Stratechery Compares Nvidia's AI Boom to 1873 Railroad Bond Crash as Hyperscaler Debt Surges

Nvidia's position depends on a bunch of things staying true simultaneously: LLM dominance, no China full-stack alternative, and no real robotics competition. They're already pushing into robotics as a backup, but China will build its own stack regardless, and that puts Europe in an awkward spot of choosing Nvidia for lock-in or for security.

Even in the West, Nvidia's grip is loosening. I'm seeing way more posts about people running big models on AMD hardware now. AMD's software story is still rough compared to CUDA though, so patching vllm for one or two models is doable but it's not a full replacement yet.

If you want paid deep analysis instead of free narrative pieces, look at SemiAnalysis, they've got free and paid tiers aimed mostly at investors.

This whole thing is the stock market phase of Universal Paperclips lol.

The classic investment trap: everyone gets the first order assumption right (compute demand grows) but blows the second order assumption (it grows every single year forever). Demand can persist without increasing yearly, and that alone could turn these bonds into a huge burden for Nvidia.

Compute needed for the same output quality has dropped roughly 90% every 18 months for about 5 years. So you could see >100x more LLM inference in 5 years but only 2x more chip demand, or you could hit some emergent capability wall that suddenly spikes demand. Nobody actually knows which way it breaks.

Worth noting AMD just bought Taalas and works with Cerebras a lot, ASIC vendors are apparently outgunning Nvidia's chips by an order of magnitude in some workloads.

Each hyperscaler dropped something like $250B on infra in the last year, which means they need to be generating around $20B a year in AI profit just to cover the cost of that cash. We're nowhere near that, cash flows are drying up and everyone's taking on debt to cover capex. Google is cash flow negative for the first time in its public history, Amazon too.

Nothing goes up and to the right forever. Building a business model on 'this time is different' always ends up finding a storm eventually.

Focusing on hyperscalers is kind of missing the point. There are still tons of small companies and individuals just starting with AI, and that customer base and revenue is largely untapped, we're still early on adoption even if tech people feel AI is 'boring' now.

Codex just hit 10 million users, which sounds big until you remember Microsoft Office has a billion users. We're nowhere close to market saturation, and even at the leading edge the models are still slow and need a lot of hand holding.

Ben's wrong here, the compute demand backlog is basically mythical and will collapse: it's driven by circular investment between the same players, and there's too much capital chasing returns that can't realistically be hit given the barriers involved.

Sure, but the real question isn't whether there's a correction coming, it's when. Could be 2 months, could be 20 years, nobody including Ben actually knows the timing.