Anthropic’s IPO Leak, OpenAI’s Dots vs. Meta’s Muse, Visual Turing Test

Summary

Alex and Ranjan spent most of the episode unpacking Anthropic’s leaked pre-IPO numbers: rapid revenue growth, but also extraordinary decade-long infrastructure and compute commitments that could pressure the company, its hyperscaler partners, and the broader AI trade if growth slows. They debated whether those commitments create real bankruptcy risk or whether counterparties would renegotiate before letting the system break. From there, they shifted to AI agents, with Alex giving a detailed hands-on review of OpenAI’s Dots and arguing that frontier models matter for higher-quality work automation, while Ranjan countered that most agent tasks are finite workflow problems that do not require frontier intelligence and may increasingly run on cheaper open-weight models. The final stretch focused on Tavus/Griffin-style AI video avatars, which the hosts found impressive but concerning because latency is still the main bottleneck and realistic faces could supercharge scams, catfishing, and spam.

Notes

• Reuters reportedly leaked most of Anthropic’s pre-IPO S-1, giving the hosts hard financial numbers instead of ARR-style marketing metrics.
• Anthropic’s revenue was framed as scaling from $400 million in 2024 to $4.6 billion in 2025.
• The standout number was a planned $518 billion of compute and infrastructure spending over the next decade.
• Alex emphasized that 80% of those commitments are binding and noncancelable.
• Ranjan immediately wanted the missing H1 2026 filing data, arguing the leak stops short of the numbers that will matter in the public S-1.
• Ranjan said the filing still leaves Anthropic with roughly $8 billion in net loss and about $2 of loss for every $1 of revenue.
• Alex said Anthropic had been described as targeting roughly $100 billion annualized revenue by year-end, which would make the spending commitments look more manageable.
• Alex argued that failing to grow into the commitments could push Anthropic into bankruptcy and hit its counterparties.
• Ranjan pushed back, saying the commitments are so large that lawyers and incentives would likely prevent a true collapse.
• Alex named the biggest counterparties as Alphabet, Amazon, and Microsoft, and linked the issue to hyperscaler cloud growth.
• He also tied Anthropic/OpenAI demand to cloud and chip investors, especially Nvidia, and to public-market concentration.
• Ranjan highlighted customer concentration as a major risk factor, saying nearly a quarter of revenue came from just two customers.
• Alex argued that if the S-1 disappoints, the market’s AI valuation premium could come under pressure.
• Ranjan said a weak filing could reverse the current rotation into AI names and force investors to question extreme valuations.
• The hosts argued that the whole market is unusually dependent on AI winners and their capex cycle.
• Alex said OpenAI’s Dots agent can run continuously and sequence tasks across his laptop and browser.
• He said it handled invoice work by finding the file in Drive, drafting a Gmail message, and writing the note.
• The travel-planning flow still required a final human click to complete the booking.
• Alex said Meta’s Muse is not a strong enterprise challenger because enterprise buyers want frontier intelligence.
• Ranjan argued that the core question is trust, because these agent products are beginning to look interchangeable.
• Ranjan defined enterprise as large, complex organizations with security and approval workflows, not just general work tools.
• Alex said Dots succeeded at signing a speaker release in his email after Claude Cowork would not do it.
• Ranjan argued many of the showcased tasks do not require frontier intelligence and are better treated as smaller text-processing problems.
• He warned that always-on inbox mining creates engagement pressure and can become overbearing.
• Alex said Muse can be useful, but it crossed a line by proactively sending him a news alert he did not ask for.
• Ranjan said proactive behavior is magical when timely, but bad when it surfaces irrelevant actions like a late dinner-reservation check.
• Alex said the newest OpenAI and Anthropic models feel much deeper and more thoughtful than prior versions and competitors.
• He added that they diagnose story drafts better and understand what is missing when fed articles and interviews.
• Ranjan said enterprises are shifting away from frontier-only models because the tasks are bounded and the cost tradeoff matters.
• He also said open-weight models are now only about 3 to 6 months behind frontier systems, versus 12 to 18 months before.
• Alex introduced Tavus’s Griffin as a lifelike avatar that can hear, see, speak, and react on live video calls.
• The company claim was that nearly half of a short test thought they were talking to a human.
• Ranjan said the avatar bottleneck is mainly latency rather than pure visual fidelity.
• He called the demo very impressive and also terrifying.
• Alex said realistic avatars would be especially useful to scammers and pig-butchering operations.
• They discussed catfishing and pig-butchering as obvious abuse cases for humanlike AI avatars.
• Ranjan joked that the next phase is “Shrimp Jesus” avatars that can be talked to.
• Alex predicted one company will ship with guardrails, others will ship without them, and one avatar could become a kind of religion.