Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit with CEOs of Cerebras & Black Forest Labs
In a Nutshell
Cerebras CEO Andrew Feldman reports that AI data center construction is consuming more power than all prior decades combined, with $25B in advance orders from OpenAI, Google, Microsoft, and others, while reasoning models running for hours to days on unlimited tokens are already delivering superhuman trend analysis and problem-solving. AGI has been achieved by any pre-2020 definition—every benchmark passed, the Turing test exceeded—yet the real constraint is now coordination and deployment rather than raw intelligence. Black Forest Labs CEO Robin Rombach shows that latent diffusion models are moving from unpredictable image generators to controllable multimodal systems already used by Scorsese for storyboarding, with the same models fine-tuned in hours to control robots, enabling both high-end film production and physical AI.
These notes were generated by AI and may contain inaccuracies.
Andrew Feldman, CEO and founder of Cerebras, returned to discuss the unprecedented scale of AI infrastructure development. The conversation compares the current mobilization to the Great Wall of China, the pyramids, and the war effort in terms of capital, time, and human resources dedicated to building AI systems.
Cerebras customers are constructing data centers that will consume more power in the coming years than the previous 50 years combined. Individual buildings the size of football fields are being built with power feeds exceeding those of midsize cities. These facilities are under construction across the United States, Canada, the Nordics, Paris and France, the Middle East, Kazakhstan, Tajikistan, Georgia, and Armenia. Every country and every U.S. state is participating in this buildout.
The primary buyers of this capacity include OpenAI, Anthropic, SpaceX, SpaceX AI, Google, Microsoft, and AWS. These organizations are placing orders for chips before manufacturing is complete, creating a $25 billion backlog at Cerebras. Demand is booked in advance rather than speculative, which is described as highly unusual compared to typical technology cycles.
The discussion addresses whether current token consumption creates genuine value. The position is that massive value is being generated alongside significant experimentation. The analogy is drawn to early AWS adoption, where engineers were encouraged to experiment freely with credit cards, resulting in both useful applications and wasted spend. Similarly, Costco shopping behavior evolved from browsing every aisle to strategic purchasing. Enterprises are moving from unconstrained token access to more disciplined allocation based on demonstrated productivity.
Sign in to read the full notes
Get access to AI-generated notes, topic timestamps, and more.