20VC: NVIDIA Bonanza: Buys Poolside & Invests in Mercor and Perplexity | Anthropic's $30TRN Revenue Assumption & OpenAI Confirms IPO | Why Customer Service, Defence and Robotics are Overinflated

Summary

The episode centers on Nvidia’s multi-pronged AI push: buying or backing companies across model, infrastructure, and application layers, including Poolside, Mercor, and Perplexity. The hosts argue this is rational ecosystem strategy, especially if open-source and open-weight models increase total chip demand, but they also stress how extreme frontier-model capital intensity creates forced outcomes and “fail-up” acquisitions. A second major thread is competitive positioning between OpenAI and Anthropic, with the panel arguing OpenAI’s slower growth and weaker relative rank may force a 2027 IPO, while Anthropic’s code focus and margin improvement look stronger. They also debate whether AI revenue is being overestimated in areas like support, defense, and humanoids, and conclude that agent trust, security, and enterprise budget discipline are still major constraints.

Notes

• Nvidia is presented as moving across the AI stack at once: Poolside, Mercor, Perplexity, and OpenAI-related financing are treated as one strategic pattern.
• Poolside is described as a model-layer transaction: Nvidia pays to license the model factory and also adds equity capital.
• The engineering transfer of 109 people to Nemotron is part of the deal structure, not just financial sponsorship.
• Poolside reportedly failed to raise enough capital for its next compute step, especially the GPU buy needed for scaling.
• The panel frames frontier AI as capital-wall driven: even good teams can be forced into acquisition or licensing by compute economics.
• Rory says Nvidia’s open-source support is economically complementary because more open-source share means more chip demand.
• Jason argues a 15x outcome is not enough for a seed investor when entry prices are already high.
• Rory counters that even a failed venture bet can still generate a rich return if it returns 15x.
• Frontier models are described as winner-take-most, with trillion-dollar closed-source winners versus smaller open-weight winners.
• Rory says Nvidia’s Mercor investment is less obviously accretive to chip demand than Poolside because it does not directly translate into more GPU sales.
• Jason cites Kroll-style analysis suggesting gross margins above 30% may no longer move M&A pricing much in agentic software.
• Jason argues Nvidia may have a de facto ecosystem budget and can spend it aggressively to support customers and partners.
• Rory notes Nvidia’s cash flow has surged from roughly $4 billion to $50 billion, with very high gross profitability and a large cash pile.
• Rory distinguishes three deal types in Nvidia’s AI strategy: buying adjacency, direct investing, and vendor financing.
• Rory estimates combined OpenAI and Anthropic revenue at roughly $100 billion and frontier training spend at $5 billion to $10 billion.
• Harry pushes back that several AI vendors already spend billions on training-like efforts, implying a bigger market than Rory’s base case.
• The key pricing question becomes what multiple the market assigns to recurring frontier-model spend, not just the revenue pool size.
• Cursor is used as a precedent for a negative-gross-margin product improving economics over time and reaching a much larger valuation.
• Rory cautions that not every negative-margin startup will recover, but Anthropic’s margin turnaround shows the pattern can work.
• Jason’s investing heuristic is to ask whether the company can win if everything goes right.
• Rory says OpenAI’s IPO messaging was forced by slower first-half growth and Anthropic’s stronger trajectory.
• OpenAI is now described as number two, which weakens its pricing power and strategic options versus Anthropic.
• Jason says a multi-model world makes life harder for the number-two player, especially when buyers can route to cheaper alternatives.
• Jason says open weights and open source increase pressure on the number-two frontier player, which cannot compete mainly on price.
• Rory says open-weight models will process most tokens, but frontier models will still capture most revenue.
• More open-weight competitors are expected to compress frontier gross margins, especially if one is backed by Nvidia.
• Rory says Anthropic and OpenAI will have to go public in 2027 if current trajectories hold.
• Rory says OpenAI has less room to maneuver because rivals are already ahead, profitable, and public.
• Jason says OpenAI’s differentiated mission is less clear now than it once was.
• Harry highlights ChatGPT’s consumer brand reach and suggests OpenAI could evolve into an ad business.
• Rory says Anthropic’s code focus is the most valuable compute wedge.
• Jason argues consumer AI is heavily subsidized for power users and therefore a weak standalone business model.
• Rory says consumer AI could still become large if the brand remains dominant and ad economics improve.
• Rory calls Anthropic’s $30 trillion TAM claim delusional because it effectively equals the entire U.S. GDP.
• Hugging Face is framed as a strategic enterprise asset for buyers wanting an open-model counterweight to frontier labs.
• Rory estimates Hugging Face revenue at roughly $150 million and says a large company would value it as a strategic counterbalance.
• Jason warns an acquirer should not alter Hugging Face, or the marketplace would break.
• Harry notes Citadel unwound most of Leopold Aschenbrenner’s AI trade, implying the position had already been de-risked.
• Rory says leverage and vendor financing require being right at every step, not only in the long run.
• Jason says the AI trade is still strong, citing a basket he describes as AI on steroids up 56.46% on the year.
• Jason mentions elite-status spillovers in Korea’s memory sector, where Samsung engineers are seen as highly desirable.
• Jason says mediocre Dogpatch apartments now rent for about $10,000 a month, underscoring AI-era local inflation.
• Rory says California captured roughly three-quarters of total AI dollars, with OpenAI and Anthropic alone around 60%.
• Rory expects AI wealth to keep driving local prices up, rather than reverting fully.
• Jason says the AI cycle is still early, implying more upside remains.
• Rory’s supply-side view is that Nvidia and hyperscalers will not blink; the constraint is capital or demand.
• Rory says public-market IPOs for Anthropic and OpenAI could keep the cycle going by turning private growth into public capital.
• Rory says Anthropic has even discussed $200 billion of GAAP revenue in the 2027 context, which he views as a slowdown from current rates.
• Jason describes a token-addiction cycle: performative spending first, then budget backlash.
• Rory says intelligence should be treated like capital: priced, allocated, and managed.
• Rory says 2027 is the year enterprises must decide whether to let AI run and hope the ROI shows up.
• Rory says productivity gains will not be allowed to reduce EPS; spend must be offset elsewhere.
• Jason says CFOs are mostly worried about retention because AI firms can poach top people with stock and comp.
• Jason says losing 30% of the company to an AI leader like Harvey would be fatal.
• Rory says token budgets are now a real labor-cost substitute inside companies.
• Jason says both the best and worst users are now addicted to tokens, with the best wanting many agents running continuously.
• Rory says the pace of diffusion across industries matters as much as the model quality itself.
• Jason says Stripe benefits from every agentic product because payments routing naturally defaults to it.
• Rory says AI-driven growth can produce a public-stock-like upside for private companies with diversified core businesses.
• Jason says AI leaders like OpenAI and Databricks would reshape the public software leaderboard if they were public.
• Rory says Stripe is already behaving like a public company while remaining private.
• Jason says current agents still cannot be trusted with sensitive work.
• Jason says open-weight models have fewer guardrails and can be used for harmful instruction-following.
• Rory says agents may fit repetitive enterprise workflows better than idiosyncratic personal tasks.
• Rory says personal productivity is real but niche because many users do not care about inbox perfection.
• Customer support is described as a likely commoditized AI category.
• Rory says defense can matter strategically while still producing poor venture returns because procurement favors portfolios and consolidation.
• Rory says humanoid robotics is real, but the category is smaller than people think compared with focused robotics.
• Jason says classic customer support and CX will collapse into broader agent surfaces within 24 months.
• Jason says venture money cannot simply turn law firms or accounting firms into the next big AI exits.
• Jason says the AI era has unlocked unusually high founder creativity and makes formerly weak categories viable again.