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

Summary

The episode opened with a long political debate over New York City primary results, which the hosts framed as a socialist/DSA sweep and a sign that the Democratic Party is being reorganized around populist left politics, Israel/Palestine, and social-media-driven radicalization. Gavin Baker, Chamath Palihapitiya, Travis Kalanick, and the hosts argued over the role of NGOs, education, age-gating, and anonymity online, with repeated claims that institutions and incentives are pushing voters toward extremism. The second half turned to AI competition and infrastructure: China’s open-weight models were described as rapidly closing the gap through distillation and cheap deployment, while Micron’s earnings were treated as evidence that DRAM/HBM is now the key AI bottleneck. The episode finished on infrastructure economics, including modular “megapods,” distributed inference vs. training, orbital compute, and IPO/valuation speculation around Anthropic and Cerebras.

Notes

• The episode opens with Jason framing the show around a “socialist sweep in New York City” and bringing on Gavin Baker and Travis Kalanick.
• Jason says Polymarket had the Mamdani sweep at only 26% before election day.
• Chamath argues AI is the biggest economic equalizer of their lifetime and says the industry has miscommunicated its benefits.
• He links the Internet and Google to knowledge capture, and AI to turning knowledge into expertise.
• Sacks characterizes the DSA as radical and says it is taking over the Democratic Party.
• He continues by listing a constitutional overhaul: abolish the Electoral College, replace the President and Supreme Court structure, and shift toward public ownership and proportional representation.
• Sacks says the DSA sees the Democratic Party mainly as a ballot-access vehicle.
• Travis frames truth and justice as society’s immune system.
• He then says communism is latent in everyone and becomes dangerous when ecosystems normalize laziness and getting something for nothing.
• Gavin says the DSA reflects both Democratic decline and the rise of downwardly mobile wealthy white liberals.
• He says NGOs and outsourced government funding are a major part of the corruption dynamic, especially around homelessness spending.
• Chamath says social-media age bans in countries like Canada, the UK, Australia, and Florida may help slow youth radicalization.
• Travis counters that under-16 bans are really a path to adult de-anonymization and censorship.
• Gavin agrees that under-16 bans could be good only if anonymity and free speech are preserved.
• Gavin cites a study he says is uncontested: electing a Republican DA reduces all-cause mortality for young Black men by 7%.
• He also argues progressive education policy can be harmful, citing removal of math in California and repeat-offender crime concentration.
• Chamath adds a crime concentration claim, saying 0.1% of people commit 70% of crimes.
• Sacks says the DSA’s reach extends beyond New York and points to Los Angeles ballot harvesting as a test case.
• He says younger Democrats are far more anti-Israel and cites 80% disapproval among Democrats.
• Chamath says people increasingly conflate Jews, Israelis, and Netanyahu, which he calls unfortunate.
• Gavin says Israel has a major PR problem and needs younger, social-media-fluent spokespeople.
• He says some parents became more politically activated during COVID after hearing what their children were being taught.
• Jason later argues the Democratic Party is being taken over by billionaire-funded socialists.
• Jason shifts to Chinese open-source AI and says GLM 5.2 is a frontier-class, MIT-licensed model with a huge context window.
• He says the model scored 51 on Artificial Analysis Intelligence and is 85% cheaper than GPT-5.5 for similar performance.
• Gavin describes distillation as large device farms querying frontier APIs and harvesting reasoning traces to train cheaper models.
• He says the future is composable models, with enterprises using a council of LLMs rather than a single vendor.
• Gavin warns frontier model companies to think carefully about ASICs and the incentives they create for Nvidia.
• Sacks says GLM 5.2 is now the best open-weight model for coding, software engineering, and long-context agent work.
• He says Z AI claims the GLM5 family was trained entirely on Huawei Ascend 910B chips.
• Sacks says Chinese companies are packaging these models as “AI in a box” and selling them globally at lower cost.
• Jason says Micron’s revenue jumped from $9B to $42B year over year and guidance rose to $50B.
• Gavin says DRAM, especially HBM DRAM, is the most important AI bottleneck because memory capacity and bandwidth set model performance.
• He says Micron’s new supply agreements include floor prices above prior cycle peaks and cover about half the company’s revenue.
• Gavin estimates DRAM may absorb 30% to 40% of hyperscaler capex next year.
• Sacks explains HBM as stacked DRAM dies, with HBM3 cited at eight dies and newer stacks moving to 12 and 16.
• Gavin says a 1 GW data center needs roughly $35B in semiconductors and $25B in power/cooling equipment.
• He also argues orbital compute may work if launch costs fall to around $5B and Starship becomes rapidly reusable.
• Chamath says data-center demand is effectively infinite while supply is meager, and that many data centers are already contested.
• Jason reads a recent trademark filing for “Megapod” as modular AI data-center hardware.
• He says the likely interpretation is a giant battery-pack-like unit with GPUs at Supercharger stations.
• Chamath says compute deployment is constrained by land, zoning, power, and basis-of-design specs.
• He says those basis-of-design documents span 50 to 500 pages.
• Chamath says the industry is relaxing liquid-cooling requirements and reviving air-cooled, prefab modules.
• He describes shipping-container-style megapods that can be prefabricated in warehouses and deployed in roughly 90 days.
• Travis says the main obstacle to non-data-center real estate is security and access control.
• Chamath says the valuable use case is contributing hardware to distributed training or lower-SLA inference pools.
• Travis says distributed training falls apart when GPUs are not physically adjacent, with at least an order-of-magnitude efficiency loss.
• Gavin adds that distributed inference is different: prefill is memory-capacity bound and decode is memory-bandwidth bound.
• He explains inference as a split between prefill and decode, with decode as a memory-bandwidth problem.
• Gavin argues this will keep GPUs useful for 7 to 12 years and make AI capex more financeable.
• The group sees distributed inference as strategically important for the American open-source and free-speech ecosystem.
• Jason says Targon lets users rent H100-class GPUs for roughly $3 to $4 per hour.
• He also speculates Tesla could bundle GPUs into Powerwalls and pair them with Starlink.
• Jason asks about Anthropic, OpenAI backlog, and public-market absorption of AI IPOs.
• Gavin says Anthropic would be worth about $3 trillion today.
• He says Anthropic could end the year above $100 billion in revenue and be highly profitable on inference.
• He argues the public markets can absorb the AI transition because the capital is shifting from private to public, not newly created.
• Gavin gives a detailed timeline for a Cerebras/OpenAI contract: wafer order, chip fabrication, serverization, and energizing.
• He says the key investor question is how many megawatts can be brought online, since monetization per gigawatt is known.
• Travis recalls Uber’s auction-style round, which started around 9 or 10 and cleared at 17.