The Fight Over Open Source AI, Anthropic's $1.5B Payout, NYC Socialists: Evictions = Violence?
Summary
A U.S. ban on Chinese open-source models would impose a domestic “token tax” while the rest of the world keeps using cheaper intelligence. Sacks said he had “good authority” that the White House had made no decision, then called action against open source a “tragic mistake” that would punish American developers rather than stop Chinese distillation. Chamath’s investor translation: forcing Coca-Cola to buy AI inputs priced 50-100 times above foreign alternatives would distort corporate costs, frontier-lab revenue, and ultimately the stock market.
The panel’s central valuation call was that foundation models are commoditizing faster than capital markets expected, shifting durable value toward applications, clouds, and chips. Chamath argued that models match published performance within weeks and that closed alternatives are mispriced at 25-50 times open models; Friedberg compared the model layer to browsers and web servers, where open source let value accrue to the internet. “You can buy 50-cent Coke or $50 Coke.”
Sacks rejected both the Kimi K3 panic and the obituary for American frontier labs. He said Kimi K3 performs strongly on front-end web coding but is neither materially cheaper to run nor caught up across every dimension; he still put the U.S. lead at six months and cited unreleased GPT-6.0 as “blowing the doors off.” Anthropic reportedly rose from $10 billion to more than $70 billion of ARR, while third-party estimates showed OpenAI moving from $33 billion in May to $41.3 billion in July.
The practical threshold is not whether open models win every benchmark, but whether many models can handle 95% of real work. Jason said startups are already shifting workloads to GLM-5.2, Grok, Nemotron, and locally hosted models, creating IPO and margin headwinds for Anthropic and OpenAI. Chamath’s reconciliation: frontier labs can still create “trillions and trillions” in applications, but assigning large terminal value to the foundational-model layer is “a mathematical mistake.”
The fight hinges on whether learning from outputs is equivalent to stealing weights, and the panel said it is not. Friedberg compared distillation with Google benchmarking millions of Yahoo and Microsoft searches; Jason drew the distinction between stolen proprietary weights and outputs used to derive new weights, with Friedberg and Sacks agreeing that stolen weights would be theft while learning from outputs is a different act. The contradiction is “IP for me, but not for thee.”
Anthropic’s $1.5 billion settlement addressed pirated books, not the unresolved legality of training on lawfully obtained content. The episode cited seven million downloaded books, roughly 500,000 covered books, $3,000 per book, $101 million for lawyers, and 91% author participation. Sacks warned that calling unwanted distillation “IP theft” could be a fatal own goal in Anthropic’s other copyright cases; Jason instead urged AI companies to reserve 10% of revenue for licensing and settlements.
The market punished unprecedented AI CapEx, but the panel treated Google’s spending as the clearest public-market way to own model proliferation. Google Cloud was said to be growing 82% year over year toward a $100 billion run rate, with total CapEx forecast at $195-$205 billion; Chamath gave management the benefit of the doubt after a 32% long-run return on invested capital. On New York housing, the same cost-mechanism framing prevailed: constrain screening, eviction, and rent-setting without adding supply, and landlords may demand 3X rent, prepayment, or leave units empty.
Deep dive
1. Washington has not decided, but an open-source ban would “backfire badly”
Jason framed Kimi K3 as a second DeepSeek moment: Moonshot AI’s open-source model was described as performing on par with Opus 4.8 and GPT-5.6 at roughly 50% lower cost. Axios had reported that Washington was considering a ban, Wired said Commerce Secretary Howard Lutnick opposed one, and Polymarket’s 2026 probability had risen from 22% to 45%.
Sacks’s inside read was narrower: “There is no decision by the White House to ban open source models.” He said the administration is discussing Chinese distillation and hearing competing views, while President Trump’s lighter-regulation instincts have helped the U.S. lead the AI race.
His own position was categorical: restricting the open-source ecosystem would be a “tragic mistake” that cuts American developers off from public-domain tools available everywhere else. Whatever Washington concludes about Chinese conduct, “you cannot punish American developers for it.” Jason later summarized the consequence as putting the country on an island of overpriced AI.
2. Distillation is easiest to stop at the source
Chamath defined distillation plainly: query another model, observe its answer, and use the output to train yours; repeat that tens of millions of times and the result is trillions of question-answer pairs. His proposed control was KYC—identity checks, bounded credit cards, and other friction that would reduce anonymous account farming but also slow revenue growth.
Sacks argued that Anthropic’s claimed “industrial scale” abuse should be easier, not harder, to detect. If distillation is a national-security threat, the frontier labs are best placed to block it before “the horse is out of the barn”; asking Washington to ban competitors while preserving rapid sign-ups reverses the responsibility.
Jason supplied the evasion mechanism: waves of accounts created by students or workers in places including Manila and India, resold through dark-web channels, and operated through American IP addresses. The panel accepted that this may violate terms of service, but separated deceptive account creation from the broader claim that learning from outputs is intellectual-property theft.
3. A protected duopoly would impose a “token tax”
Chamath saw regulatory intervention as a valuation-preservation strategy for frontier labs whose models are being matched within weeks. If comparable models are available at 25-50 times lower prices, investors will eventually move terminal value toward applications above the model and infrastructure—cloud and chips—below it.
His Coca-Cola example carried the market mechanism: ban open source and an American enterprise may be forced to buy an AI input costing 50-100 times its foreign competitors’ alternative. That cost worsens the enterprise’s valuation, while Anthropic’s and OpenAI’s newly protected revenue also deserves a lower multiple because it is regulatory, geographically limited, and no longer market-clearing.
Sacks compressed the argument into a “token tax” imposed by a government-enforced duopoly. Chamath’s conclusion was even sharper: intervention would “tank the stock market. Period. Not debatable,” though the distribution of losses between buyers and token sellers would depend on the policy.
4. Model outputs are not model weights
Friedberg described distillation as ordinary competitive engineering: carmakers inspect rival cars, while early Google submitted millions of searches to Yahoo and Microsoft and compared result sets to improve ranking. That was benchmarking—not entering a competitor’s servers and stealing its algorithm.
Jason supplied the technical boundary he thought policymakers miss. Model weights are the numerical parameters—the software itself—and stealing Anthropic’s proprietary weights would be theft; using its service, studying outputs, and deriving independent weights is a different act, even if fake accounts create a terms-of-service violation. Friedberg and Sacks agreed with that distinction.
The contradiction matters because OpenAI and Anthropic argue they may learn from publishers’ outputs against creators’ wishes. Sacks said The New York Times accuses OpenAI of scraping its site at nonhuman scale to derive new weights—the same conceptual defense American labs now resist when Chinese labs learn from them.
5. Open source moves value out of the gatekeepers
Friedberg’s enforcement point was almost physical: an open model is downloadable software that can run offline, like a book already sitting on someone’s computer. Restricting use after publication would create an ugly free-speech and enforcement problem, absent a copyright judgment reached through normal legal process.
His internet analogy began with Netscape’s proprietary browser and server, then Mozilla’s Firefox, Google’s open-source Chromium, and Apache’s HTTP server. Once anyone could create a website without paying a gatekeeper, value accrued to Google, eBay, Etsy, Amazon, small businesses, and the wider network—not the few companies controlling entry.
Applied to AI, open models could prevent gains from concentrating in “two or three or four companies and their small group of billionaire shareholders.” Friedberg’s call was to let models proliferate, pursue actual software theft separately, and allow a million AI-enabled enterprises to capture the productivity upside.
6. Kimi K3 did not erase the American frontier
Sacks initially worried Kimi K3 meant China had caught up with a much cheaper frontier model, but changed his view as operating-cost details emerged. He said it was not significantly cheaper to run and that its exceptional arena score covered front-end coding and web development—not the full range of model capabilities.
His remaining estimate was that the U.S. is still six months ahead, with substantially better models inside American labs. He cited reports that Sam Altman would discuss GPT-6.0 in Washington and described it as “blowing the doors off,” arguing that the right policy is to “let our horses run.”
Revenue was his reality check: Anthropic began the year around $10 billion of ARR and was said to exceed $70 billion midway through, against an internal forecast of $100 billion. Third-party estimates put OpenAI’s run rate at $33 billion in May and $41.3 billion in July, with exit-ARR expectations reportedly raised from $60 billion toward $75 billion.
Sacks therefore expects both open and closed models to win. Open source offers control, customization, ownership, and data sovereignty but requires more work; closed vendors supply harnesses, connectors, support, and enterprise agreements. The frontier labs’ attempt to draw a government foul looked to him like a basketball “flop,” not evidence that their business had already collapsed.
7. Ninety-five percent of work is enough to compress terminal value
Jason pushed the strongest bear case: startups that recently spent hundreds of thousands of dollars per quarter with frontier labs are moving toward GLM-5.2, self-trained models, and local deployment. In his own comparisons through Perplexity Computer, Grok, Nemotron, GLM-5.2, and Claude often produced results he could not distinguish.
His conclusion was a “non-zero chance” that open source derails or delays Anthropic’s and OpenAI’s IPOs through margin compression and stranded infrastructure spending. Startups matter because enterprise buyers eventually copy their architecture, while improving hosting intermediaries are removing the implementation burden that historically protected closed vendors.
Chamath refined rather than fully endorsed the claim: it is not that open models perform 95% of bleeding-edge tasks, but that “95% of the tasks can be done by many different models.” That alone makes the model layer commodity-like even if premium systems retain the difficult final 5%.
The frontier labs’ answer, in his view, is vertical integration: use superior private models to build life-sciences, cybersecurity, and other applications, perhaps retaining a next-generation model exclusively for internal products. They can still create “trillions and trillions of enterprise value,” but assigning substantial terminal value to access at the foundational layer is “a mathematical mistake.”
8. Chinese forks have already become American products
Sacks’s first example was Thinking Machines’ model, described as the best American open-source model and bootstrapped through distillation from Kimi K2.5. If Anthropic broadly taints the Chinese base as stolen IP without an evidentiary process, American derivative work becomes collateral damage.
Cursor’s Composer 2 supplied the next step: it reportedly started with Kimi K2.5, then post-trained on Cursor’s proprietary coding data. That is the open-source loop—take public weights, fork them, add proprietary work, and produce a distinct American product.
Once those weights are downloaded and run in an American data center, Sacks stressed, “No packets are going back to China. No data is going back to China.” A blanket ban would therefore put “a dagger through the heart” of an American ecosystem competing directly with the closed labs requesting protection.
9. China’s long game may be to commoditize bits and own molecules
Friedberg zoomed out from model competition to economic structure. The U.S. accumulated trillions through intellectual property, services, and converting “one bit to another,” while outsourcing the manufacturing capacity that converts physical molecules into useful goods.
If open-source AI flattens the value of knowledge and services, he argued, the residual prize sits in electricity and molecule conversion. His figures were stark: roughly one terawatt of U.S. electricity production versus China heading toward eight, and 10 billion square feet of U.S. manufacturing capacity versus China’s 200 billion.
That gives China, in his framing, 20 times the manufacturing footprint and eight times the power-generation capacity. Friedberg presented commoditizing global knowledge and services, then retaining value through physical production, as a possible 1-, 2-, or 3-decade strategy.
10. Anthropic’s $1.5 billion settlement punishes piracy, not yet training
Jason summarized the settlement as the largest U.S. copyright settlement: Anthropic had downloaded seven million books from pirate sites, roughly 500,000 books were covered, authors would receive about $3,000 per book, lawyers $101 million, and 91% of eligible authors had claimed shares.
Sacks’s nuance was that the settlement does not resolve whether training itself is fair use. Anthropic was exposed because it acquired stolen copies without buying even one; had it purchased a copy of each book, fair use could have remained its defense, though that doctrine is still being litigated.
Friedberg tested the boundary with Jason’s book: if an AI never reads the protected text but learns from public reviews, commentary, and metadata, has copyright been infringed? His answer was that copyright clearly prohibits reproducing text as one’s own, but “knowledge can’t be contained”; reading and transforming diffuse knowledge is different from copying expression.
Jason accepted the review example but drew a harder line when a trained product directly competes with the copyright owner, pointing to Thomson Reuters versus Ross, The New York Times, and music litigation among roughly 150 major cases. His proposal was for AI companies to pool 10% of revenue for licenses, settlements, and continuing access to new material.
11. “IP theft” could turn Anthropic’s argument against itself
Sacks said Anthropic coined “industrial scale distillation attacks” in a February blog post but did not call the practice IP theft there. Its original public rationale focused on Chinese models losing guardrails and creating national-security risk—a framing he said failed to gain enough policy traction.
Escalating to IP theft creates the hypocrisy: Anthropic wants the right to train on creators’ output while denying others the right to train on its output. Sacks asked whether this could be a “fatal mistake,” because publishers can now cite the lab’s own principle and claim its entire product rests on uncompensated work.
His preferred formulation would target fake accounts and proxies as deceptive business practices that violate service terms, while leaving fair use untouched. Jason agreed that this would have been more legally coherent; instead, Anthropic risks being “hoisted on their own petard” in litigation involving billions of dollars and perhaps the viability of its products.
The startup backlash shows the spillover: Sacks cited Garry Tan and 200 startups writing a letter warning that an IP-theft label could taint every American model derived from Chinese weights. Jason nevertheless maintained that publishers should coordinate, separate search crawling from AI crawling, and use collective refusal to try to force a licensing settlement.
12. Capital markets punished CapEx that the panel wanted to own
The reported numbers were historic: Google Cloud grew 82% year over year toward a $100 billion run rate, while Google’s annual CapEx forecast was $195-$205 billion. Tesla’s CapEx rose 140% and was expected to reach $25 billion; at taping, Google was down 7% and Tesla 14% after both reported negative free cash flow—Google’s first negative free-cash-flow period since going public.
Chamath’s bull case rested on Google’s roughly 32% average return on invested capital over 25 years. A company that compounds capital at that rate deserves latitude during a buildout, even if one unverified comparison suggested its CapEx would equal 20% of the annual U.S. military budget.
Model fragmentation strengthens Google because it can monetize silicon, GCP, Search, YouTube, advertising, and applications without picking a single model winner. Friedberg called Google the best public-market AI holding: the downside case still leaves world-class infrastructure selling other people’s models, while Google also has Waymo, roughly 10% of SpaceX, and a good chunk of Anthropic.
SpaceX was said to trade around $1.5 trillion, 30% below its first-day close after going public at $2 trillion, with staged lockups creating pressure. Apple provided the contrast: about $755 billion in buybacks and $140 billion in dividends over a decade—a $900 billion “do-no-harm” allocation that Jason argued left potential ambition on the table, though Chamath noted shareholders could reinvest the returned cash.
13. New York’s tenant protections could raise rents and empty units
Jason described Mayor Zohran Mamdani’s proposal as barring landlords from charging application fees for credit checks, allowing either a credit check or a 40-times-rent income test but not both, recognizing tenant unions, and accompanying a one-year rent freeze. The triggering language came from an activist who said the administration was ending tolerance for “the violence of evictions.”
Friedberg answered through private-property rights, quoting John Quincy Adams: “Property must be secured or liberty cannot exist.” His chain was that moral condemnation of owners legitimizes confiscatory controls; weakened ownership produces anarchy, competing groups consolidate power, and temporary anarchy eventually hardens into tyranny.
Sacks focused on tenants rather than landlords. Buildings without reliable rent lose maintenance funding, while non-evictable residents who create noise, damage, smells, intoxication, or disorder trap elderly and working-class neighbors in decline. Affluent progressives can afford “luxury beliefs” about public order because they do not depend on the affected apartments, parks, buses, or subways.
Chamath and Jason returned to supply: Austin data, Chamath said, shows that looser permitting adds units and each unit pushes rent down. If New York instead restricts screening and eviction, landlords may start at three or four times the rent, demand a year’s prepayment, sell, or leave units vacant; Jason cited reports of 50,000 “ghost apartments,” predicting that the intended affordability policy shrinks supply.