Full Transcript
Loading transcript…
AI Bubble Burst? Companies are seeing the problems with AI!
Watch with subtitles, summary & AI chat
Add the free Subkun extension — works directly on YouTube.
- Watch
- Subtitles
- Summary
- Ask AI
Business leaders, entrepreneurs, and tech workers managing AI budgets.
TL;DR
Companies are facing massive AI costs due to 'token maxing,' where employees overuse expensive models on simple tasks. The video explains five methods to cut AI bills by up to 90%, including model routing and caching, and highlights business opportunities from this crisis.
Key Takeaways
- Token maxing describes the phenomenon where companies incentivize heavy AI usage, only to face skyrocketing costs as agents consume 5-30 times more tokens per task than simple chatbots.
- Per developer token consumption rose 18-fold in nine months not due to price hikes but because agentic AI loops reread files and spawn sub-agents, multiplying token volume per person.
- Rewarding token volume as a productivity metric is flawed; it measures effort, not results, akin to paying a taxi driver more for taking the longest route.
- The big model fallacy—using the most expensive AI for every task—is the single costliest architectural mistake; switching defaults to cheaper models for simple tasks can slash bills.
- Model routing automatically directs easy queries to cheap models and complex ones to frontier models, cutting costs by up to 80% without users noticing any difference in quality.
- Caching avoids paying twice for identical work by storing and instantly returning responses to repeated queries, dramatically reducing costs for high-volume, repetitive workloads.
- Keeping conversation context lean—starting fresh chats, attaching only relevant files, and using projects—prevents paying to reread entire histories for trivial follow-up questions.
- Owning hardware for routine tasks and using open-source models locally can eliminate per-token costs for predictable workloads, offering a path to near-zero marginal AI expense.
In This Video
- 00:00The Token Maxing Crisis
Meta's leaderboard and Uber's budget blowout reveal the AI spending crisis.
- 01:25What Is Token Maxing?
Token maxing is the wasteful overuse of expensive AI models.
- 01:49Why Costs Exploded
Frontier models got expensive and agents burn 5-30x more tokens.
- 03:24The Incentive Problem
Companies rewarded token volume, not results, fueling waste.
- 04:33Five Methods to Cut Costs
Change defaults, route models, cache, lean context, own hardware.
- 06:26Caching and Context Tricks
Cache responses and keep chat history lean to slash bills.
- 08:09Opportunities in the Chaos
The crisis creates valuable skills and new business opportunities.
Questions & Answers
What is token maxing?
Token maxing is when employees use excessive AI tokens, often encouraged by leaderboards, leading to huge bills. It's like paying a taxi driver for taking the longest route.
Why are AI costs exploding in 2026?
Two reasons: frontier models got expensive (up to $50 per million tokens) and AI agents burn 5-30x more tokens per task than chatbots, causing per-developer token use to rise 18x in nine months.
How to reduce AI costs by 90%?
Five methods: change default model, model routing, caching, keep context lean, and run small models on your own hardware. Stacked, they can cut bills by up to 90% without losing quality.
What is model routing for AI?
Model routing automatically sends each request to the cheapest suitable model, like a hospital triage. 80% of requests are easy and don't need expensive models, slashing costs.
What is token maxing?
Token maxing is when employees use excessive AI tokens, often encouraged by leaderboards, leading to massive bills. It's like paying a taxi driver for taking the longest route.
How to cut AI costs without losing quality?
Use five methods: change default model, model routing, caching, keep context lean, and own hardware for boring tasks. Stacked, they can cut bills by up to 90%.
Key Terms
- token maxing — The practice of using excessive AI tokens, often encouraged by leaderboards, leading to skyrocketing costs.
- model routing — A system that automatically sends each AI request to the cheapest suitable model, reducing costs by up to 80%.
- caching — Storing and reusing previous AI responses to avoid paying for repeated identical requests.
- big model fallacy — The mistaken belief that the most expensive AI model is always needed, when cheaper models suffice for many tasks.