Hyperscalar capital spending for 2026 is going to be in the close neighborhood of $700B, which is over 2% of US GDP. That is about 3x the GDP spend of peak Apollo program in the 1960s, and about the same as the telecom/fiber buildout of the late 90s and the railroad buildout of the 19th century (both followed by a collapse). And there just isn't that much revenue coming into the system, and there aren't the productivity gains coming out of it. When 95% of corporate AI initiatives are still failing, the value proposition isn't there. And if you try to look at something like Microsoft's reported $37B in AI revenue a lot of that is really internal spend from leasing compute to OpenAI, which it partially owns. The real revenue coming into the AI industry is likely well under $100B this year, and the productivity gains to end consumers is likely much less. So if you think a few $10B/yr here or there is "serious dough", it just isn't enough to fill the gap. And OpenAI should burn through $14B this year, up by a factor of 3x over last year. Anthropic has a projected revenue for 2026 of $26B and is running around cash flow neutral, but that doesn't approach the $700B spending gap. And that is with accounting that depreciates GPUs on 5-6 year schedule instead of the more realistic 2-3 year schedule--so Anthropic may kick the can down the road a bit, but in 2-3 years they'll still be depreciating GPUs that they're throwing away and having to replace (of course this may be WHY Anthropic is leasing compute from xAI since then that accounting hit falls on xAI instead of Anthropic).
In 10 years, we probably will have $700B/yr in productivity gains and revenue from LLMs, but we're not going to be able to sustain $700B/yr in capital spending until we get there. And the problem is much worse than the fiber buildout of the late 90s. Fiber built out in 1998 was still usable 10 years later. The GPUs that are being built out today are going to be obsolete trash in 3 years.