Featuring
With Jason taking the week off, the notorious threesome moved from Ukraine and a possible market turn to downtown San Francisco's office wreck, WeWork's bankruptcy setup, and Biden's sprawling AI order. The sharpest exchange asked whether bargain-basement offices can create their own demand; the most revealing one forced Chamath to defend a much narrower version of AI regulation than Washington delivered. Chamath had the best episode because he kept dragging attractive spreadsheets back to the people and incentives required to make them real.
Spice rack
Will cheap San Francisco offices bring employers back, or has demand changed for good?
Original point: A cheap building does not create a tenant: even free space would not persuade him to relocate a distributed workforce back to San Francisco while the move raises costs and the city remains unattractive to employees.
What everyone argued
Chamath Palihapitiya
Chamath says the spreadsheet case mistakes price for demand. His company went remote, employees dispersed around the country, and moving them back would add office costs and relocation friction; he says he would reject even free San Francisco space under current conditions.
David Sacks
Sacks concedes that the demand picture is unclear, then argues distressed buildings offer asymmetric upside if rates fall and the city responds to budget pressure by cutting barriers, improving public safety, and treating business as a partner. He frames the bet over five to ten years, not the next lease cycle.
David Friedberg
Friedberg argues the market can normalize toward pre-2010 rents after lenders and owners recognize losses. He expects lower prices to make offices economical again and tells Chamath that, once repricing occurs, companies like his will return.
Winner circle
Chamath wins the medium-term round. Friedberg correctly saw that losses had to be recognized, and Sacks correctly treated distressed property as a conditional option on rates and reform. But through 2026, pricing reset faster than occupancy: demand improved without coming close to filling the hole, which is exactly the distinction Chamath insisted on.
Commentary
Chamath Palihapitiya
Assumptions and fact checks
Even near-zero office rent would not overcome the cost of relocating a distributed workforce or employees' resistance to returning to San Francisco.
Why it mattersRent is only one component of return-to-office cost. Hiring constraints, relocation, commuting, taxes, safety perceptions, and employee preferences can dominate the lease price, so cheaper space alone cannot guarantee demand.
His own company's reluctance is representative of enough employers to keep citywide demand weak.
Why it mattersThe mechanism is strong, but tenant demand is uneven. AI companies were expanding in San Francisco by 2026 even while older and non-prime office stock remained deeply impaired.
David Sacks
Sacks earns credit for admitting that demand is the leap of faith. Once cheap prices, lower rates, safer streets, faster permits, and returning tenants are all required, however, the thesis is no longer simply 'buy below replacement cost'; it is a multi-leg turnaround bet.
Assumptions and fact checks
Budget shortfalls will force San Francisco to reform taxes, permitting, public safety, and its relationship with business.
Why it mattersFiscal pressure can produce reform, but it can also produce borrowing, taxes, service cuts, or temporary fixes. San Francisco did pursue permitting and conversion initiatives, yet the causal path to broad office demand remains incomplete.
Distressed office purchases offer three- or fourfold upside over a five-to-ten-year horizon.
Why it mattersLow basis creates optionality, but returns depend on occupancy, tenant improvements, financing, operating costs, and the quality of each building. Land-value pricing does not itself cap future capital calls.
David Friedberg
Friedberg explains how a market clears, but not why the old quantity of office use returns after it clears. His best case is for selected buildings at a new basis, not for citywide mean reversion.
Assumptions and fact checks
Once rents and asset values return toward pre-2010 levels, employers will again find San Francisco offices compelling.
Why it mattersLower occupancy cost helps, but remote-work practices, labor geography, safety, taxes, and building quality can keep demand below its old equilibrium. Prime space and AI tenants may recover while commodity offices do not.
The commercial-real-estate loan book was broadly unmarked and losses would flow quickly through bank balance sheets.
Why it mattersAccounting treatment differs by loan status, classification, reserves, collateral, and bank. Market-value impairment can be real without every loan being carried identically or forcing an immediate realized write-down.
U.S. commercial banks held about $3 trillion in commercial-real-estate loans in late 2023.
CheckThe Federal Reserve's H.8 release for November 2023 put commercial-real-estate loans at commercial banks at approximately $2.9 trillion, making the on-air $3 trillion figure a fair rounding.
Should AI rules police harmful outputs, or test selected high-risk systems before release?
Original point: Government should punish concrete harms such as fraud and impersonation instead of supervising model architecture, training methods, or scale before a crime occurs.
What everyone argued
Chamath Palihapitiya
Chamath agrees that Biden's order is incoherent and says most AI should be treated as ordinary software. He preserves a narrow exception: specific high-risk businesses should enter an FDA-like sandbox before launch, with the use cases defined in advance rather than every large model swept into a general regime.
David Sacks
Sacks says precautionary AI rules invite incumbents to shape barriers against open-source challengers. He predicts a thicket of agency rules will make software companies beg for one federal regulator, converting permissionless software into an FDA- or FCC-style industry.
David Friedberg
Friedberg argues for output-based law: prosecute fraud, impersonation, and other concrete harms, but do not regulate shifting model sizes or development methods. He says technical thresholds will age quickly and could push model development toward other countries.
Winner circle
Chamath wins narrowly on argument quality, not because Biden's order survived—it did not. He conceded the order's defects, treated most AI as software, and preserved only a bounded prevention case. Friedberg's output-first rule is the right default, but his factual overstatement of the order and failure to handle irreversible harms leave Chamath's narrower framework standing.
Commentary
Chamath Palihapitiya
Chamath wins by shrinking his claim when the evidence demands it. His missing work is the most important work: naming which uses enter the sandbox and preventing a narrow gate from expanding into a general license to write software.
Assumptions and fact checks
A small set of high-risk AI businesses can be defined clearly enough for pre-market sandbox review without pulling ordinary software into the regime.
Why it mattersSector-specific review is plausible where harms and endpoints are concrete, but model reuse and fast capability changes make boundaries difficult. The proposal needs explicit covered uses, risk thresholds, deadlines, and an appeal path.
The executive order's breadth mainly reflected influential people inserting favored provisions.
Why it mattersThe order plainly aggregated many policy goals, but the transcript offers no evidence tracing particular provisions to particular lobbyists or advisers.
The order proposed reporting when a foreign person used a U.S. infrastructure-as-a-service provider to train a large AI model capable of malicious cyber activity.
CheckSection 4.2 directs Commerce to propose rules requiring U.S. IaaS providers to report certain foreign-person transactions involving large models with potential malicious cyber capabilities. The on-air leap to ordinary H-1B employees using an employer's systems is not established by that provision.
David Sacks
Sacks offers the strongest objection to Chamath's sandbox: the gatekeeper may protect incumbents more reliably than the public. He weakens it by presenting a plausible incentive as a proven conspiracy and a possible bureaucratic endpoint as inevitable.
Assumptions and fact checks
Large AI companies supported the order chiefly to achieve regulatory capture and suppress open-source entrants.
Why it mattersIncumbents can benefit from compliance costs, but motive requires evidence. Firms may simultaneously seek safety rules, liability clarity, national-security controls, and competitive advantage.
Fragmented agency action would inevitably produce a single federal software commission.
Why it mattersThe forecast did not materialize before the order was rescinded, and Congress did not create such a commission by August 2026. Fragmentation can persist, be narrowed by courts, or be replaced without a new regulator.
David Friedberg
Friedberg has the cleanest default rule and the weakest reading of the actual order. His case works best as a presumption—regulate concrete harms unless a defined catastrophic risk justifies prevention—not as proof that pre-release review is never warranted.
Assumptions and fact checks
Existing laws applied after harm occurs are sufficient for nearly all AI risks.
Why it mattersOutput-based enforcement avoids freezing technology and works for familiar harms, but ex post remedies can be inadequate where damage is irreversible, attribution is hard, or the affected party cannot be made whole.
Model-development rules would materially shift AI leadership from the United States to India, China, or Singapore.
Why it mattersCompliance burdens can affect location and investment, but leadership also depends on chips, energy, talent, capital, market access, and export controls. The transcript does not isolate regulation's effect.
Executive Order 14110 set a maximum parameter count and empowered agency chief AI officers to regulate all private model builders above it.
CheckThe order described dual-use foundation models partly by broad characteristics including tens of billions of parameters, but its reporting thresholds were based on training compute and specified risk conditions. Agency chief AI officers were assigned responsibility for their agencies' AI use, not blanket regulation of every private developer.
The order required every piece of private AI-generated content to be watermarked so the government could track and audit it.
CheckThe order directed Commerce to develop standards and guidance for authenticating content and detecting synthetic content, and directed agencies to address official government content. It did not itself impose a universal watermark-and-track mandate on all private AI content.
Chamath asks the question the bargain hunters need to answer: who actually signs the lease? His case would be stronger with market-wide tenant data, but it correctly refuses to let replacement cost stand in for willingness to occupy.