The compute crisis, visualized
- https://www.youtube.com/watch?v=jB2iKoBSPyo
- Original title: Elon won after all
Microsoft, Google and Anthropic are all openly capacity-constrained — they literally cannot make more money because they lack compute. Theo uses the bizarre headlines (Anthropic paying SpaceX ~$1B/month, Google ~$920M/month for spare xAI compute) to walk through the entire supply chain bottleneck behind the AI buildout, and why SpaceX/OpenAI ended up the unlikely winners.
The argument is that compute is gated by a chain of bottlenecks where any single one acts as a filter on the whole:
- TSMC fabrication — the only company that can fab leading silicon; Apple's long-term deals keep it insulated while Nvidia's slice keeps growing. New fab capacity takes 8–10 years to stand up, so allocation is sold years in advance.
- High-bandwidth memory — only Micron, Samsung, SK Hynix make it; demand for GPU RAM is so high they've reallocated away from consumer. Micron is even shutting down its Crucial consumer brand. AI needs huge RAM (even smaller models need 100GB+), so prices have skyrocketed.
- Hard drives — Western Digital sold out all of 2026 by February; 16TB drives that were ~$170 refurb in 2024 now run $360+. No drives → no point buying GPUs.
- Power — US commercial/industrial electricity use now exceeds residential; the grid isn't growing fast enough, and China adds more capacity per year than the US has since the '90s.
Why not just make more? Because every layer must scale together — 10x'ing TSMC just shifts the constraint to HBM, then drives, then power, and betting wrong on one layer wastes billions. SpaceX/xAI over-bought compute early (Colossus cost ~$3–4B, now rents for ~$1B/month, paying for itself in months) and Elon's Tesla battery/generator deals solved power; OpenAI similarly bet early on scaling laws — which is why their rate limits are more generous than Anthropic's. Anthropic and Google were conservative and are now paying competitors. The real winner is Nvidia, which can't manufacture enough to meet demand. Theo's practical takeaway: RAM, SSDs, GPUs and drives are not getting cheaper any time soon.