Chamath Palihapitiya
· Chamath Palihapitiya
· July 19, 2026 at 15:06
· ⏱ 6 min read
| Read on Substack ↗
Summary
Chamath Palihapitiya argues that CEOs must aggressively manage AI token costs by routing tasks to cheaper models and bringing intelligence in-house to protect proprietary data, otherwise bloated AI spend will hit earnings. He highlights that open-weight models are rapidly closing the gap with frontier models on most tasks, threatening the business model of closed-AI providers and shifting the cost burden of the $1.4T infrastructure buildout away from enterprises.
•Token costs at Chamath's own firm doubled every ~45 days while incremental productivity gain was only 5-10%.
•Cheap models are now 80-95% as good as frontier models on most tasks; enterprises should route tasks through a control plane to avoid paying premium for unnecessary capability.
•Moonshot AI's Kimi K3 (2.8T parameters) topped the Frontend Code Arena at $3/$15 per million tokens vs Claude Fable 5's $10/$50, and will release full open weights making it the largest open-weight model.
•Thinking Machines Lab's Inkling (975B MoE, 41B active) under Apache license achieved 84.7% accuracy on a financial document triage task with fine-tuning, vs 78.2% for the best frontier model at ~1/14th the cost.
•Prime Intellect raised $130M Series A at $1B valuation, has $100M ARR from decentralized compute and open-source training software; a case study with Ramp showed their post-trained model beating Opus at spreadsheet search while running 27% faster and cheaper.
•TerraFirma (ex-SpaceX founders) raised $115M to retrofit construction machinery into robots, enabling one operator to manage 3-5 machines and achieving up to 300% productivity improvement.