Unlocking the Future of AI: Trends, Infrastructure, and Tokenomics
summary
Claims that inference costs fall 100x over two years while GPU demand rises are backed by concrete numbers: DeepSeek causes inference traffic to double quarter‑over‑quarter with another 50% uptick, Cursor hits ~$200M ARR, and ARC AGI1 costs >$1M for ~16 hours of compute. Technical fixes include disaggregated prefill, moving KV cache to storage to exploit a 5–15 minute cache window, and NVIDIA’s Leptton acquisition to enable an open GPU‑hour marketplace; elasticity and assembly‑line inference deliver real cost/latency tradeoffs that demonstrably work.
AI-written summary · about
chapters (14)
- 0:00The state of AI economics
- 4:38The trilemma of AI inference
- 7:07Latency and architectural trade-offs
- 8:57DeepSeek's operational lessons
- 11:02Marketplace and performance metrics
- 13:01The AI token factory economics
- 16:11Efficiency through user density
- 17:50Optimizing the inference triad
- 19:38Context length and memory constraints
- 21:32Long context economic impact
- 22:15Jevons' paradox in AI
- 23:20Disaggregated serving innovations
- 24:59Architecting for profitability
- 27:28Future of token pricing
related talks
22mCurrent State of AI
28mThe AI Tipping Point
20mOperating Frontier Intelligence at Scale
40mFueling the AI Revolution with Gaming
24mClosing Keynote
20mFireworks AI: Go from Hype to High - Value AI
06mAI Inference at the Edge: Beyond the Hype
24mFIRESIDE CHAT: The Future of AI Infra: The Opportunities and Challenges That Await Us On Our Journey