20VC: Deepseek Raises $50BN | Wall St's $725BN AI Question | The Rise of Open Source & How it Threatens OpenAI & Anthropic | OpenAI Builds it's Own Chip: Jalapeno | The Death of Moats & The New AI Software Winners

Summary

Harry Stebbings, Jason, and Rory frame the current AI market as a battle over talent, cost, and sovereignty rather than just model benchmarks. They discuss Google losing prominent AI researchers, arguing that elite engineers increasingly choose environments with research freedom and the ability to ship, not just pay. The conversation then widens to open source, especially China-linked models like DeepSeek and Zhipu, which the speakers see as both a deflationary force on pricing and a sovereignty strategy for non-US markets. They also stress that infrastructure constraints are showing up in memory and DRAM pricing, while hyperscaler capex is racing far ahead of AI revenue. The final sections focus on enterprise ROI, agentic workflows, and OpenAI’s custom Jalapeno chip, which they debate as a possible cost-defense move against open-source pressure and a sign that the middle of the AI market is getting squeezed.

Notes

• Google is framed as losing two major AI scientists while Anthropic and OpenAI win on research environment and shipping velocity.
• The talent war is described as about freedom to pursue research, not just compensation.
• The speakers suggest Google remains strong, but frontier innovation is still viewed as led by OpenAI and Anthropic.
• Open source is treated as a major deflationary force on closed-source AI pricing and margins.
• The number-three vendor is seen as the most exposed to open-source pricing pressure.
• China and Europe are discussed as trying to avoid dependence on US frontier models for sovereignty reasons.
• Anthropic and OpenAI are said to be inaccessible in China and Hong Kong without VPN workarounds.
• DeepSeek is described as intentionally constrained inside China, including no web search and different training data.
• The DeepSeek deal is interpreted as state-controlled because the government keeps voting rights.
• Anthropic is said to have delayed model shipping until US government concerns were satisfied.
• Zhipu AI’s GLM 5.2 is cited as beating GPT 5.5 on coding benchmarks.
• AI infrastructure demand is now visibly lifting memory prices.
• DRAM contract prices are said to have risen 90% to 95% in Q1.
• AI capex is crowding out other uses of capital through the price mechanism.
• Goldman Sachs is cited projecting $7.6 trillion in cumulative AI capex from 2026 to 2031.
• The hyperscalers are described as spending roughly $700 billion a year on capex.
• Current AI revenue is still far below the spend required to justify that capex.
• Capex intensity is now said to exceed free cash flow for the Mag 7, forcing borrowing.
• The AI demand curve is described as effectively infinite if intelligence were free.
• Price, not model capability, is described as the real governor of adoption.
• AI usage is moving from token-maxing toward efficiency as buyers become more cost-sensitive.
• Anthropic and OpenAI are viewed as running a mixed business with subsidized free users and profitable enterprise inference.
• Enterprise inference is described as carrying 40% to 70% gross margins.
• The episode argues the key 2027 enterprise question will be ROI rather than raw usage.
• Jason says enterprises will ration tokens by demonstrated business impact.
• Rory argues the economics may require roughly $1 trillion of revenue once electricity is included.
• He estimates 7% to 8% of the labor force would need to be replaced by tokens for the math to work.
• Jason says productivity gains can be erased by a parity tax if competitors adopt the same tools.
• Rory extends that point: if everyone adopts AI, costs fall broadly but profits may not improve.
• Jason says his team built an AI VP of Finance in China that already outperforms humans on that workflow.
• The workflow spans quoting, contract generation, Salesforce updates, billing, collections, and treasury tools.
• Jason says the new premium skill is being a master of agents.
• Prompt engineering is framed as less important because many prompts are now simple enough for non-experts to write.
• Jason says a mediocre prompt was enough to build a finance director agent from China.
• The Menlo fund discussion emphasizes that headline fund size need not equal actual deployment.
• The fund can be supplemented with SPVs and deal-by-deal sidecars.
• Kalshi is presented as a prediction market with regulatory risk, but also regulatory arbitrage.
• Accenture is described as exposed because AI can compress the labor content of systems integration.
• AI systems integration firms can undercut legacy consultancies by doing the same work with fewer people.
• LLM-driven migrations are described as moat destroyers.
• If a consultancy bills by bodies, AI reduces the need for headcount and breaks the model.
• Jason says newer startup teams should be smaller, higher-paid, office-based, and intensely focused.
• Startup execution is portrayed as a chain of endless sprints rather than a long marathon.
• OpenAI’s Jalapeno is introduced as a custom inference chip co-developed with Broadcom.
• Rory estimates inference consumes 50% to 60% of revenue and GPUs are more than half of capex.
• Jason says open source gets more important as token demand rises and pricing tightens.
• Rory argues OpenAI and Anthropic should focus on demand capture, not chip vertical integration.
• Rory says their current model depends on having outsourced roughly $300 billion of capex to others.
• Jason says the Jalapeno decision was made in an earlier era of abundance.
• OpenAI and Anthropic are now seen as exposed to a more serious open-source threat.
• Closed-source vendors are said to need a middle tier of models, not just flagship and tiny offerings.
• Software buyers are looking for mid-priced, high-quality intelligence without frontier pricing.