
The AI buffet is getting expensive
Chamath Palihapitiya took one look at rising LLM token costs and said the quiet part out loud: AI demand is running hotter than the infrastructure built to support it. If you’re paying more to run models, that usually means the system is getting stretched — kind of like trying to stream four 4K movies on a Wi-Fi router from 2012.
The stat bouncing around here is ugly in a very useful way: average token costs have climbed to $2.12 per million tokens, up 12% in a week and 65% since the end of February. That’s not just a nerdy pricing update. It’s a signal that inference demand may be outrunning available compute, and markets tend to notice when scarcity shows up wearing an AI badge.
Who gets paid when the pipes clog?
If compute is the bottleneck, investors usually start circling the companies that can relieve the pressure:
- Nvidia and AMD if customers keep buying accelerators like they’re limited-edition sneakers.
- Broadcom and custom-chip players if the market keeps leaning into specialized silicon and networking gear.
- CoreWeave, Alphabet, and Oracle if scarce GPU cloud capacity keeps commanding premium pricing.
- TSMC if advanced-node manufacturing and CoWoS packaging stay red-hot.
- Micron if high-bandwidth memory remains the AI version of bottled water in a blackout.
That’s the odd twist: when AI gets more popular, the real winner may be the stuff hidden under the hood — the chips, memory, cloud racks, and packaging that make the whole circus run.
Big picture
This isn’t a clean “AI is good” or “AI is bad” headline. It’s more like: the demand curve is still screaming higher, but the infrastructure bill is coming due. If token costs keep rising, the market may have to stop treating AI as a software story and start pricing it like an industrial supply chain.
