Podcast
20VC- Sam Altman Offers Trump 5% of OpenAI- Fool or Genius? | Alex Karp Sounds the Alarm- Enterprises Fear Frontier Models & Questionable ROI of AI | the Rise of Chinese Open Source- Deepseek Building Own Chips
Startup Funding | The Pitch
- Washington Makes Frontier AI A Preapproval Problem
- Washington now requires pre-approval and deeper oversight for frontier AI, shifting software from free launch to regulated release.
- Rory warns this reverses the internet-era hands-off approach and could entangle firms in politics and ownership debates. (Time 0:04:43)
- Giving Government Equity Invites Bigger Political Stakes
- Sam Altman’s 5% stake proposal to the US government is framed as alignment but risks political escalation and loss of control.
- Rory argues the gesture invites regulatory appetite to grow from 5% to far larger demands like taxation or board influence. (Time 0:06:50)
- Use Small Equity Stakes To Buy Strategic Alignment
- Consider small dilution if it buys political alignment and boardroom access with a powerful partner.
- Jason notes 5% can create outsized alignment benefits similar to strategic stakes like Klaviyo-Shopify deals. (Time 0:09:03)
- Equity Does Not Equal Political Alignment
- Ownership alone doesn’t guarantee alignment: Microsoft’s 30% in OpenAI hasn’t created harmony.
- Rory compares to TARP and warns political actors will exploit narratives of AI’s societal impact to demand more control. (Time 0:11:10)
- Founders Are Less Dilution Sensitive In AI Era
- Massive dilution is normalized in AI funding; founders often accept many small rounds, reducing sensitivity to each round’s dilution.
- Jason notes multiple tiny rounds add up and seed investors effectively see much higher real entry prices. (Time 0:19:03)
- Enterprises Worry About ROI And Data Risk
- Large enterprises are skeptical of frontier models on ROI and data exposure, creating demand for privacy-first alternatives.
- Alex Karp’s CNBC remarks spurred Palantir’s stock jump by pitching trust and data control to enterprises. (Time 0:28:16)
- Turning Excess Compute Into A Cloud Business Is A Levered Bet
- Meta and SpaceX turning excess compute into cloud services signals a market willing to buy spare capacity, but risks oversupply if demand slows.
- Rory warns upside if demand holds, downside if compute demand softens and sellers flood market. (Time 0:33:01)
- NVIDIA Is Financing Demand With Contingent Hardware Deals
- NVIDIA’s ‘Compute Now, Pay Later’ shifts risk to Nvidia via revenue-sharing and backstops, expanding neo-clouds but creating contingent liabilities.
- Rory explains accounting recognition and warns debooking risks if customers fail or demand drops. (Time 0:41:00)
- Some Model Teams See Owning Silicon As Strategic
- Building custom chips for models is increasingly argued as defensible to own compute and optimize efficiency.
- Rory softened his prior view after acknowledging model-specific silicon can materially improve performance versus general GPUs. (Time 0:45:27)
- Chinese Generative Video Has Proven Monetization Pathways
- Chinese AI video models like Kling have proven paid consumer demand, charging quickly and monetizing better than some Western efforts.
- Rory notes 30-second video generation costs ~$1.30–$2 in GPU, making monetization viable for paying users. (Time 0:50:27)
- China’s Open Source Models Will Flourish Behind The Firewall
- China is accelerating in open source models and will fill gaps where Western frontier models are blocked inside China.
- Jason observed ChatGPT/Claude access restrictions spur local alternatives, driven by domain expertise and huge domestic demand. (Time 0:52:13)
- Match Model Tier To Problem Complexity
- Use frontier models for unknown problems and open source for commoditized tasks to save cost and time.
- Rory and Jason share that frontier models solved complex algorithms in minutes whereas open models required hours and extra cost. (Time 0:58:10)
- Customer Experience Is A Clear Early ROI Use Case
- CX use cases show clear ROI per resolution, making them early sustainable AI bets with measurable cost-per-resolution economics.
- Rory explains moving from 30% to 65% resolution at ~$1 creates viable customer ROI, then further gains cost more. (Time 1:00:46)
- Enterprise Adoption Requires Services Not Just Models
- Microsoft and Amazon embedding engineers to help enterprises adopt AI reflects that models alone won’t drive enterprise outcomes.
- Rory compares this to IBM-era services: vendors with enterprise trust will act as integrators and change managers for adoption. (Time 1:02:44)
- Join Startups With Near Term Liquidity Signals
- For employees pick startups with credible near-term liquidity signals like prior tender offers or clear runway to make them tender-worthy.
- Rory advises joining firms that can become unicorns within 1–3 years and checking whether the company plans tender offers. (Time 1:10:09)