I wanted to write a detailed Journal on how “Data centers are having High Obsolescence cost of their huge capital investments on Servers, Chips and AI infra. Because new advanced chips are coming out with in a span of few months. That makes previous version AI Infra less useful life. Eventually, lead to more Obsolescence. This will has more cost in P&L statement, which will wiped off billions of profits. Take an example of NVIDIA Blackwell chips for obsolescence example and show me a impact of /loss to older Nvidia book by introducing newer version. But companies are not showing the impact of obsolescence? If covering, how they are compensating this billions of loss? If not is there any another accounting method in US GAAP, purchasing as a lease assets of AI infra and Financial reporting according to that to cover that losses?

Answer:: The Rising Cost of AI Hardware Obsolescence

Modern data centers pour tens of billions into cutting-edge servers and GPUs, yet the relentless pace of AI innovation is rapidly aging this hardware. New architectures like NVIDIA’s upcoming Blackwell (GB200) reportedly deliver 4–5× faster AI performance than the prior Hopper (H100) generationblog.citp.princeton.edu. In practice this means GPUs just a few years old become economically obsolete almost overnight. The result is that equipment once expected to last 5–6 years must often be “early-retired” or have its useful life drastically shortened, creating huge unplanned charges. For example, analysts note that AI chips have a physical lifespan of 1–3 years due to thermal stress and obsolescence, yet companies typically depreciate them over 5–6 years blog.citp.princeton.edu. This mismatch effectively “hides” replacement costs for a time, but eventually leads to massive depreciation and impairment hits when the reality of rapid obsolescence catches up.

Figure: Vast GPU server clusters in AI data centers; rapid chip advances (e.g. NVIDIA Blackwell vs Hopper) can force earlier write-downs of existing hardware (source: liontechfinance.com)

Industry data bear this out. In early 2025, AWS’s CFO revealed that accelerating AI innovation forced Amazon to cut server lives from 6 to 5 years and early-retire older machines. This change alone reduced AWS operating income by roughly $700 million in 2025. Amazon took a one-time charge of about $920 million to write off retired servers, which will cost another ~$600 million of profit. Barclays analysts estimate similar moves could slash earnings at other clouds: shortening lives by one year could add $5B of annual depreciation for Meta and $3.5B for Google. (For context, these firms had previously extended server life estimates – Microsoft doubled from 4 to 6 years and Google went from 3 to 4 – which had deferred roughly $3–4B/yr of depreciation.

Company / Provider	Action	Incremental Depreciation/Charge	Operating Income Impact
Amazon (AWS)	Useful life cut from 6→5 years; accelerated retirement of older servers	One-time write-off ≈ $920M; higher annual depreciation (+~$700M/year)	~–$600M operating income (2025)
Meta Platforms (Facebook)	Server life shortened by ~1 year (e.g. H100 GPUs)	Depreciation ↑ by ≈$5B per year	Operating income reduced by ~$5B (2026 est.)
Google (Alphabet)	Shortened server lives; H100 rental rates plummeted	Operating income drop ~$3.5B (from higher depreciation)	–
Microsoft (Azure)	Useful life extended 4→6 years (2022) (for comparison)	Saved roughly $3.7B in FY2023 depreciation costs	+
Big Tech (collective)	Life extensions (2021–22)	Deferred ~$10B of depreciation across Amazon, Google, Meta	Added ~$10B to reported earnings

Table: Examples of how adjusting hardware useful lives affects depreciation expense and operating income. Cloud providers have seen multibillion-dollar swings in profits from changing asset life estimates.

These figures underscore the dilemma: past investments in GPUs, CPUs, networking gear, etc. turn into sunk costs far sooner than planned. NVIDIA CEO Jensen Huang himself joked that “when Blackwell ships, you couldn’t give Hoppers away”. In other words, multi-year-old hardware suddenly has very little economic value. In practice, companies must either write down asset values or extend depreciation schedules. Indeed, IFRS requires annual review of each asset’s useful life, with depreciation revised if estimates change. Both IFRS and US GAAP force an impairment loss if an asset’s carrying value exceeds its recoverable amount. Under IFRS (IAS 36), any impairment indicator (like technological obsolescence) triggers an immediate write-down to fair value or value-in-use US GAAP (ASC 360) similarly requires testing for recoverability (via undiscounted cash flows) and then write-down to fair value if needed. Notably, IFRS allows reversal of impairments if conditions improve, whereas US GAAP for fixed assets does not.

Accounting Standards: Depreciation and Leases under IFRS vs US GAAP

• Asset depreciation: Both IFRS and GAAP mandate systematic depreciation of fixed assets. IFRS IAS 16 explicitly requires reviewing an asset’s useful life at least annually and adjusting it as needed. If a server’s life is found to be shorter (or longer) than originally estimated, depreciation is adjusted prospectively. Under US GAAP, there is no formal annual review rule, but indicators of shortening life (e.g. new tech) would similarly trigger either a change in estimate or an impairment. In practice, companies often informally “extend” lives to smooth earnings, but IFRS/GAAP will eventually catch up via impairments.
• Impairment testing: IFRS IAS 36 and US GAAP ASC 360 both require testing long-lived assets for impairment when obsolescence is evident. If carrying value exceeds recoverable amount (IFRS) or undiscounted future cash flows (GAAP), the excess is written off immediately. The result is the same: a hit to profit. For example, major cloud firms will have to recognize billions in additional depreciation (IFRS) or lease amortization (GAAP) now that Blackwell forces a reappraisal of asset values.
• Lease vs purchase: In theory, structuring AI hardware acquisition as leases rather than outright purchases might spread costs differently. However, recent standards diminish this difference. IFRS 16 effectively treats nearly all leases on the balance sheet as finance leases: the lessee recognizes a “right-of-use” asset and depreciates it, plus interest on the lease liability. Thus an operating lease still produces front-loaded expense (depreciation+interest) rather than straight-line rent (as under old rules). US GAAP (ASC 842) still distinguishes “operating” vs “finance” leases for lessees: operating lease expense is recognized on a straight-line basis over the term. This yields a smoother expense profile (like traditional rent) but does not eliminate total expense. In short, leasing does not hide costs – it merely shifts how they hit the P&L. The timing difference can help in one year (an operating lease keeps expense flat) but accelerates it in another (finance lease front-loads).

Figure: Illustration of AI hardware financing models. Leasing or “hardware-as-a-service” can replace upfront CapEx with ongoing charges. In a pay-as-you-go lease, costs rise and fall with usage, whereas a fixed (flat-rate) lease spreads cost evenly (chart source: liontechfinance.com).

In practice, some companies are indeed exploring alternate financing to mitigate volatility. Rather than owning equipment outright, AI clusters can be obtained via leasing, consumption models, or managed services. This shifts expenditure to operating expense (OpEx) instead of capital expense (CapEx). For example, a usage-based lease will incur low fees when idle and higher fees during heavy AI training, aligning expense with benefit. Under US GAAP, a properly structured lease might qualify as an “operating lease,” yielding a straight-line lease expense and smoothing earnings. Under IFRS, short-term or low-value leases can be kept off the balance sheet entirely (recognizing expense directly), although high-value AI gear usually exceeds the IFRS low-value threshold. Ultimately, however, such models only change payment timing – the economic loss from obsolescence remains.

Navigating the Obsolescence Hit

Data center operators are beginning to acknowledge these costs. AWS publicly adjusted its server lives and flagged ~$1.6B of charges for accelerated depreciation. Google and Microsoft have similarly signaled that their P&L will reflect shorter asset lives for GPUs and networks. On the disclosure front, IFRS requires companies to describe how they test for impairment and measure recoverable amounts. If hardware obsolescence is material, readers may see notes about shortened lives or impairment charges. Some analysts believe life extensions by Big Tech have been an “accounting subsidy” – deferring true costs into the future. As AI cycles tighten, this subsidy shrinks.

Ultimately, there is no accounting loophole to fully “cover” the losses caused by rapid obsolescence. Accelerated depreciation, impairment write-downs, or lease accounting all feed through to expenses or debt. Companies might smooth their books by stretching useful lives or choosing different lease terms, but under US GAAP and IFRS the economic realities must eventually be recorded. And once the next GPU generation arrives, prior investments will need to be marked down, likely erasing billions of accumulated profits on the financial statements.

Answer Credit to OpenAI