Price:

China loses a cloud route to advanced AI chips

Nov 22, 2026Summary from 5 podcasts.
  • Commerce’s rule targets Chinese firms’ remote access to advanced chips through data centers abroad.
  • Data centers in five countries need licenses, and existing facilities have 90 days to apply.
  • Hosts see hardware controls as a practical chokepoint, but debate whether they preserve the U.S. advantage.

The Commerce Department has tightened export controls to close a route Chinese AI companies used to reach advanced U.S. chips: renting computing power from data centers outside China. The restriction targets remote access, not just direct chip shipments, making cloud facilities part of the effort to limit Chinese labs’ access to leading-edge hardware.

Under the new rule, data centers in Singapore, Malaysia, Thailand, Indonesia and Brazil must obtain export licenses for advanced AI chips. Existing facilities have 90 days to apply. Shailesh Chitnis, a global business writer for The Economist, said the measure closes a loophole that had allowed Chinese firms to rent chips through cloud providers in places including Malaysia and Singapore.

The change is already affecting plans and investment. On The AI Daily Brief, Nathaniel Whittemore reported that Chinese AI company Moonshot canceled a $150 million deal for access to Nvidia chips through an Indonesian data center. Tencent president Martin Lau warned the controls could delay model releases, while Alibaba CEO Eddie Wu said they would affect products and services planned for next year. TSMC and Samsung also suspended some chip deliveries to Chinese customers, according to the show.

The rule reflects a deeper argument in Washington over whether selling chips to China helps the United States or its rival. David Sanger, on The Daily, said Nvidia CEO Jensen Huang argued that continued sales could make China dependent on American technology. Scott Bessent’s camp favored restricting access to preserve the U.S. lead. Three export-control bills are also under negotiation in Congress as part of the National Defense Authorization Act.

The controls may make advanced computing harder for Chinese companies to obtain, but they cannot settle the competition. Sanger noted that China is working to develop domestic chip alternatives. On Moonshots, Dave Blunden argued that software models are difficult to contain, while hardware and data centers offer more practical points of control. The rule puts that strategy into practice, and raises the cost for Chinese firms trying to catch up.

Source Intelligence

- Deep dive into what was said in the episodes

Anthropic Goes Public. Markets Keep Selling. | Rabbit Hole Recap 429 • Nov 20

  • The hosts said new US export controls could bar overseas GPU-cloud providers from serving Chinese firms with models running on US-made chips; Trump's threat to tax data centers built abroad added to fears of central planning and uncertain investment rules.
Also discussed on this episode: (15)

BTC Markets (4)

  • The hosts describe Bitcoin as caught in thin, apathetic markets: they link the drop to Binance unwinding positions, while ETF selling and liquidity rotating toward Anthropic remain possibilities rather than confirmed causes.
  • Bitcoin traded near $59,420 as network activity looked subdued: mempool counts varied sharply by source, fees were low, and the next difficulty adjustment was projected upward despite a recent downward correction.
  • The hosts said Bitcoin was 408 days past its prior price peak, near historical bear-market duration ranges; they watched $57,717 and a $55,000-$58,000 retest zone, with Strategy's average cost near $75,000.
  • The hosts framed Bitcoin as a potential safe haven if central banks keep weakening fiat currencies, but said the immediate flight to safety was benefiting the US dollar as investors exited riskier assets.

ETFs (1)

  • Galaxy Research reported record advisor and wealth-platform exposure to Bitcoin ETPs, while James Check said Bitcoin settlement value had risen since summer even as price fell.

Models (1)

  • The hosts called Anthropic's public-market valuation difficult to justify against its revenue and Samsung's scale; they argued that commoditizing AI tokens could leave frontier labs with thin margins unless they monetize major scientific or health breakthroughs.

AI Infrastructure (1)

  • The hosts questioned NVIDIA's leadership change and argued the AI trade faced political and strategic uncertainty; one host cited Travis Calacanis's claim that lobbyists tied to Anthropic helped push Jensen out, and linked Jensen's support for open-weight models to the dispute.

Markets (1)

  • The hosts described a broad risk-off move: the dollar strengthened as equities, crypto, and Bitcoin ETFs sold off, and they expected volatility to remain elevated into December.

Macro (1)

  • The hosts said US housing had stalled as mortgage rates near 7% deterred buyers and homeowners with cheaper loans from moving; they debated whether the Fed would cut rates as energy prices eased and markets weakened.

Censorship (1)

  • The hosts questioned alarming reports about dirty bombs, cyberattacks, and grid failure, but argued that crises can expand government surveillance and strengthen pushes for tokenization, digital IDs, and central bank digital currencies.

Privacy (1)

  • The hosts said third-party breaches expose risks beyond a company's own systems, citing leaks affecting Chase customers and Uber drivers; they argued that self-custody also depends on users' devices, technical skill, and ability to patch vulnerabilities.

Payments (1)

  • Block's Square data showed repeat Bitcoin spending at merchants, especially food and drink businesses; the host used the findings to argue that tax and accounting complexity supports a de minimis exemption for payments.

Protocol (1)

  • Mikhail Kudinov and Jonas Nick advanced Shrinks, a proposed signature scheme intended to offer quantum resistance and an alternative to current schemes; the hosts described the work as research that still needs review and a path to protocol adoption.

Regulation (1)

  • The hosts expected Democratic lawmakers to target crypto figures associated with Trump and urged Bitcoiners to distinguish Bitcoin from the wider crypto sector; they saw party divisions, DSA organizing, and social-media campaigning reshaping the political fight.

Media (1)

  • A video from the American Communist Party of New York invoked the 1971 break between the dollar and gold; the hosts saw it as evidence that political creators across the left and right are adapting their messaging to current social-media formats.

Why You Should Try Gemini Neon (Even if You Love Fable) • Nov 20

  • A new Commerce Department rule closes remote access to AI chips through data centers in third countries, cutting off a route Chinese labs used legally. Tencent and Alibaba warned of delayed products and model releases, while Moonshot canceled a planned Nvidia-chip deal in Indonesia.
Also discussed on this episode: (14)

Markets (2)

  • Anthropic’s IPO raised $107 billion and made it the world’s seventh-largest publicly traded company. Shares closed above the IPO price despite a weak market, giving the company a successful debut and substantial funding for future model development.
  • The host described volatile trading and mixed investor demand: retail buyers traded both sides, while Sequoia, Iconiq, Lightspeed, and Altimeter bought shares. Anna Wong questioned whether AI demand creates durable pricing power; the Nasdaq 100 fell more than 2%.

Safety (3)

  • Dario Amodei framed Anthropic’s public-market debut around medical progress and safety, saying powerful AI could compress a century of scientific and medical advances into a decade. He pledged to pace training and deployment until the company had confidence systems remained safe and under human control.
  • The White House replaced cybersecurity director Sean Cairncross with Kristen Davies amid conflicting accounts of his departure. Cairncross had helped shape the administration’s voluntary frontier-model testing policy; the host said the change could signal a shift in AI and cyber policy.
  • JPMorgan Chase disclosed a vendor cyberattack that may have exposed customer names and email addresses, separate from a banking outage. The European Commission said a hack exposed staff data and disabled payroll and other systems; officials attributed it to the open-source Artex agentic hacking tool.

Regulation (1)

  • Anthropic’s founders retain control of shareholder votes through a golden share that directs 51% of votes through a governance council. Ann Lipton said the Long-Term Benefit Trust could use a small group of directors to keep the company on mission, but questioned how likely that check was.

Chips (1)

  • AMD shares fell after Hunterbrook Media alleged that a blacklisted Chinese server maker continued receiving AMD firmware and that China and Hong Kong account for as much as 20% of AMD revenue. AMD denied the report; the host said firmware alone did not prove dealings with the company.

China (1)

  • Export controls are squeezing Chinese AI investment as companies face weaker monetization and limited computing access. Tencent’s revenue growth missed expectations, while Alibaba reported stronger cloud growth but forecast a slowdown as it increases use of in-house chips.

Models (5)

  • Google released Gemini 4 Neon as its fourth model in two months, after Argon, Carbon, and Hydrogen. Users praised its writing and personality after initially dismissing another rapid release; DeepMind researcher Tianfu Fu described Neon as a smaller model positioned near Opus and Sol.
  • Gemini 4 Neon led several model benchmarks: Artificial Analysis scored it 61 on the Intelligence Index and made it the Coding Agent Index leader. It beat Opus 5.5 by more than seven points on Cursor Bench and 14 points on Automation Bench, while coding results against Opus were close.
  • Artificial Analysis found Neon used more tokens per task than Carbon but finished faster; its task cost sat in the middle of the tested range. Daria Anutmaz reported building a playable tower-defense game quickly, while Pavel Hurin’s bug-hunt test put Neon first.
  • The host found Neon useful because its answers often approach problems differently from Claude and OpenAI models, even when it is not clearly better. Neon helped reshape an oil-episode outline that other models kept turning into a sequence of facts.
  • The host said Neon can reverse its position after mild pushback, making extended brainstorming unreliable despite its value as a source of fresh perspectives. Google’s agent harnesses also lag Anthropic and OpenAI, and Neon can explain coding work too technically.

Agents (1)

  • Apple is developing a personal agent that could browse, use apps, run commands, write and debug code, and choose tools across a user’s devices. Mark Gurman reported that Apple is also making third-party computer control harder on Macs, raising concerns about competition and consumer choice.

How Business AI Use is Changing • Nov 17

  • A new Commerce Department rule requires data centers in five countries to obtain export licenses for advanced AI chips, targeting remote access by Chinese labs. Existing facilities have 90 days to apply, while Treasury and Commerce publicly diverged over the rule.
Also discussed on this episode: (16)

Big Tech (2)

  • NVIDIA’s board is focused on an independent investigation into Jensen Huang and is not discussing his return, despite a petition and shareholder pressure. Donald Trump said he would not object to Huang returning and disputed claims that national security was compromised.
  • The Trump administration plans to designate NVIDIA, Amazon, Microsoft, OpenAI, and Anthropic as national champions, with expedited export approvals, government help selling abroad, and access to federal financing. The program aims to bundle chips, models, and other AI products for international sales.

AI Infrastructure (1)

  • Sam Altman says AI demand and revenue can sustain infrastructure investment without Jensen Huang personally directing NVIDIA’s financing. NVIDIA guaranteed $100 billion for OpenAI’s Ohio facility, while reports say Huang had already begun demanding upfront chip payments and pulling back commitments.

Safety (2)

  • OpenAI’s Apollo Research pilot embeds six evaluators for four months to test whether internal models undermine alignment controls. Apollo can inspect reasoning traces, reinforcement-learning rewards, and model checkpoints, but OpenAI can withhold information, restrict staff access, and redact confidential material from reports.
  • Apollo Research says OpenAI’s access and disclosure terms fall short of credible independence; Neil Nanda and Stephen Adler also criticized limits on evaluator access and employee reporting. The host argues external standards and genuinely independent third parties are needed, rather than access terms set by the labs themselves.

Banking (1)

  • JPMorgan Chase traced a widespread outage affecting bank logins, ATMs, and transfers to a software bug in on-premises hardware, and said it was not a cyberattack. Vivek Wadhwa and other commentators questioned the denials after outages also hit Microsoft, the FAA, Workday, and major banks.

Enterprise (9)

  • Three enterprise AI studies from Databricks, Arcana Research, and Ramp point to a buyer’s market: companies have more model choices and are becoming more willing to switch providers. The host cautions that the samples are limited, particularly Ramp’s tech-forward customer base and Arcana’s Global 2000 survey.
  • Ramp found token use rose much faster than spending, while average token prices at OpenAI and Anthropic fell. The host says cheaper competitors contribute to price pressure, alongside provider efforts to make models more efficient.
  • Databricks found companies increasingly use multiple model providers, and Arcana found substantial willingness to switch from OpenAI and GitHub Copilot. The host argues enterprises should track cost per successful task, not token price alone, as routing tools make model switching easier.
  • Arcana found AI had become an eight-figure budget item among surveyed Global 2000 leaders, and most expected spending to rise in 2026. Yet two-thirds lacked visibility into model and API costs, leaving many companies less equipped to manage their budgets or negotiate effectively.
  • AI spending is highly concentrated: Ramp’s top 1% of firms spend roughly 625 times the median per employee, while Databricks found a 281-fold gap between median and 99th-percentile users. Coding dominates current spend, with engineering teams accounting for most Ramp expenditure.
  • Anthropic leads several enterprise indicators: more Ramp businesses pay for Anthropic than OpenAI, most proprietary-model tokens on Unity Gateway go to Anthropic, and Arcana respondents most often named it an exciting vendor. Customer-service use is also growing across Ramp and Arcana respondents.
  • Ramp token-spending growth slowed through September before rising in October, while Arcana respondents still expected 20% to 40% annual increases. The host says October’s pace would imply roughly 22% annual growth if sustained, broadly within Arcana’s forecast.
  • Frontier-model spending rose on Ramp even as Databricks’ token share for Fable fell sharply; meanwhile, Opus gained share. The host reads the split as a possible product-specific shift, with Opus gaining ground rather than evidence of declining interest in frontier models overall.
  • The host says enterprises have more negotiating leverage as model competition intensifies, provided they build the infrastructure to switch providers. Whether today’s highest-spending firms become closer to the norm will shape AI market size and whether infrastructure spending produces returns.

Agents (1)

  • Arcana ranked autonomous agents fourth among enterprise AI pilots and plans but last among tools actually deployed. The host says agents remain aspirational for many average enterprises, despite expectations that 2026 will be remembered as the year agents became practical.

5 Questions That Could Reshape AI • Nov 16

  • Congress may tighten chip export controls through three bills under negotiation for the National Defense Authorization Act. The proposals address routes such as remote access to advanced chips through Southeast Asian data centers; the host said OpenAI and Anthropic now support tougher controls.
Also discussed on this episode: (12)

Markets (3)

  • Anthropic’s IPO has become a test of AI’s public-market case: reported demand is far beyond the deal size, while investors still lack a clear way to value a company with enormous compute costs.
  • Anthropic’s quarterly reports could become the clearest public evidence of whether AI demand can repay infrastructure costs. Kim Forrest called the company a likely poster child for business AI, while Mike Smith said the industry needs unprecedented revenue growth within two to three years.
  • Anthropic’s IPO roadshow included a claim that the company could reach $1 trillion in annual recurring revenue by late 2028 and a pitch that it could displace hyperscalers, labs, and other industries. The host described that ambition as part of Anthropic’s winner-take-all strategy.

AI Infrastructure (1)

  • Dario Amodei called compute Anthropic’s deepest business dilemma: the company is buying ahead to avoid shortages, but its spending could outrun the revenue AI generates. He said buying too little previously cost customers, research cycles, and growth.

Models (3)

  • Anthropic has accelerated product releases, moving from a productivity suite to a video-meeting app after Opus 4.5 enabled faster feature shipping. David Sacks warned that the company treats recursive self-improvement as a business advantage, while its roadshow promised AI-driven cures for major diseases.
  • A rival model release could steal attention from Anthropic’s IPO. Token Gremlin claimed OpenAI’s GPT 6.2 Astra was near release, while Alibaba’s Qwen 4 Max advertised a million-token context window and strong benchmark results that the host said looked weaker in real use.
  • OpenAI’s frontier-model gap has reached a reported record, feeding claims that the company has fallen behind Anthropic. The host expects a release soon, but OpenAI employee Rune said new models need federal safety approval and can trigger export-control action.

Big Tech (2)

  • A campaign to restore Jensen Huang gained support from shareholders, employees, and technology executives, with allies lobbying the White House. Gene Munster raised his estimate of Huang’s return sharply; Balaji Srinivasan argued shareholders could force a board change under NVIDIA’s bylaws.
  • Microsoft and FAA outages last week were reportedly unrelated to AI, but the host said many Americans assumed AI was responsible. The outages showed how quickly AI has become the default explanation for failures, even when simpler causes are offered.

China (2)

  • The Huang scandal’s spy narrative remains contested: the host said David Sacks amplified reporting that CIA and Justice Department officials were unsure Alice Huang was a Chinese spy. SNL focused on the sex-scandal angle and market fallout, leaving the chip-security allegations largely out of its satire.
  • Treasury Secretary Scott Bessent questioned whether NVIDIA knowingly let chips reach China through shell companies. SemiAnalysis’s Dylan Patel alleged that more than 1.5 gigawatts of NVIDIA chips in China violate U.S. law; NVIDIA’s earnings call was expected to face questions about those shipments.

Regulation (1)

  • Howard Lutnick faces scrutiny over alleged delays in adding Megaspeed to the entity list and unusually close Commerce Department ties to NVIDIA. The New York Times reported NVIDIA’s general counsel inserted comments into draft policies; Lutnick agreed to testify before the Senate Banking Committee.

What Trump, Putin and Xi Left on the Table • Nov 20

  • David Sanger describes a split over chip exports: Jensen Huang argued that selling U.S. chips would make China dependent on American technology, while Scott Bessent’s camp favored restricting access to preserve the U.S. lead. Trump had lifted Biden-era controls after Huang and Commerce Secretary Howard Lutnick pressed for sales.
  • David Sanger says Chinese firms accessed advanced U.S. chips through data centers in Singapore and Malaysia, financed by Chinese companies and connected to China by high-speed links. The administration revised a rule to restrict remote access through such facilities.
Also discussed on this episode: (12)

Diplomacy (3)

  • David Sanger says Xi Jinping, Donald Trump, and Vladimir Putin used their first joint meeting to project power, but the roughly 80-minute summit produced no announced agreement. Its Yalta imagery echoed the 1945 meeting, which lasted eight days.
  • David Sanger says Trump wanted arms control, Putin wanted to discuss Ukraine, and Xi initially resisted the meeting before embracing his role as host and convener.
  • David Sanger describes Trump as deferential toward Xi and Putin at the APEC summit, where Trump met each leader separately before the three-way meeting; the White House provided no briefings during the trip.

AI Infrastructure (4)

  • David Sanger says the leaders left artificial intelligence off the announced agenda, despite urgent questions about regulation, national security, and the risks of a fast-moving technology.
  • David Sanger warns that advanced AI could help states or terrorist groups exploit flaws in electric grids, nuclear plants, and water systems, or assist in developing bioweapons. He also sees risks when AI becomes embedded in everyday products controlled by foreign companies or governments.
  • David Sanger argues that AI’s speed and reach are outpacing political decision-making, and that geopolitical power may come to depend more on effective AI use than on the size of a country’s nuclear arsenal.
  • Trump later told NBC that he and Xi discussed AI and agreed on minimal guardrails for superintelligent systems, but said they did not define what those guardrails would be.

China (1)

  • Natalie Kittrowev raised Jensen Huang’s reported resignation from NVIDIA after he admitted a relationship with a woman U.S. intelligence believed was a Chinese spy. David Sanger said Trump mentioned Huang to Xi at dinner, but no response from Xi was known.

Chips (1)

  • David Sanger says China has struggled to manufacture the most advanced chips because production depends on specialized equipment, including machines largely made in the Netherlands and Japan. Biden-era agreements restricted shipments of that equipment to China.

Models (1)

  • David Sanger says Chinese developers are pursuing smaller open-weight models that can be adapted for specific uses and integrated into products, rather than relying only on the largest closed models. He argues that broad adoption could give China leverage in markets for everyday goods.

Startups (1)

  • The closing report said Anthropic’s seven founders were each worth about $25 billion after a surge that eased investor concerns about the AI industry and created a new class of wealthy Americans.

Safety (1)

  • The closing report said White House cybersecurity official Sean Cairncross was leaving after technical outages disrupted email, banking transactions, and commercial air traffic. Officials said the outages were not cyberattacks, but had not clearly explained their cause.

Democrats Promised Lower Prices. What Can They Actually Do? • Nov 17

  • The Trump administration said it would close a loophole letting Chinese companies access advanced AI chips, including Nvidia chips, by renting them from data centers in Southeast Asia and elsewhere. The proposed rule aims to prevent remote access that bypasses restrictions on chip purchases.
Also discussed on this episode: (13)

Energy (3)

  • Rebecca Elliott says the Iran war has raised fuel prices by disrupting oil shipments through the Strait of Hormuz and damaging refineries. Even a formal end to hostilities may not lower prices quickly if mines, drone threats, shipping costs, and refinery repairs keep supply constrained.
  • Elliott says Congress could suspend the 18-cent federal gasoline tax or restrict fuel exports, but neither is a reliable fix. Export limits might lower prices briefly, then prompt producers to cut output and push prices higher over time.
  • Elliott says US oil production prevents the shortages seen in some other countries, but American fuel prices still track a global market. If the war ends and Strait of Hormuz shipping becomes safer, new production in the United States and South America could create an oil glut and bring prices down quickly.

Health (3)

  • Margot Sanger-Katz says expiring Obamacare subsidies have made coverage more expensive for people buying their own insurance, while Republican Medicaid changes are expected to leave millions uninsured and strain hospitals and state budgets. Democrats plan votes to restore the subsidies and reverse the Medicaid cuts.
  • Sanger-Katz says Senate Democrats could use reconciliation to pass healthcare legislation with a simple majority, rather than the 60 votes normally needed to overcome a filibuster. President Trump could still veto the bills, and Democrats do not have a veto-proof majority.
  • Sanger-Katz sees room for bipartisan action on prescription drug costs, including a proposal to cap monthly insulin payments at $35; TrumpRx and Trump’s deals with drugmakers could create common ground. She says Medicare for all faces a steep path under Trump and within a divided Democratic Party, though Bernie Sanders plans to press the issue as Senate Health Committee chairman.

Markets (1)

  • Ben Casselman says housing costs reflect a long-running shortage, especially in coastal cities with strong job markets. A major housing bill passed this year may encourage some construction, but local building rules, homeowner opposition, and the scale of the shortage limit what federal policy can accomplish quickly.

Trade (1)

  • Casselman says tariffs raised prices for some goods and cost households hundreds to a couple thousand dollars a year, though tariffs were not a major driver of inflation. He expects Democrats may find Republican support to roll back tariffs on Canada, but lower tariffs would not guarantee that companies reduce prices.

Labor (1)

  • Casselman says raising the federal minimum wage could quickly lift pay for millions of workers, and state and local increases have not produced the severe job losses critics predicted. Democrats still face disputes over the target wage, phase-in period, and small-business exemptions, as well as the prospect of a Trump veto.

AI Infrastructure (2)

  • Casselman argues that data centers affect affordability through electricity demand, higher construction-material and labor costs, and the interest-rate pressure tied to inflation. A bipartisan congressional bill would ease electrical-infrastructure construction and make data centers pay more of the grid costs they create.
  • Casselman says some economists see a case for slowing the AI data-center boom: its rapid buildout raises prices and interest rates and could create a bubble. A slower pace, they argue, might make growth more sustainable without causing a broader economic collapse.

Elections (1)

  • Casselman says voters punished Democrats in 2024 and Republicans in 2026 over affordability, making the issue likely to remain central in 2028. Democrats may prioritize visible, near-term measures such as restoring healthcare subsidies or raising wages, while longer-term housing and infrastructure policies take years to affect costs.

Chips (1)

  • Donald Trump said he would be open to Jensen Huang returning as Nvidia’s leader after Huang resigned following his admission of an affair with a woman US officials had identified as a Chinese intelligence asset. The episode raised questions about Huang’s handling of a potential national-security threat.

Anthropic’s $1.4T IPO: The Biggest Wealth Event in Human History? • Nov 19

  • The United States and nine partner countries signed the San Francisco Declaration on Secure Foundations of Enduring Prosperity, while a proposed Commerce rule would require licenses for advanced AI chip shipments to five countries. Alex sees a widening US-China split; Dave argues software models are hard to contain and hardware controls matter more.
Also discussed on this episode: (20)

China (1)

  • Donald Trump arrived in China, where Xi Jinping greeted him on the tarmac; Vladimir Putin was also in China for trilateral meetings. The hosts described the visit as a major-power contest over influence.

Chips (2)

  • NVIDIA reported $115.8 billion in revenue, $107 billion in data-center revenue and $65 billion in net income, yet its stock fell 1.5%. The host said the board had sought a succession plan for 10 years; Alex expects activist investors to challenge its governance.
  • Dave argues NVIDIA's cash-funded investments around its ecosystem are not inherently fraudulent, while Alex warns that reported circular financing could unwind if Jensen Huang leaves. Dave says NVIDIA must expand beyond GPUs as hyperscalers build competing chips and AI threatens CUDA's moat.

Markets (1)

  • The host says Anthropic's IPO offers 825 million shares at $70 to $78, with five-times demand and a possible $1.4 trillion valuation. The filing lists $17 billion in third-quarter revenue and an $81 billion annualized run rate as of October 31.

VC (1)

  • Anthropic's IPO could give average employees shares worth $72 million, while OpenAI insiders' average holdings were estimated at $126 million. Dave says newly liquid AI fortunes could fund another wave of startups; he put current venture capital above $100 billion to $150 billion, versus $20 billion to $40 billion for much of his career.

Safety (2)

  • Microsoft email outages, an FAA ground stop, bank service disruptions and a confirmed Optimizely exploit hit within six days. Alex saw no evidence of one coordinated attack; the host and Dave emphasized how a small number of aging, connected systems can disrupt whole sectors.
  • Alex says AI can find vulnerabilities at scale while cyber defense still patches systems one port at a time; he cites Linus Torvalds's view that AI-driven Linux kernel patch discoveries may have passed their peak. The speakers warned of false attribution and urged companies to replace legacy software and hardware.

Space (2)

  • SpaceX caught Starship Ship 42 after a 22.5-hour flight and deployment of 26 Starlink V3 satellites. Peter estimates mature reusable launches could cost $27 to $53 per kilogram; Elon Musk later cited $500,000 in propellant and a 200-ton payload, implying tickets below $1,000 to orbit.
  • Alex compares reusable Starship flights to packet-switched networks and shipping containers: interchangeable launches could carry cargo that refuels and assembles in orbit, replacing one-off, monolithic missions. Salim says precise landing control could enable applications well beyond rockets.

AI Infrastructure (2)

  • Salim says SpaceX's Memphis-area data-center builds create an organizational learning advantage by compressing construction cycles. SemiAnalysis reports 150 megawatts online in 122 days at Colossus 1 and 210 megawatts in 91 days at Colossus 2, far faster than typical projects.
  • Peter says AI can help manage construction by sequencing work, reading job-site video and coordinating robots; he expects two-to-three-times gains, while Alex argues 10x could be the floor. The discussion frames AI as a tool for redesigning physical operations, not only software work.

Models (2)

  • Alex prefers Opus 5.5 for its cost and responsiveness but says Fable 5.2 leads marginally on broader leaderboards; he found OpenAI's GPT 6.2 models improved math performance but noted reports of doom loops. The panel says generalist models may now handle robotics tasks without dedicated robotics models.
  • Dave says MIT CSAIL's Weight Space Programming research explores one neural network directly changing another's parameters, including through modular updates or added expert modules. He argues that reusing learned knowledge matters as frontier-model training costs rise.

Digital Sovereignty (1)

  • Palmer Luckey's company Slab proposes ocean platforms with a thousand-year design life, initially as a place for industries constrained by state and local rules. Alex says Starlink could connect seasteads and ocean data centers could supply their economic purpose; Salim warns about corrosion and erosion.

Agents (1)

  • Salim expects customer service and operations to shift first from task supervision to outcome supervision because resolution, satisfaction and retention are measurable. He says a pilot cut one company's process from 20 days to 38 minutes using agents and organizational redesign.

Education (1)

  • Peter argues education should develop mindset, purpose, AI fluency and entrepreneurship rather than primarily prepare students for jobs. He says AI gives more people the ability to identify problems and build solutions, making agency over one's future a central educational goal.

Health (1)

  • Alex says frontier models have learned biological sequences and may support antiviral discovery; he cites MDL001 as a broad-spectrum antiviral with promising animal-model results. He sees broader progress coming from models trained across biology, alongside specific drug and regulatory applications.

Startups (1)

  • Salim advises Indian AI founders to build applications around local needs instead of competing with Silicon Valley on foundation models. He highlights voice agents for small businesses that can serve customers across India's many languages and scripts.

Longevity (1)

  • Alex argues longevity would not let dictators preserve power unchanged for centuries: complex societies force leaders to adapt, much as long-lived software must be continually patched. The question cited Xi Jinping and Vladimir Putin discussing predictions that people could live to 150.

Robotics (1)

  • Peter argues robotics and AI could turn waste dumps into resource-refining facilities that sort metals, plastics and organic matter; Dave says sorting after disposal is now feasible. The discussion also cites seven grams of plastic in the average human brain as a measure of microplastic exposure.

Transchipment: Huang and the AI aftershocks • Nov 17

  • Shailesh Chitnis says the Commerce Department tightened controls on advanced-chip shipments and closed a loophole that let Chinese AI firms rent leading-edge chips through cloud providers in Malaysia, Singapore, and Australia.
Also discussed on this episode: (12)

Big Tech (3)

  • Shailesh Chitnis says Jensen Huang’s supporters cite NVIDIA’s stock decline and argue his ability to make bold bets and mobilize the supply chain is hard to replace. NVIDIA was valued at $5 trillion before the scandal, and Huang had targeted $10 trillion.
  • NVIDIA’s quarterly results will reflect the period when Huang was CEO, and investors expect another blowout quarter. Shailesh Chitnis says they will judge interim CEO Colette Kress on talent retention, competition, and the fallout from the alleged spy scandal.
  • Shailesh Chitnis says Scott Bessent wants NVIDIA’s next CEO to come from outside the company, arguing its culture is broken. Huang is due to testify behind closed doors in Congress, while Democrats have called for further hearings next year.

AI Infrastructure (1)

  • Shailesh Chitnis says NVIDIA’s financing of new cloud providers and AI labs, which then bought its chips, unsettled investors. Its gross margins were almost 75%, far above industry norms, so some industry figures see more competition and lower prices as a useful correction.

Immigration (5)

  • Duncan Robinson says UK net migration is expected to fall to the tens of thousands, its lowest level in about 30 years, after governments tightened visa and settlement rules. The sudden reversal followed a post-lockdown surge and a broad liberalisation of work and study visas.
  • UK net migration reached 944,000 in 2023, then became a political liability after successive governments failed to deliver promised reductions. Duncan Robinson says the foreign-born share of the population rose from about one in ten in 2004, when the EU expanded to include Poland, to one in five today.
  • Duncan Robinson says immigration remains politically volatile even as numbers fall: some voters oppose any level, and Channel small-boat crossings can dominate debate when overall arrivals shrink. He says the sharp drop could also reopen liberal arguments for free movement with the EU.
  • Duncan Robinson says the economic case for immigration rests partly on higher tax receipts and improved fiscal forecasts, but the latest arrivals included many low-wage care workers. He says liberals also became more immigration-sceptical when arrivals reached 1.4 million in a single year.
  • Duncan Robinson says forecasts range from a brief dip below zero to two or three years of net emigration, with some projections putting departures in the tens of thousands annually. He notes that Britain was a country of emigration for much of the 20th century.

Media (3)

  • Catherine Nixey says Robert Pattinson will play Bond in the 26th official film, a notable choice because the role has usually gone to less established actors. The franchise includes 25 films that have grossed an estimated $23 billion, and about half the world has seen at least one.
  • Catherine Nixey says Bond must serve the era that created Ian Fleming, the era of the new film, and the wartime world of British naval intelligence that shaped Fleming’s characters. Fleming began writing the first Bond novel in 1952 and drew on tough commando types he had encountered.
  • Catherine Nixey argues that Bond’s sexism creates a narrow path for reinvention: modernising the character risks alienating traditionalists, while preserving him risks clashing with current norms. She contrasts Fleming’s era with a present-day MI6 led by a woman and recalls Judi Dench’s M calling Bond a sexist, misogynist dinosaur.