AI Just Hit a Tipping Point | Raoul Pal the Journey Man

The speakers describe AI costs falling sharply: Pal says 10 million Deep Seek V4 Flash tokens cost about 50 cents, versus $6,000 a few years earlier, and that Opus is roughly four times more token-efficient than its predecessor. He also says token efficiency is reducing his spending above his subscription, while characterizing the broader cost decline and adoption ramp as unusually fast.

The guest links increasingly capable agents to substantial compute needs: he cites future million-chip, $10 billion training runs that could support a million agents, alongside examples of 10,000 agents working on a Navier–Stokes problem and 900 Anthropic agents identifying a possible CRISPR mechanism. Pal predicts that hosting agents in cloud virtual machines could make CPUs and RAM a bottleneck, with AMD and Intel among the suppliers and demand potentially exceeding supply.

Pal describes AI as both a productivity opportunity and a cost challenge for enterprises: he says technology companies are seeing earnings per employee rise as headcount stays flat or shrinks, and argues that greater token efficiency could make AI useful across enterprises. But he also says Harvey's margins fell to minus 50% as law firms used more tokens with improved models, increasing the amount it pays providers such as Claude.

The guest cites abrupt capability gains in both reasoning and multi-agent work: OpenAI reportedly raised performance on Frontier Math from 10% to 60% within a few days, while examples included agents tackling a Navier–Stokes problem and identifying a possible CRISPR mechanism. Separately, speakers said Minax H3 cut the time to generate 15 seconds of video from four minutes to eight seconds, and described some AI video as Hollywood-level based on their personal assessment.