When the accountants come for the hyperscalers

What is the useful life of a GPU and why does that matter?

The Economist recently published an article with the scary headline The $4trn accounting puzzle at the heart of the AI cloud. The subhead reads A beancounter’s look at the hyperscalers’ balance-sheets¹.

So what’s going on here?

The Economist is essentially arguing that the AI boom is riding on accounting assumptions as much as silicon. The hypescalers are extending the useful life of their servers on the balance sheet (six years instead of four or five), which lowers annual depreciation expenses and inflates operating profit.

But the economic reality is that Nvidia is refreshing chips yearly, and Jensen Huang, its CEO, quipped that Hoppers will become nearly worthless once Blackwell ships. If the true economic life of an AI accelerator is closer to two or three years, then the hyperscalers’ reported earnings are overtstated, potentially by trillions in market value.

This all sounds very bad.

But there is nuance to be had

There are a few key points worth teasing apart.

1. Accounting vs Economic Reality

Depreciation schedules are always approximations. They smooth capital spending into steady expense recognition. But in AI, hardware obsolescence is accelerating. A six year depreciation schedule for a GPU is like assuming your iPhone 6 is still a competitive product in 2025. If investors are pricing Microsoft, Alphabet, Amazon, Meta, and Oracle on the assumption that their server fleets will be useful for half a decade, they may be misled.

2. The Leverage of a Small Adjustment

A 5-10% hit to EPS doesn’t sound catastrophic, but because these companies are so highly valued and trade at elevated multiples, even small accounting shifts ripple into trillions of market cap. That’s the article’s $4trn puzzle: change one accounting line item, and you can vaporize the equivalent of Germany’s GDP in perceived wealth.

3. Why the Hyperscalers Stretch Lifetimes

Extending useful life isn’t necessarily fraudulent; it’s a hedge. Not every AI workload needs the latest H100/B100/Rubin chip. Older GPUs can still run inference, storage, or traditional cloud tasks. By shifting old accelerators down-market, the companies extract residual value. In that sense, the six year schedule is aspirational recycling of hardware, not a pure fiction.

4. The Broader Problem

This debate shows how fragile the AI trillions are when framed through accounting. Cloud profits depend less on actual customer revenues today than on how CFOs classify and amortize sunk GPU costs. It’s reminiscent of the telecom boom in the late 1990s, when fiber build outs were capitalized and amortized over heroic time horizons. The network capacity was real, but the profits were paper.

5. Strategic Angle

For Nvidia, this dynamic is perfect. Its customers extend depreciation to flatter their income statements, while needing to buy fresh hardware annually to stay competitive. Nvidia sells into both the real demand curve and the accounting illusion. The hyperscalers get squeezed between the physics of silicon and the optics of EPS.

So is The Economist correct?

In a sense, yes, they are correct. The $4 trillion puzzle is not about fraud, but about the mismatch between finance and physics. If chip obsolescence is accelerating to 2-3 years, capital intensity will crush margins unless demand, and pricing power in AI cloud, ramps equally fast. The market is implicitly betting that software/services revenues will mask this churn. If not, we’re looking at a balance sheet trap where trillions in “value” are just depreciation assumptions in drag.

This raises a deeper question: is AI cloud a utility business, like power plants, where long asset lives justify stead returns, or is it a consumer electronics business, where assets expire every upgrade cycle? The answer determines whether the trillion dollar valuations are sustainable or not.

OK, but GPUs can be used for things other than model training and inference

This is the crux: GPUs don’t turn into e-waste once they’re superseded. The hyperscalers already have a tiered ecosystem of workloads. Cutting-edge chips (H100, B100, Rubin, etc.) go to frontier model training. Slightly older ones get pushed to inference. Older still handle recommendation systems, analytics, video transcoding, ad targeting, or even conventional compute tasks. From a systems engineering view, the value curve is not binary. Obsolescence is gradual, not a cliff.

So why do the accountants raise the alarm?

1. The Economic Life Test

Depreciation rules are meant to reflect useful economic life, which is the period where the asset is contributing revenue at roughly the level assumed when it was purchased. Chips age badly because speed increases every generation. If last year’s GPU is 50% slower, its economic contribution to top-line growth collapses, even if it still hums away at something else. Accountants say the marginal revenue per chip falls faster than the six-year schedule implies.

2. The Re-purposing Caveat

The hyperscalers argue that these assets don’t go dark. A degraded A100/Hopper still has resale value internally, in which its duty shifts to inference or externally, in which it is sold to resellers or crypto miners. That makes six years plausible as an average life across the entire fleet. The problem is, the mix is skewing: a growing fraction of spend is on bleeding-edge accelerators dedicated to workloads that demand constant refresh. Those assets probably don’t yield six years of high-margin revenue.

3. Margins vs Cash Flow

From a cash flow perspective, this is less existential than the bean counters make it out to be. The hyperscalers have the money. They just prefer to recognize expenses slowly. Even if the economic life is shorter, the servers still throw off cash as long as there’s demand for older workloads. The issue is reported profitability, not survival.

4. The Strategic Signal

What’s really at stake is the story told to investors. Stretching depreciation lives sense the signal: This is utility-like capex, with stable returns. If reality is closer to consumer electronics churn, then the valuation multiple should be lower. That’s why short-sellers like Jim Chanos hammers the point. Changing that narrative even slightly can swing hundreds of billions in market cap.

So, is it really the problem the bean counters suggest?

• As a solvency/liquidity issue: No. Hyperscalers generate enormous free cash flow, and GPUs don’t evaporate in value after 24 months.

• As an EPS/accounting optics issue: Yes. If investors are paying utility-style multiples while the assets actually churn like smartphones, then the market’s pricing is fragile.

In other words, it’s not about whether the chips still work. It’s about whether the rate of obsolescence in frontier workloads is compatible with the story Wall Street is telling itself.

---

If you enjoy this newsletter, consider sharing it with a colleague.

I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

• follow me on Twitter,

• connect with me on LinkedIn, or

• send an email to dave [at] davefriedman dot co. (Not .com!)

1. The hyphen in balance sheets is, as best I can tell, a peculiar British affectation.