The Huge Short

The AI Bubble is now 17 times bigger than the dot-com bubble and four times bigger than the subprime mortgage crisis and Michael Burry has just put 80% of his chips on red.

For those of you trying to hedge, let me know your strategy because NVDA value is larger than 6 INDUSTRIES in the S&P 500 (pictured below). From a practitioner's standpoint, this picture doesn't even show the whole story. As I see company after company integrate with close-source AI models made by US companies up to their eyeballs in investor funding. We're witnessing an unprecedented level of dependency. Indeed, healthcare, finance, logistics, education are all creating mission critical workflows with APIs that could vanish the moment the roulette wheel stops.

The optimists will tell you the infrastructure isn't wasted and that all the new data centres and GPUs can pivot to something else. However, they're missing the forest for the trees. Blackwell chips aren't Swiss Army knives; they're purpose-built for transformer architectures and the specific computational patterns of current AI models. More critically, the vast majority of that compute capacity is burning through training runs, not inference. When the money dries up these companies won’t have the cash flow to keep the lights on, let alone pivot to some mythical Plan B.

Burry's recent move has sent a ripple through the market, with AI shares taking their first meaningful hit in months. But if history teaches us anything, it's that euphoria has a short memory. The hype train appears to be merely pausing at a signal light, steam still billowing, ready to barrel forward at mere mention of Grok-5 in a tweet from Elon. But I caveat this prediction; no one can truly predict the market. Something Burry well knows is that the market has a perverse talent for staying irrational longer than most can stay solvent.

One thing that keeps me up at night (for real): when this bubble bursts, we won't have a safety net. If open-source models (not the Chinese) haven't caught up to the long context state-of-the-art closed source ones, we're looking at an extinction-level event for a large number of companies. There will be no substitute, no fallback, no gradual transition. One day you'll have your business, the next day you'll get a 503 error that never resolves. This is why my strategy, and the one I'm putting into practise, is simple: build everything assuming the closed-source rug will be pulled tomorrow. Use these tools but architect for their absence.

At least when the dust settles, all those laid off for AI may have jobs again. If not, perhaps the vacant data centres could be used to host a VR version of GTA 6.

10 November 2025