI Infrastructure Disposition Assessment
Whether you are refreshing on schedule or reassessing capacity, tell us about the AI infrastructure you are retiring or considering retiring. A specialist will discuss real recovery value, secure data destruction, and disposition strategy for your situation. No commitment, no spam. Prefer to talk directly? Call 585-406-1249 or email info@roctelecom.com.
"*" indicates required fields
Related reading:
- Why AI Is Shortening Your Hardware Lifecycles (and What to Do About It)
- How GPU Decommissioning Differs from Standard Server Retirement
- Data Center Asset Recovery: The 2026 Strategic Guide
- The 10 Data Center Liquidation Mistakes That Drain Your IT Budget
- GPU Export Controls and ITAD: What ECCN 3A090 Means When You Retire and Resell AI Hardware
- Hyperscale ITAD vs Enterprise ITAD: What’s Actually Different in 2026
- The Northern Virginia Data Center Market: A 2026 Overview
