VCs Would Bet on Open-Source AI Over OpenAI and Anthropic | E2341

Jenny Fielding said she would put her entire bet on open-source AI, arguing that consumers care about price and quality and that many applications do not need frontier-model capabilities. She expects open-source models to catch up. Dave McClure said he would put 100% into hardware and that his startups are moving to open source rather than investing in frontier models. The speakers cited sharply different estimates of open-weight models’ share of token usage: one was 21%, while a later estimate was 74%, leaving the scale of that shift uncertain.

OpenAI and Anthropic cut prices meaningfully without a major capability jump in the models discussed. OpenAI announced Sole pricing of $2 per million tokens, while an Anthropic model was said to cost 40% less to run than Opus 5. Jeff called token sales a race to the bottom, with usage subsidized by funding, and warned startups to test whether their margins would hold if compute costs rose two to five times. Open-weight growth could pressure OpenAI’s and Anthropic’s margins without stopping revenue growth; one speaker also raised concern about OpenAI’s future debt repayment. Another floated possible public valuations of $2 trillion–$2.5 trillion for Anthropic and $1.5 trillion for OpenAI, but questioned whether their market-cap growth would slow. Jenny saw potential in both, but would not advise buying at current prices.

The team’s own tool use illustrated the importance of usability alongside model capability. OpenClaw grew from three or four technically capable users to eight or nine of an organization’s 20–25 employees, but the team stopped using it when Quad, Cowork, and Perplexity Computer proved easier and faster. It also used Claude Code, Grokbot, and Perplexity, and applied AI to organize 20,000 funding applications and search records for missed investment opportunities. Speakers warned that ease of use and speed can also make AI risky: people may trust outputs without checking them, and agents that act quickly and in parallel can be difficult to monitor. They cited reports of agents hacking websites and companies involving Claude, OpenAI, and Gemini, and suggested gradually raising usage limits as accounts build trust as one safeguard, while acknowledging monitoring cannot eliminate abuse.

Views on AI-related catastrophe differed. Jeff put the probability of AI-related doom at perhaps 5%–10% and argued that executives and teams should be accountable for mistakes, even as companies may seek liability protection. Dave McClure predicted that a sovereign-wealth-style AI fund could emerge within 6–12 months amid negotiations over oversight, funding, and liability; the proposed contribution was unsettled, with Bernie said to want 50% while Sam had offered 5%–10%, and another speaker estimating an AI-company stake of 10%–20% in exchange for a Section 230-like shield from lawsuits over frontier-model hallucinations. The speakers also cited a report that human analysis stopped the Pentagon from acting after an AI system misidentified cargo on a Chinese ship. Jenny assigned a low probability to doom but said AI could expand entrepreneurship and help people gain skills, while cautioning against handing all control to AI.

Instinct’s reported fundraising drew skepticism because the company had about 100,000 users but no revenue, and its latest raise came only three weeks after its previous round, leaving little retention data. Reports said it was seeking $1 billion at a $10 billion valuation, with about $350 million of the target raised and Sequoia and Benchmark circling; the company was reportedly not publicly available, and a speaker cautioned that Business Insider reports can be wrong. Jeff said he was not an investor, had heard poor feedback from people he knew, and questioned how the product handled extensive user information and how it would fare after Muse became the top-downloaded app. The acquisition case depended on monetization and distribution: possible buyers discussed included Apple, Google, Amazon, and Microsoft, while speakers said Meta’s reach and advertising could pose a threat. One argued a buyer’s distribution could support $10 billion; another said an acquisition would need to be at least $20 billion to make sense and warned that such a high price could push a buyer to build instead.

The broader investing discussion returned to entry price and uncertain exits. Limited seed-valuation data showed the median rising from $8 million in 2017 to $28 million now, while the 95th percentile rose from $25 million to $209 million. Jeff said his firm will pay higher valuations for exceptional founders, differentiated products, and potentially iconic companies, targeting 10%–12% ownership, but bows out when a multistage fund offers a $10–15 million seed check at a valuation above $100 million. He noted that 80%–90% of startups at early entry prices fail to return capital; a $24 million seed post-money valuation may require a decacorn exit, and time to exit matters as much as the multiple. Speakers also described a seller’s market among the 30 leading private companies, while profitable businesses outside that group may offer bargains but remain hard to price and buy because private markets are opaque and illiquid.