A company can report a $1 billion net loss in the same year its core business runs at a 60% profit margin. That isn't a hypothetical — it's what CoreWeave, one of the largest GPU cloud providers, booked in 2025. The gap between "operationally thriving" and "billion-dollar loss" comes down to a single number that most people never think about: how many years a company claims a piece of equipment will last.
That number is called useful life, and the assumption behind it is currently a multi-billion-dollar argument playing out across the entire AI industry. It's also one of the most misunderstood — and most consequential — decisions any business makes about its own equipment, whether that equipment is a $300,000 server rack or a $3,000 laptop.
The $30 Million Question: How Long Does a GPU Server Last?
When a business buys equipment, it doesn't expense the full cost the day it's purchased. Instead, accounting rules spread that cost — depreciation — over the years the equipment is expected to remain useful. Buy a $120,000 delivery van you expect to run for 6 years, and you recognize roughly $20,000 of expense per year, not $120,000 all at once.
The catch: "expected to remain useful" is an estimate, not a fact. And in the GPU cloud business — companies that buy racks of Nvidia chips and rent out computing power to AI labs — that estimate varies wildly by company.
CoreWeave depreciates its GPU servers over 6 years. Lambda uses 5 years. Nebius uses 4. On paper, that sounds like a rounding difference. In dollars, it isn't. A $300,000 eight-GPU server generates about $50,000 of annual depreciation expense on a 6-year schedule, versus $75,000 on a 4-year schedule. Multiply that gap across a 10,000-GPU fleet, and the choice of depreciation schedule alone can swing reported annual expenses by roughly $30 million — with zero difference in the actual hardware, the actual revenue, or the actual cash in the bank.
That's the paradox at the center of the AI infrastructure boom: the same business, the same machines, the same customers, can look wildly profitable or hemorrhaging cash depending entirely on an assumption buried in a financial-statement footnote.
Why a "Profitable" Company Can Report a Massive Loss
CoreWeave is the clearest example. In 2025, the company generated $5.13 billion in revenue and roughly $3.1 billion in adjusted EBITDA — a margin near 60%, which would make almost any business owner happy. Yet CoreWeave's net loss for the year topped $1 billion.
The difference is two large non-cash and financing items: about $2.45 billion in depreciation and roughly $1.2 billion in interest expense on the debt used to buy the hardware in the first place. Depreciation and interest are real economic costs, but they aren't cash leaving the building that day — they're the accounting system's way of spreading a big upfront purchase (and the debt used to finance it) across future years.
None of this means CoreWeave's numbers are fake. It means that "profitable" and "reports a profit" are two different questions, and the bridge between them runs directly through management's estimate of how long its equipment will last.
The Bigger Fight: Michael Burry's $176 Billion Claim
The GPU cloud upstarts aren't the only ones under scrutiny — the hyperscalers are too. Investor Michael Burry (known for correctly betting against the 2008 housing bubble) has argued that Meta, Amazon, Microsoft, Google, and Oracle are depreciating their Nvidia GPUs over 5 to 6 years, when the real economic life — given Nvidia's shift to releasing a new, meaningfully faster chip architecture roughly every year instead of every two — is closer to 2 to 3 years.
Burry's math: if the true useful life is 3 years instead of 6, the industry is understating depreciation expense by $50–60 billion a year, compounding to an estimated $176 billion of overstated profits across 2026–2028. His argument isn't that the accounting is illegal — it's that longer useful-life assumptions make current profits look better than the underlying economics support, because expensive chips may be technologically obsolete for cutting-edge work well before the depreciation schedule says they're "used up."
The defense from the hyperscalers is what's sometimes called the value-cascade theory: a GPU doesn't retire when a newer model ships. It gets redeployed — from training frontier models, to running inference, to handling lighter internal workloads — extending its genuinely useful economic life even after it's no longer state-of-the-art.
What makes this more than a theoretical debate is that real companies have already moved in opposite directions on it. In 2025, Amazon shortened the estimated useful life of a subset of its servers from 6 years to 5, explicitly citing the accelerating pace of AI hardware development — and took a $920 million accelerated depreciation charge as a result. In the same period, Meta did the opposite: it extended its server useful lives to 5.5 years, which reduced reported depreciation expense by $2.9 billion. Same underlying technology cycle, same industry, opposite conclusions — which tells you how much genuine judgment (and incentive) is baked into this one number.
The Other Risks Hiding Behind the Depreciation Line
Depreciation assumptions aren't the only soft spot in the GPU cloud business model, and they're worth understanding if you're evaluating any capital-intensive company — as a customer, an investor, or just a curious reader:
- Thin operating margins. Break-even analysis on debt-financed GPU fleets suggests providers need to clear roughly $1-plus per GPU-hour just to cover costs, against market rental rates that have fluctuated between about $1.70 and $2.35. That's a real margin, but not a huge cushion if rental prices soften.
- Customer concentration. Microsoft alone represented 67% of CoreWeave's 2025 revenue. A business built this dependent on one customer carries risk that a healthy income statement doesn't show.
- Residual value risk. GPUs are often used as loan collateral. Every time Nvidia ships a faster chip, the resale value of the previous generation drops, which weakens the collateral backing the debt used to buy it.
- Utilization dependency. An idle GPU still depreciates and still accrues interest on the debt that paid for it — it just doesn't generate revenue while doing so. Utilization rate is arguably more important to real profitability than the headline margin.
What This Means If You Run a Business (Not a GPU Cloud)
You don't need a 10,000-GPU fleet for this lesson to apply. Every business that buys equipment — a laptop, a delivery vehicle, a espresso machine, a CNC router — makes the same kind of useful-life estimate, just at a much smaller scale. And the same distortions are available to you, for better or worse.
Under U.S. tax rules for 2026, the Section 179 deduction lets you immediately expense up to $2,560,000 of qualifying equipment (phasing out above $4,090,000 in purchases), and bonus depreciation sits at 100% for qualifying property placed in service after January 19, 2025. In plain terms: for most small businesses, current law already lets you write off equipment fast, which sidesteps a lot of the "which useful life should I pick" debate for tax purposes.
But your books — the records you use to actually understand your business, separate from your tax return — are a different story, and this is where the GPU cloud lesson matters most:
- Match your useful-life estimate to reality, not to the number that makes this quarter look best. If you consistently replace laptops every 3 years, don't depreciate them over 5 just because it smooths out expenses. The IRS's own class-life tables (Publication 946) are a reasonable anchor — adjust from there only when you have a specific reason (how you actually use the equipment, your maintenance history, manufacturer guidance).
- Document your reasoning once, and revisit it only when the facts change — a new model releasing isn't automatically a reason to change your estimate; a real shift in how long you actually keep and use equipment is.
- Watch for the "flip-flop" tell. If a business changes its useful-life estimates in whichever direction makes the current period's numbers look better — like the accelerated Amazon/Meta example above, moving in opposite directions in the same industry — that's a signal to look closer, whether it's your own bookkeeping habits or a company you're evaluating.
- Separate cash flow from profit in your own head. A business (yours included) can be cash-flow healthy while showing an accounting loss, or vice versa. Depreciation and interest are the two line items most responsible for that gap — know which one is driving your numbers before you panic or celebrate.
The GPU cloud story is really a magnified version of a decision every equipment-owning business makes: how long do you honestly expect this thing to last? Get that estimate right, and your financial statements tell you the truth about your business. Get it wrong — deliberately or not — and you can end up managing a fiction instead of a company.
Keep Your Depreciation Assumptions Honest and Auditable
Depreciation schedules are exactly the kind of decision that benefits from a clear, permanent record of what you assumed and why — not a number buried in spreadsheet formatting that's easy to quietly change later. Beancount.io gives you plain-text accounting with full version history, so every adjustment to a useful-life estimate, every depreciation entry, and every reason behind it is transparent and traceable over time — no black boxes, no vendor lock-in. Get started for free and keep your books as auditable as the assumptions behind them deserve to be.