All-In: Anthropic’s $2T IPO, Zuck’s AI Manifesto, Nvidia’s $500B AI Bet, Grok’s Comeback

概览

This episode centers on whether the AI boom is becoming a durable economic buildout or an overheated market narrative. The hosts and guest Gavin Baker discuss Anthropic’s reported $2 trillion IPO target, its extraordinary revenue ramp, and whether demand, compute, energy, and market share can support continued growth.

A major thread is the debate between centralized frontier AI and decentralized/open-source AI. Zuckerberg’s AI manifesto becomes a springboard for arguments that AI is too important to centralize, while Anthropic and effective altruism are criticized for favoring tighter control and regulation.

The discussion then moves into AI infrastructure finance: Nvidia’s effort to make GPU compute an investable, asset-backed category, the risks of overbuild, and why Anthropic’s public financials could become a key signal for the whole AI supply chain. The final sections cover Amazon’s delivery subcontractor model, Grok’s comeback, and a possible private-equity bid for Workday as a signal for software valuations.

分段落总结

[00:00] Opening and Rahm Emanuel Interview Recap

[事实] Jason opens with David Sacks and Gavin Baker, noting that the crew is smaller but the docket is long.

[事实] The hosts briefly discuss a recent All-In interview with Rahm Emanuel, describing it as spicy and noting Emanuel’s experience with Obama, Clinton, Chicago, and Japan.

[事实] Sacks frames Emanuel as a throwback to the Clinton-Obama Democratic Party and argues that such moderates face difficulty in a Democratic Party influenced by Democratic socialists.

[02:41] Anthropic’s Reported $2 Trillion IPO

[事实] Jason says Financial Times reported that Anthropic is targeting a $2 trillion IPO and that investors expect the company could go public in October.

[事实] He cites expectations that Anthropic could end the year at a $100 billion to $120 billion annualized run rate, which he calls an unprecedented Silicon Valley revenue ramp.

[事实] The hosts discuss Anthropic as still having a top AI model while Grok and OpenAI are catching up, and Jason notes reported talks for Anthropic to buy Decart for $6 billion.

[推测] The segment frames Anthropic’s IPO as a test of whether AI revenue growth can justify valuations normally reserved for the largest technology platforms.

[04:40] Gavin Baker on Anthropic’s Growth and IPO Pricing

[事实] Gavin says he is not currently an Anthropic shareholder and generally avoids investing in Elon Musk competitors.

[事实] He calls Anthropic’s rumored numbers exceptional, even while saying the company may be losing share to OpenAI, open source, and Grok.

[事实] Gavin argues that the overall AI market is expanding so quickly that Anthropic can still grow rapidly despite potential share loss.

[事实] He is skeptical of the $2 trillion leak, suggesting IPO valuation leaks can come from bankers jockeying for position or trying to shape expectations.

[09:09] Sacks on Revenue Ramp, TAM, and Compute Constraints

[事实] Sacks says Anthropic has reportedly grown around 10x year over year for several years, but asks whether the market and physical infrastructure can support another 10x.

[事实] He argues demand for AI tokens can keep growing as agents expand into more contexts, but compute and energy may become limiting factors.

[事实] Sacks says Anthropic made a prescient bet on coding, partly by seeing strong usage from products built on its platform, and that OpenAI has since pivoted harder into coding.

[事实] He says he has heard OpenAI’s recent growth accelerated to over 20% month over month.

[12:12] Anthropic Maximalism and Knowledge Work TAM

[事实] Gavin says he has heard that Dario Amodei expressed extreme confidence about Anthropic’s future scale, including a vision where Anthropic could become uniquely important among private companies.

[事实] Gavin says he would take the under on Anthropic becoming the only private company in the world and warns that such language sounds hubristic.

[事实] He estimates the knowledge-work opportunity in the tens of trillions of dollars and says AI so far appears more like a growth accelerator than broad labor substitution.

[事实] He notes some evidence of negative impact for very young workers entering the workforce, especially in coding.

[16:21] Data Centers, Energy, and Physical Buildout

[事实] Gavin emphasizes that data centers are physical infrastructure projects involving heat, remote sites, turbines, labor coordination, and real-world constraints.

[事实] He says turbine capacity is a bottleneck, but companies such as Caterpillar, Cummins, GE Vernova, and Siemens Energy are expanding capacity.

[事实] The hosts discuss Texas energy audits for large data-center users and argue that operators with their own energy solutions should be able to proceed.

[事实] Sacks and Gavin reject the media narrative that data centers broadly consume excessive water or raise electricity costs, though they acknowledge noise as a legitimate local issue.

[推测] Their argument is that political and regulatory resistance, more than technical impossibility, could slow the AI infrastructure buildout.

[24:18] Open Source as a Pricing Headwind for Anthropic

[事实] Jason says he thinks Anthropic’s growth will slow considerably next year and says it will not reach $1 trillion in revenue next year.

[事实] He cites cheaper open models, including GLM and Zuckerberg’s open-source push, as a reason enterprises and startups may shift away from expensive frontier tokens.

[事实] Jason argues that corporate America has already embraced open source over the past 20 years because it reduces dependence on large vendors.

[推测] This segment introduces the central economic threat to Anthropic: frontier-model pricing power may erode if open models become good enough for most use cases.

[25:44] Why Open Source Matters for AI Power

[事实] Gavin says Anthropic’s AI and its Constitution are designed around what Anthropic believes is best for humanity.

[事实] He argues that open source creates a richer variety and diversity of AIs, which he sees as better for humans and for America.

[事实] He says Grok’s pursuit of objective truth is valuable and that history may judge Elon favorably for it.

[27:31] Zuckerberg’s AI Manifesto

[事实] Jason introduces Mark Zuckerberg’s 6,500-word essay, “The Future Is for Everyone: The Path to a Positive AI Future.”

[事实] Jason summarizes Zuckerberg’s position as optimistic, open-source-oriented, pro-personal AI agent, pro-tutor, and focused on abundance.

[事实] Sacks says he agreed with much of the essay, especially its critique of people rushing to build a future they describe as dangerous or dystopian.

[事实] Sacks says the key debate is not only open versus closed, but centralized versus decentralized AI power.

[33:24] Too Dangerous to Distribute vs. Too Dangerous to Centralize

[事实] Gavin says Anthropic and effective altruists believe AI is too dangerous to distribute.

[事实] He contrasts that with Zuckerberg, Elon, and Jensen Huang, whom he says believe AI is too dangerous to centralize.

[事实] Gavin says he wants an AI that looks out for him as an individual, rather than an AI aligned to Dario Amodei’s view of the public good.

[推测] The hosts treat AI decentralization as both a civil-liberties issue and a geopolitical strategy.

[35:22] AI, Competition, and the Nuclear-Weapons Analogy

[事实] Sacks argues effective altruists see competition as a race to the bottom, while he believes competition brings out the best.

[事实] Gavin compares the right to have one’s own AI to the right to bear arms, saying he wants an AI aligned with his values.

[事实] The hosts reject the analogy between AI and nuclear weapons, arguing that AI is primarily a consumer and business technology.

[事实] Gavin says banning open source or centralizing AI control would cause America to lose to China.

[39:40] Jobs, Capability Growth, and Meta’s Distillation Hypocrisy

[事实] Jason praises Zuckerberg’s framing that AI does not have to automate faster than it increases individual capabilities or demand for new skills.

[事实] Jason connects Zuckerberg’s view to Jensen Huang’s line that people will not lose jobs to AI, but to people using AI.

[事实] Jason criticizes Zuckerberg for defending the right to learn from observable content while Meta has sued startups that tried to index or use Facebook and Instagram social-graph data.

[推测] Jason’s critique is that Meta’s open-data stance may be more convenient when Meta is the AI model builder than when outsiders want access to Meta’s own networks.

[42:21] Regulation Could Undermine Anthropic’s Own Business

[事实] Sacks argues that Anthropic’s pricing power depends on staying about six months ahead of open models.

[事实] He says FAA- or FDA-style regulatory approval for model releases would destroy that time advantage.

[事实] Sacks says a regulatory slowdown would also help Chinese competitors catch up, hurting both Anthropic and U.S. competitiveness.

[事实] He still believes some customers will pay a premium for true frontier intelligence, similar to customers paying more for Apple over Android.

[45:10] CEOs Going Direct on X

[事实] Jason notes that executives including Nikesh Arora, Jensen Huang, and Mark Zuckerberg have recently become more active with long-form posts on X.

[事实] Gavin says public figures and CEOs are at risk if they do not have their own authentic voice and brand on X.

[事实] The hosts praise Zuckerberg’s more direct public posture and joke about his recent public persona and physical activities.

[推测] The hosts see X as becoming an important venue for unfiltered executive communication and narrative control.

[49:13] Open Source Can Increase Frontier Token Value

[事实] Gavin argues that open-source models can make frontier models more valuable because a top model can orchestrate cheaper, less capable models.

[事实] He gives a scenario where frontier tokens capture most economic value while open-source tokens represent most usage volume.

[事实] Jason asks Gavin and Sacks for Anthropic revenue expectations if 2026 ends at $100 billion.

[事实] Gavin declines to take the over on $1 trillion, while Jason and Sacks discuss a $400 billion to $500 billion exit ARR range as plausible.

[53:15] Why Anthropic Opposes Open Models

[事实] Sacks says Anthropic’s hostility to open models may come from the need to slow commoditization if its own growth slows.

[事实] He argues global compute governance is unrealistic because China would not slow down.

[事实] Gavin says humans remain extremely computationally efficient compared with AI systems that require massive data-center power.

[推测] The hosts imply that human labor may stay valuable partly because biological intelligence is energy-efficient relative to current AI infrastructure.

[56:40] All-In Summit Promotion

[事实] Jason promotes the All-In Summit in Los Angeles from September 13 to 15.

[事实] He lists announced or possible participants including Jensen Huang, Satya Nadella, Jared Isaacman, Steve Hilton, Spencer Pratt, Abigail Shrier, Gavin Baker, and Blake Scholl.

[事实] He says the event will include AI and robotics demos, a gifting suite, wellness offerings, DJs, and parties.

[58:14] Nvidia’s $500 Billion AI Compute Financing Plan

[事实] Jason says Nvidia is partnering with Goldman, BlackRock, and others to raise $500 billion for AI compute.

[事实] He describes the idea as treating GPUs as financeable, income-producing assets where customers borrow to buy Nvidia systems and repay loans from compute revenue.

[事实] Gavin says large asset managers and banks would not participate unless they believed Nvidia GPU compute was a financeable asset.

[事实] Gavin argues Nvidia’s flexible compute benefits from the diversity of model architectures across major Chinese open-source models.

[61:40] Residual Value Guarantees and GPU Lifespan

[事实] Gavin says Nvidia can support financing by offering residual value guarantees, effectively guaranteeing that GPUs can be rented at a certain rate after several years.

[事实] He says Nvidia has unusually strong telemetry into compute supply and demand, allowing it to price those guarantees intelligently.

[事实] Jason raises the risk that GPU lifespan assumptions could be wrong.

[事实] Gavin says CoreWeave has indicated that older Ampere GPUs from 2020 can still be rented profitably for 2029, implying a nine-year useful life.

[63:06] Dark GPUs and Overbuild Risk

[事实] Sacks says the biggest risk is not demand disappearing, but a glut of compute similar to dark fiber after the dot-com crash.

[事实] He argues political headwinds and difficulty building data centers may actually protect against oversupply.

[事实] Gavin says Nvidia could receive revenue share above a guaranteed floor, making the structure similar to royalties.

[推测] The risk framework is that AI infrastructure can fail if capacity is built faster than profitable demand, but physical and political friction may act as a brake.

[65:17] Nvidia as AI’s Financing Layer

[事实] Sacks says Jensen Huang is alleviating a financing constraint that could otherwise limit the AI buildout.

[事实] He compares Nvidia’s role to creating a line of credit through banks and private equity firms for downstream compute buyers.

[事实] The hosts say this is not circular financing in the way some critics worry, but asset-backed financing based on expected GPU cash flows.

[事实] Gavin describes Nvidia as becoming something like the central bank of AI.

[67:16] GPU Financing Compared to Aircraft Financing

[事实] Sacks compares GPU financing to aircraft financing, where lenders underwrite not only the buyer’s creditworthiness but also the value of the airplane itself.

[事实] He says Nvidia would need standardization and reference designs to make GPU assets easier for Wall Street to finance and securitize.

[事实] Jason says physical buildout constraints and major buyers such as Anthropic and OpenAI act as governors on the system.

[69:00] Anthropic IPO as Industry Signal

[事实] Sacks says Anthropic’s IPO is important because quarterly earnings would become a key signal for the entire AI industry.

[事实] He says skeptics argue AI is a bubble and CapEx lacks ROI, so public Anthropic numbers would help test the demand signal.

[事实] Sacks says if Anthropic slows down due to lack of demand, there could be a pile-up across the AI supply chain.

[事实] Gavin says if Anthropic slows because it is being passed by OpenAI, SpaceX/xAI, or open source, the impact would be less severe than a demand collapse.

[71:52] Token Profitability and AI Spend

[事实] Gavin says, based on public tea leaves and signals, Anthropic is generating cash and profitable.

[事实] He says macro and value investors are wrong if they assume AI tokens are broadly subsidized and unprofitable.

[事实] Jason argues that U.S. companies could spend 5% to 10% of employee salaries on AI tokens to make workers more efficient, implying $500 billion to $1 trillion of possible U.S. AI spend.

[推测] The hosts present AI token demand as analogous to prior enterprise spending on SaaS, computers, and phones.

[74:28] Amazon DSP Lawsuit and Mamdani’s Position

[事实] Jason says New Jersey’s attorney general sued Amazon over its delivery service partner model, arguing Amazon avoids responsibility by hiring drivers through subcontractors.

[事实] He says Zohran Mamdani supported requiring Amazon to hire drivers directly under municipal law.

[事实] Jason says the issue involves liability, union bargaining, labor obligations, and whether Amazon is using subcontracting to insulate itself.

[事实] Sacks says Amazon argues that ending the DSP model would raise delivery costs, slow service, threaten contractor jobs, and reduce flexibility during demand spikes.

[78:05] Capitalism, Worker Treatment, and Amazon’s Responsibility

[事实] Sacks supports freedom of contract and says the DSP market structure evolved because workers, DSP owners, and Amazon all benefit.

[事实] Gavin says Amazon should show a “Mamdani fee” as a line item if policy changes raise consumer costs.

[事实] Jason argues Amazon got too clever and should proactively make more delivery drivers full-time Amazon employees with better pay, benefits, and equity.

[事实] Jason compares the opportunity to Howard Schultz making Starbucks jobs worth having through wages and benefits.

[82:07] Amazon, Uber, and the Cost Debate

[事实] Sacks compares the Amazon DSP issue to past criticism of Uber’s contractor model.

[事实] Jason says Uber is different because it carries major insurance and lawsuit exposure for drivers.

[事实] Jason argues Amazon should keep some DSPs for holiday surge capacity but employ a larger core driver base directly.

[事实] Sacks counters that requiring direct employment could add about $5.20 per package, cost households $664 per year, and put roughly 5,000 jobs at risk.

[87:40] Grok 4.6 and Grokbot

[事实] Sacks asks whether Grok 4.6 and Grokbot prove the frontier model market is more fluid than an Anthropic-OpenAI duopoly.

[事实] Gavin says Grok 4.6 looks disruptive on price and capability, citing cursor benchmarks, a Databricks evaluation, and positive comments from DHH.

[事实] Jason and Gavin attribute Grok’s catch-up to Elon bringing in Cursor-related talent and SpaceX/xAI operators.

[事实] Gavin says Grok 4.7 is expected in a few weeks and should be more capable.

[91:34] Grokbot and Personalized AI

[事实] Gavin says Grokbot feels important because it makes personalized AI easier and more democratic.

[事实] The hosts compare Grokbot to tools from Claude and Perplexity, while saying Grokbot feels easier to use.

[事实] Jason says this connects to Zuckerberg’s manifesto idea of creating bots for everyone.

[推测] The segment positions Grokbot as a consumer-access layer, not just another benchmark-driven model release.

[92:19] Elon Hosting Anthropic While Competing with It

[事实] Jason notes the tension that Elon is hosting Anthropic compute while Grok is increasingly competing with Anthropic.

[事实] Sacks says Elon’s compute clusters give him a call option to train a frontier model and a put option to sell compute if that fails.

[事实] Jason says reports indicate Elon has the right to take compute back from Anthropic, and Sacks says both sides have a 90-day cancellation clause.

[事实] Sacks says if compute remains supply constrained, that dynamic is advantageous for Elon because he can charge close to spot pricing or reroute compute to his own models.

[93:57] Grok’s Role in SpaceX/xAI Valuation

[事实] Gavin says investors often have nuanced views on Starlink, direct-to-cell, orbital compute, and terrestrial compute, but many do not discuss Grok.

[事实] He says Grok is Pareto-dominant on multiple measures and should be considered when thinking about SpaceX/xAI value.

[事实] The hosts acknowledge they are talking their book because they have been early investors in SpaceX.

[推测] Gavin’s point is that Grok may represent underappreciated upside in the broader Elon ecosystem.

[95:24] Workday, Silver Lake, and Software’s Private-Equity Bid

[事实] Gavin says breaking news during the show suggested Silver Lake may be buying Workday, and Workday was up 17%.

[事实] He says software had been in a difficult market for 12 to 18 months and that a private-equity bid could change the investing landscape.

[事实] Gavin argues open-source AI is a godsend for the American software industry because it reduces dependence on a few dominant frontier model providers.

[事实] The hosts compare this to Bending Spoons acquiring software businesses, cutting headcount, and making remaining employees AI-first.

[98:12] Closing

[事实] Jason says Gavin is tied with Brad Gerstner in guest ratings for “fifth bestie.”

[事实] The episode closes with jokes about future guests, dinner plans, and the date August 14.

[事实] A brief postscript says they need to get merch.

播客点评/总结

[推测] This episode is most valuable for listeners tracking the economic layer of AI: revenue ramps, compute supply, energy constraints, financing structures, and how open-source models pressure frontier-model pricing. The discussion is especially strong when Gavin and Sacks connect model demand to GPU financing and infrastructure bottlenecks.

[推测] The episode’s main limitation is that several claims are presented as insider signals, market chatter, or personal interpretation rather than verified public data. The hosts often phrase bold forecasts confidently, so listeners should separate “the host says” from independently confirmed facts.

[推测] The liveliest parts are the open-source versus centralized-AI debate and the Grok comeback discussion, where the hosts tie technical benchmarks to ideology, geopolitics, and market structure. The Amazon DSP debate also adds useful contrast by showing real disagreement among the hosts.

[推测] This is best suited for listeners interested in AI investing, tech policy, infrastructure finance, and Silicon Valley strategy. It is less suited for someone seeking a neutral primer, because the episode is opinion-heavy and strongly pro-decentralization, pro-open-source, and pro-AI-buildout.