The Hyperscaler Depreciation Gap: What It Means for Buyers

Microsoft, Alphabet, Amazon and Meta will spend a combined $650–700 billion on infrastructure in 2026, most of it on GPUs, servers, and the data centers that house them. Over the same trailing four quarters, those same four companies reported roughly $149 billion in depreciation and amortization — about a third of what they’re spending. That $285 billion annual gap isn’t a bookkeeping curiosity. It is being closed, quietly, by extending how long hyperscalers claim their servers and GPUs will last — and the assumptions baked into those schedules are now showing up directly in colocation pricing, GPU-as-a-service contracts, and the lease terms buyers are being asked to sign.

Why the Gap Exists — and How It’s Being Closed

Over the past three years, every major hyperscaler has quietly extended the useful life it assigns to servers and networking gear. Microsoft moved from 4 to 6 years starting in its 2023 fiscal year. Alphabet made a similar move to 6 years in January 2023. Amazon extended servers from 5 to 6 years in 2024, then partially reversed course in January 2025 — cutting a subset back to 5 years and taking roughly $920 million in accelerated depreciation on equipment retired early, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning,” per its 10-K. Meta extended most of its server and network fleet to 5 to 5.5 years in the same filing season, trimming an estimated $2.9 billion from 2025 depreciation.

Each of these moves is legal, disclosed, and prospective — meaning no restatement, just a lower depreciation charge going forward under both US GAAP and IFRS. But stacked together, independent analysts Steve Cooper and Dennis Jullens, writing in The Footnotes Analyst, estimate the cohort’s reported depreciation for 2026 runs tens of billions below what a 4-to-5-year life would produce. Investor Michael Burry has gone further, arguing in late 2025 that accelerator fleets have a real economic life closer to 2–3 years and that the industry understates depreciation by roughly $176 billion between 2026 and 2028 — a figure the Footnotes Analyst piece says it cannot independently verify from public filings, while agreeing the underlying forecasting question is real.

Why a Longer Assumed Life Isn’t Free for the Buyer

Here’s the part that matters if you’re not a hyperscaler shareholder but a buyer of capacity: whoever books a 6-year useful life on a GPU is implicitly pricing the hardware behind your lease, your cloud commit, or your GPUaaS contract as if it stays commercially competitive for six years. Nvidia’s release cadence says otherwise. H100 (2022) gave way to Blackwell (2024), then Blackwell Ultra (2025), with Vera Rubin targeted for 2026 and Vera Rubin Ultra for 2027. Microsoft CEO Satya Nadella has openly cited this cadence as a reason to favor faster, shorter build cycles rather than get “stuck with four or five years of depreciation on one generation.”

When the assumed useful life (what shows up in the accounting, and often in the pricing model) runs longer than the competitive life (how long that hardware stays attractive relative to the newest silicon), someone eats the difference. That can show up as a mid-contract price increase, a refresh fee buried in a renewal, or — most commonly — a buyer discovering in year four of a five-year colocation or GPUaaS term that their rack is running two generations behind, with no contractual path to an upgrade that doesn’t cost extra.

Where This Shows Up in Your Contract

This isn’t abstract for colocation and GPU-capacity buyers specifically, because the market structure has changed. Colocation providers are increasingly moving beyond powered shell into bare-metal and GPU-as-a-Service offerings, which DatacenterDynamics has noted means the GPUs themselves — not just the facility — now sit on the provider’s balance sheet, usually financed. CoreWeave’s $7.5 billion debt raise in 2024, secured largely by its GPU fleet and customer contracts, is the template: a special-purpose vehicle buys the hardware, debt is sized against a combination of customer contracts and an assumed residual value at maturity (typically 3–6 years out), and equity investors target a levered return on top. A three-year-old H100 can still resell for roughly 60% of its original price today — but that residual assumption gets used three separate times in the deal (lease pricing, debt sizing, and equity return modeling), and all three depend on Nvidia’s pricing discipline holding and util­ization staying high. If either slips, the “optimistic residual gets counted three times,” as one analysis of the colo financing structure put it, and the gap lands somewhere — most likely in your renewal pricing.

For buyers weighing a build-vs-buy-vs-colocate decision, this depreciation question is now a real input, not a footnote. A self-build gives you control over your own refresh cadence and lets you choose your own (honest) useful-life assumption in your TCO model. A colocation or GPUaaS contract transfers that decision to a provider whose pricing may assume a longer, more optimistic hardware life than the market will actually support — and you won’t find that assumption disclosed anywhere in the contract itself.

The Scale of What’s at Stake

The numbers underline why this isn’t a rounding error. Microsoft’s trailing-four-quarter infrastructure spend ran around $97 billion against roughly $10 billion of quarterly infrastructure depreciation in its most recent quarter alone. Meta spent $19 billion in a single quarter against $6 billion of depreciation and amortization. Amazon’s capital expenditure actually exceeded its operating cash flow in a recent quarter — a first since 2022 — while its 2026 capex guidance sits near $200 billion. Across the four companies, debt issuance tied to AI infrastructure has accelerated sharply: roughly $97 billion in AI-linked bonds priced in the fourth quarter of 2025 alone, with Meta’s Hyperion joint venture with Blue Owl Capital raising about $27 billion off Meta’s own balance sheet in October 2025. None of this capital is waiting around for a depreciation debate to resolve — it is already funding the hardware sitting in racks buyers are signing up to use.

The Counter-Argument Buyers Shouldn’t Ignore

It would be a mistake to treat every useful-life extension as a sign of trouble. Older GPUs genuinely get a second life running inference workloads after being retired from frontier training clusters, which supports a longer blended life for the fleet as a whole rather than any single chip. The Footnotes Analyst’s own framing is that this is “the real issue” but “the wrong debate” — a legitimate forecasting and componentization question under accounting standards like ASC 360 and IAS 16, not evidence of fraud, as Burry’s “one of the more common frauds of the modern era” framing implies. Residual value data — that 60% resale figure on a three-year-old H100 — also cuts against the most bearish claims. Buyers should read the controversy as a signal to ask sharper questions, not as proof that every provider’s pricing is built on sand.

What to Actually Do About It

Before signing a multi-year colocation, GPUaaS, or large cloud commit, ask the provider directly what hardware generation and refresh cadence its pricing assumes, and get in writing whether the price is fixed for the term or includes a refresh or technology-true-up clause. Model your own total cost of ownership using a 3-to-4-year competitive life for GPU-dense infrastructure even if the provider’s accounting uses 5 or 6 — the gap between those two numbers is exactly the risk you’re otherwise absorbing silently. And if you’re negotiating a long-term deal, push for contract language that specifies who bears the cost of a hardware refresh if the market moves faster than the depreciation schedule assumes — because right now, in most contracts, that party is you.

The depreciation debate playing out in hyperscaler earnings calls isn’t just an accounting story for equity analysts. It is quietly setting the economic assumptions underneath every GPU lease, colocation contract, and cloud commitment signed in 2026. Buyers who ask about the hardware-life assumption behind their pricing — and negotiate accordingly — will be in a far stronger position than those who find out the hard way at renewal. For a structured framework to evaluate these tradeoffs before your next contract, see the TechInfraHub Data Center Buyer’s Toolkit.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →