Let’s talk charts. But not stock charts. Let’s look at the depreciation chart for a $40,000 AI chip.
The big tech firms—the hyperscalers—want you to believe this is a gentle 5- or 7-year slope. This is the story they are selling investors.
The reality? The data? The real chart shows a 24-month cliff.
This disconnect isn’t just an accounting quirk. It’s a massive, hidden risk. And it’s at the center of a major financial debate about whether the profits we’re seeing from the AI boom are even real.
The great disconnect: Accounting vs reality
I’m not the only one spotting this. Michael Burry, the investor famed for “The Big Short,” has been public about this, calling it “one of the more common frauds of the modern era”.
The game is simple: buy billions in hardware that you know has a realistic 2-3 year product cycle, but tell the market you’re going to depreciate it over 5-6 years. What’s the result? You understate your annual expenses, which artificially inflates your earnings and makes your profits look fantastic.
This isn’t just creative accounting; it’s a willful distortion of the asset’s true economic life.
The 24-month “Obsolescence clock”
The market, as always, tells the truth. NVIDIA sets the pace. You don’t even need inside information; just look at their product announcements:
- A100 (Ampere): Q2 2020.
- H100 (Hopper): Q3 2022.
- B100/B200 (Blackwell): Q4 2024 (Est.).
This 24-month “Obsolescence Clock” is the real functional life of a peak AI chip. But it gets worse.
The 2-year functional clock is now converging with the 1-3 year physical clock. Forget 5-7 years; a Google architect has stated on the record that these GPUs, running 24/7 for AI, are physically failing in just one to three years. The 5-7 year accounting schedule is pure fantasy.
The “Great filter”: Why value collapses
Value in this space doesn’t drift down. It falls off a cliff. This is what I call the “Great Filter.”
A new-generation chip isn’t just “faster”; it introduces a qualitative feature that the previous one simply lacks. The H100 (Hopper) chip has a dedicated Transformer Engine and supports the FP8 data format. The A100 does not.
The result? The H100 isn’t 20% faster. For the all-important LLM inference workload, it’s up to 30x faster.
The day the H100 was benchmarked, the A100’s peak value evaporated. It was instantly obsolete for top-tier work. That is the 24-month cliff.
The hyperscaler excuse and the Myth of “$0”
Now, the hyperscalers will argue, “But we still use them!” And they do. They justify the long depreciation schedule with “systematic repurposing”. The chip “cascades down” their internal ecosystem, from S-tier training to standard inference, and finally to running some low-level internal analytic.
That’s a luxury only a mega-cap giant has. For everyone else, the cliff is real.
But “obsolete” does not mean “worthless.” This is key. The chip’s value doesn’t go to $0. It finds a “Legacy Floor.”
- An 8-year-old Tesla V100 (2017) still sells for $3,500-$6,000.
- A 5-year-old A100 80GB (2020) still commands $9,500-$14,000.
It finds a new life, but at a 50-70% discount from its peak.
The investor takeaway: A $176 Billion risk
Here is the takeaway for your portfolio. This accounting trick is how the AI bubble is being inflated.
Burry projects this practice will allow tech giants to overstate earnings by a staggering $176 billion between 2026 and 2028. He’s targeting companies like Meta and Oracle, projecting their 2028 earnings could be inflated by over 20%.
The smart money is catching on. Barclays is already cutting 2025 earnings forecasts for these hyperscalers by as much as 10%, flagging depreciation as the “not-so-hidden” cost of AI.
This AI capital spending boom looks just like the railroad and telecom booms of the past—massive overinvestment, hyped-up profits, and, ultimately, poor shareholder returns.
The real three-stage value lifecycle
Here is the real lifecycle. Forget the accounting.
Stage 1: The “Peak” (Years 0-2): Value is at or above MSRP.
- Stage 2: The “Viable” (Years 2-4): The successor is announced. The chip is instantly, functionally obsolete. Value collapses 50-70% to its new “inference” price.
- Stage 3: The “Legacy” (Years 4+): Depreciation halts. The value stabilizes at a “Legacy Floor” of 10-20% of its original price.
The risk for investors is simple: you are mistaking a 5-7 year accounting fantasy for the 24-month technological reality. And when the bill for all this spending truly hits the books, the earnings you’re counting on won’t be there.
Source: https://www.fxstreet.com/news/the-ai-accounting-fraud-a-176-billion-risk-hiding-in-plain-sight-202511121444
