OpenAI vs Anthropic IPOs, Anthropic $3T, Zuck's Price War, China Ends Open Source?, Trump Accounts

2026-07-11 - 101 min - source - Read full transcript

Jason Calacanis (host)Chamath Palihapitiya (host)David Sacks (host)Brad Gerstner

Key insights

Gavin Baker's forecast that Anthropic could trade at $3 trillion if it went public today anchors the panel's IPO expectations.

Baker, another All-In guest, predicted two weeks earlier that Anthropic would end 2026 with over $100 billion in revenue and be 'very profitable.' Brad Gerstner adds that at $100B+ ARR, Anthropic's forward run rate next year could exceed $100B again, and that Altimeter would be 'a buyer at scale and at size' in both the Anthropic and OpenAI IPOs based on what is known today.

ai-trillion-dollar-ipos

SpaceX's IPO is treated as a deliberate blueprint for how Anthropic and OpenAI should structure their own trillion-dollar debuts.

Gerstner calls the SpaceX IPO 'textbook': it raised $75 billion at a $1.75 trillion valuation, is now up about 25% and trading near $2 trillion on roughly $35 billion of forward revenue, and pioneered a staged lockup release tied to milestones plus early index inclusion despite historical volatility concerns. He expects Anthropic and OpenAI to have 'gone to school' on that structure.

ai-trillion-dollar-ipos

Chamath argues frontier labs should IPO now, before an industry-wide token-spend reckoning becomes visible to public markets.

He recounts his own CTO reporting that token costs are doubling every 45 days while downstream productivity gains are roughly 5% at best, meaning true ROI on AI spend is close to flat. Chamath frames this as a problem 'everybody in the next three or four years will for sure go through,' and argues getting a public listing done before that reckoning 'seeps into the water table' is what allows a company to exit at a high price.

enterprise-ai-token-economics

Using Fable 5's own answer, Chamath estimates AI's actual contribution to S&P 500 earnings growth is close to zero once Nvidia's chip revenue is excluded.

He asked Fable what share of S&P 500 EPS growth since 2004 came from AI and got roughly 50%, but that figure included Nvidia's chip sales to AI buyers. Isolating the S&P 493 (excluding the AI infrastructure names), EPS growth was about 9%, and after attributing most of that to pricing power and buybacks, the actual AI-driven ROI came out to roughly 0-2%, which he expects sophisticated investors to start demanding companies justify.

enterprise-ai-token-economics

Despite predictions that cheap open-source models would erode frontier-lab revenue, enterprise dollar share of wallet is moving toward the closed frontier labs, not away from them.

Sacks cites data showing open source's share of enterprise AI spend fell from about 19% to 11% year over year, even as usage of both categories is rising. He argues most enterprises lack the technical capability to build the token-routing middleware needed to exploit cheaper open models (unlike Coinbase or DoorDash, which built it themselves), so 'the spirit is willing, but the flesh is weak' - they default to the frontier labs because it is simply the easiest, most convenient choice.

open-source-vs-frontier-ai

Chamath floats a non-consensus hypothesis: model intelligence may not be converging at all, and the frontier-to-commodity gap could widen rather than close.

He notes that while benchmark scores across labs appear to be converging, revenue distribution is not - only two companies (Anthropic and OpenAI) register meaningful token revenue. His argument: if superintelligence becomes self-recursive, a smarter model generates more revenue, which buys more compute, which builds an even better model, extending the lead rather than closing it as agentic tasks get more complex over the next two to three years.

open-source-vs-frontier-ai

China may be moving to restrict its own labs' open-source releases, following the same 'stay open until you catch the frontier, then close' pattern OpenAI and Meta already followed.

Per two Reuters reports, Chinese regulators reportedly met with Alibaba, ByteDance, and Zhipu about limiting overseas access to top Chinese models, citing national-security concerns about IP leaks and models being 'weaponized' against Chinese interests. Sacks argues this parallels Sam Altman taking OpenAI from nonprofit/open to for-profit/closed, and Meta quietly stepping back from Llama's original open strategy, once a lab gets close enough to the frontier that capturing value outweighs the developer-goodwill benefits of staying open.

open-source-vs-frontier-ai

Zuckerberg is reframed mid-episode as pivoting Meta from an open-source strategy to a direct price war against frontier labs.

Jason and Sacks note Zuckerberg tweeted more about the new Spark 1.1 model launch (via his old college handle @finkd) than he had in his entire prior X history, positioning it as matching frontier quality at roughly one one-hundredth the cost - which the panel reads as an admission that the original 'scorch the earth with open source' game-theory play had underperformed.

open-source-vs-frontier-ai

Chamath argues US electricity capacity, not chips or software, is the real bottleneck on AI scaling, citing a shortfall equivalent to three additional Californias' worth of power by 2050.

His team's internal analysis projects the US will be roughly three entire California's-worth of energy short of expected load growth by 2050, even before accounting for AI data center demand, arguing the country needs more nuclear, solar, and battery buildout, which is itself a regulatory bottleneck.

ai-energy-bottleneck

Taiwan's chip production is exposed to an acute energy vulnerability: it holds only two to three weeks of LNG reserves.

Citing a Wall Street Journal report, Chamath notes that because Taiwan (source of most advanced chips) runs largely on imported LNG with only weeks of reserve, a Chinese blockade could halt Taiwan's energy supply almost immediately, making the AI buildout's chip supply chain directly hostage to an energy chokepoint separate from the compute race itself.

ai-energy-bottleneck

Brad Gerstner details how Trump accounts (Invest America accounts) work as a hybrid philanthropic and retirement-savings platform seeded at birth.

Every child born in the US gets a $1,000-seeded, privately owned S&P 500 investment account; family, employers (up to $2,500/year tax-free), and philanthropists can add up to $5,000/year combined per child, tax-free, until the recipient turns 18, at which point the funds can roll into an IRA or Roth IRA. Gerstner says over 1.5 million accounts were created and over $1 billion deposited in the first 24 hours after the July 4 launch, with a goal of auto-enrolling all 50-70 million eligible US children within 90 days.

trump-accounts-wealth-building

Gerstner frames Trump accounts as potentially the largest direct philanthropic platform in US history, targeting $100 billion raised in the first 12 months.

Early philanthropic commitments include Michael and Susan Dell ($6B+ across 25 million lower- and middle-income children), SpaceX president Gwen Shotwell ($350M in SpaceX shares to lower-income kids), Micron ($250M in employer matches), and Gerstner personally committing $100 million for all children under five in Indiana. He argues this structure - donations flowing directly into privately owned citizen accounts rather than through NGO overhead - is more efficient than traditional philanthropy.

trump-accounts-wealth-building

Media referenced

Wall Street Journal article on Taiwan's LNG reserves - article - Chamath cites the piece as a wake-up call: Taiwan holds only two to three weeks of LNG reserves, so a Chinese blockade would knock out the energy that runs the world's leading chip fabs almost immediately.
Reuters reports on China restricting overseas access to Chinese AI models - article - Two anonymously sourced Reuters reports, published about 15 minutes apart, describe Chinese regulators meeting with Alibaba, ByteDance, and Zhipu about limiting outside access to top Chinese open and closed models, citing national-security and IP-leak concerns.
Mark Zuckerberg tweet announcing Meta Spark 1.1 - other - Zuckerberg (posting from his old college handle @finkd) announces a new agentic coding model at roughly one one-hundredth the cost of competitors, which the panel reads as Meta pivoting from an open-source play to a direct price war.
Jesse Zhang tweet on frontier-lab share of wallet - other - Cited by Brad Gerstner as evidence that even as commodity open-source tokens proliferate, the dollar share of enterprise AI spend is still shifting toward the closed frontier labs, not away from them.
Praveen (Uber CTO) X posts on agentic pods - other - Uber's CTO describes embedding engineers inside non-engineering departments (legal, HR, procurement) to find AI-driven ROI outside of pure coding, after reining in runaway developer token spend.
Andy Fang (DoorDash CTO) X post on open-weight code review - other - DoorDash publishes benchmarks showing it can safely route lower-stakes code review work to the open-weight Kimi K2.6 model while reserving Anthropic's frontier model for the hardest tasks.
Ali (Databricks) post on AI harness optimization - other - Databricks' founder describes cutting inference costs roughly in half by optimizing the harness (memory, skills, scaffolding) around a model, independent of which underlying model is used.
Decagon founder blog post on mature vs. immature AI use cases - article - Cited by Sacks: open, post-trained models work well once a use case is well understood (e.g. customer support), but immature, undefined use cases still need the most powerful general frontier model available.
Nikesh Aroura post on model fungibility - other - Cited as evidence enterprises want to hot-swap between cheap models freely, but no one has yet solved making memory, context, and history portable across models to make that practical.

Companies

Anthropic - Confidentially filed for IPO on June 1; Gavin Baker's prediction that it would trade at $3T if public today anchors the episode's IPO discussion; also central to the open-source-vs-frontier debate as one of the two dominant revenue labs.
OpenAI - Discussed as Anthropic's IPO rival, rumored to be exiting 2026 near $70B ARR versus Anthropic's rumored $100B+, with GPT-6 expected within 30 days; seen as having more IPO complexity due to its corporate restructuring.
SpaceX - Its trillion-dollar IPO (raised $75B at $1.75T) is treated as the blueprint Anthropic and OpenAI are expected to follow on pricing, lockup structure, and index inclusion; trading roughly at its IPO price, around $150/share, at a $2T market cap.
Meta - Sacks and Jason frame Zuckerberg's Spark 1.1 launch as him abandoning an earlier scorched-earth open-source strategy in favor of a direct price war, undercutting frontier-lab pricing by roughly 100x.
Uber - CTO Praveen's public breakdown of curbing runaway developer token spend and instead embedding engineers into non-engineering departments is cited as a model for finding real enterprise AI ROI.
DoorDash - CTO Andy Fang publishes benchmarks showing safe routing of lower-stakes code review to the open-weight Kimi 2.6 model while reserving Anthropic's frontier model for the hardest tasks.
Databricks - Founder Ali describes a custom multi-model, multi-harness routing layer (OmniGent) that cut inference costs roughly in half by optimizing the harness rather than switching models.
Nvidia - Cited as a company whose own next-generation chip design work is now AI-assisted end to end ("the machine is building the machine"), and as the likely beneficiary, via hardware/hosting revenue, of open-source inference that never shows up as lab revenue.
Alibaba / ByteDance / Zhipu (GLM) - The three Chinese labs reportedly summoned by Chinese regulators over restricting overseas access to their top models; Zhipu's GLM 5.2 is repeatedly cited as the open-source model closing the gap with US frontier labs.
Moonshot AI - Referred to in the transcript as 'mythos'; described as leaving watermarks in GLM 5.2 that reveal it was trained via distillation from US frontier model outputs.
Micron - Contributed $250M (up to $1,000 per employee) to Trump/Invest America accounts as an employer-matching example.
Robinhood - Vlad Tenev and team helped build the Trump accounts app's consumer product experience.
11 Labs - CEO Mati confirms the company is a major frontier-model customer but is also building its own proprietary voice models to reduce dependence on and competitive exposure to frontier labs.
Lovable - CEO Anton tells Jason the vibe-coding company grew from $100M to $600M in revenue over two years while also building proprietary models to reduce frontier-lab dependence.
Coinbase - Cited by Sacks as one of the few enterprises technically capable of building its own token-routing middleware to shift work between frontier and cheaper models.

Techniques and frameworks

Model fungibility / intelligent token routing - The practice of dynamically routing AI tasks to the cheapest capable model, discussed as technically difficult because it requires making memory, context, and conversation history portable across different models.
AI harness optimization - Databricks' finding that restructuring the scaffolding (memory, skills, tool access) around a model can roughly halve token costs for the same underlying model, distinct from switching models entirely.
Distillation (AI training technique) - Training a new model on another model's outputs; cited as the mechanism by which GLM 5.2 reportedly absorbed US frontier-model behavior, and as China's stated justification for tightening control over its own labs.
Trump accounts / Invest America Act - A federally enabled, privately owned S&P 500 investment account seeded with $1,000 at birth, allowing up to $5,000/year in contributions from family, employers, or philanthropists, tax-free compounding until age 18, then rollover into an IRA or Roth IRA.

Summary

The episode opens with Bestie Brad Gerstner filling in for a vacationing Friedberg, and quickly moves into the year's dominant markets story: the trillion-dollar IPO wave. With SpaceX now trading near its $2 trillion IPO price, the panel treats its listing as a deliberate blueprint - staged lockup releases, early index inclusion, a $75 billion raise - that Anthropic and OpenAI are expected to follow when they go public, likely within six to nine months. Gerstner cites Gavin Baker's prediction that Anthropic could trade at $3 trillion today given its rumored $100 billion-plus 2026 revenue run rate, positioning Anthropic ahead of OpenAI (rumored near $70 billion) on the revenue trajectory but with OpenAI carrying more IPO complexity due to its corporate restructuring.

The conversation's center of gravity is the growing tension between AI revenue growth and unproven ROI. Chamath opens with a pointed anecdote from his own portfolio company: token costs doubling every 45 days against roughly flat downstream productivity gains, an imbalance he expects most enterprises to hit within a few years. Using Fable 5 to strip Nvidia's chip revenue out of S&P 500 earnings growth since 2004, he estimates AI's actual contribution to enterprise earnings is closer to 0-2% - a gap he expects sophisticated investors to start demanding companies explain. Yet despite this ROI uncertainty and predictions that cheap open-source and Chinese models (like Zhipu's GLM 5.2) would erode frontier-lab pricing power, Sacks presents data showing the opposite: open source's share of enterprise AI spend actually fell from about 19% to 11% year over year, because most enterprises lack the technical sophistication to build the routing middleware needed to exploit cheaper models - only a handful, like Coinbase and DoorDash, have managed it. Chamath goes further, floating a non-consensus hypothesis that intelligence might not be converging at all: if frontier model improvement becomes self-recursive, the gap between the top labs and everyone else could widen rather than close as agentic tasks grow more complex.

A parallel thread tracks the open-source-versus-closed dynamic playing out globally. Two Reuters reports suggest China may restrict its own labs' overseas model access, which Sacks frames as the same "stay open until you catch the frontier, then close" pattern OpenAI and (to a lesser extent) Meta already followed. Zuckerberg's surprise pivot mid-week - tweeting more about the new Spark 1.1 model, priced at roughly one one-hundredth of frontier cost, than in his entire prior X history - is read by the panel as an admission that Meta's earlier "scorch the earth with open source" strategy underperformed and is being replaced with a direct price war. Chamath closes the technology segment with a structural argument that the real long-term bottleneck isn't chips or model quality but energy: his team's analysis projects the US will be roughly three Californias' worth of power short of 2050 load growth, while Taiwan's chip fabs run on only two to three weeks of LNG reserves, tying the entire AI buildout to acute energy chokepoints in both countries.

The back half of the episode shifts entirely to Gerstner's other project: Trump accounts (Invest America accounts), a federally enabled program giving every US child a privately owned, $1,000-seeded S&P 500 investment account at birth. Gerstner walks through the mechanics - up to $5,000/year in tax-free contributions from family, employers, or philanthropists, tax-free compounding until age 18, then rollover into an IRA or Roth IRA - and reports over 1.5 million accounts created and $1 billion deposited within 24 hours of the July 4 launch, alongside major philanthropic commitments from Michael and Susan Dell, SpaceX's Gwen Shotwell, and Micron. Jason and Sacks push back on the "TDS" criticism that some parents are refusing to participate simply because of the program's name, framing the accounts as a bipartisan, capitalism-forward alternative to both traditional Social Security and progressive redistribution proposals, while Gerstner sets a goal of auto-enrolling all 50-70 million eligible children within 90 days and raising $100 billion in philanthropic contributions within the first year.

Notable Quotes

"Right now, our token costs are doubling every 45 days... My upside is essentially flat." - Chamath Palihapitiya, relaying his CTO's assessment

"Anyone who's saying that these closed models are going to lose or are somehow losing, you're just not seeing it in the data." - David Sacks

"The non-consensus argument might be that intelligence is not converging at all... the smarter your model gets, the more revenue you get, the more compute you can buy, the better the model is that you can build." - Chamath Palihapitiya

"If a Trump account had been maxed out, and you have the standard market rate of return that we've had for the past 30 years, then by age 28, that kid will be a millionaire." - Brad Gerstner

"We are about three entire California's worth of energy short, and that's just assuming regular consumption of devices and cars, fridges, televisions, and computers." - Chamath Palihapitiya