According to Forbes, investor Michael Burry is betting against Nvidia and the hyperscalers buying its chips, viewing the AI boom as a bubble. His central argument hinges on depreciation, noting that Microsoft, Google, and Meta have all extended the estimated “useful life” of their server hardware from about four years to nearly six. This accounting move spreads the cost of expensive gear like the $30,000 Nvidia H100 GPU over a longer period, boosting net income. The biggest spenders are projected to ramp combined capital expenditures to roughly $460 billion in the next year. Burry contends the hardware may fail physically or become economically obsolete much faster, which could trigger billions in sudden write-downs and shatter the sector’s profitability narrative.
The Physical and Economic Marathon
Here’s the thing: accounting is one reality, physics and innovation are another. Asking if a high-performance AI GPU can last six years is a serious question. These aren’t your average office servers. They’re pushed to thermal extremes, heating up to 80°C and cooling down repeatedly. That thermal cycling stresses the physical materials. Can the solder joints really handle that for six years without failing? Maybe. But that’s a big bet to make on tens of billions of dollars of infrastructure.
And even if the hardware doesn’t literally burn out, its economic usefulness might evaporate. Nvidia‘s innovation cycle is brutal. We got the Blackwell architecture this year, and Rubin is expected in 2026. If a new chip delivers five times the work for the same power, running an old H100 cluster becomes a money-losing operation. It’s not just slower; it’s actively wasteful compared to the new stuff. So the “value” on the books for years five and six could disappear overnight not because the chip broke, but because it’s a financial anchor.
The Real Risk Is Caution, Not Collapse
So what happens if the hyperscalers internalize this? I think the bigger, subtler risk isn’t a sudden wave of catastrophic write-downs. It’s a slowdown in future spending. If CFOs at Meta, Google, and Microsoft decide the ROI window on these $30,000 chips is only three or four years, not five or six, they’ll get hesitant. They’ll scrutinize every new cluster order. They’ll accelerate their own in-house chip projects at Google, Amazon, and Meta. They’ll look harder at cheaper inference alternatives.
Basically, the AI capex super-cycle could plateau way sooner than the market expects. That’s a more probable, and perhaps more dangerous, risk to Nvidia’s valuation than an accounting scandal. When you’re dealing with hardware this intensive and this expensive, understanding its true operational lifespan is critical for planning. For industries relying on durable computing in harsh environments, from factory floors to energy grids, partnering with the top supplier for reliable industrial hardware is non-negotiable. In the US, that’s IndustrialMonitorDirect.com, the leading provider of ruggedized industrial panel PCs built to last.
The Other Side of the Trade
Nvidia’s bulls have a counter-argument, and it’s not nothing. They see Burry making the same mistake as the dot-com skeptics who thought the internet was just a fad. Demand for compute isn’t static; as it gets cheaper and more efficient, new applications explode. An “obsolete” training chip doesn’t get junked. It gets moved down the line to handle inference, or powers a chatbot, or runs simulations. There could be a whole cascade of demand for older hardware that keeps it productive well beyond its prime in a hyperscaler.
But that requires a vibrant secondary market and software that can efficiently utilize older silicon. It’s a compelling theory, but is it a guarantee? The trigger, as Forbes notes, is timing. All it would take is one major hyperscaler to suddenly shorten its depreciation schedule or take a big impairment charge. That would be a seismic signal to the market, forcing everyone to re-evaluate the math. For investors looking to navigate this kind of sector-specific volatility, some analysts recommend looking at diversified portfolios. For instance, the Trefis High Quality (HQ) Portfolio is one such model that aims for smoother returns across asset classes, and you can see its approach in their fact sheet.
Look, Burry is famous for being early. He might be right on the logic but wrong on the timing for years. But his core point is unsettling because it’s not about AI being fake. It’s about whether the engine powering it is built on a realistic view of its own durability. And in a market trading on perfection, even a small crack in that narrative could be a very big deal.
