# How Long Do AI GPUs Last? The Depreciation Debate in Numbers > What CoreWeave, Nebius, Alphabet, Meta, Amazon, Microsoft and Oracle assume in their 10-Ks, how much one year of useful life moves profit, a worked straight-line example, and the evidence on both sides of the argument. Source: https://www.exent.ai/resources/gpu-depreciation-debate · Reviewed by David, Senior Analyst, Exent · Published Nov 3, 2026 · Updated Nov 3, 2026 ## Key takeaways - In their accounts, AI servers last five to six years: CoreWeave and Alphabet use 6 years, Meta 5.5, Nebius and IREN 5, Oracle 6, Microsoft a 2–6 year range. How long GPUs stay economically useful is the open question. - The choice moves profit by billions: Meta’s move to 5.5 years cut 2025 depreciation by $2.92B, while Amazon’s cut of some servers from 6 to 5 years added $1.4B, citing the pace of AI development. - Worked example: a $1B GPU cluster costs $167M a year in depreciation over 6 years and $333M over 3. If its real life is 3 years, $500M of book value is left when it stops earning. - Critics such as investor Michael Burry argue longer lives flatter earnings; Nebius’s Q2 2026 letter reported older-generation GPU prices more than 30% higher than in Q1, evidence the other way for now. - For a neocloud the stakes are larger than for a hyperscaler: at Jun 30, 2026, property and equipment was $46.7B of CoreWeave’s $77.1B of assets. ## How long do AI GPUs last? In company accounts, five to six years; in the market, nobody knows yet, because the current GPU generations have not been around that long. A GPU can keep running for years after purchase, so the real question is how long it earns enough rent to justify its cost while new generations keep arriving (NVIDIA said in August 2026 that its Vera Rubin platform was ramping into full production, with racks running at partners including CoreWeave and Nebius). Accountants call the first number the useful life; investors care about the second, the economic life. Scope: how listed AI cloud companies and hyperscalers depreciate servers and GPUs in their SEC filings, and why it matters for earnings, debt and stock analysis. It does not cover gaming GPUs, tax depreciation or resale prices of individual chips. Background on the companies is in [what is a neocloud](https://www.exent.ai/resources/what-is-a-neocloud). ## What useful life do CoreWeave, Meta, Amazon and others use? CoreWeave, Alphabet and Meta lengthened server lives between 2023 and 2025; Amazon lengthened in 2024 and then shortened part of its fleet in 2025. From each company’s latest annual report: | Company | Useful life for servers / GPUs | Last change | Disclosed effect | Filing | | --- | --- | --- | --- | --- | | CoreWeave | 6 years (technology equipment) | From 5 to 6 years, Jan 1, 2023 | $20M lower expenses in 2023 | 10-K FY2025 | | Alphabet | Generally 6 years | To 6 years in 2023 | $3.9B lower depreciation, $3.0B higher net income in 2023 | 10-K FY2025 / FY2023 | | Oracle | 6 years (servers and networking) | None disclosed | Depreciation $7.6B in FY26 vs $3.9B in FY25 (growth, not a change) | 10-K FY2026 | | Meta | 5.5 years (most servers and network assets) | Lengthened to 5.5 years, Jan 1, 2025 | $2.92B lower depreciation, $2.59B higher net income in 2025 | 10-K FY2025 | | Nebius | 5 years | From 4 to 5 years, Jan 1, 2026 | ~$167.6M lower depreciation expected in 2026 | 20-F FY2025 | | IREN | 5 years (HPC hardware) | Not stated | Not stated | 10-K FY2026 | | Amazon | 5 to 6 years | A subset shortened from 6 to 5 years, Jan 1, 2025 | $1.4B higher depreciation, $1.0B lower net income in 2025 | 10-K FY2025 | | Microsoft | 2 to 6 years (servers and network equipment) | None disclosed | Not applicable | 10-K FY2026 | Amazon gave its reason in the filing: the increased pace of technology development, particularly in AI and machine learning. Meta and Alphabet said their changes followed internal assessments of their servers. Both directions are allowed under US accounting rules; a useful life is a management estimate that auditors review, and a change applies going forward rather than restating past years. ## How much does useful life change profit? A worked example Every year of useful life you add lowers the yearly expense, and the effect is largest at short lives. Illustrative: a $1B GPU cluster depreciated straight line to zero (cost ÷ years): - - - - Now suppose the company books 6 years but the cluster stops earning after 3. Years 1–3: depreciation of $1,000M ÷ 6 = $166.7M a year, $500M in total. At the end of year 3 the book value is $1,000M − $500M = $500M for equipment that no longer earns its keep. That $500M eventually has to be written off, usually as an impairment. Profit was not lost; it was shown earlier than it was earned. If the cluster earned, say, $400M of revenue a year (illustrative), pre-tax profit after depreciation would look like $233M a year on a 6-year life and $67M on a 3-year life: the same business, a 3.5× difference in reported profit. ## What is Michael Burry’s GPU depreciation argument? Burry argues that hyperscalers extended server lives while buying chips on a two-to-three-year product cycle, which in his view overstates earnings. In a November 10, 2025 post he wrote that heavy chip spending “should not result in the extension of useful lives of compute equipment” (as quoted by [CNBC](https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html)), and CNBC reported his estimate of about $176B of understated depreciation across 2026–2028. That is one investor’s opinion, not an audit finding. The same concern appears in research form: Gartner said on October 14, 2026 that over $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030. ## What is the evidence that GPUs last longer? The best current evidence is rental pricing for older chips, which has held up in 2026. Nebius’s Q2 2026 shareholder letter reported more than 30% higher pricing on older-generation GPUs than in Q1, and Nebius itself lengthened its GPU life from four to five years on January 1, 2026, citing usage patterns and utilization commitments. A second argument: older GPUs move from training frontier models to inference and smaller jobs, so they keep earning at a lower rate instead of stopping. A third: many contracts now run about as long as the accounting life (IREN’s Microsoft contract has a five-year average term, and IREN depreciates HPC hardware over five years), which ties revenue to the depreciation period. None of this proves a six-year economic life; it shows that, so far, demand has outrun supply. ## Why does depreciation matter more for neoclouds than for Big Tech? Because GPUs are most of what a neocloud owns, and its lenders often hold them as collateral. At June 30, 2026, CoreWeave’s property and equipment, net, was $46.7B of $77.1B of total assets; Nebius’s depreciation and amortization was $259.7M in Q2 2026, 45% of revenue. A hyperscaler that shortens server lives takes a hit to earnings but keeps its other businesses; a neocloud whose GPUs lose value faster than its accounts assume can also see the collateral behind its loans shrink. Who carries the chip risk depends on the model: | Who owns the GPUs | Examples | What a shorter economic life would hit | | --- | --- | --- | | GPU cloud | CoreWeave, Nebius, IREN | Earnings, impairments, value of GPU-backed loans | | Hyperscaler | Microsoft, Alphabet, Amazon, Meta, Oracle | Earnings; offset by other businesses | | Landlord | Applied Digital, TeraWulf, Hut 8 | Mostly the tenant, unless the tenant fails | Debt and cash figures for each company are in [neocloud stocks compared](https://www.exent.ai/compare/neocloud-stocks-compared); the stock pages are [CRWV](https://www.exent.ai/stocks/crwv), [NBIS](https://www.exent.ai/stocks/nbis) and [ORCL](https://www.exent.ai/stocks/orcl). ## How can you check depreciation risk in a filing? - Find the useful-life table in the property and equipment note of the 10-K, and any “change in accounting estimate” with its dollar effect. - Compare depreciation to revenue each quarter: a rising share means new capacity is not yet earning; a falling share can mean longer lives. - Read impairments, and what they relate to: IREN booked $450.4M of non-cash impairments in its June 2026 quarter, primarily from decommissioning bitcoin-mining hardware as sites shifted to AI. This is not evidence that its AI GPUs became obsolete. - Match contract length to useful life: a 2-year contract on a 6-year asset leaves 4 years of re-leasing risk. - Look at older-GPU pricing in shareholder letters and calls; it is the earliest sign of the economic life changing. When this page’s worry is wrong: if older GPUs keep re-leasing at similar prices, a six-year life is reasonable and the bear case fades. When the reassurance is wrong: a sudden slowdown in AI demand or a much more efficient chip would cut older-GPU prices first, and the accounting would lag. Limitations: company disclosures are annual, pricing data is selective, and no public dataset tracks GPU rental prices across the industry. The wider valuation question is in [are we in a bubble](https://www.exent.ai/resources/are-we-in-a-bubble), and the supplier view in [should I buy Nvidia stock now](https://www.exent.ai/resources/should-i-buy-nvidia-stock-now). ## Add market context with Exent Market intel Market intel provides market context, cross-asset analysis, an events calendar and regular written notes. These can inform your review of an existing plan; they do not reliably predict crashes or guarantee better outcomes. 1. An events calendar with earnings dates, so you know when the next 10-K or 10-Q with useful-life notes arrives. 2. Regular written notes that read cross-asset ratios such as high-beta vs low-volatility stocks, the risk appetite that decides how forgiving the market is about accounting choices. 3. A reading of US market risk (Risk ON, Neutral or Risk OFF) built on four pillars, with an email when the reading changes. Exent is our product: [Exent Market intel](https://www.exent.ai/intel) ## Frequently asked questions ### What is the useful life of a GPU for depreciation? Five to six years at most large AI companies: CoreWeave and Alphabet use 6 years, Meta 5.5, Nebius and IREN 5 (10-K and 20-F filings, 2026). Microsoft uses a range of 2 to 6 years for servers and network equipment. ### How does CoreWeave depreciate its GPUs? Straight line over 6 years. CoreWeave changed the useful life of its data-center computing equipment from five to six years effective Jan 1, 2023, according to its FY2025 10-K. ### Why did Amazon shorten server useful life? Amazon cut the useful life of a subset of servers and networking equipment from six to five years from Jan 1, 2025, citing the increased pace of technology development in AI and machine learning. It added $1.4B of depreciation in 2025. ### How much did Meta’s useful-life change add to profit? Meta’s move to 5.5 years for most servers and network assets from Jan 1, 2025 reduced 2025 depreciation by $2.92B and raised net income by $2.59B, or $1.00 per diluted share. ### What did Michael Burry say about GPU depreciation? In November 2025 he argued that hyperscalers extended server lives while buying chips on a two-to-three-year cycle, overstating earnings; CNBC reported his estimate at about $176B of understated depreciation over 2026–2028. It is an opinion, not an accounting finding. ### Do older GPUs still make money? For now, yes. Nebius reported more than 30% higher pricing on older-generation GPUs in Q2 2026 than in Q1. Whether that lasts depends on demand staying above supply. ### What happens if GPUs must be written off faster? Depreciation rises or an impairment is booked, lowering reported profit; for a GPU cloud it can also reduce the collateral behind GPU-backed loans. In a $1B illustration, a 6-year life that proves to be 3 years leaves $500M of book value to write off. ### Is longer depreciation an accounting trick? Not by itself. Useful life is a management estimate reviewed by auditors and disclosed with its dollar effect; the debate is whether the estimate matches how long the chips keep earning. ## Sources 1. [Form 10-K for the year ended December 31, 2025](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm) — CoreWeave (SEC), 2026-03-02 2. [CoreWeave Reports Strong Second Quarter 2026 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm) — CoreWeave (SEC), 2026-08-11 3. [Annual report for 2025 (Form 20-F)](https://www.sec.gov/Archives/edgar/data/1513845/000110465926052948/nbis-20251231x20f.htm) — Nebius Group (SEC), 2026-04-30 4. [Letter to shareholders, Q2 2026 (Form 6-K, Exhibit 99.2)](https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-2.htm) — Nebius Group (SEC), 2026-08-12 5. [Nebius reports second quarter 2026 financial results (Form 6-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1_ex99-1.htm) — Nebius Group (SEC), 2026-08-12 6. [Form 10-K for the year ended December 31, 2025](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm) — Alphabet (SEC), 2026-02-05 7. [Form 10-K for the year ended December 31, 2023](https://www.sec.gov/Archives/edgar/data/1652044/000165204424000022/goog-20231231.htm) — Alphabet (SEC) 8. [Form 10-K for the year ended December 31, 2025](https://www.sec.gov/Archives/edgar/data/1326801/000162828026003942/meta-20251231.htm) — Meta Platforms (SEC), 2026-01-29 9. [Form 10-K for the year ended December 31, 2025](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm) — Amazon (SEC), 2026-02-06 10. [Form 10-K for the fiscal year ended June 30, 2026](https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm) — Microsoft (SEC), 2026-07-29 11. [Form 10-K for the fiscal year ended May 31, 2026](https://www.sec.gov/Archives/edgar/data/1341439/000119312526277521/orcl-20260531.htm) — Oracle (SEC), 2026-06-22 12. [Form 10-K for the fiscal year ended June 30, 2026](https://www.sec.gov/Archives/edgar/data/1878848/000187884826000052/iren-20260630.htm) — IREN (SEC), 2026-08-27 13. [IREN Reports FY26 Results (Form 8-K, Exhibit 99.1)](https://www.sec.gov/Archives/edgar/data/1878848/000187884826000051/irenreportsfy26results.htm) — IREN (SEC), 2026-08-27 14. [Big Short investor Michael Burry accuses AI hyperscalers of artificially boosting earnings](https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html) — CNBC, 2025-11-11 15. [Gartner says over $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030](https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030) — Gartner, 2026-10-14 16. [NVIDIA announces financial results for second quarter fiscal 2027](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm) — NVIDIA (SEC), 2026-08-26 --- Educational content, not investment advice. Past performance does not guarantee future results.