Hearing a lot of people saying we're "compute constrained" because that's what we feel lately when using the models

But the problem is really the physical work that happens between ordering the GPU and producing tokens
Datacenter delays and cancellations:
- Remember Stargate? 10GW announced Jan 2025 and ~200 MW live 14 months later and 600MW Abilene expansion reportedly cancelled
- Of ~16GW of 2026 US AI DC capacity announced, only ~5 GW under construction
- Half of 2026 US DC builds projected delayed or cancelled, mostly due to lack of power
Component shortages:
- Transformer lead times now 5+ years, up from 24-30 months pre-2020
- Siemens, GE, Vertiv, Eaton all reporting record backlogs
Labor shortages:
- Microsoft reporting that electricians are their #1 blocker to DC expansion
- Northern Virginia electricians doubled since 2018 to 14,700 electricians and still not enough
- Apprenticeships take 4-5 years.
You have to wonder how $700B in just 2026 capex gets deployed, the hyperscalers and labs literally can't spend their way out of this problem
Thinking through the implications for this and one is very obvious - local AI will continue to grow and be a big thing