We scan new podcasts and send you the top 5 insights daily.
Nicolai Tangen says a new AI model can take six months to build and may be obsolete in five, making depreciation and model lifespan central to valuing AI companies. He extends the same question to chips: how long can they be used, and how should investors depreciate them?
Historically, infrastructure from tech bubbles (e.g., fiber optic cables) had long-term value for second-wave investors. AI's core infrastructure, GPUs, has a short 2-3 year shelf life, creating a unique and devastating "depreciation bomb" risk for investors caught in the hype cycle.
The relentless pace of new AI models, which perform best on the latest hardware, drastically shortens the effective lifespan of GPUs. This changes the traditional 6-year depreciation model and complicates the financial calculus for building data centers versus renting cloud capacity.
The long-term risk for the AI infrastructure boom is its rapid pace of obsolescence, with replacement cycles estimated at just five years. Companies must generate earnings from current investments quickly enough to fund the next wave of upgrades, or risk being forced to finance functionally obsolete assets.
The hosts challenge the conventional accounting of AI training runs as R&D (OpEx). They propose viewing a trained model as a capital asset (CapEx) with a multi-year lifespan, capable of generating revenue like a profitable mini-company. This re-framing is critical for valuation, as a company could have a long tail of profitable legacy models serving niche user bases.
While the industry standard is a six-year depreciation for data center hardware, analyst Dylan Patel warns this is risky for GPUs. Rapid annual performance gains from new models could render older chips economically useless long before they physically fail.
According to George Hotz, trained AI models are the fastest depreciating assets ever created. A state-of-the-art model that cost $100M to train can be surpassed in months, making its value plummet. This economic reality suggests that withholding models for "safety" also serves to generate hype before its competitive edge disappears.
The useful life of an AI chip isn't a fixed period. It ends only when a new generation offers such a significant performance and efficiency boost that it becomes more economical to replace fully paid-off, older hardware. Slower generational improvements mean longer depreciation cycles.
The AI landscape is uniquely challenging due to the rapid depreciation of both models (new ones top leaderboards weekly) and hardware (Nvidia launched three new SKUs in one year). This creates a constant, complex management burden, justifying the need for platforms that abstract away these choices.
The massive capital expenditure to train a frontier AI model becomes nearly worthless in months as competitors release superior models. This makes trained models a uniquely fast-depreciating asset, creating immense pressure on labs to monetize quickly through API access or investor hype before their technological advantage evaporates completely.
Unlike the railroad or fiber optic booms which created assets with multi-decade utility, today's AI infrastructure investment is in chips with a short useful life. Because they become obsolete quickly due to efficiency gains, they're more like perishable goods ('bananas') than permanent infrastructure, changing the long-term value calculation of this capex cycle.