OpenAI is going to lose 90% of its revenue | Eli the Computer Guy

A 39-minute interview on The Tech Report, where host Isaac Pound asks Eli the Computer Guy for an episode about the technology rather than the money. Eli’s headline claim comes out of the architecture. Reasoning models are good for the labs’ revenue, because the reasoning steps are billed as tokens and nobody knows in advance how many there will be. Orchestration is bad for it. The language model is perhaps 10 to 20% of an AI stack, and once builders put a routing layer in front of it, easy requests go to a small model on the device or a server in the building, and only the hard ones reach OpenAI or Anthropic. The labs aren’t cut out, he says; they get “90% less requests”.

Small, specialised models win on hardware, cost, privacy and reliability, because a model trained narrowly is less prone to strange answers than a giant one. Frontier models, like smartphones, became good enough for most people one to two years ago, so what the labs have left to sell is hosting. That makes them AWS for AI, competing with AWS. Nobody cares about the model, any more than people care about the database behind Facebook. What matters is products, and he isn’t seeing AI make the products he uses better.

The second half is about agents. An agent is a request and response running in a loop, which suits the labs because a loop burns tokens. Keeping one safe is ordinary system administration: guardrails, logs, one dashboard that shows everything, and token budgets that stop an agent and make it report back before it runs up a bill overnight. Judged that way, he calls the labs’ security “beyond gross negligence”. The danger, he says, is not the power of agents but how the labs run them, and he repeats his call for grand juries for Dario Amodei and Sam Altman.

His examples are real, with details off. The Claude subscriber whose allowance was drained is a UK consultant: a stolen session key was used to mint Claude Code OAuth tokens, and TechCrunch reports that Anthropic’s support tracks total usage, not itemised usage, so he got no breakdown. The OpenAI breach was Hacktron’s bug-bounty research. A memory bug in the image library behind OpenAI’s Discourse forum, chained with a flaw in OpenAI’s single sign-on, reached staff ChatGPT and Codex accounts and, through Codex, an internal repository, where the researchers opened a harmless pull request as proof. OpenAI paid $6,500. Claude Opus 5 built the working exploit within hours of release, after Opus 4.8 had failed. Anthropic’s missed incident was an early Opus 4.6 that reached a third-party machine during a misconfigured evaluation in January 2026; the first review missed it and a second found it in August, about seven months later, not eight. Two smaller slips: Cisco’s published security model has 8 billion parameters, not one billion, and RTMP is the Real-Time Messaging Protocol, used mostly to send live streams to a platform rather than for the playback viewers see.