AI Models AI Capex

Bernie interviewed Claude recently (https://www.youtube.com/watch?v=h3AtWdeu_G0). The one-shotted responses were predictably a little sycophantic, but some responses were a little surprising. Made me think I'm kinda doing something similar in a more permissive timeframe, I have a harness over Claude that "persistently" studies AI spending, it does deep research, responds to my inputs, persists notes, writes code, etc.

It's not really "introspection" or "Cogito, ergo sum", but I found that over tens of millions of tokens of iterations the system was able to create something novel. I am showing off one of the dashboards it created based on the deep research and Python modelling conducted:

https://ai-capex-sens-wxc4b.ondigitalocean.app/

---

TLDR; AI hype is one thing, but how else would you actually go pedal to the metal spending the GDP of a middle income country on new AI data centers? Basically, theres hundreds of billions of dollars in cloud business from enterprise contracts, boring IT stuff but its where a lot of the money in tech comes from and its real. Beyond the pontification about superintelligence is a similar reality for AI in the eyes of the hyperscalers that scaled the cloud business in the past, hundreds of billions to trillions in business from enterprise contracts (cloud replaced on-prem servers, AI replaces or augments knowledge workers).

The model estimates AI revenue from automatable knowledge-work spend, stress it with adoption/pricing assumptions, and run it against capex + operating + financing costs under delays/shocks. In the base case, it doesn’t break even by 2032, but of course there's other sources of revenue so this isn't all that doomer or surprising that its risky to spend so much. However, its nice to have numbers and sensitivity to back that conclusion up instead of hand waving it.

There's a more finance-y writeup: https://substack.com/home/post/p-190893117 (also researched and written by AI).

---

There's been a lot of AI mania over vibe coding, but the main conclusion I got from running this harness is we can in fact use AI harnesses to simply think and make better decisions too. In this case, the AI is reducing the burden of injecting rigor into writing and decision making.