Pick up an Nvidia H100 server today and it will cost you somewhere between $250,000 and $400,000. Now ask a simple question: how many years should that server last before it's worthless?
Get that number wrong by even a year or two, and you can swing a company's reported profit by hundreds of millions of dollars — without a single dollar of cash actually changing hands. That's not a hypothetical. It's exactly what's playing out right now inside the "neocloud" industry, the wave of GPU-rental companies that have popped up to feed the AI boom, and the answer each one gives is quietly one of the most consequential accounting decisions in corporate America.
What a Neocloud Actually Is
A neocloud is a company that buys massive fleets of GPU servers — often financed with debt secured against the hardware itself and against contracted customer revenue — and rents that computing power out by the hour to AI labs, startups, and enterprises that don't want to buy their own chips.
The business model sounds simple: buy hardware, rent it out, collect the spread. But the accounting underneath it is anything but simple, because it hinges entirely on one assumption — how long the hardware will remain useful before it needs to be replaced.
The Same Server, Three Different Answers
Here's where it gets interesting. Three of the best-known neoclouds have each picked a different answer to "how long does a GPU server last":
- CoreWeave depreciates its GPU equipment over 6 years
- Lambda uses a 5-year schedule
- Nebius uses a more conservative 4-year schedule
Same category of hardware, same underlying Nvidia chips, three different useful-life estimates — and each one produces a materially different income statement.
Do the math on a single $300,000 server:
- Over 6 years (straight-line), that's about $50,000 a year in depreciation expense
- Over 4 years, that's about $75,000 a year
That $25,000-per-server gap doesn't sound like much until you scale it. Across a fleet of 10,000 GPUs, the difference between a conservative and an aggressive depreciation schedule works out to roughly $30 million a year in reported expense — on physically identical hardware doing physically identical work. One company's income statement can show tens of millions more in profit than another's, purely because of a bookkeeping estimate, not because the business actually performed better.
A Quick Refresher on How Depreciation Actually Works
If you haven't touched accounting since a college survey course, here's the short version. When a business buys a long-lived asset — a server, a delivery van, a commercial oven — it generally can't deduct the full purchase price as an expense in the year it's bought. Instead, the cost gets spread across the years the asset is expected to be useful, through periodic depreciation charges. Two variables drive the size of each charge:
- Useful life — how many years (or units of output) the asset is expected to remain productive
- Residual (salvage) value — what the asset is estimated to be worth once that useful life ends, which gets subtracted from the depreciable base
The simplest method, straight-line depreciation, just divides (cost minus residual value) evenly across the useful life. A $300,000 server with zero assumed residual value, depreciated straight-line over 6 years, produces a $50,000 annual charge; over 4 years, $75,000. Some companies use accelerated methods that front-load even more expense into the early years, on the theory that new technology loses most of its value fast and slows down later — which is arguably a more realistic model for something as fast-moving as GPU hardware.
Crucially, depreciation is a non-cash expense. The cash left the business back when the server was purchased (or as loan payments come due, if it was financed). Depreciation is simply the accounting mechanism for recognizing that spend over time instead of dumping it all into one year's income statement. That's exactly why it's such fertile ground for aggressive assumptions: nobody is writing an actual check for "depreciation" this month, so a longer useful-life estimate costs nothing today and quietly inflates reported profit for years.
Why This Isn't Just an Accounting Footnote
This matters because depreciation schedules aren't neutral. A longer schedule spreads the cost thinner, which makes near-term profit look bigger. A shorter schedule front-loads the expense, which looks more conservative but drags reported earnings down in the early years.
CoreWeave is a good illustration of how wide that gap can be between "the business is doing fine" and "the business is profitable." In its most recent fiscal year, CoreWeave posted roughly 60% adjusted EBITDA margins — genuinely strong — while simultaneously reporting a net loss of about $1.17 billion. The bridge between those two numbers included roughly $2.45 billion in depreciation (a non-cash charge) and about $1.2 billion in interest expense on the debt used to buy the hardware (a very real cash charge). Revenue was around $5.1 billion against roughly $21.4 billion in debt, with two-thirds of that revenue coming from a single customer.
None of that means the depreciation number is "wrong." It means the useful-life estimate is doing an enormous amount of work in determining what "profitable" even means for these companies.
The $176 Billion Question
The debate over GPU useful life escalated in late 2025 when investor Michael Burry — famous for his early call on the 2008 mortgage crisis — publicly argued that major AI infrastructure spenders, including Meta, Amazon, Microsoft, Google, and Oracle, were depreciating their Nvidia GPUs over five to six years when the real economic life is closer to two or three. His estimate: roughly $176 billion of understated depreciation, and therefore overstated profit, across the industry between 2026 and 2028. He specifically flagged Oracle and Meta as having profits potentially overstated by around 27% and 21%, respectively, by 2028.
The counterargument from the companies themselves is that a GPU's useful life doesn't end when it's no longer fast enough for cutting-edge model training. Instead, it "cascades" — aging hardware gets reassigned from the most demanding workloads (training frontier models) down to less glamorous but still highly profitable work (inference, smaller models, batch processing). A four-year-old GPU that's useless for training the next frontier model can still make money running inference for years afterward.
Who's right matters enormously for anyone reading these companies' financial statements, because it determines whether reported profits reflect real economic performance or an optimistic bet dressed up as an accounting policy.
The Lesson for Every Business Owner, Not Just AI Giants
You don't need a fleet of GPUs to run into this exact problem. Any business that buys equipment — a bakery's ovens, a landscaper's trucks, a dentist's imaging equipment, a software company's servers — faces the identical question: how many years of useful life do I assign to this asset, and how does that choice change what my books say about profitability?
Take a concrete example. Say a small web agency buys a $40,000 rack of servers to run client sites and internal tools. Depreciate it over 5 years and the annual charge is $8,000. Depreciate it over 3 years — a defensible call, since server hardware genuinely ages fast — and the charge jumps to about $13,333 a year. In year one, that's a roughly $5,300 difference in reported profit, purely from an estimate, with zero difference in the cash actually spent or the work actually delivered. If that agency is applying for a loan, courting an investor, or simply trying to understand whether last quarter was actually a good quarter, the schedule they picked is quietly shaping the answer.
This is also exactly the kind of decision that interacts with tax law, not just financial reporting. In the U.S., businesses often have a choice between depreciating an asset over its financial "book" useful life for their own management reporting, while using IRS-prescribed MACRS schedules (or accelerated options like Section 179 expensing or bonus depreciation) for tax purposes. Those two schedules frequently diverge on purpose — a business might depreciate a server over 5 years on its internal books while writing it off much faster for tax purposes, which is perfectly legal but means "profit" on your books and "taxable income" on your return are never quite the same number. Understanding that gap, instead of being surprised by it every April, is one of the quieter benefits of clean bookkeeping.
A few things worth internalizing from the neocloud story:
- Depreciation is an estimate, not a fact. Two honest, reasonable people can look at the same asset and pick different useful lives. That's not fraud — it's judgment. But it means you should understand why your schedule is set the way it is, not just accept a default.
- EBITDA and net income tell different stories on purpose. A business can be cash-flow healthy (strong EBITDA) while showing a net loss because of heavy depreciation and interest from debt-financed growth. Neither number alone tells you the whole picture — you need both.
- Fast-depreciating assets deserve shorter schedules; durable ones deserve longer ones. If you're buying technology that becomes obsolete quickly, an aggressive (shorter) depreciation schedule is the more honest one, even if it makes this year's numbers look worse.
- Utilization and residual value matter as much as the schedule. A GPU sitting idle depreciates the same on paper whether or not it's earning revenue. The real health check isn't the depreciation line — it's whether the asset is actually generating enough cash to justify what you paid for it.
Keep Your Depreciation Schedules Honest and Transparent
Whether you're running a single food truck or scaling a software company, the assumptions behind your depreciation schedule directly shape what your financial statements say about how well the business is really doing — and vague or inconsistent record-keeping makes those assumptions impossible to audit, even by yourself. Beancount.io gives you plain-text accounting where every asset, every depreciation entry, and every assumption behind it lives in a transparent, version-controlled ledger instead of a black box. Get started for free and see exactly how your numbers add up, no guesswork required.