chamath palihapitiya

Karpathy leading recursive self-improvement at Anthropic (model improves itself during forward pass). Combined with continual learning (human-like experience learning), could enable 10x yearly model quality gains — new Moore's Law. Composer 2.5 proves 3 weeks RL on proprietary data + massive compute yields Pareto-dominant model.

Anthropic's fear-based messaging has catalyzed bipartisan political opposition to data centers (Abbott, Shapiro) just as rising Treasury yields increase cost of capital; frontier labs with non-investment-grade balance sheets cannot fund required compute, creating a self-inflicted existential funding gap.

Governors Abbott (TX) and Shapiro (PA) restricting data centers reflects constituent rage at unaffordability and tech oligarch wealth, not genuine safety concerns; the 'guillotine' metaphor signals political risk to AI capex if industry doesn't address job displacement fears and wealth distribution.

Anthropic and OpenAI growth is 100% supply-constrained; 9 GW of planned data center power faces 50% protest rates, making energy the critical gating factor for AI monetization.

All AI misses (OpenAI, Anthropic) are 100% due to power supply constraints, not demand; transformers, gas turbines, grid infrastructure backlogs worsening; less than half of announced gigawatt projects actually building due to red tape and supply chain; this structurally benefits hyperscalers who control grid access.

Open-source models (Kimi K3, DeepSeek, Llama) match closed models within weeks of release, eliminating sustained advantage; closed labs (Anthropic, OpenAI) face margin compression and seek regulatory capture to ban open source, but banning would hurt US developers and tank stock market by forcing 50-100x AI cost premium vs rest of world.

Open-source models (Nvidia Nemotron, GLM, etc.) now match closed frontier models for 95% of enterprise tasks. When wrapped with control planes like 8090, they deliver 16.4x cost reduction, making AI sovereignty economically compelling versus renting intelligence from labs that compete with you.

Benioff, Chamath, and Freeberg converge: Chinese models already match US on weaker hardware; restricting H100/Blackwell only accelerates Huawei. Selling chips locks in Nvidia standard, funds US R&D lead, and reduces conflict via interdependence. Chamath: "Let Jensen win. I'd rather he win than Huawei win."