Cerebras CEO warns AI sovereignty concerns driving governments away from Chinese models
"I think in the US we need more domestic open source models. We need to give the world a choice. If they want to run open source right now, it's OSS 12B or Chinese models. I think OpenAI made a good call releasing OSS 12B some months back. That was a good open-source model. But I think giving them more power might be sort of. The government's actions with regard to fable and 56 where they said, 'Oh, whoa, let's think and then we can act.' I think particularly here in Europe was a bit of a wakeup call."
About this episode
Jason Calacanis hosts two in-depth conversations exploring the unprecedented buildout of AI infrastructure and its creative applications. The episode features Andrew Feldman, CEO and founder of Cerebras, who reveals his company has accumulated a $25 billion backlog for AI inference chips as global demand dramatically outpaces supply. Feldman describes an infrastructure mobilization unlike anything in peacetime history, with data centers being constructed worldwide from Texas to Kazakhstan that will consume more power in the next few years than the previous 50 years combined. Individual facilities now draw electricity comparable to midsize cities, as companies like OpenAI, Anthropic, Google, Microsoft, and AWS compete for capacity. Cerebras has broken through traditional Moore's Law constraints, achieving performance improvements significantly beyond the historic doubling-every-18-months trajectory. Feldman discusses the shift from simple AI tasks to complex reasoning models that require exponentially more computational power, explaining how unlimited tokens enable unlimited reasoning when given sufficient time. He addresses growing concerns about AI sovereignty, noting that governments and regulated industries increasingly prefer domestic open-source alternatives to avoid dependency on either US tech giants or Chinese models. The conversation explores the balance between innovation velocity and responsible deployment, with Feldman defending measured government review of powerful new models while warning against excessive regulation. In the second segment, Robin Rombach, CEO of Black Forest Labs, discusses his company's work on open-source and proprietary image and video generation models. He reveals that legendary director Martin Scorsese is actively using their AI tools to visualize scenes for upcoming films, describing sessions where Scorsese iterated on images of Eastern European village settings. Rombach explains the evolution from simple text-to-image systems to sophisticated multimodal models that combine images, video, and audio while predicting actions for robotic applications. He sees the technology converging toward world action models that can both create synthetic content and control robots in the physical environment. The episode captures a pivotal moment as AI transitions from experimental technology to production infrastructure reshaping global power consumption, geopolitics, creative industries, and the nature of human-computer interaction.
Key takeaways
- Cerebras has accumulated a $25 billion backlog for AI chips as OpenAI, Anthropic, Google, Microsoft, and AWS chase existing demand rather than speculative future needs
- Global AI data center buildout will consume more power in the next several years than the previous 50 years, with individual facilities drawing electricity comparable to midsize cities
- Cerebras broke Moore's Law trajectory with new chip architecture, projecting performance improvements significantly beyond traditional 18-month doubling cycle for AI inference workloads
- Nations including Kazakhstan, Tajikistan, Armenia, and Georgia are building major AI infrastructure as every country seeks participation in the technology race
- AI sovereignty concerns are driving governments and regulated industries toward domestic open-source models rather than dependence on US tech giants or Chinese alternatives
- Martin Scorsese is actively using Black Forest Labs' AI image generation tools to visualize scenes for upcoming films during pre-production development
- AI reasoning models now consume massive token volumes internally, making inference speed critical as models can run for 24-48 hours to achieve weeks of equivalent human thinking