By OSCAR MACKERETH
Since the release of ChatGPT in November 2022, the buildout of AI infrastructure has become a ‘must participate’ trend for Mega Cap Tech, driven by the transformative potential of artificial intelligence to reshape industries and economies.
Already, the ‘big four’ hyperscalers (Amazon (AWS), Microsoft (Azure), Google (GCP) and Meta) have invested hundreds of billions of dollars into the technology through a blend of equity investments, capital expenditure, and R&D spend, with guidance for continued growth 2025. Company executives have repeatedly emphasised the size of the AI opportunity, to the point of suggesting an inability or preference to over-index on spending, rather than risk of being under-invested in the technology.
Figure 1: Cerno Capital, Bloomberg
This has prompted a surge in capex, with hyperscaler expenditure rising from ~US$150bn to US$230bn between 2023 and 2024, with over US$300bn expected for 2025. The majority of this is expected to be dedicated to AI infrastructure in the form of data centres, where industry operators have suggested that GPUs account for 60-80% of total cost of ownership. The majority of GPU spend is currently attributable to Nvidia GPUs and accompanying hardware, which have become essential commodities for those businesses looking to train and run Large Language Models (LLMs).
Outside of the ‘must-participate’ nature of AI, this spend acceleration is at least partially attributable to the cadence of product releases, which have solidified Nvidia as a leader in accelerated computing. In 2023, Stanford’s Human-Centred Ai Group evidenced a 7,000x rise in GPU performance since 2003, with an accompanying 5,600 improvement in price per performance over the same period[i].
Figure 2: Stanford’s Human-Centred AI
Similarly, over the last eight years and five architectures, Nvidia GPUs have increased LLM inference energy-efficiency by ~45,000 times[ii], with an architecture change driving an average 92.5% reduction in joules consumed per token.
The scale of these capacity improvements in GPU architectures provide exponential gains in AI performance and efficiencies. Jensen Huang (the CEO of Nvidia) went as far as suggesting that these gains are significant enough to the total cost of ownership for its data centre customers, that using competitor chips for free would still be damaging to competitiveness[iii].
Thus, the pace of Nvidia innovation has created a product-led growth flywheel in which each new architecture offers sufficient performance gain to force customers to adopt them to remain competitive, creating a cycle of necessity-driven capex lock-in. Whilst there are some competition concerns around ASICS and novel architectures, this trend of innovation driven spend looks set to remain; Huang[iv] proclaims his company’s GPUs as progressing above the historical rate of Moore’s Law. Conveniently, this has served to justify an accelerated product cadence with new architectures being moved from biennial to annual releases.
Figure 3: Nvidia, Cerno Capital
Therefore, through the necessity of AI participation and accompanying pre-requisite of accumulating Nvidia’s latest chipsets, the capital intensity of the hyperscalers has been driven well above historical norms. Hyperscaler guidance points to a collective >US$300bn of capital expenditure in 2025, representing 23.3% of forecasted revenues – well above the historical median of 13.9%.
Capex, Depreciation and the Optics of Profitability
Some market participants have expressed concerns over the ROI of this spend, drawing parallels between the scale and sentiment of the infrastructure buildout to historical bubbles, such as the telecom infrastructure buildout in the late 1990s and early 2000s. However, enthusiasts are eager to highlight a key difference: today’s hyperscalers maintain some of the largest balance sheets in history and are investing in an already profitable business line. Their conviction is found in the thesis that extensive capex in AI data centres will eventually pay dividends, as the emerge as the foundations of a new digital economy.
Given the rapid acceleration of AI revenues in the last year, company executives have been quick to affirm this view, commemorating the financial success of these business lines whilst simultaneously justifying the need for continued investment. Microsoft CEO Satya Nadella, for example, celebrated Microsoft’s inference business as the fastest line in the business history to reach a US$10bn revenue run rate, with growth actively being constrained by a lack of available compute capacity[v]. Similarly, Amazon’s AWS AI business was labelled a ‘once-in-a-lifetime opportunity[vi], rapidly achieved a ‘multibillion-dollar annualised revenue run’ whilst necessitating continued investment in infrastructure assets ‘in advance of when we can monetise it with customers using the resources’[vii].
What must be acknowledged is that this investment has coincided with notable changes in accounting. Interestingly, these changes have been enacted in near unison: all four hyperscalers have extended the useful lives of their servers and networking equipment.
Historically, hyperscalers depreciated servers over ~3 years, however, in 2020 they began pushing those timelines out. Amazon led the way, extending server depreciation from 3 to 4 years in 2020, after observing longer physical use[viii]. In 2021, the business further extended servers to 5 years and networking gear from 5 to 6 years, citing internal efficiency improvements that ‘lower stress on the hardware and extends the useful life’[ix].
Between 2021-2022, Microsoft[x], Alphabet[xi] and Meta[xii] followed suit, collectively lifting the useful lives of their server equipment to 4 years. In the year following, Microsoft[xiii] and Alphabet [xiv]further extended the depreciable lives for server equipment to 6 years, and Meta[xv] to 5 years, essentially moving in lockstep with Amazon.
To date, the hyperscalers have benefited from front-loaded profits whereby capitalised GPUs can be deployed to generate training and inference revenues before significant depreciation expenses have to be recorded. Typically, the longer-term unit economics of a GPU would depend on the generated revenue exceeding the expenses recorded against it, with a positive ROI requiring revenue per unit to grow at least at the same rate (or faster than) depreciation per unit over time. Therefore, as these companies embark of a historical investment cycle, whilst simultaneously extending the depreciation period, the hurdle rate required to achieve positive economic value is mechanically lowered. This benefits these companies by smoothing earnings and protecting operating margins.
Figure 4: Cerno Capital, Bloomberg
Whilst exact figures are not available, estimating ~55% of capex as dedicated to GPU assets, with a 10% salvage value, we can see that the extension of the useful lives of data centre assets from 3 to 6 years would have reduced collective data-centre depreciation expenses from US$39bn to US$21bn in 2024. Moreover, with a guided ~US$300+ billion of Capex in 2025, the estimated depreciation figure would be reduced from some US$51bn to US$28bn – a 46% saving. Clearly this policy, whilst receiving little scrutiny, has significant consequences for the profitability of this infrastructure buildout.
Correlation or Coincidence?
The elongation of useful lives for data centre assets raises a critical question: are these adjustments a reflection of genuine economic and technological realities? Alternatively phrased: are these policies a lever by which hyperscalers are enhancing the optics of their investment programs amid rising investor concerns over the return on AI initiatives? Regardless of perspective, the underlying issue remains the apparent tension between operational pragmatism and financial presentation. Do these extensions reflect genuine improvements in hardware longevity?
The proposed rationale for extending the useful lives of AI data centre assets is rooted in a triad of factors: enhanced operational efficiency, maturity of infrastructure, and a long-term investment strategy that seeks to align capital expenditure with the extended revenue-generating potential of AI and cloud services. Executives argue that continuous improvements in software and data-centre operations mitigate wear and tear on hardware, while the scale and optimisation of these infrastructures enable them to capitalise on efficiency gains, thereby justifying a longer useful life[xvi],[xvii].
At Amazon, for instance, the CFO has been explicit in linking operational improvements with extended asset longevity. He stated, ‘we continue to refine our software to run more efficiently on the hardware. This then lowers stress on the hardware and extends the useful life’[xviii], suggesting that technological refinements make their infrastructure more durable and economically viable for a longer period. Similarly, Microsoft’s Satya Nadella noted; ‘Our data centres, networks, racks, and silicon are all coming together as a complete system to drive new efficiencies to power both the cloud workloads of today and the next generation AI workloads’[xix].
Through these narratives, executives seek to frame longer useful lives as the natural outcome of their domain‐specific expertise, rather than as a financial manoeuvre. They argue that continuous software refinements and operational improvements validate extending depreciation periods. Yet, despite this compelling narrative, these justifications appear to clash with the realities of the current AI hardware landscape.
Firstly, recent GPU cycles illustrate that economic life is decoupled from physical durability. For example, the extended utility of A100 GPUs during H100 supply chain delays masked their underlying obsolescence. Once H100s became widely available, A100s saw steep value erosion[xx]. With Blackwell chips expected to enter circulation without similar supply bottlenecks, the economic lifecycle of H100s may compress significantly. Thus, contrary to the notion of prolonged utility through operational optimisation, the pace of product cadence and supply availability may in fact accelerate depreciation timelines – even as hyperscalers move in the opposite direction.
Secondly, those supply and demand observations used to justify elevated capex should simultaneously evidence the higher expected utilisation for data centre GPUs, as capacity is built into pent-up demand. However, high utilisation rates generally result in faster hardware degradation[xxi], thus, shortening the asset’s actual useful live. Hyperscalers undoubtedly have invested significant capital into optimising data centre-maintenance which should help to buffer some of this impact. However, as all of them have identified the majority of their capex as AI-specific hardware[xxii],[xxiii],[xxiv],[xxv], this short-lived asset should take up a greater proportion of the data-centre asset base-a shift that logically implies a shorter useful life on a gross basis.
Thirdly, this concern is further amplified by the accelerated product cadence observed in the AI space. Driven by NVIDIA’s annual release cycle, a leading-edge GPU in a traditional three-year depreciation cycle will now be three generations old by the time it is replaced. Moving to a five- or six-year depreciation schedule therefore suggests that by replacement, the hardware would be five or six generations behind. This view is supported by Satya Nadella’s recent commentary: ‘On inference, we have typically seen more than 2x price performance gain for every hardware generation and more than 10x for every model generation due to software optimisations’[xxvi]. As such, given the current cadence of product improvements, the economic obsolescence of AI hardware appears likely to occur significantly before than its physical failure.
Finally, it is important to recognise that the current phase of GPU deployment-particularly for large model training-is arguably more akin to research and development than to conventional commercial activity. As such, applying traditional depreciation schedules imposes a veneer of financial regularity over a highly uncertain, research-led process. Unlike a factory asset with predictable utility, the value of GPU clusters may violently appreciate or depreciate, based on breakthroughs or breakdowns in scaling laws and AI paradigms. This binary risk profile suggests that the true useful life of these assets is unknowable ex-ante, rendering current accounting assumptions vulnerable to swift obsolescence or write-down.
As such, the asset optimisation strategies presented by the hyperscalers appear to be credible and underpinned by tangible operational advancements. However, when examined within the context of the current progress in AI hardware, the extended depreciation schedules appear increasingly contradictory to reality.
Interviews with industry operators suggest that the unified extension of depreciation schedules across hyperscalers may partly reflect a form of “safety in numbers” rationalisation, where no single firm risks elevated investor scrutiny by deviating from the group norm. While not fraudulent per se, this dynamic raises the possibility of accounting optimism, wherein depreciation assumptions are calibrated for an optimistic scenario in a fundamentally uncertain technological reality. If asset lives continue to be extended based on overly hopeful research assumptions – which may not materialise – then future impairments could reveal a significant mispricing of risk.
Indicators of Financial Strain
Investors should be alert to the potential risks inherent in extended depreciation policies, which may not accurately reflect the physical realities of rapidly evolving AI hardware. Central to this concern is the divergence between the book value of assets-prolonged by extended depreciation, and their real-world physical performance which is subject to accelerated technological obsolescence. Over time, this misalignment may compel companies to reverse their depreciation policies or record impairment charges when the equipment fails to support the intended workload, thereby reducing operating profit. Such adjustments have the potential to disrupt reported earnings and negatively impact long-term profitability by forcing the recognition of costs in future periods.
One potential illustration of these risks is highlighted by Amazon’s recent adjustments. As the originator of the depreciation extension trend, the company’s assets have had the longest time-weighted exposure out of any of the hyperscalers. Perhaps coincidently, the company recently recognised an ‘increased pace of technology development, particularly in the area of artificial intelligence and machine learning’[xxvii]. As a result, the company has partially walked back its previous depreciation extension by retiring a portion of existing assets ahead of schedule, whilst reducing the useful life of a subset of servers and networking equipment from six to five years. We estimate that this adjustment will have a gross impact of approximately US$2.22bn on operating income between Q4FY24 and FY25 – comprising a US$920mn early retirement expense in Q4FY24, an additional US$600mn over FY25, and a further US$700mn increase in depreciation expense from FY25 onwards. Measures to offset these effects include extending the useful life of heavy equipment from 10 to 13 years, which is projected to increase operating profit by around US$900mn, further illustrating the complexity of re-aligning financial reporting with operational realities.
Therefore, on a risk basis we suggest that investors monitor two leading indicators.
First, the frequency and magnitude of adjustments to depreciation policies, such as early retirements or shortened useful lives: Frequent revisions may indicate that the original estimates of asset longevity were overly optimistic, while large adjustments suggest that the operational performance of the hardware is not meeting prior assumptions. Such changes can serve as forward-looking indicators of future impairment charges or accelerated depreciation, both of which could materially affect operating profit if past adjustments are ultimately reversed.
Second, any structural divergence between net income and free cash flow: Extended depreciation policies tend to suppress non‐cash expenses and inflate net-income, even as cash outlays for new AI infrastructure remain significant. A widening gap between reported earnings and free cash flow in the short term is indicative of an investment cycle. However, a structural degradation of free cash flow margins may indicate that the financial statements do not fully reflect the true cost of maintaining and replacing AI hardware. This divergence, when persistent, could act as a leading indicator that accounting practices are being used to enhance profitability metrics without corresponding improvements in cash generation.
Implications for investors
The acceleration in AI data centre capex stands out even against the backdrop of a Tech industry renowned for rapid reinvestment. Powered by the dual narratives of AI as both a ‘must participate’ and ‘winner takes most’ market, the hyperscalers have embarked on expansive infrastructure programs that mirrors historical technology bubbles. However, unlike in the .com bubble, these technology leaders are highly profitable and well-positioned financially to weather missteps if demand proves slower to materialise.
Still, market participants should remain attentive to the evolving interplay of operational demands and accounting treatments. Prolonged depreciation timelines may legitimately reflect improved server efficiencies. Conversely, they may partially obscure the true cost of staying at the cutting edge of AI. Ultimately, whilst we have insufficient domain expertise or insight to accurately call out the ‘correct’ depreciation timeline, we do believe that these depreciation policies are not granular enough. Should hardware cycles tighten again, these adjusted lifespans appear to represent risks of future restatements and write-offs, which could cast doubt on the earnings quality.
The prudent approach is to recognise the inherent risk in these extended schedules, while acknowledging the historical evidence that hyper-competitive technology sectors often see accelerated obsolescence well before an asset’s ‘official’ expiry date. Against this backdrop, heightened vigilance around reported profitability and capex trends is warranted, ensuring stakeholders distinguish between genuine operational strength and what might merely be an artefact of evolving accounting practices.
Within this context, Cerno Capital portfolios continue to reflect our view that ownership of general-purpose AI architectures confers limited economic moat. The commercial value of large language models is most effectively realised when paired with proprietary and large-scale datasets, from which differentiated insights can be derived and monetised.
Illustrative holdings within our portfolios include Experian’s credit scoring systems, RELX’s legal research tools, IQVIA’s healthcare-grade AI platforms, and Dassault Systèmes’ digital design software. Each of these businesses is highly leveraged to the evolution of AI technologies and positioned to benefit from accelerated value delivery to end clients – achieved with materially lower capital intensity and without the financial drag of a rapidly depreciating asset base.
Footnotes:
[i] https://aiindex.stanford.edu/wp-content/uploads/2023/04/HAI_AI-Index-Report_2023.pdf
[ii] https://blogs.nvidia.com/blog/accelerated-ai-energy-efficiency/
[iii]https://www.youtube.com/watch?v=cEg8cOx7USk&ab_channel=StanfordInstituteforEconomicPolicyResearch%28SIEPR%29
[iv] https://techcrunch.com/2025/01/07/nvidia-ceo-says-his-ai-chips-are-improving-faster-than-moores-law/
[v] Microsoft: Q1 2025 earnings call
[vi] Amazon: Q3 2024 earnings call
[vii] Amazon: Q3 2024 earnings call
[viii] Amazon: Q4 2019 earnings call
[ix] Amazon: Q4 2019 earnings call
[x] Microsoft: Q1 2021 Earnings Call
[xi] Alphabet: Q1 2021 Earnings Call
[xii] Meta: Q4 2021 Earnings Call
[xiii] Microsoft: Q4 2022 Earnings Call
[xiv] Alphabet: Q4 2022 Earnings Call
[xv] Meta: Q2 2022 Earnings Call
[xvi] Microsoft: Q4 2022 earnings call
[xvii] Alphabet (Google): Q3 2023 earnings call
[xviii] Amazon: Q4 2021 earnings call
[xix] Microsoft: Q2 2025 earnings call
[xx] Price plummets 70%: How the AI computing power rental bubble burst? – ChainCatcher
[xxi] Datacenter GPU service life can be surprisingly short – only one to three years is expected according to unnamed Google architect | Tom’s Hardware
[xxii] Amazon: Q4 2024 earnings call
[xxiii] Microsoft: Q2 FY2025 earnings call
[xxiv] Alphabet: Q4 FY2024 earnings call
[xxv] Meta: Q4 2024 earnings call
[xxvi] Microsoft, Q2 2025 earnings call
[xxvii] Amazon: Q4 2024 Earnings call
Taken from the Cerno Global Leaders Investment Report Q1 2025. For a copy of the full report, please contact [email protected]
Disclaimer
This document is issued by CERNO CAPITAL PARTNERS LLP and is for private circulation only. The information contained in this document is strictly confidential and does not constitute an offer to sell or the solicitation of any offer to buy any securities and or derivatives and may not be reproduced, distributed or published by any recipient for any purpose without the prior written consent of CERNO CAPITAL PARTNERS LLP. The value of investments and any income generated may go down as well as up and is not guaranteed. You may not get back the amount originally invested. Past performance is not necessarily a guide to future performance. Changes in exchange rates may have an adverse effect on the value, price or income of investments. There are also additional risks associated with investments in emerging or developing markets. The information and opinions contained in this document are for background purposes only, and do not purport to be full or complete. Nor does this document constitute investment advice. No representation, warranty, or undertaking, express or limited, is given as to the accuracy or completeness of the information or opinions contained in this document by CERNO CAPITAL PARTNERS LLP, its partners or employees and no liability is accepted by such persons for the accuracy or completeness of any such information or opinion. As such, no reliance may be placed for any purpose on the information and opinions contained in this document.