Socialists Sweep NYC, China Catches Up in Coding, AI Memory Crunch, Micron's Blowout Quarter

What It Covers

Chamath, Sacks, Travis Kalanick, and Gavin Baker analyze the DSA's sweep of three New York congressional primaries, China's GLM 5.2 open-source model matching Anthropic's Claude Opus 4.8 performance at 85% lower cost, Micron's revenue quadrupling to $42B annually, and the emerging modular data center hardware race shaping AI infrastructure economics.

Key Questions Answered

• •DSA Political Strategy: The Democratic Socialists of America explicitly use the Democratic Party as a ballot access vehicle while building independent organizational infrastructure. Their co-chair stated publicly they caucus with Democrats only when useful and view the establishment as an obstacle. Three incumbents lost in New York primaries where DSA-backed candidates outperformed with younger, college-educated, higher-income voters — the demographic that can financially afford socialist ideology.
• •China's AI Distillation Playbook: China's GLM 5.2 model scores 51 points on the Artificial Analysis Intelligence Index — the highest of any open-weight model — by systematically harvesting reasoning traces from US frontier model APIs through masked accounts at scale. This distillation process feeds back into reinforcement learning, enabling near-frontier performance at a fraction of training cost. The model was reportedly trained entirely on Huawei Ascend 910B chips, signaling meaningful indigenous silicon progress.
• •Composable AI Architecture: Enterprises should route roughly 85% of queries to open-weight models hosted on proprietary data, reserving only the hardest tasks for frontier models like GPT or Claude. This composable approach — what Andrej Karpathy calls a "council of models" — delivers Pareto-dominant outcomes while dramatically reducing inference costs. Open-source models shift economic value from frontier lab margins to infrastructure providers without reducing overall AI capability or investment returns.
• •HBM Memory as the Critical Bottleneck: High-bandwidth memory DRAM represents 30–40% of all hyperscaler capital expenditure and is the single most constrained resource in AI infrastructure. Only three companies globally — Micron, SK Hynix, and Samsung — manufacture HBM. Micron's entire 2026 supply sold out in advance, driving revenue from $9B to $42B year-over-year. Consumer electronics prices are rising as AI data centers outcompete smartphones, gaming consoles, and laptops for available DRAM supply.
• •Orbital Compute Economics: Building a one-gigawatt terrestrial data center costs approximately $35B in semiconductors plus $25B in power and cooling infrastructure, with the latter figure being inflationary due to human labor costs. Once Starship achieves full reusability, launching equivalent compute capacity into orbit costs an estimated $5B, putting total orbital deployment at roughly $40B versus $60B+ terrestrially. The gap widens further as land entitlements and power access become increasingly constrained on the ground.
• •AI as Economic Equalizer: AI converts the internet's stored knowledge into actionable expertise accessible to every individual without gatekeeping. The practical effect is that any person gains access to a co-founder-level strategic and technical thinking partner at zero marginal cost. The failure to communicate this framing clearly has allowed anti-AI narratives — funded in part by safety-focused labs seeking regulatory moats — to dominate public perception and fuel political backlash in congressional races.
• •IPO Pricing Discipline: Companies going public should use Dutch auction mechanics to clear price rather than banker-managed book-building, which optimizes for fees over accuracy. Cerebras broke its deal price within days of IPO, triggering price-insensitive selling from institutional managers who exit any stock below deal price regardless of fundamentals. Gavin Baker estimates Anthropic would trade at approximately $3T as a public company based on projected revenue exceeding $100B annually and 85% gross margins on inference-dominated revenue.