OpenAI is going to lose 90% of its revenue | Eli the Computer Guy
Summary
Host Isaac and guest Eli the Computer Guy dissect the AI tech stack, explaining how reasoning models inflate revenue while orchestration layers route tasks to efficient local models. They argue that as specialized, smaller models become viable, frontier labs will lose 90% of their request volume, shifting their role to cloud infrastructure providers. The discussion also covers the security negligence of major AI firms, highlighting missing log files and OAuth vulnerabilities, and advises using token limits to prevent costly agent infinite loops.
Topics discussed
Sponsor: Perk for travel and spend management
Concerns over AI company revenue and architecture
Introduction and basics of AI reasoning
Pattern matching vs reasoning models and token billing
The AI stack and the orchestration layer
How orchestration layers impact AI company revenue
Routing requests to local and enterprise models
Mixture of Experts (MoE) vs orchestration layers
Value of niche, specialized AI models
Running AI on average systems and the 'I don't know' problem
Specificity of training data and model reliability
Compute resources and the value of hosted infrastructure
Frontier models as smartphones and business viability
Evaluating AI: Tech stack vs business and NoSQL analogy
The AI revolution and the need for constant explanation
Sponsors: ActiveCampaign and Granger
AI integration into products and user experience
Agentic AI: Loops, autonomy, and token consumption
AI alignment, guardrails, and secure containment
Sponsors: Perk and Kqu Business Phone
Security failures: Anthropic and OpenAI log access issues
OAuth vulnerabilities and ethical hacking of AI labs
Agent containment, infinite loops, and sysadmin skills
Managing token usage and preventing infinite loops
Vibe coding, whiteboards, and the role of engineers
Conditional statements and coding best practices
Sponsor: Perk for travel and spend management
Listen ad-free on Castria