Analysis · AI Infrastructure · August 17, 2026
Nebius Gave Its Servers One More Year. The Accounting Change Was Worth $75.7 Million.
On August 12, Nebius Group reported second-quarter revenue of $582.3 million, up 454% from a year earlier, and its stock rose 34.14% that day. Filed the same morning, on page F-9 of the company’s Form 6-K, was a paragraph almost nobody trades on. Nebius had reassessed how long its servers last, concluded the answer was five years rather than four, and applied the change from January 1. For the first half of 2026, that single revision reduced depreciation expense by $86.1 million and raised net income by $75.7 million.
Nothing about this is improper. Changing an accounting estimate is permitted under U.S. GAAP, it is applied prospectively rather than by restating the past, and Nebius disclosed it plainly — which is the only reason it can be written about here. But it is a reminder of something structural about the AI infrastructure trade: an enormous share of the sector’s reported profitability now rests on an assumption about how long a graphics processor stays useful, and the companies building the same data centers with the same chips do not agree on the answer.
They are not even moving in the same direction. Meta extended its server lives to 5.5 years in January 2025 and booked $2.92 billion less depreciation as a result. Amazon, in the same month, shortened the life of part of its fleet from six years to five and took a $1.0 billion hit to net income — citing, explicitly, the pace of AI development. Alphabet depreciates servers over six years. Nebius has now landed on five.
1. What the filing actually says
The disclosure appears in the summary of significant accounting policies. Nebius notes that nothing material changed from its 2025 annual report “except for the update to the estimated useful lives of servers and network equipment,” then explains:
“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 and concluded that the estimated useful lives of such assets should be extended from four to five years.”
The disclosed effects are precise. For the three months ended June 30, depreciation expense fell $43.0 million and the net loss narrowed by $34.1 million. For the six months, depreciation fell $86.1 million and net income rose $75.7 million.
Those numbers are large relative to what Nebius reported. The company recorded $259.7 million of depreciation and amortization in the second quarter; without the change it would have been roughly $302.7 million. The reported loss from operations was $175.9 million; on the old four-year basis it would have been about $218.9 million — a reported operating loss roughly 20% smaller than it would otherwise have looked.
Second quarter 2026, in millions of U.S. dollars. The restated column adds back the company’s own disclosed effect of the change in estimate — $43.0m of depreciation and $34.1m of net loss. It is arithmetic on Nebius’s disclosure, not an alternative accounting treatment.
2. Why one year moves so much money
The leverage here comes from the size of the capital base. Nebius spent $8.13 billion on property, equipment and intangible assets in the first half of 2026 alone — up from $1.05 billion in the same period of 2025, a nearly eightfold increase. Straight-line depreciation divides that number by an assumed life. Every additional year in the denominator lowers the annual charge, and the effect compounds across a fleet that is being expanded faster than it is being retired.
The arithmetic is unforgiving in both directions. Depreciated over four years, the first half’s capex alone implies about $2.03 billion of annual expense. Over five years, roughly $1.63 billion. That $406 million annual gap is the entire difference between the two assumptions, on six months of spending.
Illustrative. Applies simple straight-line depreciation to the $8,130.3m of property, equipment and intangibles Nebius purchased in the six months to June 30, 2026, with no salvage value and no assumption about timing of placement in service. Real depreciation schedules differ; this isolates the sensitivity to the life assumption alone.
3. Four companies, four answers
What makes this more than a single-company footnote is the dispersion. I pulled the most recent annual filing for each of the large operators I either hold or track, and read the property and equipment policy directly. The assumptions do not converge.
| Company | Assumed useful life | Most recent change | Effective | Disclosed effect |
|---|---|---|---|---|
| Nebius (NBIS) | 5 years | Extended, 4 → 5 | Jan 1, 2026 | H1 2026: depreciation −$86.1m; net income +$75.7m |
| Meta (META) | 5 – 5.5 years | Extended, to 5.5 | Jan 1, 2025 | FY2025: depreciation −$2.92b; net income +$2.59b (+$1.00 / diluted share) |
| Amazon (AMZN) | 5 – 6 years | Shortened, subset 6 → 5 | Jan 1, 2025 | FY2025: depreciation +$1.4b; net income −$1.0b (−$0.10 / share) |
| Alphabet (GOOGL) | 6 years | No change disclosed in FY2025 | — | Not separately quantified in FY2025 10-K |
Where a company discloses a range, the bar shows the range. Arrows indicate the direction of the most recent change in estimate.
4. Amazon is the interesting one
Almost every version of this story told over the past two years has been about companies extending lives and flattering earnings. Amazon is the counterexample that keeps the argument honest. Effective January 1, 2024 it moved servers from five years to six — the flattering direction. Then, effective January 1, 2025, it went back the other way for part of the fleet, and said why in its 10-K:
“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.”
That is the same industry condition — rapid AI hardware development — being read as a reason to depreciate faster at one company in the same month Meta read the environment as a reason to depreciate slower. Both cannot be describing the same economic reality about the same class of equipment. One of these assumptions is closer to right, and the filings do not tell you which.
5. The case that longer lives are correct
It would be lazy to treat every extension as earnings management, and I want to state the other side properly, because I hold two of these companies.
Old accelerators keep earning
A chip stops being state of the art long before it stops being useful. Prior-generation accelerators get pushed down the stack to inference, fine-tuning and internal workloads rather than scrapped. If those hours are still billable, a longer life genuinely matches the revenue the asset produces — which is precisely what depreciation is supposed to do.
Operators have real utilization data
Nebius says its reassessment was “based on updated information and usage patterns obtained.” These companies observe failure rates and retirement behavior across enormous fleets. An outsider reasoning from product-cycle length has strictly less information than an operator reading its own maintenance records.
It is disclosed, not hidden
Every figure in this article came from the companies’ own filings, quantified by them, in the places where such things belong. That is the system working. The risk is not concealment; it is that readers price the headline and never open the footnote.
There is also a scoreboard forming, and so far it does not support the alarmed reading. Impairment of property and equipment is the line where an over-optimistic life assumption eventually surfaces — the company admits the asset is not worth what the schedule says. Meta booked $237 million of such impairments in FY2025, down from $288 million in 2024 and $738 million in 2023, even as its server fleet and its assumed lives both grew. That is the opposite of what the bear case predicts.
Amazon’s equivalent signal is not cleanly readable. Its FY2025 filing bundles asset impairments together with legal settlements, tax disputes and severance inside a $2.4 billion fourth-quarter charge without separating the equipment portion, so the company that shortened its lives does not give you a clean impairment number to check the decision against. That is a limitation of the disclosure, not evidence either way — but it is worth knowing that the check is unavailable.
6. What this means for the book
I own three of the four companies in the table, and the exposure is not evenly distributed.
| Position | Weight | Life assumption | Note |
|---|---|---|---|
| Meta (META) | 8.42% | 5 – 5.5 yrs | Largest single exposure to an extended-life assumption in the book |
| Alphabet (GOOGL) | 3.49% | 6 yrs | The longest assumption of the four, unchanged in FY2025 |
| GraniteShares 2× NBIS (NBIL) | 2.06% | 5 yrs (via NBIS) | Leveraged — roughly 2× daily exposure to the same underlying |
| Nebius (NBIS) | 1.24% | 5 yrs | Direct position |
Two things follow. First, roughly 13.2% of the book sits directly in companies whose reported earnings are measurably sensitive to this one assumption, with Meta alone accounting for most of it. Second — and this is the part my own Risk X-Ray was built to catch — the Nebius exposure is doubled up. Holding NBIS at 1.24% alongside a 2× leveraged NBIS fund at 2.06% means the economic exposure to Nebius is closer to 5.4% than to 1.24%. That is the exact pattern the tool flags, in my own account, on the one name in the group that just changed its assumption.
What I am not claiming
I am not alleging that any of these companies acted improperly, and nothing here suggests the numbers are wrong. Changes in accounting estimate are legal, ordinary, and were disclosed and quantified by the companies themselves — which is the only reason this article could be written. Nor am I claiming to know the right useful life for an H100 or a Blackwell rack. I am claiming something narrower: that the assumption is large enough to move reported profit by billions, that four sophisticated operators disagree about it, that one of them is moving in the opposite direction from the rest, and that a reader who stops at the revenue line will not see any of it.
What would change my read
WatchThree things I am specifically waiting for
1. Impairment lines. If property and equipment impairments at Meta or Amazon step up materially from FY2025 levels, that is the market telling us the longer assumptions were wrong, in the one place it cannot be argued with. 2. A reversal. If any company that extended lives shortens them back, as Amazon did, the extension thesis weakens considerably. 3. Nebius’s next filing. Nebius is depreciating a fleet that is growing far faster than it is aging; the full-year effect of this change will be larger than the $75.7 million booked so far, and I want to see it quantified against a full year of the $8.1 billion-and-rising capital base.
The AI infrastructure trade has been argued mostly on demand — whether the tokens, the contracts and the power exist to justify the spending. That argument is covered elsewhere on this site, in The AI Capex Reality Check and Data Center Alley. This is a quieter question, and in some ways a more tractable one, because it does not require forecasting anything. The filings already tell you what each company assumed. They just do not agree.