GPU Depreciation Lives and Their Effect on Reported AI-Sector Profits

Property and equipment is stated at cost less accumulated depreciation and depreciated using the straight-line method over the shorter of the estimated useful life of the asset or the lease term. The estimated useful lives of our property and equipment are generally as follows: software developed or acquired for internal use, three years ; servers and network equipment, two to six years ; buildings and improvements, five to 15 years ; leasehold improvements, three to 15 years ; and furniture and equipment, one to 10 years . Land is not depreciated.

Microsoft MSFT · FY Jun 30 6 yrs 4→6 (FY2023), held FY2023: +$3.7B operating income / +$3.0B net income / +$0.40 EPS. [ FY2023 10-K ] Extended, held

Property and Equipment Property and equipment is comprised of technical infrastructure, office space, corporate and other assets currently in service, and assets not yet in service. Technical infrastructure includes data center land, buildings and leasehold improvements, and servers and network equipment. Office space includes office land, buildings, and leasehold improvements. Assets not yet in service are those that are not ready for their intended use, including data center buildings and servers in the process of construction or assembly. Property and equipment are stated at cost less accumulated depreciation. Depreciation commences once assets are ready for their intended use and is recorded using the straight-line method over the estimated useful lives of the assets, which we regularly evaluate for factors such as technological obsolescence and our planned use and utilization. We depreciate data center and office buildings over periods of seven to 40 years. We depreciate servers and network equipment generally over a period of six years . We depreciate corporate and other assets over periods of two to 25 years. We depreciate leasehold improvements over the shorter of the remaining lease term or the estimated useful lives of the assets. Land is not depreciated.

In January 2021, we completed an assessment of the useful lives of our servers and network equipment and adjusted the estimated useful life of our servers from three years to four years and the estimated useful life of certain network equipment from three years to five years … The effect of this change in estimate was a reduction in depreciation expense of $721 million and $1.6 billion and an increase in net income of $561 million and $1.2 billion, or $0.84 and $1.81 per basic and $0.83 and $1.78 per diluted share, for the three and six months ended June 30, 2021, respectively.

Alphabet Servers (from 4y) & certain network equipment (from 5y) → 6 years Jan 2023 ~$3.4B lower 2023 depreciation est. 2022 10-K (Feb 3, 2023)

The estimated useful lives as of December 31, 2025, are as follows: Property and equipment Estimated useful life Buildings Lesser of forty years or the remaining life of the underlying building Servers and networking equipment Five to six years (1) Heavy equipment Ten to thirteen years (2) Other equipment Three to ten years

Effective January 1, 2024, we changed our estimate of the useful lives for our servers from five to six years

Amazon's FY2025 10-K, filed 2026-02-06: Effective January 1, 2025 we changed our estimate of the useful lives of a subset of our servers and networking equipment from six years to five years. The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.

Property and equipment, including finance leases, are depreciated and stated at cost less accumulated depreciation. Depreciation is computed using the straight-line method over the estimated useful lives of the assets or the remaining lease term, whichever is shorter. The estimated useful lives of property and equipment and amortization periods of finance lease right-of-use (ROU) assets as of December 31, 2025 are described below: Property and Equipment Useful Life/ Amortization period Servers and network assets Five to 5.5 years Buildings 25 to 30 years Equipment and other One to 25 years Finance lease right-of-use assets Five to 20 years Leasehold improvements Lesser of estimated useful life or remaining lease term

The estimated useful lives of property and equipment and amortization periods of finance lease right-of-use (ROU) assets as of December 31, 2024 are described below: Property and Equipment Useful Life/ Amortization period Servers and network assets Four to Five years Buildings 25 to 30 years Equipment and other One to 25 years Finance lease right-of-use assets Five to 20 years Leasehold improvements Lesser of estimated useful life or remaining lease term _______________________ (1) Effective January 2025, the useful lives of certain servers and network assets are extended to5.5 years.

In January 2025, we completed an assessment of the useful lives of property and equipment, which resulted in an increase in the estimated useful lives of most servers and network assets to 5.5 years, effective January 1, 2025.

Comprised primarily of servers and networking equipment with estimated useful life of six years.

The estimated useful lives of the Company's property and equipment are as follows: Technology equipment 6 years Software 3 - 6 years Data center equipment and leasehold improvements Shorter of remaining lease term or estimated useful life of up to 12 years Furniture, fixtures, and other assets 3 - 5 years

Effective January 1, 2023, the Company changed its estimate of the useful life for its computing equipment utilized in data centers from five to six years , reflecting continuous advancements in hardware performance, software optimization, and data center design improvements. The effects of this change in estimate for the year ended December 31, 2023 on computing equipment that was included in property and equipment, net on the consolidated balance sheets as of December 31, 2023 was a reduction in total expenses of $ 20 million. The per share impact of the change in estimate was a $ 0.10 increase for the year ended December 31, 2023.

Server and network equipment 4.0 years Infrastructure systems and equipment 3.0 - 10.0 years Office furniture and equipment 3.0 years Buildings 20.0 years Other property and equipment 2.0 - 10.0 years

Our main assets, GPUs, are currently depreciated over four years. In January 2026, the Company completed an assessment of the useful lives of servers and network equipment based on updated information and usage patterns obtained subsequent to year-end and concluded that the estimated useful lives of such assets should be extended from four to five years. Management is considering the application of this change in accounting estimate prospectively beginning January 1, 2026. Based on the servers and network equipment placed in service as of December 31, 2025, we expect this change in accounting estimate will reduce the depreciation expenses for fiscal year 2026 by approximately $167.6 million. The actual impact in future periods will depend on the timing and level of future capital expenditures and asset retirements. We will continue to evaluate the appropriateness of our estimated useful lives and may revise them in future periods if warranted. Any such revisions will be accounted for prospectively.

The estimated useful lives of the Group’s property, plant and equipment are generally as follows: Useful life (in years) Buildings 20 Other PPE 3 - 10 Mining hardware 4 HPC hardware 5 Leasehold improvements Lesser of the life of the lease or the useful life of the improvement

It cannot tell you the useful life of an H200, a B200, or an MI355X. No filing in this set isolates GPUs from CPUs, networking, or, in Microsoft’s case, software. It cannot tell you Meta’s pre-2025 life; the 10-K states the destination, 5.5 years, and not the prior figure. It cannot tell you tax depreciation, lease accounting, or residual value in a particular contract. CoreWeave’s $0.10 per-share effect is a 2023 figure. Amazon’s $1.4 billion is the effect of a subset, for one year, on assets already owned plus those acquired in 2025, not a run-rate for 2027 purchases. Oracle is absent. A download of the Oracle annual report path we tried returned a 404, so Oracle’s life is not in this comparison. Lives change. These are the notes as filed for the periods named, read on October 2, 2026.

Are Amazon and Meta actually comparable? No, and the gaps run deeper than the headline invites. Amazon's change covers "a subset of our servers and networking equipment." Meta's covers "most servers and network assets." Neither sizes the subset or the share "most" represents. The bases differ too: Meta computed its effect on assets "placed in service as of December 31, 2024," while Amazon's includes equipment acquired during 2025. And "servers" means the whole fleet, not just accelerators, so this is no clean read on GPU life. What survives is narrow and worth having: two companies buying from the same vendors in the same quarter told their auditors different things about how long the gear lasts, and both were signed off.

It is also important to separate server-level changes companies actually made from a GPU-specific story that is now being told about them. When companies such as Amazon and Alphabet extended useful lives from four to five or six years in 2022-2024, those changes were typically framed and modeled at the level of servers and networking equipment as a class , in data centers that were in many cases heavily CPU-dominated. The policies were likely not a specific decision to “extend the life of GPUs”; they were broad server-fleet assumptions that later were applied to a growing mix of CPU and GPU configurations. Thus, in my view, the debate should be less about “you extended GPU lives,” and more “you did not revisit GPU-based server lives once the fleet mix shifted toward GPUs.”

Change in Accounting Estimate In July 2022, we completed an assessment of the useful lives of our server and network equipment. Due to investments in software that increased efficiencies in how we operate our server and network equipment, as well as advances in technology, we determined we should increase the estimated useful lives of both server and network equipment from four years to six years. This change in accounting estimate was effective beginning fiscal year 2023. Based on the carrying amount of server and network equipment included in property and equipment, net as of June 30, 2022, the effect of this change in estimate for fiscal year 2023 was an increase in operating income of $3.7 billion and net income of $3.0 billion, or $0.40 per both basic and diluted share.

Gross margin 146,052 135,620 8% Operating income 88,523 83,383 6% Net income 72,361 72,738 (1)% Diluted earnings per share 9.68 9.65 0%

Change in Useful Lives of Our Server and Network Equipment In January 2023, we completed an assessment of the useful lives of our servers and network equipment and adjusted the estimated useful life of our servers from four years to six years and the estimated useful life of certain network equipment from five years to six years. This change in accounting estimate was effective beginning in fiscal year 2023, and the effect was a reduction in depreciation expense of $983 million and $3.9 billion and an increase in net income of $765 million and $3.0 billion, or $0.06 and $0.24 per basic and $0.06 and $0.24 per diluted share for the three and twelve months ended December 31, 2023, respectively.

Based on the servers and network assets placed in service as of December 31, 2024, the financial impact of this change in estimate included a reduction in depreciation expense of $ 2.92 billion and an increase in net income of $ 2.59 billion, or $ 1.00 per diluted share, for the year ended December 31, 2025.

We review the useful lives of equipment on an ongoing basis. Effective January 1, 2025 we changed our estimate of the useful lives of a subset of our servers and networking equipment from six years to five years . The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning. The effect of this change in estimate for the year ended December 31, 2025, based on servers and networking equipment that were included in “Property and equipment, net” as of December 31, 2024 and those acquired during the year ended December 31, 2025, was an increase in depreciation and amortization expense of $ 1.4 billion and a reduction in net income of $ 1.0 billion, or $ 0.10 per basic share and $ 0.10 per diluted share, which primarily impacted our AWS segment.

AWS Net sales $ 90,757 $ 107,556 $ 128,725 Operating expenses 66,126 67,722 83,119 Operating income $ 24,631 $ 39,834 $ 45,606 Consolidated Net sales $ 574,785 $ 637,959 $ 716,924 Operating expenses 537,933 569,366 636,949 Operating income 36,852 68,593 79,975 Total non-operating income 705 21 17,336 Provision for income taxes ( 7,120 ) ( 9,265 ) ( 19,087 ) Equity-method investment activity, net of tax ( 12 ) ( 101 ) ( 554 ) Net income $ 30,425 $ 59,248 $ 77,670

Amazon AMZN · FY Dec 31 6 yrs (subset 5) 4→5 (Jan 2022) · 5→6 (Jan 2024) · 6→5 subset (Jan 2025) Jan-2022: +$2.8B benefit to net loss ($0.28 EPS). 5→6 (FY2024): +$2.5B net income ($0.23 EPS). 6→5 subset: -$0.7B 2025 op income + ~$920M Q4’24 accelerated depreciation. [ FY2024 10-K ] Extended, then reversed

Oracle ORCL · FY May 31 6 yrs 4→5 (FY2023) · 5→6 (FY2025), held FY2025 (5→6): -$733M opex / +$573M net income / +$0.20-0.21 EPS. (FY2023 4→5: -$434M opex.) [ FY2025 10-K ] Extended, held

In the year each company extended its server lives, the change lifted net income by a strikingly consistent ~4 to 4.6%: Microsoft +$3.0B (4.1%), Alphabet +$3.0B (4.1%), Amazon +$2.5B (4.2%), Meta +$2.59B (4.3%), Oracle +$0.57B (4.6%). Real money, and a real single-year boost to reported earnings per share. But notice what that is and is not. It is a one-time, disclosed tailwind in a single year, around 4 to 5% of profit. It is not a claim that earnings are fabricated, and it is an order of magnitude smaller than the cumulative figures circulating in the bubble debate.

During fiscal years 2023, 2022, and 2021, depreciation expense was $ 11.0 billion, $ 12.6 billion, and $ 9.3 billion, respectively. Depreciation expense declined in fiscal year 2023 due to the change in estimated useful lives of our server and network equipment.

We completed our most recent servers and networking equipment useful life study in Q4 2024, and are changing the useful lives of a subset of our servers and networking equipment, effective January 1, 2025, from six years to five years. For those assets included in “Property and equipment, net” as of December 31, 2024, whose useful life will change from six years to five years, we anticipate a decrease in 2025 operating income of approximately $0.7 billion.

In Q4 2024, we completed a useful life study for certain types of heavy equipment and are increasing the useful life from ten years to thirteen years for such equipment effective January 1, 2025. Based on heavy equipment included in “Property and equipment, net” as of December 31, 2024, we estimate an increase in 2025 operating income of approximately $0.9 billion, which will be recorded primarily in “Fulfillment” and impact our North America and International segments.

Taken together, however, the two adjustments essentially cancel out each other’s effect on operating income.

Effective January 1, 2025 we changed our estimate of the useful lives of a subset of our servers and networking equipment from six years to five years. The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning. The effect of this change in estimate for Q3 2025, based on servers and networking equipment that were included in “Property and equipment, net” as of June 30, 2025 and those acquired during the three months ended September 30, 2025, was an increase in depreciation and amortization expense of $392 million and a reduction in net income of $298 million, or $0.03 per basic share and $0.03 per diluted share , which primarily impacted our AWS segment. The effect of this change in estimate for the nine months ended September 30, 2025, based on servers and networking equipment that were included in “Property and equipment, net” as of December 31, 2024 and those acquired during the nine months ended September 30, 2025, was an increase in depreciation and amortization expense of $889 million and a reduction in net income of $677 million, or $0.06 per basic share and $0.06 per diluted share , which primarily impacted our AWS segment.

The major hyperscalers normalized on a six-year server life around 2023, and most still use five to six years for the bulk of their data-center hardware in 2026. The shift was deliberate and lucrative. When Alphabet extended its server life to six years for fiscal 2023, it told investors the change would reduce depreciation by roughly $3.4 billion for the year, money that flowed straight to operating income. Microsoft and the others moved in the same direction over the same window, and a six-year assumption became the de facto industry standard for cloud infrastructure. But 2025 and 2026 introduced a split. Amazon went the other way for its AI fleet. Effective January 1, 2025, it shortened the useful life of a subset of its servers and networking equipment from six years to five. That single change increased depreciation and amortization expense by about $217 million in Q1 2025 alone and cut net income by roughly $162 million; by Q2 2025 the quarterly hit had grown to about $280 million of added D&A. Amazon had also taken roughly $920 million in accelerated depreciation in Q4 2024 as it retired some equipment early, and guided to about a $0.6 billion reduction in 2025 operating income from the move. Meta, meanwhile, pushed in the opposite direction, extending the useful life of certain servers and network assets to about 5.5 years and booking an estimated $2.9 billion reduction in depreciation expense, equal to nearly 4% of its estimated annual pre-tax profit. So within the same 12 months, one giant shortened its assumption and took a charge while another lengthened its assumption and harvested a gain. That divergence is exactly what makes the depreciation question impossible to wave away as settled accounting.

Company Line item as named in the filing Stated life Change effective 2023-2025 $ effect Amazon Servers and networking equipment Five to six years 2024: 5 to 6. 2025: 6 to 5 (subset) Yes, both Meta Servers and network assets Five to 5.5 years 2025: up to 5.5 (most) Yes Alphabet Servers and network equipment Generally six years Jan 2023: up to six Yes, in its FY2023 10-K Microsoft Servers and network equipment Two to six years None n/a Oracle Computer, network, machinery and equipment Six years (footnote) None n/a CoreWeave Technology equipment Six years Jan 2023: 5 to 6 Yes Equinix Core systems 3 to 40 years "certain revisions" in 2024 and 2025 No IBM Information technology equipment 1.5 to 6 years None n/a Tesla AI infrastructure (owned data centers) 5 to 30 years None n/a Dell Computer and other equipment 3 to 5 years None n/a Applied Digital Networking equipment, electrical equipment, software 3 to 5 years None n/a Digital Realty Machinery and equipment 7 to 15 years None n/a Nvidia No server line (general P&E) Two to seven years None n/a Apple None Not disclosed None n/a

Straight-line depreciation with zero salvage is cost divided by life. On $1,000,000 that is $500,000 a year at two years, $200,000 at five, $181,818.18 at 5.5, and $166,666.67 at six. The two-year charge is exactly three times the six-year charge, because 1/2 divided by 1/6 is 3. Moving from six years to five years adds $33,333.33 a year. Moving from six years to 5.5 years adds $15,151.52 a year. Those four rates are the decision content of the filings.

Depreciation expense on property and equipment was $ 6.00 billion and $ 4.28 billion for the three months ended June 30, 2026 and 2025, respectively, and $ 11.67 billion and $ 8.12 billion for the six months ended June 30, 2026 and 2025, respectively. Within property and equipment, our servers and network assets depreciation expense was $ 4.62 billion and $ 3.12 billion for the three months ended June 30, 2026 and 2025, respectively, and $ 9.01 billion and $ 5.74 billion for the six months ended June 30, 2026 and 2025, respectively.

Income from operations for 2025 was $83.28 billion, an increase of $13.90 billion, or 20%, compared to 2024, driven by an increase in advertising revenue, partially offset by an increase in costs and expenses.

Property, plant and equipment, net consisted of the following: (Dollars in millions) Estimated Useful Life August 31, 2026 May 31, 2026 Computer, network, machinery and equipment 1-6 years $ 77,755 $ 59,634 Buildings and improvements 1-40 years 25,539 21,263 Furniture, fixtures and other 5-15 years 427 452 Land - 1,329 1,329 Construction in progress - 48,546 39,973 Total property, plant and equipment 1-40 years 153,596 122,651 Accumulated depreciation (25,751) (22,694) Total property, plant and equipment, net $ 127,845 $ 99,957

Total operating expenses 46,751 39,721 37,608 Operating income 20,606 17,678 15,353 Interest expense (4,599) (3,578) (3,514) Non-operating income (expenses), net 3,547 60 (98) Income before income taxes 19,554 14,160 11,741 Provision for income taxes 2,467 1,717 1,274 Net income $ 17,087 $ 12,443 $ 10,467 Preferred stock dividends 103 - - Net income available to common shareholders $ 16,984 $ 12,443 $ 10,467 Earnings per share attributable to common shareholders: Basic $ 5.94 $ 4.46 $ 3.82 Diluted $ 5.83 $ 4.34 $ 3.71

Total operating expenses 12,617 10,649 Operating income 6,728 4,277 Interest expense ( 1,428 ) ( 923 ) Non-operating income, net 307 73 Income before income taxes 5,607 3,427 Provision for income taxes 847 500 Net income $ 4,760 $ 2,927 Preferred stock dividends 81 - Net income available to common shareholders $ 4,679 $ 2,927 Earnings per share attributable to common shareholders: Basic $ 1.58 $ 1.04 Diluted $ 1.56 $ 1.01

Property and equipment, net, co nsisted of the following (in millions): As of December 31, 2025 As of June 30, 2026 Technical infrastructure (1) $ 203,679 $ 247,177 Office space 48,348 50,635 Corporate and other assets 14,463 6,498 Property and equipment, in service 266,490 304,310 Less: accumulated depreciation ( 98,485 ) ( 105,912 ) Add: assets not yet in service 78,592 122,814 Property and equipment, net $ 246,597 $ 321,212 (1) As of December 31, 2025 and June 30, 2026, approximately 60 % of technical infrastructure assets were comprised of servers and network equipment. The remaining balance was comprised of data center land and buildings and related assets.

Operating income $ 112,390 $ 129,039 $ 16,649 15 % Operating margin 32 % 32 % 0 % Other income (expense), net $ 7,425 $ 29,787 $ 22,362 301 % Net income $ 100,118 $ 132,170 $ 32,052 32 % Diluted net income per share (1) $ 8.04 $ 10.81 $ 2.77 34 %

Property and equipment, at cost, consisted of the following (in millions): December 31, 2024 2025 Gross property and equipment (1): Land and buildings $ 123,039 $ 155,121 Servers and networking equipment 113,156 172,492 Heavy equipment 52,228 65,545 Other equipment 53,509 63,376 Other assets 5,487 5,819 Construction in progress 46,636 71,745 Gross property and equipment 394,055 534,098 Total accumulated depreciation and amortization (1) 141,390 177,073 Total property and equipment, net $ 252,665 $ 357,025

Operating income by segment is as follows (in millions): Year Ended December 31, 2024 2025 Operating Income North America $ 24,967 $ 29,619 International 3,792 4,750 AWS 39,834 45,606 Consolidated $ 68,593 $ 79,975

Property and equipment, net, consisted of the following (in millions): December 31, 2025 December 31, 2024 Technology equipment $ 20,903 $ 9,146 Software 802 140 Data center equipment and leasehold improvements 2,842 384 Furniture, fixtures, and other assets 18 9 Construction in progress 9,376 3,201 Total property and equipment 33,941 12,880 Less: accumulated depreciation and amortization ( 3,384 ) ( 965 ) Total property and equipment, net $ 30,557 $ 11,915

Depreciation and amortization on property and equipment was $ 2.4 billion, $ 861 million, and $ 101 million for the years ended December 31, 2025, 2024, and 2023, respectively.

Net loss and comprehensive loss $ ( 1,167 ) $ ( 863 ) $ ( 594 ) Net loss attributable to common stockholders, basic $ ( 1,196 ) $ ( 937 ) $ ( 594 ) Net loss attributable to common stockholders, diluted $ ( 1,223 ) $ ( 937 ) $ ( 594 )

Net income was $60.46 billion, with diluted earnings per share (EPS) of $23.49 for the year ended December 31, 2025.

Technology and infrastructure expense for the three months ended June 30, 2026 increased by $837 million, or 125%, compared to the three months ended June 30, 2025. This increase was primarily attributable to an increase in depreciation and amortization of approximately $752 million, from $537 million for the three months ended June 30, 2025, to approximately $1.3 billion for the three months ended June 30, 2026. These increases in depreciation and amortization were related to investments in our platform and servers, switches, and other networking equipment fixed assets within our infrastructure that were placed in service. Technology and infrastructure expense for the six months ended June 30, 2026 increased by $1.5 billion, or 126%, compared to the six months ended June 30, 2025. This increase was primarily attributable to an increase in depreciation and amortization of approximately $1.4 billion, from $967 million for the six months ended June 30, 2025, to approximately $2.4 billion for the six months ended June 30, 2026. These increases in depreciation and amortization were related to investments in our platform and servers, switches, and other networking equipment fixed assets within our infrastructure that were placed in service.

The following table presents our historical consolidated results of continuing operations for the periods indicated: Year ended December 31, 2023 2024 2025 (in millions of U.S. dollars) Revenues 9.8 91.5 529.8 Operating costs and expenses: Cost of revenues 19.6 43.7 166.2 Product development 87.1 114.8 177.3 Sales, general and administrative 159.5 255.5 380.1 Depreciation and amortization 29.3 77.1 417.9 Total operating costs and expenses 295.5 491.1 1,141.5 Loss from operations (285.7) (399.6) (611.7)

Microsoft’s property note, in millions, shows servers, network equipment, and software at a gross carrying amount of $215,874 million on June 30, 2026, against $132,836 million a year earlier. The gross balance rose by $83,038 million. That line includes software, whose disclosed life is three years, and servers, whose disclosed life is two to six years. Dividing $215,874 million by any single life produces a fake depreciation number. We do not do it.

Scenario Useful life assumption 2026-2028 cumulative impact (≈) Bear (Burry’s harsher version) 3 years for AI hardware $280-320 billion Base case (this analysis) 4 years matching pre-extension policy ~$200 billion Burry published number Mixed 3-4 years $176 billion Amazon-aligned 5 years across the board $80-100 billion Status quo 5-6 years (current disclosed) $0 - already in earnings

The structure: Affected base entering 2026: approximately $467 billion in net property, plant, and equipment across the four extending hyperscalers (Microsoft, Google, Meta, Oracle), in the categories where useful life was extended - primarily servers, networking, and some structured cabling. Projected 2026-2028 capex: approximately $1.34 trillion in additional spending across the same categories, based on each company’s most recent guidance and analyst consensus. Affected capex share: roughly 60%. This excludes buildings, land, leasehold improvements, and general data center fitouts (which retain their original lives, typically 15-40 years), and includes server hardware, networking equipment, and the GPU and CPU components subject to the extension. Annual delta: approximately 6.82% of the affected base. This is the difference between the disclosed schedules (5-6 years) and the 4-year counterfactual matching pre-2022 policy. Three-year cumulative: roughly $228 billion. Rounded conservatively to ~$200 billion to absorb estimation error.

Burry’s $176 billion figure comes from re-depreciating the hyperscalers’ projected GPU capex over a three-year life instead of the five-to-six years they use, then summing the difference for 2026 through 2028. The arithmetic is not exotic. Take a $10 billion tranche of GPUs. On a six-year schedule it generates about $1.67 billion of annual depreciation; on a three-year schedule it generates about $3.33 billion. The roughly $1.8 billion gap between those two numbers is profit that exists under the longer assumption and vanishes under the shorter one. Scale that to the hundreds of billions in AI capex the industry is committing, and the analysis lands at roughly $50-60 billion of understated depreciation per year, or about $176 billion cumulatively across the three-year window.

SA estimate - the forward D&A run-rate, stated assumptions: take the $433.9B trailing-4Q cohort alone, assume ~85% is depreciable equipment and shorter-life assets on a blended ~6-year straight-line schedule and the rest is 25-40-year shell, and that single year of spend adds roughly $65-75B of annual depreciation once fully in service - on top of the ~$149B already being recognized, and before the larger 2026 cohort implied by ~$700B of guidance. If server lives are economically closer to four years than six (the Amazon-direction scenario), the same cohort implies $95-110B. This is an illustration of operating leverage in reverse, not a forecast of reported figures: actual recognition depends on in-service timing, mix, and future policy changes.

During the years ended December 31, 2024 and 2025, we spent $52.5 billion and $91.4 billion on capital expenditures, respectively. In 2026, we expect to significantly increase , relative to 2025, our i nvestment in our technical infrastructure, including servers and network equipment, and data centers.

Depreciation of our property and equipment commences when such assets are ready for their intended use. For the years ended December 31, 2024 and 2025, our depreciation on property and equipment was $15.3 billion and $21.1 billion, respectively.

Cash capital expenditures were $77.7 billion, and $128.3 billion in 2024 and 2025, which primarily reflect investments in technology infrastructure (the majority of which is to support AWS business growth) and in additional capacity to support our fulfillment network, both of which we expect to increase in 2026.

Depreciation and amortization expense on property and equipment was $ 30.2 billion, $ 32.1 billion, and $ 41.9 billion which includes amortization of property and equipment acquired under finance leases of $ 5.9 billion, $ 3.9 billion, and $ 3.3 billion for 2023, 2024, and 2025.

Additions to property and equipment ( 115,948 ) ( 64,551 ) ( 44,477 )

NOTE 6 - PROPERTY AND EQUIPMENT The components of property and equipment were as follows: (In millions) June 30, 2026 2025 Land $ 10,546 $ 9,338 Buildings and improvements 182,749 137,921 Leasehold improvements 16,348 12,117 Servers, network equipment, and software 215,874 132,836 Furniture and equipment 6,250 6,407 Total, at cost 431,767 298,619 Accumulated depreciation ( 118,691 ) ( 93,653 ) Total, net $ 313,076 $ 204,966 During fiscal years 2026, 2025, and 2024, depreciation expense was $ 34.3 billion, $ 22.0 billion, and $ 15.2 billion, respectively.

Cash used for capital expenditures increased from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026 primarily due to the expansion of our data centers. We expect this upward trend to continue during fiscal 2027 and in the following fiscal years as we increase our existing data center capacity and establish data centers in new geographic locations in order to meet current and expected customer demand for our cloud offerings.

Property, plant and equipment, net consisted of the following: Estimated May 31, (Dollars in millions) Useful Life 2026 2025 Computer, network, machinery and equipment 1-6 years $ 59,634 $ 30,345 Buildings and improvements 1-40 years 21,263 10,881 Furniture, fixtures and other 5-15 years 452 466 Land - 1,329 1,352 Construction in progress - 39,973 16,510 Total property, plant and equipment 1-40 years 122,651 59,554 Accumulated depreciation (22,694) (16,032) Total property, plant and equipment, net $ 99,957 $ 43,522 Comprised primarily of servers and networking equipment with estimated useful life of six years. Comprised primarily of servers, networking equipment and leasehold improvements to be deployed at our data centers. Depreciation expense on property, plant and equipment in fiscal 2026, 2025 and 2024 were $7.6 billion, $3.9 billion and $3.1 billion, respectively.

Cash used in investing activities during 2025 mostly consisted of $69.69 billion of purchases of property and equipment as we continued to invest in servers, data centers, and network infrastructure, and $18.33 billion of purchases of non-marketable equity investments, and $10.05 billion of net purchases of marketable securities.

Depreciation of our property and equipment commences when such assets are ready for their intended use. For the six months ended June 30, 2025 and 2026, our depreciation on property and equipment was $9.5 billion and $13.6 billion, respectively.

Depreciation and amortization expense on property and equipment, including corporate property and equipment, are allocated to all segments based on usage. Total depreciation and amortization expense, by segment, is as follows (in millions): Three Months Ended June 30, Six Months Ended June 30, 2025 2026 2025 2026 North America $ 3,742 $ 4,500 $ 7,272 $ 8,780 International 1,180 1,293 2,316 2,570 AWS 4,844 8,076 9,234 15,353 Consolidated $ 9,766 $ 13,869 $ 18,822 $ 26,703

Depreciation expenses on property, plant and equipment during the three months ended August 31, 2026 and 2025 were $3.2 billion and $1.4 billion, respectively.

Property, plant and equipment, net consisted of the following: (Dollars in millions) Estimated Useful Life August 31, 2026 May 31, 2026 Computer, network, machinery and equipment 1 - 6 years (1) $ 77,755 $ 59,634 Buildings and improvements 1 - 40 years 25,539 21,263 Furniture, fixtures and other 5 - 15 years 427 452 Land - 1,329 1,329 Construction in progress (2) - 48,546 39,973 Total property, plant and equipment 1 - 40 years 153,596 122,651 Accumulated depreciation ( 25,751 ) ( 22,694 ) Total property, plant and equipment, net $ 127,845 $ 99,957

Property and equipment, at cost, consisted of the following (in millions): December 31, 2024 2025 Gross property and equipment (1): Land and buildings $ 123,039 $ 155,121 Servers and networking equipment 113,156 172,492 Heavy equipment 52,228 65,545 Other equipment 53,509 63,376 Other assets 5,487 5,819 Construction in progress 46,636 71,745 Gross property and equipment 394,055 534,098 Total accumulated depreciation and amortization (1) 141,390 177,073 Total property and equipment, net $ 252,665 $ 357,025 __________________ (1)Includes the original cost and accumulated depreciation of fully-depreciated assets.

Property and equipment, net consists of the following (in millions): December 31, 2025 2024 Land $ 3,687 $ 2,561 Servers and network assets 98,040 68,397 Buildings 55,568 47,076 Leasehold improvements 8,346 7,293 Equipment and other 9,377 7,150 Finance lease right-of-use assets 8,187 5,384 Construction in progress 50,521 26,802 Property and equipment, gross 233,726 164,663 Less: Accumulated depreciation ( 57,326 ) ( 43,317 ) Property and equipment, net $ 176,400 $ 121,346

The four US hyperscalers - Microsoft, Alphabet, Amazon, and Meta - purchased $433.9B of property and equipment in the four quarters through March 2026 (SA estimate: sum of reported cash-flow figures), against roughly $149B of reported depreciation over the same span. Quarterly combined capex reached $129.8B in 1Q26, up 80% year over year, and company guidance for calendar 2026 sums to roughly $700B of intended spend. The capex-to-depreciation gap is structural: depreciation recognizes spend over 5-6 year server schedules and 25-40 year building schedules, so today's income statements carry only a fraction of today's build-out.

Depreciation is capex on a delay. A dollar of server spend hits the cash-flow statement today and the income statement in roughly equal slices over the next five to six years; a dollar of data-center shell spreads over 25-40 years. When capex triples in three years, reported depreciation mechanically lags - and reported margins look better than the underlying economics until the wave catches up.

Construction in progress (CIP) signals future depreciation pressure. Hyperscalers have more than $100 billion in assets classified as CIP, including those related to new data center expansion. According to GAAP, these balances are not depreciated until the assets are placed in service, meaning today’s depreciation expense reflects past investment cycles and may not fully reflect the current wave of AI-driven infrastructure spending. Given the materiality of AI-related CAPEX spending and a relatively long period required to place some of these assets in service, the CIP balances are increasingly significant for forecasting depreciation trends.

The change, disclosed in the social media giant’s earnings materials on Jan 29, extends the so-called useful life period of certain servers and networking assets to five and a half years, from the four to five years it previously used. While that may sound trivial, its impact on earnings will be sizeable given the heavy spending on these relatively short-lived assets. By Meta’s reckoning, the shift is expected to reduce the company’s depreciation expense by US$2.9 billion in 2025, which would, on its own, amount to almost 4 per cent of the estimated pre-tax profits for the year. With Meta planning to spend as much as 75 per cent more this year on capital expenditures to build out its AI capabilities, the effect will be even bigger in 2026.

Rubin vs Blackwell vs Hopper is a comparison of three NVIDIA datacenter generations at three different points in their life. Hopper (H100, H200) is mature and widely rentable, Blackwell (B200, B300, GB200, GB300) is the current volume generation, and Rubin is the next one, which NVIDIA says is in full production and ramping as of its August 26, 2026 earnings release but is not something Aquanode rents.

Vera Rubin NVL72 production is ramping up with racks running at partners CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius.

Independent comparison site GPU Insights, which tracks rack-scale AI fabrics, notes that as of August 2026 Nvidia’s Rubin/NVL72 platform was already in production and shipping to hyperscalers since July 2026, while AMD’s Helios remained at reference-design status, with initial shipments expected in the second half of 2026. On paper, AMD’s per-GPU memory advantage (432GB versus Nvidia’s reported roughly 288GB) is real and significant. But paper specs and shipped silicon are different things, and Nvidia’s roughly six-month production head start is the single biggest variable that could blunt AMD’s memory-capacity pitch before it reaches enough customers to matter.

NVIDIA confirmed at GTC 2025 and reaffirmed across CES 2026, GTC 2026, and Computex 2026 that it has moved to an annual architecture cadence: Blackwell now, Rubin H2 2026, Rubin Ultra H2 2027, Feynman 2028, Rosa Feynman 2029-2030.

NVIDIA has announced plans to upgrade its AI accelerators every year, revealing that it is developing Blackwell Ultra chips for 2025 and a next-generation platform called Rubin for 2026.

AMD is extending its annual cadence of CPU, GPU, networking and rack-scale innovation through 2030. The company shared new details on its roadmaps, including:

December 16, 2024 Announcement This Release Note announces General Availability of Trillium AKA v6e. Trillium is the 6th generation and latest Cloud TPU.

March 31, 2026 Feature Generally available: TPU7x is generally available (GA). TPU7x is the first release within the Ironwood family, Google Cloud's seventh generation TPU. TPU7x supports large-scale AI training and inference, providing performance and cost-effectiveness for demanding workloads such as large language (LLMs), mixture of experts (MoEs), and diffusion models.

Performance gains continue to compound and shatter records for AI accelerator capabilities. For the Visual Language Model (VLM) test, the best per-accelerator server scenario result for this benchmark round improved by 2.99x from v6.0 just six months ago. For the Deepseek R1 test, the best per-accelerator server scenario result was 5.7x better compared to v5.1 one year ago. This performance gain translates into more capabilities for end customers, potentially serving more users or enabling more sophisticated inference use cases.

Where B200 leads: Blackwell dominates all tasks it submitted for. On GPT-OSS 120B, the B200 is roughly 2.5x the per-GPU throughput of the H200. The key driver is the combination of FP4 precision support and 8.0 TB/s HBM3e memory bandwidth versus the H200's 4.8 TB/s. For memory-bandwidth-bound workloads like large LLMs, doubling bandwidth nearly doubles throughput when the model fits in VRAM. The NVL72 system interconnect (NVLink 5 at 1.8 TB/s bidirectional per GPU) also lets the system run the 120B model across GPUs without the bandwidth penalty you see with InfiniBand-connected nodes.

GB300 NVL72 delivers up to 10x lower cost per million tokens than H200 NVL8 (Figure 6, below).

Compared with NVIDIA GB200 NVL72, Vera Rubin NVL72 delivers up to 10x more tokens per megawatt and one-tenth the cost per million tokens, providing more intelligence within the same power footprint.

Figure 1. Pareto frontiers illustrating the 10x generational uplift in agentic inference performance of the Vera Rubin platform (internal 2T MoE workload)

Although Microsoft plans to build AI infrastructure aggressively, CEO Satya Nadella said this week that his company is trying to space out its AI chip purchases and not overinvest in a single generation of processors. He added that the biggest competitor for any new Nvidia AI chip is its predecessor. "One of the biggest learnings we had even with Nvidia is that their pace increased in terms of their migrations," Nadella said. "That was a big factor. I didn't want to go get stuck with four or five years of depreciation on one generation."

There are a number of ways AI chips could depreciate before six years. They could wear out and break, or they could become obsolete as newer GPUs are released. They could still be useful for running certain workloads, but with much worse economics. Nvidia CEO Jensen Huang has implied as much. When Nvidia announced a new Blackwell chip earlier this year, he joked that the value of its predecessor, the Hopper, would deteriorate. "When Blackwell starts shipping in volume, you couldn't give Hoppers away," Huang said in March at Nvidia's AI conference. "There are circumstances where Hopper is fine," he continued. "Not many." Nvidia now releases new AI chips on an annual basis, versus the two-year cadence it had before. Advanced Micro Devices , its closest GPU competitor, followed suit.

For example, industry supply capacity for AI accelerators, including graphics processing units, as well as memory devices, is competitive, and we at times have to accept less favorable terms with suppliers to minimize supply constraints. While this permits us to secure cloud infrastructure capacity, it has increased excess and obsolescence risk of such hardware products and could adversely impact our profitability and cash flows. We expect these factors will continue to impact us in the future.

H100 1-year GPU rental contract pricing has shot up almost 40% from a low of $1.70/hr/GPU in October 2025 to $2.35/hr/GPU by March 2026.

On-Demand GPU rental capacity is sold out across all GPU types - those that have locked up on-demand instances are not willing to relinquish this capacity back into the pool despite recent price hikes.

Customers are fighting to pay $14/hr/GPU for p6-b200 spot instances in AWS, some Neocloud Giants no longer sell single nodes, H100s are getting renewed at the exact same rate they were signed at 2-3 years ago and some H100 contracts are being renewed for 4 years though 2028.

Before late 2025, the prevailing expectation across the ecosystem was that Hopper (i.e. H100 and H200) rental prices would drop considerably as Blackwell deployments ramped given the latter’s much lower cost of compute. Instead, the opposite happened in late 2025: demand for H100s was holding firm, and in many cases, strengthening .

Our index is constructed from direct survey data across a pool of 100+ market participants including Neocloud providers, buyers and sellers of compute that is captured every month to determine a representative range (25 th to 75 th percentile) for GPU rental contracts. We validate these pricing levels with transaction data as well as by arranging a few transactions ourselves, connecting buyers and sellers of compute within our network.

The H100 (released March 2023) has experienced the most dramatic secondary market pricing volatility of any data center GPU. Used and refurbished units traded as high as $50,000 per GPU during mid-2024 scarcity, then dropped sharply as supply increased and buyer power returned. Retail prices held steady at $25,000-$40,000 through this period, masking the secondary market turbulence underneath. Cloud rental pricing followed a similarly dramatic arc-dropping from roughly $7-$10/hr at launch in 2023 to about $2-$4/hr by late 2025, with some providers dipping below $2/hr on spot. Silicon Data recorded a short-lived rebound in early 2026, so the trajectory has been volatile rather than one-way down. AWS reportedly cut H100 pricing by about 44% in June 2025, helping trigger a broader market reset.

AIMultiple’s Cloud GPU Rental Price Index, updated August 25, 2026, found that the median H100 rental price across its tracked cohort now sits around $3.38 an hour, down from more than $7 an hour in early 2024. That is a genuine price collapse of over half in just two and a half years, and it reflects the normal pattern for an aging chip: more supply comes online, more providers compete for the same customers, and rates fall.

Reading the trend: H100 on-demand rates fell from the $8-9/hr seen at the 2023 shortage peak to the ~$2-3/hr range as supply caught up and H200/B200 arrived. Each new generation tends to push the previous one's price down rather than raising the overall $/GPU-hour, so an index expressed in $/hour understates the fall in cost per unit of compute.

That year-over-year climb on an aging chip is the detail worth sitting with. Normally, GPU rental prices fall as a chip ages and newer silicon takes over premium workloads. Instead, H100 pricing rose through most of 2026 even as H200 and Blackwell chips became available, because demand for any working GPU capacity, new or old, has outpaced the industry’s ability to add supply.

As of October 12, 2026, the median on-demand price has been flat over the past 90 days, at about $3.50 per GPU per hour, though it is about 13% above where it was a year ago.

As of October 12, 2026, the median on-demand price is $3.50 per GPU per hour across 38 providers with a priced on-demand config.

As of Oct 11, 2026 2.83 USD / GPU-hour 0.7% (7D)

Current Market Pricing Latest available pricing across all contract terms ($/hr/GPU) SKU Period Semi Analysis Spot-Contract Composite Index On-Demand 1m 3m 6m 1y* 18m* 2y* 3y* 4y* 5y* B200 Apr 2026 $ 3.68 B300 Apr 2026 - GB200 Aug 2025 - GB300 Apr 2026 - H100 Apr 2026 $ 2.82 $2.10-2.70 H200 Apr 2026 - MI300 Nov 2025 - MI325 Nov 2025 - MI355 Nov 2025 -

H200 Bloomberg ticker: -- View $3.44 4.2% (7D) A100 Bloomberg ticker: SDA100RT View $1.53 0.7% (7D) B200 Bloomberg ticker: SDB200RT View $5.93 1.0% (7D) B300 Bloomberg ticker: -- View $6.65 2.5% (7D) MI300X Bloomberg ticker: -- View $2.64 0.0% (7D)

At 70% fleet utilization, four-year amortization of a fully burdened H100 server comes to $1.12/hr, and energy plus facility adds another $0.33/hr - against a $3.99 on-demand price at Lambda. The residual $2.54/hr - 64% of the price - is demand premium: scarcity rent, not cost recovery (SA estimate).

Bucket 1 - Hardware amortization. We take the GPU street price (SA estimates: H100 SXM5 $28,000, H200 $38,000, B200 $40,000), apply a ~1.4x server-overhead multiplier to capture the host system the GPU cannot run without - chassis, CPUs, system memory, storage, NICs, in-node switching (SA estimate) - then amortize straight-line over a 4-year refresh cycle with a 30% end-of-life residual value, spread across utilized hours only: hardware $/hr = (street price x 1.4) x (1 - 0.30) / (4 x 8,760 x utilization)

Depreciation life is the second-biggest lever. We use a 4-year refresh cycle. On a 3-year life, the H100 hardware bucket rises from $1.12 to $1.49/hr; on 5 years it falls to $0.90/hr. Hyperscalers have publicly stretched accounting lives to 5-6 years; AI-native operators often model 3.

The 30% residual is generic, not SKU-specific. Our H100-specific residual curve puts a 4-year-old unit at 15% of original value, which would raise the H100 hardware bucket to about $1.36/hr. We use the fleet-level 30% default for cross-SKU consistency.

Cost of capital is excluded. Straight-line amortization ignores the time value of money. Applying our 8% WACC default as a capital-recovery annuity raises the H100 hardware bucket ~34% (to about $1.50/hr) and shaves the premium accordingly - it does not change any qualitative conclusion.

Real-world evidence supports longer useful lives than the 2-3 year skeptic estimate. Azure retired K80/P100/P40-powered VMs in August/September 2023-GPUs launched between 2014 and 2016, implying 7-9 year useful service lives. Azure retired V100-powered NCv3-series VMs in September 2025, approximately 7.5 years after the V100's launch. The NVIDIA T4, released in September 2018, still generates rental revenue at ~$0.15/hr on Vast.ai more than 7 years later. And CoreWeave's H100s from 2022 contract expirations were immediately rebooked at 95% of original pricing.

Some of Nvidia's customers say AI chips will retain value for a long time and that customers will continue to pay for access to older processors because they'll still be useful for other tasks. CoreWeave , which buys GPUs and rents them out to clients, has used six-year depreciation cycles for its infrastructure since 2023. CoreWeave CEO Michael Intrator told CNBC this week, following quarterly earnings, that his company is being "data driven" about GPU shelf life. Intrator said that CoreWeave's Nvidia A100 chips, which were announced in 2020, are all fully booked. He also added that a batch of Nvidia H100 chips from 2022 became available because a contract expired, and they were immediately booked at 95% of their original price. "All of the data points that I'm getting are telling me that the infrastructure retains value," Intrator said.

We seek to maximize the value of our fleet of GPUs to the fullest extent possible, and do this by monetizing the infrastructure underlying expired contracts through either another contract for the remainder of its economic life or through an on-demand consumption pool. We expect a large number of GPUs to become available for the on-demand distributed compute pool as our customers renew contracts with future generations of chips. These current generation chips will be well-suited to handle inference workloads and lower intensity training workloads after contracts expire, which will lead to cost optimal solutions for our customers. As the AI revolution continues and more widespread enterprise adoption of AI workloads occurs, there will be a breadth of workload complexity levels which will need to be matched to infrastructure with varying levels of performance and cost. Some of those workloads will be less computationally intensive and can be served by older generations of infrastructure.

The strongest counterargument is the ‘value cascade’: a GPU’s job changes as it ages, so it keeps earning long after it stops being the fastest chip in the rack. In this model, a GPU spends years one and two on frontier training where peak performance is everything, moves in years three and four to high-value real-time inference where it is still plenty fast, and settles into years five and six handling cost-sensitive batch inference and analytics where latency barely matters and economics rule. The newest Blackwell parts take over training, displaced H100s slide into premium inference, and aging A100s cascade down to standard inference and ordinary compute. Each step keeps the asset generating revenue, which is the literal test for whether a longer useful life is justified.

Nobody knows yet. The GPU lifecycle for AI workloads doesn't have enough historical data to resolve the debate conclusively. Meta's published data on H100 reliability-148 GPU failures out of 419 total disruptions across 16,384 H100 GPUs over 54 days of Llama 3 405B training, implying roughly a 9% annualized failure rate under heavy utilization-adds another variable. What we do know is that legacy GPUs retain meaningful economic value far longer than the "worthless after next generation" framing suggests. The market data and deployment evidence both point toward 5-7+ years of useful economic life, with the exact duration depending on workload mix, failure rates, and the pace of software optimization.

The secondary market suggests GPUs lose value fast but not instantly: an Nvidia H100 holds roughly 75-85% of its acquisition value through 24 months, then slides to around 69% of contemporaneous new pricing by year three. That is steeper than a six-year straight-line assumption implies, but far gentler than the near-worthless-by-year-three picture Burry’s two-to-three-year framing can suggest. New H100s held a $25,000-$40,000 range from mid-2024 into early 2026, with SXM5 80GB units at the top of that band and PCIe parts near the bottom. Peak-scarcity secondary units briefly traded as high as the low $40,000s in mid-2024 before normalizing toward the mid-$20,000s by late 2025. Age is the dominant variable in the resale data. Units under a year old fetch roughly $18,000-$25,000, one-to-two-year-old units land around $12,000-$18,000, and chips past two years trade at roughly $7,000-$12,000. Warranty status moves the needle too: a unit with 12-plus months of transferable third-party coverage can command 8-12% more than an identical chip with expired support, because Nvidia’s manufacturer warranty does not transfer to secondary buyers.

Recent reporting has noted that certain AI hardware, such as Nvidia’s H100 systems, can lose a substantial portion of their resale value within the first few years, with units trading at less than half the price of new ones by year three 4 .

At the server level-the typical B2B purchase unit-a used H100 server (8-GPU) trades at roughly $150,000-$180,000. Compare that to a new B300 server at approximately $500,000. Refurbished GPU servers from vendors like Alta Technologies and NewServerLife offer 40-70% savings over new OEM systems with warranties and tested components.

Read this table against the decomposition: a 45-55% discount for a 3-year commitment prices the committed GPU-hour at roughly where the premium share compresses from ~85% toward ~65% - hyperscalers converging on today's neocloud economics, three years out. CoreWeave offering up to 60% off for terms tells you a neocloud with visibility into its own pipeline expects Hopper-class scarcity rent to largely evaporate inside the commitment window. The on-demand price is a queue price; the reserved price is the provider's actual supply forecast, with money behind it.

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.

This seemingly aberrant depreciation policy has a sizable impact on CoreWeave's bottom-line: artificially suppressing it's operating expenses and boosting it's operating income. Do note that CoreWeave also uses its sizable stash of NVIDIA GPUs as collateral to obtain debt financing. A longer depreciating period allows for a higher book value of its pledged GPUs.

The total loans available provided by the DDTL Facilities are constrained by the purchase price of assets for which the loans are being used to finance with such percentage based upon the depreciable cost of graphics processing unit ("GPU") servers.

Borrowings under this agreement were established to primarily finance capital expenditures necessary to fulfill customer contracts, including the acquisition of GPU servers and related infrastructure. Under the DDTL 2.0 Facility, additional loans may be drawn until June 2025, with an option to extend the commitment period by three months subject to lender consent. The total loans available are constrained by the purchase price of assets for which the loans are being used to finance with such percentage based upon the depreciable cost of GPU servers.

Contracted customers had a weighted average contract term of approximately four years as of December 31, 2024.

As of December 31, 2025, our committed contracts had a weighted-average contract duration of approximately five years.

Our customers often make large infrastructure commitments to us structured as take-or-pay contracts with terms that are typically two to five years in length.

Included within “GPU Hardware” above are GPUs and related infrastructure with an aggregate net carrying amount of $ 1,487.1 million as of June 30, 2026 (June 30, 2025: nil ) that are owned by IE US Hardware 3, LLC (a wholly owned subsidiary of the Company), the Company's financing subsidiary, and are pledged as collateral under the Group’s GPU Financing facility. These assets may be used only to settle obligations of IE US Hardware 3, LLC and are not available to satisfy claims of the general creditors of IREN Limited or any of its other subsidiaries until such time as the GPU Financing facility has been repaid and the security released. Depreciation on these assets is recorded on a straight-line basis over an estimated useful life of 5 years.

During the year ended June 30, 2026, the Group entered into lease financing arrangements for the acquisition of GPUs, together with related ancillary equipment. The arrangements provide financing for 100 % of the purchase price and are structured as 36-month and 24-month leases, respectively.

The maturity date for borrowings under the MFSA and Note issuance under the Note Purchase Agreement is the date falling 30 months after the funding date of the relevant borrowing under the MFSA or the funding date of the relevant Note issuance under the Note Purchase Agreement (as applicable).

Under the five-year agreement, Nebius will provide $12 billion of dedicated capacity across multiple locations, based on one of the first large-scale deployments of the NVIDIA Vera Rubin platform. Nebius will deliver this capacity starting early 2027. Furthermore, in connection with access to these NVIDIA Vera Rubin deployments, Meta has committed to purchase additional available compute capacity across certain upcoming Nebius clusters up to a total of $15 billion over a five-year period.

Given contract lengths generally range between two to five years, and the economic life of our infrastructure assets exceeds the typical contract length according to our estimates, our business model can generate attractive returns by creating opportunities to re-contract infrastructure after a customer’s initial contract expires.

Using committed contracts that were in effect as of December 31, 2024 as a basis, we anticipate that our average cash payback period, including prepayments from customers, will be approximately 2.5 years. Our cash payback period is the time we anticipate it would take to break-even on our investment in GPUs and other property and equipment through adjusted EBITDA. The cash payback period reflects how efficiently we monetize our compute resources, providing a measure of the speed at which capital invested is expected to generate returns. The calculation is performed on a per-GPU basis, starting with the estimated cash spent on property and equipment (calculated as change in total property and equipment excluding the change in construction in progress) during the quarter, divided by the number of GPUs that went into service in the period.

We continually work to upgrade and enhance our platform, solutions, and services in response to customer demand and to keep up with technological changes. Part of this process entails cycling out older components of our infrastructure and replacing them with the latest technology available. This requires us to make certain estimates with respect to the useful life of the components of our infrastructure and to maximize the value of the components of our infrastructure, including our GPUs, to the fullest extent possible. We cannot guarantee that our estimates will be accurate or that our attempts at maximizing value will be successful. Any changes to the significant assumptions underlying our estimates or to the estimates of our components' useful lives, or any inability to redeploy components of our existing infrastructure to extend past their contracted life could significantly affect our business, operating results, financial condition, and prospects.

CoreWeave is the filing in this set that actually sells GPU capacity, and even there the life is “technology equipment,” not a named GPU generation. As of December 31, 2025, CoreWeave reported 43 data centers, over 850 MW of active power, about 3.1 GW of contracted power, and $60.7 billion of remaining performance obligations, up from $15.1 billion a year earlier. Committed contracts had a weighted-average duration of approximately five years. Technology equipment is depreciated over six years. The average contract is shorter than the book life. If the hardware cannot be redeployed after the contract, the remaining book value is the risk CoreWeave describes when it says a wrong useful-life estimate can affect results. That is a disclosure, not a forecast that the hardware will be idle.

As we develop and introduce new solutions and services and add new and upgraded components of our platform (such as our expected deployment of the NVIDIA Rubin platform in the second half of 2026), we face the risk that customers may not value or be willing to adopt these newer offerings, and may forgo adopting one or more newer generations of our existing offerings. Regardless of the improved features or superior performance of the newer offerings, customers may be unwilling to adopt our platform due to design or pricing constraints, among other reasons. Even if customers choose to adopt our platform or new solutions and services that we develop, they may be slow to do so.

In addition, the equipment we deploy is subject to rapid technological change. Successive generations of GPUs and supporting infrastructure are being released on short development cycles, and customer demand may concentrate on the newest generations of hardware. As a result, the economic useful life and residual value of our deployed equipment may be shorter or lower than we anticipate, pricing for services delivered on older generations of hardware may decline, and we may be required to make substantial ongoing capital expenditures to refresh our fleet and to modify our facilities to meet the power density, cooling and other requirements of new hardware. If we do not manage technological transitions effectively, our competitive position, the utilization and value of our assets, our depreciation profile and our results of operations and financial condition could be materially and adversely affected.

250-10-45-17 A change in accounting estimate shall be accounted for in the period of change if the change affects that period only or in the period of change and future periods if the change affects both. A change in accounting estimate shall not be accounted for by restating or retrospectively adjusting amounts reported in financial statements of prior periods or by reporting pro forma amounts for prior periods. 250-10-45-18 Distinguishing between a change in an accounting principle and a change in an accounting estimate is sometimes difficult. In some cases, a change in accounting estimate is effected by a change in accounting principle. One example of this type of change is a change in method of depreciation, amortization, or depletion for long-lived, nonfinancial assets (hereinafter referred to as depreciation method). The new depreciation method is adopted in partial or complete recognition of a change in the estimated future benefits inherent in the asset, the pattern of consumption of those benefits, or the information available to the entity about those benefits. The effect of the change in accounting principle, or the method of applying it, may be inseparable from the effect of the change in accounting estimate. Changes of that type often are related to the continuing process of obtaining additional information and revising estimates and, therefore, shall be considered changes in estimates for purposes of applying this Subtopic.

ASC 250-10-50-4 The effect on income from continuing operations, net income (or other appropriate captions of changes in the applicable net assets or performance indicator), and any related per-share amounts of the current period shall be disclosed for a change in estimate that affects several future periods, such as a change in service lives of depreciable assets.

We often receive questions regarding whether the use of “new information” represents a change in estimate or an error correction. The response depends on when the information was reasonably available, when the estimate was updated for the information, how the information was interpreted, etc. For example, not changing an estimate’s underlying assumptions in a timely fashion as new information is available or circumstances change could result in an error. Additionally, misinterpreting information used to develop or support an estimate in a previously issued set of financial statements is an error.

Change in estimate. A ten-year life was set on reasonable assumptions. Usage intensified, technology moved, or maintenance history now points elsewhere. You revise forward. Nothing restates. Correction of an error. Manufacturer documentation and internal experience already indicated five years, and a ten-year life was used anyway, without support. That is not a new estimate - it is a misstatement being corrected, and prior periods restate.

And also: Although not defined, we believe the use of the term “useful economic life” in ASC 360-10-35-4 is intended to have the same meaning as “useful life,” as defined in the ASC Master Glossary. The useful life assessment of a long-lived asset is based on entity-specific assumptions about how the entity intends to use the asset , which may be different from market-participant assumptions. Accordingly, the useful life could be different than the economic life or actual physical life of the asset.

Straight-line depreciation is an industry convention, not a GAAP mandate, and may not always reflect actual economic consumption. While hyperscalers generally use straight-line depreciation for servers, GAAP allows any systematic and rational method, including accelerated depreciation, if it better reflects how an asset’s economic benefits are consumed.

We review leases, property and equipment, and intangible assets, excluding goodwill, for impairment when events or changes in circumstances indicate the carrying amount may not be recoverable. The evaluation is performed at the lowest level of identifiable cash flows independent of other assets. We measure recoverability of these assets by comparing the carrying amounts to the future undiscounted cash flows that the assets or the asset group are expected to generate. If the carrying value of the assets or asset group is not recoverable, the impairment recognized is measured as the amount by which the carrying value exceeds its fair value.

The Company reviews long-lived assets, including property and equipment, right-of-use assets and intangible assets subject to amortization, for impairment whenever events or changes in circumstances indicate that the carrying amount of the asset or asset group may not be recoverable. The recoverability of the long-lived asset or asset group is assessed by comparing the undiscounted future cash flows expected to be generated by the asset or asset group to its carrying value. If the carrying amount of a long-lived asset or asset group exceeds the expected undiscounted cash flows, an impairment loss is recognized in an amount equal to the excess of the asset or asset group's carrying value over its fair value. Fair value is determined using valuation techniques such as discounted cash flow models, market comparisons, and, where applicable, independent third-party appraisals. For the years ended December 31, 2025, 2024, and 2023, no material impairment charges were recorded.

In performing substantive procedures8 to respond to the identified and assessed risks of material misstatement associated with accounting estimates, the auditor should test an accounting estimate using one or a combination of the following approaches: 1. Test the company's process used to develop the accounting estimate (see paragraphs .09-.20 of this standard); 2. Develop an independent expectation for comparison to the company's estimate (see paragraphs .21-.26 of this standard); and 3. Evaluate audit evidence from events or transactions occurring after the measurement date related to the accounting estimate for comparison to the company's estimate (see paragraphs .27-.29 of this standard).

Examples of assumptions that ordinarily would be considered significant assumptions include those that: 1. Are sensitive to variation, such that minor changes in the assumption can cause significant changes in the estimate; 2. Are susceptible to manipulation or bias; 3. Involve unobservable data or company adjustments of observable data; or 4. Depend on the company's intent and ability to carry out specific courses of action.

Evaluating potential bias in accounting estimates includes evaluating bias in estimates individually and in aggregate. It also includes evaluating whether bias results from the cumulative effect of changes in estimates.

Dustin Madsen, vice president of the Society of Depreciation Professionals and founder of Emrydia Consulting, said depreciation is a financial estimate by management and that developments in a fast-moving industry like technology can change initial predictions. Depreciation estimates, Madsen said, generally take into account assumptions such as technological obsolescence, maintenance, historical lifespans of similar equipment and internal engineering analysis. "You're going to have to convince an auditor that what you're suggesting what its life will be is actually its life," Madsen said. "They will look at all of those factors, like your engineering data that suggests that the life of these assets is approximately six years, and they will audit that at a very detailed level."

Burry this week suggested that companies including Meta , Oracle, Microsoft, Google and Amazon are overstating the useful life of their AI chips and understating depreciation. He pegs the actual useful life of server equipment at around two to three years, and said companies are inflating their earnings as a result. Amazon and Microsoft declined to comment. Meta, Google and Oracle did not respond to requests for comment.

In a recent post on X, the “Big Short” investor warned that major technology companies may be understating depreciation by a massive $176 billion between 2026 and 2028, artificially inflating earnings and misleading investors.

According to Burry’s estimates, Oracle (ORCL) could overstate earnings by 26.9% by 2028, and Meta (META) may inflate profits by 20.8% over the same period.

Michael Burry has accused the hyperscalers of inflating earnings by roughly $176B over 2026 to 2028, putting Meta at ~21% and Oracle at ~27% overstated. That is his own forward projection, and it does not match what the filings disclose: none of these companies break out a separate GPU asset class, and the disclosed, single-year impacts run around 4 to 5% of profit, not 20-plus. The accusation points at a real lever. It overstates the size of it.

Hyperscalers justify the longer depreciation schedule by arguing for a value cascade model. They contend that older generation GPUs, once replaced in top-tier training jobs, are simply cascaded down to power less computationally intense but high-volume inference (running the model) or other tasks, where they can still generate significant economic value for years. They also cite continuous software and data center operational improvements that extend the hardware’s life and efficiency.

We side with the hyperscalers rather than Michael Burry in the depreciation debate. Data Centers existed before AI caught on in late 2022, when ChatGPT was first introduced. During 2021, there were as many as 4,000 of them in the US as a result of the rapidly increasing demand for cloud computing. Many are still operating with their original chips. The revenues and earnings of the hyperscalers continue to rise rapidly (chart).

Nvidia pushed back, arguing that customers consistently use four-to-six-year depreciable lives based on observed utilization and longevity, according to CNBC. 1

Nvidia responded to Burry’s critiques and the pile-on that came with it with a detailed memo to Wall Street analysts , disputing his claim and insisting that it does not resemble historical accounting frauds such as Enron or WorldCom

The cascade does not make Burry wrong; it makes the question conditional. If you mean the chip’s life as a frontier training accelerator, two to three years is fair. If you mean its life as a revenue-generating asset across all workloads, five to six years is defensible. The accounting standard asks for the second number, the economic-life-of-the-asset number, which is why auditors have largely signed off. The honest critique is not that the schedules are fraudulent but that blending genuinely short-lived training silicon with genuinely long-lived inference silicon into a single estimate can flatter the average.

Analysts tracking this debate believe hyperscalers will likely converge toward something closer to a five-year depreciation cycle - shorter than today's prevailing five-to-six-year assumption, but still longer than the two-to-three-year hardware-replacement pace Burry argues is the real economic reality.

The cascading-use argument has empirical support. Hyperscalers run a multi-tier infrastructure - top-tier GPUs serve frontier training for 12-18 months, then cascade to fine-tuning, then to inference, then to non-AI compute or as backup capacity. If the cascade actually extends total economic life to 5-6 years, the longer schedules are correct. The bull side cites internal utilization data showing meaningful inference workloads still running on Volta-generation hardware deployed in 2017. The counter to the counter: the cascade was true for the pre-AI fleet. It is not yet proven for the AI fleet, where each new generation is materially more capable per dollar than the prior, and where customers are willing to pay materially more for the new generation. The economics may not support the cascade in the same way.

Since changing those estimates - whether to shorten or lengthen them - requires evidence that will stand up to an external audit , for GPUs specifically, the question is whether there is already sufficient, reliable data to justify a change. Should the depreciable life for GPU servers be say, two- or three-years, or conversely is it appropriate to support keeping a five- or six-year server life for a mixed CPU/GPU environment? Reasonable people can disagree about where the estimate should land within that range - I personally think companies with a six-year useful life of GPUs would follow Amazon’s lead and reduce the useful lives of their GPUs. But GAAP anticipates exactly this kind of uncertainty: it expects estimates to evolve as data accumulates , not to be retroactively treated as errors or fraud simply because new information , new chip generations, or new economic perspectives emerge.

What does this not prove? It does not prove anyone did anything wrong. Changing a useful life is a change in accounting estimate: permitted, applied prospectively, disclosed, audited. Amazon moved its number in the direction that hurt its own earnings, the opposite of what an earnings-flattering story predicts. Nor does it prove any estimate wrong. Nobody has six years of operating history on this generation of accelerators, so every figure in that table is a forecast, and firms with different refresh cadences and resale markets should forecast differently.