Organized & Compiled: Deep Tide TechFlow

Guests: Gavin Baker, Founder and Portfolio Manager of Atreides Management, public NVDA and SpaceX bull; David Sacks, Partner at Craft Ventures, Host of All In Pod, investor in Anthropic

Hosts: Jason Calacanis (main host), David Friedberg, All In Pod

Podcast Source: All In Podcast

Original Title: Anthropic's 2T IPO, Zuck's AI Manifesto, Nvidia's 500B AI Bet, Grok's Comeback

Broadcast Date: August 14, 2026 (recorded on Thursday, released on Friday)

Conflict of Interest Statement: David Sacks is one of the investors in Anthropic; his statements regarding Anthropic in this episode are influenced by his investment position; Gavin Baker has a long-standing public position in NVDA and holds a position in SpaceX's private market. Their comments should be read in light of their respective positions.

Key Points Summary

This episode of All In is the "2+2" round table with Chamath absent, co-hosted by Jason Calacanis and David Friedberg, presenting four major AI events of the week: Anthropic's $2 trillion IPO window, Zuckerberg's 6,500-word manifesto against the “apocalyptic club,” Nvidia's $500 billion computing financing platform, and Grok 4.6's return after the xAI rebranding to SpaceXAI.

Gavin Baker, a tech stock investor appearing for the first time on the show, has long held NVDA and SpaceX publicly, and dissects each event from the perspective of “AI computing + capital structure.” Sacks often presents the same facts but extends the discussion into political philosophy: he suspects that the $2 trillion valuation is a “banking leak to set the rhythm,” and calls Anthropic's S-1 filing “the pace car for whether the entire AI industry is a bubble.”

The four topics support and undermine each other: Anthropic is going public, Zuck wants open sourcing, Nvidia aims to turn GPUs into the “Fed of AI,” and Grok is set to challenge OpenAI from a score of 61. The most valuable judgment of this episode is concentrated in one sentence: where will the “sudden stop” in AI computing expansion come from?

Highlights of Opinions

Signals of Anthropic and AI Bubble

• “If there is any sudden stop in Anthropic's quarterly report, you will see a bunch of companies crash behind it... They generate $10 billion in revenue per GW of computing power, which is the fundamental reason the entire food chain can afford current chip prices.” (Sacks)
• “I think that's exactly why Anthropic going public is so important for the market. You'll have quarterly earnings reports to look at their numbers each quarter. I guess this will be the most important signal for the entire industry... If there’s a bump in Anthropic’s quarterly report, it will be the first to show it.” (Sacks)
• “Pricing at $2 trillion may instead be a signal of a smooth roadshow and conservative pricing. If bankers lose their lead underwriting position, they'll leak information to the media to set the pace.” (Gavin)

Zuckerberg's Anti-Apocalyptic Manifesto

• “What makes me nod the most is the core contradiction he points out at the beginning: Effective Altruism and the Anthropic side are rushing to create a future they don’t even believe in, which means this worldview is broken from the start.” (Sacks)
• “AI safety should be based on a balance of power; there should be no single centralized intelligence. History shows that absolute power appearing in the guise of charity has never produced a benign result.” (Quote from Zuckerberg in "The Future Belongs to Everyone," relayed by Sacks)
• “If the folks at Anthropic truly believe the future is dangerous, the only safe method is to centralize AI in one or two companies to form a cartel, deeply bonded with the government. This is a very authoritarian idea.” (Sacks)

Nvidia's Computing Securities Platform

• “Smart asset managers like Blackstone and KKR are willing to get into this business because they believe GPU computing power is a financeable asset. Nvidia acts as an intermediary, connecting clients in need of computing power with lending institutions.” (Gavin)
• “In this structure, Nvidia basically becomes the central bank of AI, similar to the Fed. The Fed's monetary policy is also implemented through commercial banks; here, it’s implemented through these PE firms and major Wall Street banks.” (Sacks)
• “This network has a natural safety valve and a dead man's switch. The safety valve is the physics of the real world; you can't build that many data centers. The dead man's switch is the two largest clients, Anthropic and OpenAI, who can pause at any time. OpenAI has already cut its building scale from $1.4 trillion to $600 billion.” (Gavin)

Grok 4.6 and the Triangular Stand-off

• “Grok 4.6 is not the end; it’s just a relatively small model with 1.5 trillion parameters. Grok 4.7 will be released in a few weeks and will be much larger.” (Gavin)
• “The reason Elon has caught up this time involves two things: First, he secured the Cursor team, which is a well-coordinated elite; second, he overhauled the management of xAI by bringing in top talent from SpaceX.” (Summary of Gavin's judgment by the host)
• “Grok Bot is also a moment like OpenClaw, pushing AI onto everyone’s desktop.” (Host Jason)

Main Text

1. Opening: The Afterglow of SpaceX Is Still Present, This Discussion Will Be More Intense

The opening features some playful banter about the afterglow of SpaceX. Gavin Baker from Atreides Management just made a hefty profit from the SpaceX IPO; host Jason jokes, “You’re still reveling in SpaceX’s glory; this must be the biggest deal of your career, right?” Gavin doesn’t deny it, only saying that SpaceX’s story is not over yet. After a few exchanges, they dive into the main topics: Anthropic’s $2 trillion IPO, Zuckerberg’s 6,500-word article, Nvidia’s $500 billion computing platform, and Grok 4.6 returning to the table. Four players, four bets.

Jason: We record on Thursday and release on Friday. Yesterday, news came that Anthropic's announcement might break records; FT reported on Wednesday evening, $2 trillion, possible October IPO, about 6 to 8 weeks from now. PolyMarket indicates an 80% chance of Anthropic going public this year; the probability of it being the top model by year-end is 67%. The annual operating revenue is estimated to be between $100 billion and $120 billion by the end of the year. This year it's increased tenfold, at a $2 trillion valuation, the price-to-sales ratio is 16 to 20 times, which is lower than SpaceX and Palantir when they went public. Sacks, this is your old topic; you said last time it could reach $3 to $4 trillion. How do you think about it this time?

2. Anthropic Going Public: The $2 Trillion Leaked Valuation Might Be Conservative

Gavin responds first: “I am not an Anthropic shareholder at this time. The timing of Anthropic's IPO is a bit late for me—I am not used to investing in Elon’s competitors.” He explains that Elon started renting computing power to Anthropic after capacity overflow at Colossus, “He trusts Dario to some extent, as they haven't triggered Elon’s ‘evil detector.’”

On the $2 trillion figure, Gavin is skeptical: “We've seen this before with SpaceX; every investment bank strived to be the lead underwriter, and the one that lost would leak information to set the rhythm. So, this $2 trillion might just be a leak to trouble the lead underwriting bank.” His logic is: If the final pricing really is $2 trillion, that likely means the roadshow was too smooth; the pricing is instead conservative, and the company can absorb the lock-up period without severe fluctuations at the start. “If it comes out at $2 trillion, I would not be surprised if it jumps to $3 trillion on the second trading day.”

Sacks: You’ve mentioned several times this year that this revenue growth rate hasn’t been seen in Silicon Valley history. If the $100 billion annualized materializes and continues to grow tenfold next year, by the end of next year, Anthropic would be at a $1 trillion annualized rate. The market's question is whether the TAM is enough, whether the computing power is sufficient, and whether the energy is adequate. I view TAM not as a problem, nor demand, but the bottleneck might be in the physical world. But I agree with Gav's assessment, the $2 trillion figure itself should not be taken too seriously; it’s necessary to look at the specific numbers in the S-1 filing.

Jason: Sacks, you often say, “If there is any sudden stop in Anthropic's quarterly report, you will see a bunch of companies crash behind it.” I’ll ask on behalf of the audience, where will that sudden stop come from? Is it an issue with demand, pricing, or competitors catching up?

Sacks: I actually think Anthropic is the “pace car” for whether the entire AI industry is a bubble. Anthropic can generate about $10 billion in revenue per GW of computing power, which is why SpaceXAI can sell computing power to Anthropic at $50 billion per GW and why SpaceXAI can afford to buy Nvidia’s $300 billion chip procurement. The entire food chain is linked by this chain. There is a natural physical safety valve in place; you can't build that many data centers; there’s also a natural ‘dead man’s switch’—the two largest clients, OpenAI and Anthropic, can press pause at any time. OpenAI has already cut its building scale from $1.4 trillion to $600 billion.

Jason: So if Anthropic really puts the brakes on, does the industry have any space to avoid crashing?

Gavin: No space. If Dario slams the brakes because demand drops, there will definitely be a crash; the two cars have no distance between them. If it’s because OpenAI or SpaceXAI or open-source models have caught up, that’s another story. The S-1 filing for Anthropic will be very important for the market, because many macro and value investors confidently claim ‘AI is a bubble, capital expenditure is unreasonable, and there won't be an ROI.’ They actually have a fundamental erroneous assumption: they think tokens are subsidized. Anthropic is already profitable, the entire token chain is profitable. Once that's put in black and white in the S-1 document, it will flip many minds. (Sacks adds that OpenAI has already pressed that brake once, cutting its scale from $1.4 trillion to $600 billion.)

3. Zuckerberg's Anti-Apocalyptic Manifesto: Power Should Be Distributed, Not Cartelized

Jason raises the second topic: Mark Zuckerberg published a 6,500-word article this week titled “The Future Belongs to Everyone.” He rejects the OpenAI view of “AI fatalism,” directly going another route: advocating for open sourcing, equipping everyone with agents, and providing the whole world with AI teachers and limitless abundance.

Jason: Sacks, have you read it?

Sacks: I have. I agree with most of it. The core contradiction he points out at the beginning strikes back at Effective Altruism and those aligned with Anthropic, saying: if you really believe that the future means widespread unemployment, AI run amok, and AI misused, then why are you so eager to create it? This immediately penetrates the foundational contradictions. I think the real thoughts of those people are not driven by profit, but stem from a missionary mindset. They genuinely believe that this future is dangerous, that only under their 'enlightened control' can humanity be safe. This argument was clearly articulated thirty years ago by Thomas Sowell in “Vision of the Anointed”; intellectuals often believe, once they maximize their power, society will change for the better as they designed. Such experiments have never succeeded in history; in the end, it's either hollow promises or an excuse for totalitarian politics.

Sacks: Zuckerberg’s later comments hit the nail on the head. He says: “Believing AI is so dangerous that only absolute power concentration can ensure safety is a problem in itself. History shows relying on absolute power disguised as charity has never produced benign results.” The core framework of this issue isn't just about open source versus closed source, but centralized versus decentralized. What we want is maximum individual empowerment, the result of decentralization, the open models you always talk about, and data sovereignty. Anthropic and OpenAI can certainly continue their proprietary cutting-edge models, but we want everyone to be empowered. I agree with Zuckerberg here.

Gavin: I read it and agree as well. Let me compress David's words: Anthropic and the Effective Altruism side believe this technology is too dangerous and must be centrally managed, while Zuckerberg, Elon, and Jensen believe this technology is too dangerous and must be widely distributed.

Jason: If Zuckerberg comes to our table for a day, we can extract all eight of his articles; this 6,500-word piece alone plus 70,000 words could be a book. Harper Business could directly reach out.

4. Nvidia's $500 Billion Computing Platform: Securities for GPUs, Making Jensen the AI Fed

Jason: Next, Nvidia published on X, saying “AI computing power is becoming an investable asset class.” They’re teaming up with Goldman, BlackRock, Blackstone to raise $500 billion for computing power. Jensen himself wrote an article for X. The basic approach is: clients don’t pay hundreds of billions up front for chips, but rather borrow money to purchase, repaying over time as revenue comes in. Nvidia acts as the intermediary. Sacks, we talked about this two years ago.

Sacks: I saw that CNBC feature. The private equity guys put it very well: This is similar to aircraft financing; airplanes have residual value, so even if an airline goes bankrupt, financiers don’t lose out. This is asset-backed financing, not revolving credit. However, to actually accomplish this, Nvidia must standardize and provide reference designs, allowing Wall Street to bundle and securitize it into MBS-like products.

Gavin: First, let’s clarify one big point: Blackstone, KKR, Goldman, Apollo, BlackRock are coming in on the premise that they recognize GPU computing as a financeable asset. Nvidia is wise to position itself as the intermediary: it claims that its computing power has a long enough lifespan, so borrowing rates can be lower than those for other computing powers. This leads to two secondary benefits. First, China’s three open-source model companies, Quen, Kimi, DeepSeek, GLM, have different paths for architectural evolution, which actually helps Nvidia; the more diversified the architecture, the greater the demand for flexible computing power. Second, Nvidia has also offered a residual value guarantee: that in 3 to 4 years, these GPUs can still be rented out at a certain price. This guarantee can be embedded into the underlying underwriting model, further lowering financing costs. Nvidia itself only bears 25% of the risk, leaving 75% to some of the world’s smartest underwriters.

Jason: So where is the biggest risk?

Gavin: The biggest risk lies not on the demand side, but on the supply side; if there is an oversupply of computing power and everyone builds it out, we might see a GPU version of “fiber optic abandonment.” **Even so, the current political headwinds in AI, NIMBY, data center site protests, and various moral panics might actually protect against overheating.

Sacks: I think Jensen's true brilliance lies in how he positions it: he doesn’t view this as “financing for clients” but rather as “building a production line.” Originally, TAM was constrained by financing capabilities; he created a structure that dismantled that constraint. So now that Elon wants to add 6 to 8 GW of computing power next year, with capital expenditures of $300 billion to $400 billion, he has this pathway. A simple method is to get seller financing from Nvidia; based on his calculations, the payback period could be as short as one year. This means Jensen, through big banks and PE firms, has extended his capital market reach.

Sacks: This is actually very similar to a central bank role. Like the Fed, the Fed’s monetary policy also relies on commercial banks for implementation. This time, Nvidia is acting as the intermediary, with PE firms and major Wall Street banks dropping down to the downstream. So in a sense, Nvidia is now the central bank of AI, the Fed of AI.

5. Grok 4.6: Elon Catching Up to Anthropic in 6 Months, OpenAI's Dominance Narrative Crumbles

Jason: Gavin, what do you think about Grok's events this week? Grok 4.6, the Artificial Analysis composite intelligence index gave it a 61 score, leveling with OpenAI’s GPT-5.6 Sol, only a very small margin behind Anthropic’s Claude Fable 5 (62) and Opus 5 (64). The pricing is $2 per million tokens for input, $6 per million tokens for output, which is 60% cheaper than Opus 5 and GPT-5.6 Sol. Last week I was saying the leading models would revert to an Anthropic + OpenAI duopoly; do I need to change that view?

Gavin: Yes, you should change it. The leading models are actually three now. Grok 4.6 can be used on the day it goes live on Cursor, and is integrated with Vercel, Cloudflare, OpenRouter for distribution. DHH (David Heinemeier Hansson, creator of Ruby on Rails) has previously provided assessments on X, and Databricks’ internal benchmarks are also well-regarded. We’ve done several independent assessments, and it stands out in the cost-performance quadrant; for the same capabilities, it presents prices that are a fraction of its competitors.

Jason: So how did it catch up in six months?

Gavin: Two things. First, he acquired the Cursor team, which is a well-honed elite team; second, he overhauled the entire management of xAI, bringing top people over from SpaceX. Now, the 4.6 model with 1.5 trillion parameters is relatively small; Grok 4.7 will be released in a few weeks and will have parameters ramped up to 2 trillion, making it even more powerful.

Jason: Sacks, does this mean Anthropic’s lead is not as solid anymore? Since you’re also an investor in Anthropic, aren’t you worried?

Sacks: Not worried. I’m actually more concerned about the opposite; what he’s doing may actually benefit Anthropic. Elon is currently developing large cluster computing; his monetization paths are twofold: on the upswing, he trains cutting-edge models; on the downside, he sells this computing power at spot prices to leading companies. So he’s training his own model (call option) while also selling computing power to competitors (put option). Both sides are profitable. But I want to highlight a new tension; he recently signed a computing power contract with Anthropic, which contains a mutual 90-day exit clause. The result is: if the spot price rises, his exit option becomes valuable; if the spot price drops, Anthropic can renegotiate. This is how Elon plays the “Spot minus 90” pricing game.

Jason: What about agents like Hermes?

Sacks: Hermes and OpenClaw share a similar idea, putting AI agents on everyone’s desktop. This aligns with the logic of Zuckerberg's declaration about “everyone having a bot”; Anthropic’s Claude Co-work, Claude Tag, and Perplexity’s Computer are also following the same path.

6. The S-1 Document Will Derail “AI Is Not a Bubble” Arguments

Both point to the same material: the upcoming S-1 filing from Anthropic. Gavin has already mentioned that macro and value investors have a deeply entrenched erroneous assumption: they think tokens are subsidized, maintained as an illusion by cloud providers and chip manufacturers in revolving credit. However, he has heard numerous public signals indicating that Anthropic is making money, and OpenAI will soon be profitable; the entire token food chain is currently profitable. If this S-1 lays out the actual gross margins, net retention, and unit economics on the table, fund managers holding to the “AI is a bubble, capital expenditure won’t yield returns” narrative will find it hard to backtrack.

Sacks’ conclusion is more straightforward: Anthropic going public is good news for the entire AI industry. It will allow the market to see a complete, quarterly, GAAP-compliant set of numbers from a cutting-edge model company for the first time, enabling everyone to validate every 90 days whether the assessment of “AI is a bubble” still holds.

7. Workday Conclusion: $43 Billion Private Equity Buyout

The last segment of the program discusses Workday, where private equity giant Silver Lake is in negotiations near a $43 billion private buyout, corresponding to an 18% jump in stock price. Sacks comments: After the arrival of the AI era, the renewal logic for SaaS companies has been unlocked, transitioning from “charging per seat” to “charging based on the number of agents,” which includes the premium of this story in Workday's high valuation for compensation and human resources systems.