Amazon’s AI Reality Check
What a simple accounting change reveals about the true cost of AI hype.
There is a lot of talk about the huge sums being spent on AI chips and data centres by the hyperscalers, much less about the related depreciation. Longtime readers will know this is an interest of mine, and I am returning to the subject because of Amazon’s change in estimated server life in its Q1 2025 filing:
“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.
Amazon further explained:
“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 Q1 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 three months ended March 31, 2025, was an increase in depreciation and amortization expense of $217 million and a reduction in net income of $162 million, or $0.02 per basic share and $0.02 per diluted share, which primarily impacted our AWS segment.”
In Q2, the impact was noted as
“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 effect… for Q2 2025 was an increase in depreciation and amortization expense of $280 million and a reduction in net income of $217 million”
With this change, Amazon is reducing its reported earnings by c.$0.5bn in H1 2025 and by over $1bn for the full year, a number which I believe significantly understates the ongoing impact.
When I first wrote an article on this published in February, 2024 (“What do MAG7 earnings and a Snickers bar have in common? - Welcome to Financial Shrinkflation”), I pointed out that even the most conservative CFO felt under pressure to follow competitors which lengthened the server depreciation life, with one doing it more than once in a fiscal year. Here is the table from that note, highlighting the changes:
Server Life Comparison – February 2024 Position
Source: Behind the Balance Sheet from 10-K Filings
In 2024 Amazon extended server lives from five to six years, citing “continuous improvements in our hardware, software, and data center designs”. The 2025 reversal was apparently because higher performance requirements for AI workloads outweighed the prior durability benefit.
When it shortened the server life, it additionally disclosed:
“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. We expect to continue to acquire more of these server and networking assets in 2025.”
But this wasn’t the only change:
“In 2024, we also determined, primarily in the fourth quarter, to retire early certain of our servers and networking equipment. We recorded approximately $920 million of accelerated depreciation and related charges for the quarter ended December 31, 2024 related to these decisions. The accelerated depreciation will continue into 2025 and decrease operating income by approximately $0.6 billion in 2025. These two changes above are due to an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”
Amazon not only reduced the asset life but took an accelerated write-down because six years was too long an assumption of useful life.
Meta went the other way, although from a lower starting point than Amazon:
“In January 2025, we completed an assessment of the useful lives of certain servers and network assets, which resulted in an increase in their estimated useful life to 5.5 years, effective beginning fiscal year 2025. Based on the servers and network assets placed in service as of December 31, 2024, we expect this change in accounting estimate will reduce our full-year 2025 depreciation expense by approximately $2.9 billion.”
Meta Asset Schedule
Source: 2024 10-K
The 10-K notes
“Within property and equipment, our servers and network assets depreciation expenses were $11.34 billion, $7.32 billion, and $5.29 billion for the years ended December 31, 2024, 2023, and 2022, respectively.”
The pro-forma impact is a reduction from $11.3bn to $8.4bn in depreciation expense on a base of $68.4bn. Clearly some of these servers are already fully depreciated and the notional life extension from 5.1 years to 6.9 years (average gross asset/depreciation charge) reflects that.
Interestingly, the Meta asset base is $121bn vs Amazon’s at $204bn as of December and $252bn as of June). Here is the Amazon breakdown:
Amazon Asset Schedule
Source: 2024 10-K
Amazon has $185bn of equipment but that includes trucks, aircraft, a zillion robots (seriously, they already have amazing robotic tech and more to come), as well as those servers.
I am curious how Meta can improve earnings by $2.9bn by extending the server life by six months yet Amazon, whose asset base is growing by $100bn this year, only gets a c.$1bn additional charge when it reduced its server life by 12 months. Something doesn’t quite add up.
Amazon’s total depreciation and amortization charge was $24.9bn in 2024 vs $15.3bn for Meta. Note that Amazon leases a significant amount of equipment, some of which is capitalised on the balance sheet but under US GAAP this notional asset is not depreciated in the P&L, unlike under IFRS where a notional asset is capitalised and depreciated.
Oracle too changed its policy in respect of server lives:
“During the first quarter of fiscal 2025, we completed an assessment of the useful lives of our servers and networking equipment and increased the estimate of the useful lives from five years to six years, effective at the beginning of fiscal 2025. Based on the carrying value of our servers and networking equipment as of May 31, 2024, this change in accounting estimate decreased our total operating expenses by $733 million and increased our net income by $573 million, or $0.21 per basic and $0.20 per diluted share, during fiscal 2025.”
Oracle was not in my original article, but they too changed policy previously:
“During the first quarter of fiscal 2023, we completed an assessment of the useful lives of our servers and increased the estimate of the useful lives from four years to five years effective at the beginning of fiscal 2023. Based on the carrying value of our servers as of May 31, 2022, this change in accounting estimate decreased our total operating expenses by $434 million during fiscal 2023.”
I am no expert on US tax law, but these assets don’t seem to qualify for accelerated depreciation as they would in the UK, hence the additional income is taxed which surely is an argument for leaving the faster write-down in place. Either the CFOs are determined to disclose a more accurate picture or they think eps is an important variable for valuation. Perhaps there is an eps metric in their compensation?
Here is the breakdown of its assets:
Oracle Asset Schedule
Source: Oracle 2025 10-K
The company has $44bn in assets and just signed a $300bn deal - no wonder the stock went up on the day. It may be interesting to watch how a company with $30bn in computer and network equipment will manage a 10x scale-up.
Also of note is that Oracle’s depreciation benefited by $0.7bn from a relifing from 5 to 6 years on an installed base of computer and other equipment of $30bn. Amazon, whose total equipment is $219bn (or almost 7x Oracle’s) and whose gross fixed assets are $394bn (or >6.5x Oracle’s), only took a hit of c.$1bn pa in its first 6 months this year, on a similar revision albeit in the other direction.
Some of Amazon’s equipment may be fully depreciated but this looks slightly odd. I normally try to age the assets which is easier for IFRS reporters as there is more granular data on the individual asset categories.
But based on the Amazon capex profile and excluding grants (which are important for the warehouse spend but likely less relevant to the data centres), here is a very rough cut estimate of the Amazon asset age profile:
Amazon Asset Age Profile
Source: Behind the Balance Sheet from Amazon Filings
Roughly half the assets are likely less than 4 years old so perhaps half or more of the servers are fully depreciated. A $1bn additional H1 charge therefore seems low. Looking at this from a different perspective, we know the AWS share of the group assets which implies that AWS has PP&E of over $100bn in 2022 and over $200bn in 2024.
Amazon Retail may be more asset intensive than AWS – it has a lot of trucks and aircraft and I don’t know the proportion of operating leases, finance leases and property “build to suit” leases. I could do significant further investigation, but I doubt that I would be able to reconcile the Amazon data. Of course, the numbers filed in the 10-Qs must be accurate for 2025, but I doubt they will be an accurate representation of the impact on depreciation going forward. And that may be why the other cloud and server operators are not following suit. Here is my previous table updated to today, and the picture on server lives now looks like this:
Server Life Comparison
Source: Behind the Balance Sheet from 10-K Filings
Everyone except Amazon and Meta is using a life of 6 years. So the equipment they install today will still be current in 2030. Really?
One other potential issue - I wonder if the sell-side is paying much attention. I have written here before about the juniorisation of the sell-side and the need for asset management teams to beef up their training (I can help!). Here is Jim Chanos with a series of tweets on a Cantor report on Iris Energy, a combination of bitcoin miner and data centre builder – just writing that phrase means I know it has rocketed and is heading for a fall…
“The Cantor report on $IREN is indicative of the ridiculous assumptions that the sell-side is using to justify its price targets. Here, Cantor assumes a $2.9B increase in PP&E from FY 2025 just ended, to FY 2027. Average PP&E also increases $2.9B.”
Chanos continues:
“Yet annual depreciation only increases by $330M over the two years(from $181M to $511M),or almost 9-yr life! As the increase in PP&E is almost all GPU’s, this is ridiculous. 5-yr life on the increased spend and a 21% tax rate would cut his FY2027 EPS estimate over 40% to $1.80!”
H/t Edelweiss Capital
Chanos probably calculates a 9 year life, based on the net book value in the balance sheet. The 2027 depreciation estimate is $511m and the net PP&E is $4847m, and I estimate the gross PP&E at c.$6bn so the average life being assumed is actually 12 years; the bitcoin assets are subject to an accelerated depreciation schedule so some may be fully depreciated.
Even so, the Cantor depreciation estimate looks way out – I calculate their capex forecast 2025-27 as $5.5bn; even allowing for $1.0bn being under construction, that would be an incremental depreciation charge, using a 6 year life of $750m; that would take the overall charge closer to $1bn vs the Cantor estimate of half that.
I appreciate that investors in Iris Energy have little interest in the forecast depreciation charge – it’s a hype stock, but this illustrates the lack of attention to this type of issue. Premium subscribers can read on for my takeaways and conclusions, and what this might mean for some of these stocks.
This week’s sponsor has an interesting offer:
As someone who’s spent years training professional investors, I’ve seen first-hand how quickly our industry is evolving. The topic I am most often asked about these days is the role of AI in finance.
That’s why I want to tell you about the AI for Business & Finance Certificate Program, developed by Columbia Business School Executive Education and Wall Street Prep.
This 8-week online program blends academic rigour with practical, non-technical applications of AI.
You’ll learn directly from Columbia Exec Ed faculty with LIVE office hours, explore interactive case studies, and join a global peer network. This course is not just for quants, but for anyone in FP&A, risk, compliance, or investing.
Enrolment closes November 10.
Use code BTBS300 to get $300 off tuition.
