All-In latest highlights: Anthropic IPO versus OpenAI—revealing the real AI ROI, China model export restrictions, and mass shareholding for everyone

Compiled & Translated: Deep Tide TechFlow

Guest: Chamath Palihapitiya (Social Capital founder), Brad Gerstner (Altimeter Capital founder and CEO), David Sacks (Craft Ventures partner)

Host: Jason Calacanis, All-In Podcast

Podcast Source: All-In Podcast

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

Air Date: July 11, 2026

Key Takeaways

In this episode of All-In (Episode 280), Friedberg is on vacation, and Brad Gerstner fills in. The show begins with a trillion-dollar IPO race: SpaceX has already gone public with a valuation of $1.75 trillion; on June 1, Anthropic secretly filed; and OpenAI is right behind it. Gavin Baker predicts that Anthropic’s revenue could top $100 billion this year, with an IPO valuation reaching as high as $3 trillion. Brad is decisive: Altimeter will aggressively buy the IPOs of both companies.

But Chamath poured cold water. He found that the token cost for his company doubles every 45 days, while downstream productivity improves by at most 5%. He asked Claude 5 one question: How much EPS growth does AI bring to the S&P 500? The answer was 50%. But after removing the portion where Nvidia sells chips to Amazon, the actual EPS growth for S&P 493 is only 9%—most of it comes from pricing power above inflation and share buybacks, and the true AI ROI is between 0% and 2%. Chamath’s view: if it can be listed now, list now—while these numbers haven’t yet seeped into the market’s “waterline.”

The second half turns to China. Reuters reports that the CCP is considering restricting overseas access to China’s top AI models, and classifying leaked AI research as national security crimes. Sacks has discussed the topic with people in Washington and even with the White House and the Treasury, and his judgment is that China’s strategy is the same as Sam Altman’s back then: catch up by going open source, and after catching up, close the doors. He also reveals that GLM-5.2 contains distilled watermarking from U.S. frontier models, and the U.S. government is likely to move to crack down on distillation. At the very end, Brad spends nearly an hour talking about Trump Accounts— a plan to give every newborn in the U.S. $1,000 and invest it in the S&P 500. Within 24 hours of the app launching, it opened 1.5 million accounts and attracted more than $1 billion in deposits.

Highlights Summary

On IPO Timing

• Chamath: "If it can be listed now, list now—while you still have a window to sell at a high price and raise a lot of money."
• Brad: "Today, Altimeter will buy these two IPOs at scale, both in terms of size and volume."
• Brad: "Anthropic’s annualized revenue could exceed $100 billion, while SpaceX’s forward revenue is only $35 billion. Based on SpaceX’s success, this will be a phenomenon-level IPO."

On AI ROI

• Chamath: "My token cost doubles every 45 days, and downstream productivity could improve by up to 5%. My cost doubles, and the benefits basically stay the same."
• Chamath: "EPS growth for S&P 493 is 9%. The vast majority comes from pricing power above inflation, and another 3% comes from buybacks. The true AI ROI is between 0% and 2%."
• Brad: "We’ve never seen revenue growth like this, because we’ve never seen a TAM this large. Intelligence is humanity’s biggest addressable market."

On Open Source vs Closed Source

• Sacks: "Companies are willing in spirit, but their capabilities are weak. They want to move away from closed models, but they can’t."
• Sacks: "The share of open source in enterprise spending is actually declining—from 19% last year to 11% this year."
• Brad: "Whether you replace a $200-an-hour consultant with a $3 cheap model or a $15 frontier model—the price difference just doesn’t matter."

On China’s Shift on Open Source

• Sacks: "China’s strategy is: when you’re catching up, you open source; once you catch up, you close source. That’s exactly what Sam Altman did three years ago."
• Sacks: "GLM-5.2 includes distilled watermarking from Mythos. The U.S. government will crack down on distillation—that’s what should happen."
• Chamath: "For the U.S., the best thing would be for China to produce a doomsday-prophet community as well."

On Trump Accounts

• Brad: "If you’re born with $1,000, and someone matches a bit more, and you save $10 each week—by age 18 you’ll have $50,000. Invest it all into the S&P 500."
• Sacks: "If Trump Accounts is fully funded from the start, then based on market returns from the past 30 years, by age 28 that child becomes a millionaire."
• Jason: "This can replace Social Security. It replaces the pledge to donate."

Main Content

Chapter 1: A Trillion-Dollar IPO Sprint—SpaceX Sets the Example, OpenAI and Anthropic Are Ready to Take the Stage

Jason: Let’s start with IPO updates. A trillion-dollar IPO sprint is underway. SpaceX has already listed, and the trading price is basically hovering near the offering price. The pricing could be called perfect; theoretically, the remaining two are OpenAI and Anthropic. SpaceX’s stock price once surged to $200, and now it has pulled back to $150—right at the offering price. Current market cap is $2 trillion, the world’s seventh-largest company. Anthropic secretly filed its application on June 1, and Polymarket assigns a 65% probability to a listing this year. Two weeks ago, Gavin Baker said he believes Anthropic’s revenue will exceed $100 billion by the end of this year and become profitable—and if it lists now, the valuation could reach $3 trillion. Chamath, you previously said that having Elon list first was a good move—what are the odds that these two show up in Q4 this year or Q1 next year?

Chamath believes both of these are excellent businesses, but the core question is where the market-clearing price really is. That depends more on how hungry the market is for new share issuance, and at what price levels those numbers can be absorbed.

OpenAI and Anthropic are in different stages. The last information disclosed by OpenAI shows that cash burn is still very high, because the business is fragmented and relies more on the consumer side. Brad previously mentioned that Anthropic may have already become unexpectedly profitable. Chamath shared a detail: he asked his CTO about token spend, and the response was, “Currently it doubles every 45 days.” When he pressed on how much downstream productivity improves, the CTO said, “At most 5%.” Costs double, while benefits are basically flat. The CTO explained that to get the next iteration improvement, you need to consume far more tokens, because the effect has started to suffer from marginal diminishing returns.

Chamath’s judgment is: if it can be listed now, list now—while these numbers haven’t seeped into the market’s awareness yet. That’s probably the window to raise a lot of money at a high price.

As an investor in both companies, Brad offered a more optimistic take. SpaceX’s IPO can be called textbook: it raised $75 billion, valued at $1.75 trillion, with forward revenue around $35 billion, and the current stock price is already up 25%. Anthropic’s revenue, reportedly, could break $100 billion this year—if that comes true, next year’s GAAP revenue could be far above that figure. Based on SpaceX’s success, Brad believes this will be a phenomenon-level IPO. SpaceX has done pioneering work across total IPO issuance volume, pricing, liquidity, index inclusion, and lock-up arrangements—both Anthropic and OpenAI are learning from it.

Regarding the controversy around index inclusion, Brad explained that the prior rules made sense because most newly listed companies are younger, have less revenue, and weaker profitability. But SpaceX is simply too big and too important—excluding it would be unreasonable. The exchanges and index providers made adjustments: they didn’t stuff it in at the absolute peak, avoiding the common problem where a 30% post-IPO pullback lands on passive investors.

Brad also revealed the latest developments at OpenAI: revenue has rebounded to roughly $70 billion this year, and GPT-6 could be released within 30 days. While that’s only twice SpaceX’s revenue and still below the rumored $100 billion for Anthropic, as one of the two major frontier labs, listing above $1 trillion with that growth rate is reasonable. He doesn’t think there is a race between the two; both will act when the timing is right. OpenAI’s corporate structure changes are more complex, so it may move after Anthropic.

Chapter 2: Token Costs Double Every 45 Days—Is AI Return on Investment Close to Zero?

Jason: Over the past few weeks, we’ve been discussing the ROI problem of token spending. In the industry, CTOs and CEOs have started publicly responding on X. Uber’s CTO Pinen shared their approach: 99% of engineers use AI tools; over 70% of pull requests come from local or cloud agents; and engineers have built 200 agentic skills. They deploy engineers to different departments as “frontline deployment engineers,” working with department leads to map out processes. Brad, what do you think about Uber doing it this way?

Brad thinks Chamath is right about the substance; the only question is the time frame. Right now, yes, a lot of spending goes into experimental buckets, so it may not show direct ROI. But enterprise adoption of AI is still early. The addressable market is every company on Earth—bigger than anything we’ve seen before. Revenue is also not concentrated; millions of customers make rational independent decisions every day.

Brad made a bold prediction: If Anthropic’s revenue exceeds $100 billion by the end of the year, its revenue next year could grow another 3 to 5 times. From $100 billion to $300 billion, the incremental $200 billion in revenue is unimaginable in Silicon Valley history.

Chamath’s skepticism centers on whether the ROI is sustainable. He asked Claude 5 two questions. The first: How much EPS growth does AI bring to the S&P 500? The answer was 50%. But he found that this figure also includes the revenue where Nvidia sells chips to Amazon. So he asked the second question: what is EPS growth for S&P 493 (excluding Mag7)? The answer was 9%. Broken down, the overwhelming majority comes from pricing power above inflation, and another 3% comes from buybacks. The true AI ROI attributable to AI is between 0% and 2%.

Chamath thinks the enterprise side looks shiny, but the problem is that smart investors like Brad and Gavin will eventually ask companies: what is your ROI? Where exactly is the EPS improvement? If the answer is “I’m not sure,” and you also don’t have sustained pricing power, then the enterprise side becomes fragile. The consumer side becomes the safe harbor instead, because you have tens of millions of buyers, the price points are far smaller, and the differences between two sets of buyers by two orders of magnitude can spare you from ROI scrutiny.

Jason added another perspective: the unique thing about this technology is that it touches everyone inside an organization. When Excel first came out, the accounting department was excited, but HR and the marketing department didn’t feel much. AI is different: in an organization of 1,000 people, everyone uses it—each person spends 200 yuan per month, doubling to 400 yuan; compared with a 150,000 yuan annual salary, that’s only an extra 3% to 4%. The key question is: does it make that person 3 to 5 times more efficient? If so, that explains why token spending is surging.

Chapter 3: Open Source vs Closed Source—Revenue Moves Toward the Frontier, But Enterprises Want to Run Away

Jason: Sacks, CTOs are starting to discuss intelligent routing on X. First, send tasks to an open-source model; if it can’t handle them, fall back to Claude. What do you think about this trend? If you’re an investor, and when the CFO at a frontier model starts asking “can it be cheaper,” what do you think about frontier model growth?

Sacks thinks enterprise CTOs do want to shift token consumption to cheaper models. They watch token costs skyrocket and try to hit the brakes—or at least control the situation. Add the AI sovereignty issue discussed last week, and enterprises worry about handing over their core alpha to a frontier lab that might become a future competitor.

Sacks’s core judgment is: enterprises want to move away from closed models, but most don’t have the technical ability to do it. The spirit is willing, but the flesh is weak.

Coinbase and DoorDash have managed to do it. They built token routing middleware, sending frontier tasks to frontier models and non-frontier tasks to standard models. But most enterprises don’t have this capability. That’s why the wallet share of closed-source models is actually increasing. The share of open source in enterprise spending fell from 19% last year to 11% this year. Of course, that doesn’t mean usage is declining; it may just be that using open-source models only pays for hosting, not the lab, making it hard to track.

Sacks also cited the viewpoint of Decagon’s founder: when you know exactly what you need to do, using a small, cheaper open-source model is correct—but you need data and post-training. If you still don’t know what you need to do, then you want the strongest general intelligence. Mature use cases: open source. Immature use cases: frontier models.

Jason mentioned a discovery by Ali, founder of Databricks: with the same model, switching the harness (task orchestration framework) can cut costs by half. GLM-5.2 combined with a specific harness performs extremely well; task volume can be cut directly in half. Jason also shared his own experience: he built a trend-discovery agent that runs every hour, and after optimization, token consumption dropped by 80%. When tokens became cheaper, he changed the agent from running daily to running hourly, and also split a single agent into three parallel tasks. When he woke up in the morning, he found 14 tasks were already completed—it felt completely different.

Brad’s view on this is: the core debate is whether intelligence will converge. When the DeepSeek moment happened 18 months ago, the market fell by 40%. Many people thought frontier models were finished and that open source would kill them. But 18 months later, the facts are exactly the opposite. Jesse Zang’s tweet pointed out that the wallet share of frontier labs is actually rising, even as token usage is rising on both sides.

Brad proposed a counterintuitive hypothesis: Maybe intelligence doesn’t converge at all. If superintelligence becomes self-recursive, the smarter the model gets, the more money it earns. The more money it earns, the more compute it buys. The more compute it buys, the better models it builds. Over the next 2 to 3 years, the gap may not be shrinking—it may actually be widening.

Jason also mentioned that he interviewed Lovable’s CEO Anton. The product launched about 30 months ago, with revenue rising from zero to $600 million. He also asked 11Labs CEO Matti: you’re a major customer of frontier models, spending tens of millions of dollars every year—are you worried about data leakage and competition? Both said they’re building their own models. These are eight-figure or nine-figure customers. If all of them start building their own vertical models, frontier labs will feel pressure. But Chamath pushed back: 11Labs wants to build the best voice agent in the world. If the best voice capability comes from frontier labs, can it afford the cost of building and relying on a slightly inferior self-built model in a competitive market?

Chapter 4: Zuck Launches a Price War—Same Quality, One-Hundredth the Cost

Jason: This week, Meta released Spark 1.1, a very strong agentic encoding model with extremely low pricing. Zuck has been unusually active on X, posting the most tweets ever. He’s basically saying: I’ll give you the same quality, but the cost is only one percent. Brad, what do you think about Zuck’s strategy?

Brad thinks Meta previously made a misstep with its open-source strategy, but now Zuck has clearly chosen the direction of a price war. Meta also released a new model API—not just providing a model, but providing tokens as well. Competition for the U.S. is a good thing.

Brad used an analogy to explain why frontier models won’t be easily replaced. If your AI agent is replacing a $200-an-hour consultant, whether you use a $3 cheap model or a $15 frontier model—the price difference simply doesn’t matter. The real question is whether that $15 model can complete the task without making mistakes. If the job crashes halfway through, you lose both tokens and time.

Chamath disagreed. He thinks it’s like when the iPhone first came out and everyone kept upgrading because the new price was worth it. But eventually people will say, “My old phone is good enough.” When he tried Claude 5, he found that some research directions were restricted—it wouldn’t answer. Everyone reaches a “good enough” threshold at different points in time.

Chamath also shared his experience with the UN AI committee. He attended with Benioff, Jensen, and Brad Smith at a UN AI committee session co-hosted by Benioff. His observation was: there is no country in the world that isn’t drafting its own sovereign AI strategy, and no country is willing to use America’s closed-source models as the answer. Many countries would rather take an open-source model—like Nvidia’s—and build an entire stack of infrastructure around it.

Examples of sovereign AI include: the Falcon model from the UAE, Saudi Arabia’s Arabic LLM, and Japan investing $6 billion in the Neoterra alliance directly jumping to physical AI and robotics. Chamath believes that when models reach 95% to 99% of frontier capability, many countries will say “it’s good enough.” On the other hand, some companies don’t have enough profitable growth to support that level of spending, and they also lack the nerve to cut costs at large scale. It’s like the famous letter he wrote to Zuck: Zuck only executed eventually because pressure forced him to. Most companies only let problems accumulate.

Chapter 5: China Considers Restricting AI Model Exports—Open the Door After You Catch Up, Then Start Closing It

Jason: Reuters reports that the CCP is considering restricting overseas access to China’s top AI models. Two regulatory agencies have held meetings with Alibaba, ByteDance, and Z.AI (the one doing GLM-5.2) to discuss restricting overseas access to top open-source and closed-source models. They also classify leaked AI research as national security crimes, and they want to control who can invest in China’s AI labs. Sacks, last week I asked the reverse question: should the U.S. ban Chinese models? Now it’s flipped—China says it wants to restrict. What do you think about this move?

Sacks thinks the news may be somewhat exaggerated. China’s top model is ByteDance’s, and it is originally closed source. Alibaba’s Qwen was previously open source and may now be shifting to closed source. Z.AI’s GLM-5.2 was also previously open source and may now be shifting to closed source.

Sacks’s judgment is: the strategy is very clear. When you’re catching up, you go open source; when you get close to the frontier, you go closed source. Sam Altman did the exact same thing with OpenAI three years ago—moving from nonprofit to for-profit, and from open source to closed source.

The advantage of open source is that it attracts a developer community and gives you a reinforcement learning data flywheel in AI. But once you catch up, closed source captures all the value.

Sacks discussed this with policymakers in Washington and even with the White House and the Treasury. He said that among all regulatory disputes, there is one absolute consensus: do whatever it takes to stay ahead of China. From the President down, everyone is asking “How far ahead are we?” and “What do we need to do to stay ahead?” The idea that America’s frontier labs exit while China’s open-source models remain freely circulating simply doesn’t exist in Washington. He also revealed that GLM-5.2 includes a distilled watermark from Mythos, and the U.S. government will likely move to crack down on distillation.

Sacks believes what China is doing won’t affect the U.S. much. The U.S. has the ability to create open-source models—Nvidia is doing it, and Reflection is doing it too. He talked to frontier labs about why they don’t do open source, and their answer was, “Demand isn’t that big. If demand is big, we’ll do it.” For China, restricting exports may actually hurt itself more.

Chamath made a joke: for the U.S., the best thing would be for China to produce a doomsday-prophet community too, spending all day worrying about AI unemployment and existential risks. If China’s labs also start getting constrained by regulation, that would be the biggest win for the U.S.

Chapter 6: Trump Accounts—Open an S&P 500 Account for Every American Child at Birth

Jason: Brad went to Washington this week. The Trump Accounts app is already the #1 most downloaded app globally. Congratulations, Brad—four years of your hard work. Tell us what happened.

Brad explained that this has been a four-year journey. Last year, as part of the Invest America Act, it was signed into law as a bill. This year, on July 4, the app went live officially. Every newborn in the U.S. receives $1,000, deposited into a private investment account, with all of it invested in the S&P 500. The account is free for life. Within 24 hours of launch, they opened 1.5 million accounts and attracted more than $1 billion in deposits. They held a New York Stock Exchange and Nasdaq joint bell-ringing ceremony for the first time in history—right there in the White House Oval Office—where hundreds of CEOs attended. The President proposed automatically creating accounts for 50 million to 70 million minors under 18.

Sacks analyzed how powerful this mechanism is from a financial planning perspective. Each year, up to $5,000 can be contributed to the child’s account (from family and friends as well). Employers can contribute an additional $2,500 tax-free. Before age 18, enjoy tax-free compounding. After age 18, you can withdraw up to 25% to buy a home, start a business, or go to college, and the remaining amount rolls into an IRA. If you wait until the child is no longer a dependent (for example, right after graduation in a 0% tax bracket range) and then do an IRA-to-Roth IRA conversion, you can essentially turn the money into lifelong tax-free investing with almost no taxes.

Sacks did the math: If Trump Accounts is fully funded from the start, then based on market return rates from the past 30 years, by age 28 that child becomes a millionaire. If there is $200,000 to $300,000 by age 18, then by age 60 compounding could grow it to over $10 million.

There were also a series of major charitable announcements. Michael and Susan Dell donated more than $6 billion, providing $250 each to children in 25 million low- and middle-income families. SpaceX President Gwen Shotwell donated $350 million worth of SpaceX stock, earmarked for children in low-income communities. Micron donated $250 million, with up to $1,000 for each employee’s child. Brad himself donated $100 million, covering all children in Indiana.

Brad said they told the President they expect to raise $100 billion within 12 months. This would become the largest direct charitable platform in American history—no intermediaries, money goes directly into the children’s accounts, and it can’t be withdrawn before age 18. On this trajectory, in the next ten years there could be more than 100 million private investment accounts, and in the next 15 years possibly $2 to $4 trillion flowing into family accounts that previously had nothing.

From a more macro perspective, Jason wrapped it up. He said this project can replace Social Security, and it replaces the pledge to donate. In the U.S. today, only 50% of people own stocks. If Trump Accounts succeeds at scaling, that could rise to 70% to 75%. Australia is one of the happiest countries in the world largely because its superannuation system forces everyone to deposit 12% to 14% of their income into an account similar to a 401k. Trump Accounts is doing something similar, but at a more fundamental level.

Jason also especially thanked Joe Gebbia (Airbnb co-founder) for joining the government-responsible software design for this project. He said the U.S. government built excellent consumer-grade software, which is rare in history. Brad added that the team includes Michael Dell, Vlad Tenev (Robinhood CEO), Joe Gebbia, and Luke Pettit from the Treasury Department. Their goal is not just to build the best government product, but to build one of the best consumer products as well.