The AI Wealth Problem

Anthropic says its Public Benefit Fund will initially receive 5% of the company’s shares, with the board able to grant another 2% for each 50% increase in company value above a set baseline. The company says roughly 20% of its shares will ultimately be committed to the public good: its co-founders have pledged 80% of their equity, and about half of employees are donating some equity through a matching program. The fund will be a wholly owned LLC rather than a 501(c)(3), which Anthropic says preserves flexibility in how it gives. Its equity will not fund lobbying or political giving.

The fund could support nonprofits or send money directly to people affected by AI, including workers facing labor-market impacts. Anthropic says its existing safety and biomedical research, along with discounts for nonprofits and scientific institutions, will remain within the company. Nayan Ransahoff estimated that AI-related philanthropic assets could eventually support $37 billion to $100 billion in annual spending, or roughly 6% to 17% of current U.S. charitable giving of about $600 billion. That estimate assumes $370 billion in assets and a 10% annual spending rate; Ransahoff cautioned that philanthropy may not have enough capable organizations or skilled allocators to put such sums to effective use.

The fund has intensified disagreement over who should define the public good and control the distribution of AI wealth. Jeremy Flores argued that the public should decide through democratic controls, while Jane Adams warned that philanthropy can divide society into benefactors and beneficiaries. David Sacks said whether assigning 20% of Anthropic’s equity to an LLC serves the public good remains to be seen. Other critics said the fund could advance Anthropic’s political interests or divert resources from the U.S. economy; Joe Weisenthal, by contrast, criticized entrepreneurs and venture capitalists for telling Anthropic how to spend its money.

The debate also focused on alternatives to concentrating AI wealth in private institutions. Andrew Curran said OpenAI had proposed a public wealth fund that would distribute returns directly to U.S. citizens, and Aaron Burnett supported using Invest America accounts. The synopsis also reports that Anthropic was expected to stop hiring junior technical talent by mid-2027, prompting some to call its fund “privatized UBI.” One critic argued that universal basic income cannot be limited to selected populations, raising a broader question about whether company-directed philanthropy can substitute for widely shared benefits.

New York City Comptroller Mark Levine urged Anthropic to pause its IPO until it addresses AI safety and governance. He argued that the company’s White House accord was insufficient and that preparing for an IPO could distract from safety work. Anthropic’s filings devoted 80 pages to risk disclosures, more than twice the space used to describe its business. Levine called for detailed safety-audit plans and warned that super-voting shares could separate control from accountability. He also said a safety incident after listing could damage the stock, and that index inclusion could quickly expose some city pension funds to Anthropic; he acknowledged that delaying the IPO could itself create a market problem.

Anthropic’s proposed AI safety bill would allow a court to block a model’s release if it posed a foreseeable, imminent risk of more than $1 billion in damages or over 50 deaths through specified hazards, including bioweapons, loss of developer control, or cyberattack. It would require independent evaluators registered with the Commerce Department to report specified imminent risks, while barring other government bodies from suspending development, deployment, or internal use. Nate Suarez and Jeffrey Ladish argued those limits could make it harder for government to halt dangerous development. Sacks said the bill’s 10^25-FLOP frontier threshold could cover much of Silicon Valley, while Alex Boraz called it inadequate for the pace of AI development.

Signals about the cost and durability of AI investment were mixed. Charles Gasparino predicted a correction, citing debt-funded spending, while Cathie Wood argued that bubble fears suggested the market was not in one and that adoption remained early. The Philadelphia Semiconductor Index fell 12% over eight consecutive days, and the Nasdaq lost more than 5% last week; meanwhile, spot GPU rental prices had risen for 17 straight days, with no reported massive overbuilds or canceled spending. Microsoft canceled a multibillion-dollar Essex data-center contract after repeated delays, with utilities not expecting to connect the site to the grid until the 2030s. Reporting attributed the cancellation to uncertain power supply rather than weak AI demand, though it unsettled confidence in nscale’s planned IPO valuation.

OpenCode said a large-scale distillation attack affected 1.4% of its console users, apparently using stored API keys to generate heavy usage and potentially synthetic training data; users reported bills of hundreds or thousands of dollars. The attackers were not identified, and use of API credentials did not establish Chinese involvement. Anthropic revoked compromised keys it detected, OpenAI said it reset keys and refunded affected customers, and OpenCode said it was working with OpenAI, Anthropic, and Google. DAXrad argued that continuous defensive agents were needed to keep pace with attackers, while Dean Ball questioned whether broader access to autonomous cyber-capable models was worth the risk.