AI, Actually podcast

Token Management, Spend, and How to Actually Bring AI Costs Down

25.8.2026
0:00
49:05
Spol 15 sekunder tilbage
Spol 15 sekunder frem

Managing AI costs has become increasingly difficult. Uber’s CTO made headlines when he shared the company had burned through its entire 2026 AI budget in just 4 months. That's the problem this episode tackles head on. Token prices are falling for a given level of intelligence, but most companies' actual AI spend keeps climbing, because the bar for "good enough" intelligence keeps moving too. Nobody has a clean way to forecast it yet.

In this episode of AI, Actually, Jim Johnson sits down with Shanti Greene, Stew Chisam, and first-time guest Jake Barger to break down what's actually driving enterprise AI costs and what to do about it. The conversation moves from the economics of model pricing into a real case study: a 33x cost reduction on a production agent, achieved in about two weeks.

You'll get insight into:

  • Why token prices dropping doesn't mean your AI bill is dropping
  • How model routing, caching, and context management quietly drive cost
  • The step-by-step process behind a 33x cost reduction on a live enterprise agent
  • When open weight and self-hosted models actually make sense, and when they don't
  • Practical takeaways for building a repeatable cost optimization process


Follow the Gang:


Chapters
00:00     Introduction and panel overview
02:44     Token Economics: Why Costs Stay High
07:34     Model Routing, Caching & Task Fit
12:22     Continuous Optimization for AI Agents
16:51     The 33x Cost-Reduction Case Study
23:36     Context Management & Smart Escalation
29:28     Open-Weight and Self-Hosted Models
37:22     Enterprise Reliability and Deployment Tradeoffs
44:26     Closing Takeaways: Managing AI Spend


#TokenManagement #AICostOptimization #LLMTokenSpend #ModelRouting #AgenticAICosts #OpenWeightModels #SelfHostedLLM #KVCacheOptimization #EnterpriseAIROI #AIInfrastructure

Flere episoder fra "AI, Actually"