The Fight Over Open Source AI, Anthropic's $1.5B Payout, NYC Socialists: Evictions = Violence?

The White House has made no decision to ban open source AI models, with David Sacks calling any such ban a tragic mistake that would damage America's position in the AI race. Howard Lutnik at the Commerce Department has similarly said he does not want to ban Chinese models and prefers incentivizing U.S. frontier labs to develop better open source alternatives. Polymarket placed the probability of the U.S. government banning an open source model in 2026 at 45 percent, up from 22 percent just days earlier.

The central technical dispute is over distillation, which involves querying a model, observing its outputs, and using those outputs to train a competing model. Speakers drew a clear distinction between learning from model outputs, which is what distillation involves, and stealing proprietary model weights, which no one has accused Chinese labs of doing. Friedberg noted distillation is a common technique used across industries, citing Google submitting millions of queries to Yahoo and Microsoft in its early days to benchmark its own search algorithm. Chamath added that virtually everyone in the industry has distilled at some point, making it a weak basis for complaint.

Anthropic drew pointed criticism on multiple fronts. Sacks accused the company of regulatory capture, arguing that if stopping distillation were truly its objective it would push to ban Chinese access to American models rather than American access to Chinese models. He also said Anthropic has done a poor job stopping distillation despite it allegedly occurring at industrial scale through waves of accounts sold on the dark web and routed through the Philippines and India. Separately, Anthropic settled an AI copyright lawsuit for 1.5 billion dollars, described as the largest copyright settlement in U.S. history, after downloading seven million books from pirated websites to train Claude. The settlement covered 500,000 books, with lawyers receiving 101 million dollars and authors receiving 3,000 dollars per book. One speaker argued Anthropic could have avoided the exposure entirely by purchasing even one copy of each work and relying on a fair use defense.

A structural hypocrisy argument ran through the discussion. OpenAI's legal position is that deriving model weights from scraped content without creator consent is not theft, which speakers noted is structurally identical to what Chinese labs do when deriving weights from American model outputs. Speakers argued that by adopting the IP theft framing against Chinese labs, Anthropic effectively conceded its own product is built on stolen content, potentially giving content creators grounds to claim a share of its revenue. Gary Tan and approximately 200 startups wrote a letter opposing the IP theft framing on the grounds it would taint all derivative works of Chinese models and threaten the broader startup ecosystem.

On revenue, Sacks said Anthropic started the year at roughly 10 billion ARR and reached over 70 billion ARR at midyear, with an internal forecast to hit 100 billion by year end. OpenAI was expecting to exit the year at 60 billion ARR but is now forecasting closer to 75 billion. Third-party tracking data showed a recent stall in Anthropic's revenue, and well over 50 percent of tokens tracked on OpenRouter are now coming from Chinese models. Chamath cautioned that open source and in-house deployments generate what he called dark tokens that are free and do not appear on any revenue chart, making current earnings an incomplete picture. Chamath argued that open source models have reached parity for 95 percent of tasks, compressing margins and threatening the long-term value of the model API layer, while Sacks countered that the revenue trajectory of both frontier labs remains the fastest growth at scale ever seen.

The broader structural argument is that real margin capture in AI is flowing to the application and infrastructure layers, not foundational models, with cloud providers like Google as primary beneficiaries. Google was described as the best public market stock for AI exposure, owning roughly 10 percent of SpaceX and a significant stake in Anthropic, and marking up its Anthropic investment by 100 billion dollars in a single quarter. Google is free cash flow negative for the first time due to its current capital expenditure cycle, with CapEx this year equal to roughly 20 percent of the entire U.S. military budget.

New York City assemblyman Zohran Mamdani introduced legislation banning landlords from charging application fees for credit checks and legally recognizing tenant unions, then froze rents for one year. An activist at a hearing characterized evictions as violence. Sacks argued that preventing evictions harms other residents because landlords without rental income cannot maintain upkeep, and that New York is simultaneously banning credit checks, background checks, and income verification, eliminating any landlord incentive to rent at market rates. He estimated roughly 50,000 ghost apartments exist in New York City as a direct result of these regulatory disincentives. Chamath argued that landlords unable to vet tenants will set rents three to four times higher and demand multi-month prepayments to offset risk, and pointed to Austin, Tokyo, Texas, Florida, and Nevada as examples where relaxing permitting constraints produced more units and lower rents.