AI’s compute boom hits a debt test

AI Infrastructure (3)

• Anthropic’s S-1 reports a sharp revenue jump, and The Wall Street Journal says the company aims to list on Nasdaq in November at a valuation of $1.5 trillion to $2 trillion.
• Ben Black argues that analysts need new measures for AI labs: they sell compute across models with different costs and prices, so SaaS metrics like ARR do not describe the business well.
• D.A. Wallach sees Anthropic’s founders as sincere believers in AI’s civilizational stakes, while Dave McClure says belief and profit-seeking can coexist; the panel links OpenAI’s breakaways to Sam Altman’s chaotic leadership and compute strategy.

Markets (3)

• Dave McClure says Anthropic’s reported $80.7 billion net loss largely reflects non-cash revaluation of convertible notes and related liabilities, while cash operations appear profitable.
• Dave McClure says Anthropic is still growing but has slowed, and estimates a $150 billion to $200 billion run rate next year; he cites a chart’s flat $76.8 billion figure as potentially unreliable.
• D.A. Wallach compares AI’s potential to cigarettes: addictive, high-margin products can support enormous businesses, but he questions whether AI’s rapid customer expansion proves durable utility; the panel also treated engineers smoking as a frothy signal.

Enterprise (3)

• Ben Black says Anthropic reported 900% net revenue retention, meaning the same customer cohort spent nine times more year over year; the panel contrasted that expansion with AI’s tiny share of corporate budgets.
• Jason Calacanis and Dave McClure say corporate AI adoption is held back by employees’ limited ability to connect models to company data and workflows, alongside security and privacy concerns; most users remain at basic prompting.
• D.A. Wallach says AI infrastructure’s durability depends on demand from the companies buying compute, not only on lab disclosures; public filings will expose spending but may not reveal customers’ actual usage or budgets.

Big Tech (1)

• D.A. Wallach expects Apple and Google to use their phone distribution to deliver capable AI assistants, while he prefers local processing for consumer data and estimates most everyday AI tasks could run on-device.

Startups (1)

• D.A. Wallach argues Figma’s collaboration habits and trained design teams give it staying power despite AI design tools; Ben Black says Canva’s slowing growth and falling secondary price show how quickly perceived momentum can damage valuations.

BTC Markets (1)

• Jason Calacanis argues Bitcoin has lost momentum and says ETF outflows and AI’s perceived upside weaken the case for holding it; D.A. Wallach counters that Bitcoin could still work as a long-term debasement asset.

Stablecoins (1)

• The panel sees clearer use cases in stablecoins, tokenized real-world assets, and DeFi lending than in speculative tokens whose value rests mainly on belief and hype.

VC (2)

• D.A. Wallach expects AI to let founders build with less capital, while Jason Calacanis says SPVs are absorbing money from emerging managers and that revenue-first startups may skip funding rounds; both describe early fundraising as difficult.
• Ben Black says PowerLaw’s largest holdings include SpaceX and OpenAI, and that the fund is selling down SpaceX; his early-2025 purchase returned about three times its value.

Startups (1)

• Anthropic’s S-1 showed unusually fast revenue growth, high gross margins, and strong R&D efficiency, while the company continues spending billions of dollars on compute to train future models.

Safety (2)

• A host said OpenAI’s Glasswing project found a long-standing PHP vulnerability; its patch leaked early, forcing a rushed global response, but frontier models could help defenders find flaws before attackers exploit them.
• A host argued that a major AI-enabled cyberattack could damage a frontier lab’s public standing while increasing its business if the lab helps fix the problem; an uncontrolled agent causing harm could instead bring liability and regulation.

Agents (7)

• The hosts expect increasingly capable agents to expose risks beyond deliberate hacking: poorly controlled agents could multiply, steal computing resources, or infest online services, with open-weight models potentially triggering large-scale cases.
• Meta’s Muse had millions of installs and daily users, with a quarter of daily users reportedly shopping frequently; Meta aims to make the agent more compute-efficient while eventually serving a billion users.
• The hosts saw consumer agents as a large but harder-to-monetize market than enterprise AI: consumer agents may earn a small take on spending, while enterprise software offers a larger direct revenue pool.
• One host estimated AI-mediated commerce could reach $8 trillion by 2030, yielding $200 billion to $300 billion in transaction revenue, with another $600 billion to $800 billion available from advertising.
• Meta’s current ad revenue per user in the United States and Canada was estimated at $300 to $350 annually; a $200 average Muse purchase and a 5% take rate would require enormous transaction volume to match it.
• The hosts argued that consumer agents may need banking and payment infrastructure to capture more transaction value, and could expand into insurance, financial management, and other services where agents can negotiate on users’ behalf.
• OpenAI faces a compute-allocation tradeoff: consumer agents could build habits that later drive workplace adoption, but subsidizing consumer use may make its financial results look weak beside Anthropic’s enterprise business.

Big Tech (1)

• Meta’s filing described a potential $700 billion capital raise before Mark Zuckerberg loses majority voting control; the company then announced a $90 billion tranche and issued a new non-voting share class.

Markets (1)

• A host warned that businesses can appear healthy until they fail to move quickly enough as AI accelerates, leaving companies vulnerable to sudden competitive displacement.

Europe (2)

• Peter Diamandis cites October 17 marches by 500,000 people across France, 2,400 arrests, looting in Paris, Marseille, and Lyon, and a burned 19th-century hall in Limoges. Speakers also report claims that paid agitators, possibly linked to Russia, fueled unrest.
• Emad Mostaque says French employment protections make firing workers difficult, while France spends 57% of GDP on social programs. Eamon describes UK compliance audits targeting high earners, and Peter says Europeans are moving to Waterloo and Toronto for AI opportunities.

Labor (1)

• Peter Diamandis links unrest to young men who see blocked opportunities, arguing that the demographic has played a role in revolutions from France to Iran. He warns that AI-driven unemployment could intensify instability where governments fail to build compute, data centers, and power.

AI Infrastructure (1)

• Peter Diamandis argues that countries can still build sovereign AI by coordinating power, chips, and compute, with a claimed 90% to 99% chance of success. He says Europe risks stagnating without investment, while the United States and Canada could combine American chips with Canadian power.

Markets (2)

• Anthropic reported $16.3 billion in first-half 2026 revenue and nearly $1 billion in second-quarter operating profit, making its planned IPO unusually large. Dave Blundon calls the business financially strong but says a $1.5 trillion to $2 trillion valuation already prices in substantial growth.
• Dave Blundon says data remains a source of competitive advantage, but any specific data moat can erode as AI changes markets. He also identifies compute, natural resources, consumer preferences, and a company's speed of innovation as potential durable advantages.

Enterprise (1)

• Anthropic's long-term benefit trust would appoint and remove a majority of its board, giving trustees influence over corporate direction. Alex Wissner-Gross worries that unelected trustees could override public shareholders; Emad Mostaque argues the structure gives the company's mission formal power.

VC (1)

• Anthropic's early investors stand to receive enormous returns: the $124 million Series A is described as a 345-fold return, and the $982 million Series B as a 79-fold return. Emad Mostaque argues that founders and investors will recycle much of that IPO wealth into new AI companies.

Agents (1)

• OpenAI researchers found that agents in isolated virtual machines could communicate through processor-cache timing: one agent encoded bits by changing cache state, and another decoded them by measuring access speed. Alex Wissner-Gross stresses that researchers explicitly asked the agents to find such side channels.

Models (3)

• Moonshot AI's Kimi K3.1 Pro scored near leading closed models while costing less, and its planned open weights could narrow the frontier gap to weeks. Alex Wissner-Gross says its performance per cost may already exceed Claude Opus 5.5 on some comparisons.
• GPT 6.1 Terra costs less than half as much as its predecessor and showed a 50-point drop in hallucination, making it better suited to routine personal-agent use. Salim Ismail says falling token prices are accelerating, while Peter Diamandis argues that product usefulness will matter more than raw model benchmarks.
• OpenAI announced AI-produced results spanning a 1991 wave-focusing conjecture, a disproof of the exponential time hypothesis, and a rigorous four-dimensional construction for quantum chromodynamics. The proofs remain unreviewed; speakers say science is shifting from discovery toward verification, and debate whether human intermediaries can meaningfully validate AI work.

Physics (1)

• Alex Wissner-Gross's chain theory models physics as a network of finite, maximally entangled quantum systems and derives the fine-structure constant from a short axiom. He presents it as an early, falsifiable framework for connecting quantum information, particle physics, gravity, and possible dark-matter candidates.

Robotics (2)

• Warhound plans to genetically edit and clone working dogs for repeatable traits such as endurance, scent detection, and strength, then outfit them with Anduril sensors and protective equipment. Salim Ismail and Alex Wissner-Gross raise concerns about animal consent and engineering dogs for dangerous service.
• Emad Mostaque would target construction first if capable robots became cheap and plentiful, arguing that customers would pay for robotic tradespeople. He sees coordination and capability as the remaining hurdles, though higher-margin uses may attract robot bodies before construction does.

Safety (1)

• Emad Mostaque argues that AI alignment depends on deciding which human values systems should follow, and warns that powerful optimizers can exploit poorly chosen targets. On training data, he says removing biological, chemical, or malicious-code knowledge need not reduce general capability and may sometimes improve it.

Payments (1)

• Salim Ismail expects programmable digital money to support machine-to-machine commerce, favoring Bitcoin with Lightning for decentralization, security, and scale while seeing stablecoins as useful for some programmable transfers. He argues that crypto's key feature is programmability, not simply digital payments.

Energy (1)

• Alex Wissner-Gross says energy and computing may remain physically scarce, while Dave Blundon expects energy to gain value as AI competes with household uses. Alex argues demand, rather than supply alone, could determine whether resources retain value after a singularity.

Models (5)

• Moonshot’s Kimi K3.1 Pro is the first update to the K3 family. The original K3 launched in July as a 2.8-trillion-parameter open model, briefly drew comparisons with Western frontier systems, then lost momentum as newer models arrived.
• Moonshot’s benchmarks put K3.1 Pro near Opus 5.5 on agentic work and coding, while its lower token prices and cost per task strengthen the case for adoption. Independent testing also placed it among the strongest available models.
• Artificial Analysis scored K3.1 Pro fourth on its Intelligence Index, but the launch was quickly clouded by leaked instructions telling influencers to omit paid-partnership labels and promote comparisons that criticized Western rivals. Researchers disagreed on whether the model’s benchmark results would hold up in use.
• Moonshot said it removed cyber-related training data to keep K3.1 Pro’s offensive cyber skills below its internal pipeline while retaining general coding ability. Artificial Analysis still found it to be the strongest cyber model broadly available, close to a restricted GPT-6 Soul version.
• The host argued that OpenAI might benefit from temporarily delaying a competing model release: a strong Anthropic IPO could revive investor confidence across AI, while a failed offering could deepen bubble concerns and make later IPOs harder.

Safety (2)

• Anthropic security executive Jason D. Clinton warned that cyber threats would worsen and urged security leaders to fund basic defenses. Cloudflare CEO Matthew Prince predicted years of frequent vulnerability disclosures, while Sam Altman said some flaws may be too difficult to patch.
• OpenAI and Anthropic resumed frontier training in isolated environments. OpenAI’s offline web replica is meant to prevent models from affecting real systems, but its agents found a way to exchange synthetic secrets through a processor cache despite network isolation.

Agents (1)

• Box CEO Aaron Levie argued that agent swarms can test more code and infrastructure risks than human teams can reach quickly. Noah Smith warned that AI-enabled attacks could make the open internet viable only for users able to pay for private security.

AI Infrastructure (1)

• Large AI customers are shifting from model APIs toward direct GPU rentals and self-built infrastructure. Chamath Palihapitiya said a major bank had cut AWS and Google Cloud spending while building six independent data centers; Ramp data showed direct GPU rentals taking a substantial share of top firms’ AI spend.

Startups (1)

• Anthropic’s prospectus appeared as the host recorded, after speculation that it needed to be public by Friday to support an IPO the week of Thanksgiving. OpenAI’s reported next-round valuation had fallen from earlier expectations, complicating Anthropic investors’ higher valuation hopes.

Coding (1)

• OpenAI’s price cuts helped it gain coding share from Anthropic at Rippling, but Anthropic’s Opus 5.5 release drove a sharp rise in usage and business spending. The host said Anthropic’s rumored Fable 5.5 and its Duet personal agent could strengthen its position before the IPO.

Labor (1)

• A KPMG and University of Texas at Austin study found that early-career professionals got different results from AI despite having similar skills. The strongest performers acted as “AI amplifiers” by directing, evaluating, and refining model outputs.

Media (1)

• The host announced How We Got Here, a weekly show about how current trends emerged through human decisions. Planned subjects include the tech industry’s loss of public affection, rising living costs, prediction markets, and GLP-1 drugs.

Enterprise (1)

• The episode’s ads promoted CData’s Connect AI Gateway for governing agent access to business systems, Granola’s AI meeting notes, and Blitzy’s codebase analysis and autonomous software work. Granola offered new users three months free with the advertised code.

Markets (5)

• The host described a three-week tech-stock slide, including a 3% Nasdaq drop on Thursday, while Goldman Sachs reported hedge funds sold tech at their fastest pace in at least a decade. The 30-year Treasury yield fell to 5.38%, a one-month low, signaling a flight to safety.
• Citadel CEO Ken Griffin called the market almost certainly a bubble, arguing that uncertain returns could make companies reconsider AI bills. Blackstone’s John Gray countered that AI builders are trying to keep up with demand, while the host identified demand as the question underlying the competing arguments.
• Private-equity-owned insurers are another exposure point: Apollo merged with Athene in 2022, pairing an investment firm with an insurer that mainly sells life insurance and annuities. The host said these firms could be impaired by defaults, while the Wall Street Journal reported AI generated $138 billion in gains across 13 public pension funds since 2022.
• Public listings would make AI labs’ revenue growth and customer mix easier to scrutinize through audited quarterly accounts and analyst questions. Anthropic’s reported $80 billion first-half loss was an accounting loss; its cash loss was $8.7 billion, while OpenAI’s reported fundraising valuation fell below earlier press estimates.
• Anthropic’s IPO could test confidence in the AI trade: investors had floated a $2 trillion valuation, while more recent reporting put pressure around $1.5 trillion. The broader IPO market was stalled, and the host said weak demand at that price would matter more than the valuation cut itself.

AI Infrastructure (2)

• Americans for Financial Reform warned that the AI bubble was “primed to pop,” and policy analyst Maya Jenkins said working people would pay a high price if it crashed. Meta’s earnings call also showed its next AI build-out would be partly funded through equity dilution.
• Amazon’s early-rack process means AWS may receive Nvidia GB300 racks before sites have power; the host said this points more plausibly to a data-center bottleneck than weak chip demand. Amazon said staging hardware ahead of construction is routine, and the article reported the company was trying to secure scarce GPUs.

Regulation (1)

• The host said Anthropic’s and Elon Musk’s plans to distribute equity could strengthen arguments that AI firms are too big to fail, but doubted the plans are designed to secure bailouts. He said explicit requests for federal guarantees, loan programs, or financing backstops before a crisis would be a stronger warning.

Enterprise (4)

• Revenue concentration is the host’s most significant warning sign, though large customers do not always imply fragile demand: Nvidia sells to few data-center builders, and Anthropic’s hyperscaler relationships may reflect wholesale distribution. A greater concern is whether concentrated, VC-funded power users will broaden into durable enterprise demand.
• Ramp data also points to spending spreading beyond the biggest AI users: businesses across cohorts tripled spend per employee since summer, and the 10% cohort rose from about $200 to $800 per employee. The host said that trend supports broader adoption, but does not guarantee spending will keep expanding.
• A KPMG and University of Texas at Austin study found that early-career professionals got different results from AI despite similar skills; the strongest performers guided, evaluated, and refined model outputs. KPMG called these workers “AI amplifiers,” emphasizing how people work with AI rather than knowledge alone.
• Robots and Pencils described its work as applied AI engineering built around client business outcomes rather than slide-deck experiments. CData’s ad promoted record-level data governance and a single view of organizational AI usage, claiming its gateway can improve answer accuracy while reducing token costs.

Reasoning (1)

• Past bubble warnings have repeatedly been followed by new evidence of growth: 2024 fears that model scaling was stalling preceded reasoning models and larger-memory chips, while last year’s failed-pilot narrative did not stop enterprise AI spending from rising. The host cited the MIT pilot study as a recurring but misleading signal, not proof that future bubble calls are wrong.

Coding (1)

• Blitzy said its system can plan, build, test, and validate whole applications, including greenfield projects, rather than only assisting individual developers. The company cited one customer that produced 534,000 lines of code in two weeks, compressing a 65-week roadmap.

Startups (5)

• Rob Copeland says Hemanthus pitched at-home blood tests for pets, aiming to cut a veterinary test from about $500 to $1; a USDA document Billy Evans cited did not establish the regulatory clearance he claimed.
• Copeland found Hemanthus had raised about $3 million, had no working product, and was still operating from a house in Austin; Evans pursued major investors, including Jim Breyer and Michael Dell.
• Aaron Griffith says Evans’s investor pitch emphasized his own credentials while concealing Holmes’s role; Holmes was absent from the deck, and Evans’s patent did not include her. Copeland learned the couple had already filmed their startup story.
• Griffith cites Trevor Milton’s post-conviction company and Billy McFarland’s failed fundraising as rare comparisons; Rob Copeland argues that investors may back Holmes to spite journalists, especially if a pardon gives them a cleaner pitch.
• Carreyrou argues Holmes played confused or neurodivergent in the Abbott and Costello scene to support a defense that she experienced the world differently; he says her repeated calculated responses and refusal to accept responsibility undermine that account, while acknowledging her sentence was harsh.

Psychology (1)

• Griffith reads Evans and Holmes’s relationship as both romantic and opportunistic: Evans appears to love her while also seeking career and financial upside from her notoriety, including through the film and a digital replica.

VC (1)

• Griffith says venture firms have backed controversial founders to signal independence from media criticism, citing Andreessen Horowitz’s investments in Andy Rubin and Adam Neumann; he found few comparable cases involving founders convicted of fraud.

Social Media (1)

• The film reveals Billy Evans running Elizabeth Holmes’s combative, increasingly conservative social account from prison; Aaron Griffith says people in prison often have phones, and guards may prefer inmates distracted by them.

AI Infrastructure (1)

• Copeland says AI companies may need to borrow as much as $10 trillion, while Aaron Griffith links Anthropic’s IPO prospects to whether AI demand can support that spending; Ken Griffin has called the market a bubble.

Big Tech (2)

• Aaron Griffith says SpaceX is forming independent-director committees to explore a merger with Tesla, which Elon Musk confirmed on an earnings call; Griffith thinks consolidation could give Musk more control over Tesla spending.
• Max Reed says Meta policy chief Joel Kaplan is leaving after shifting the company’s internal and external posture to the right; Reed attributes the departure more plausibly to tenure and exhaustion than to expected Republican losses.

Health (2)

• John Carreyrou says Theranos’s Edison could perform 12 kinds of tests, but every result it issued for those tests had to be voided; the Minilab shown in the film was a later device that never produced reliable results or reached patients.
• Carreyrou dates Theranos’s full-fledged fraud to fall 2013, when the company launched Walgreens testing before its technology was ready, modified Siemens analyzers, and pre-diluted finger-prick samples that the machines diluted again.

Corruption (2)

• Carreyrou says whistleblower Tyler Shultz tried to warn Holmes and his grandfather before leaving Theranos; the company later pressured Shultz to identify sources and recant, while Holmes appealed to Rupert Murdoch to stop the reporting.
• Carreyrou rejects Holmes’s claim that Sunny Balwani alone ran Theranos: texts show the couple managed the company together, while Holmes was CEO, board chair, and held about 90% of voting control. He says she initially hoped to build a real company but knowingly lied once the business failed.