Jul 10, 2026 · 1h 4m
Cerebras and Black Forest Labs CEOs detail massive AI scaling and Hollywood partnerships
Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit with CEOs of Cerebras & Black Forest Labs
As AI transitions from simple prompting to complex reasoning, the physical infrastructure constraints and creative applications will dictate which technologies scale and who controls them.
- 1AI data center construction is facing unprecedented power and physical space constraints amid surging hardware demand.
- 2The transition to computationally intensive reasoning models is forcing chipmakers to bypass traditional limits of Moore's Law.
- 3Open-source generative video models are actively transforming Hollywood film production and establishing new frameworks for robotics.
Don't miss
Robin Rombach details how Black Forest Labs is partnering with director Martin Scorsese to use generative AI tools for visual brainstorming in film production.
The brief
The global AI infrastructure build-out has reached a scale comparable to history's greatest engineering marvels, driven by an insatiable demand for physical space, unprecedented power grids, and specialized silicon.
Cerebras Systems CEO Andrew Feldman reveals a massive twenty-five billion dollar backlog, arguing that the shift toward computationally intensive reasoning models is breaking Moore's Law and rewriting chip architecture.
In the creative sector, Black Forest Labs CEO Robin Rombach details how open-source video models are entering Hollywood production, highlighted by an active partnership with legendary filmmaker Martin Scorsese.
As generative models advance from simple prompting to understanding deep user intent, the industry faces critical debates over data sovereignty, the safety of recursive learning, and the role of domestic open-source software.
Ultimately, the massive capital deployment in AI hardware is transitioning from speculative experimentation to concrete applications, laying the groundwork for robotics and autonomous problem-solving.
What was said on this episode
32 statements · 25 positive · 6 negative · 1 mixed
New data centers will consume more power than Earth used over the previous fifty years.
“the next several years going to use more power than the previous 50 years on Earth took”
Listen at 2:21
Cerebras currently has a $25 billion backlog.
“we have a $25 billion backlog”
Listen at 3:51
Demand for AI data-center capacity is already booked rather than speculative.
“The demand is booked”
Listen at 4:15
AI experimentation can coexist with enormous net value creation.
“it doesn't mean that the net value isn't enormous”
Listen at 5:15
Modern AI increasingly understands user intent rather than merely following exact prompts.
“increasingly, it's understanding what your intent was”
Listen at 7:33
Cerebras inference performance will exceed twice its current trajectory within eighteen months.
“in the next 18 months, we'll be way over 2x”
Listen at 13:26
Current AI demand will use all available AI silicon.
“there is no silicon that will go unused”
Listen at 14:56
Companies should avoid complete dependence on external chip suppliers.
“you just can't be entirely dependent on other people's chips”
Listen at 16:06
Many routine enterprise tasks need reliable open-source AI rather than frontier models.
“what this needs is sort of rock solid open source capabilities”
Listen at 17:34
The United States needs more domestically developed open-source AI models.
“in the US we need more domestic open source models”
Listen at 18:47
Governments may reasonably require staged deployment of sufficiently dangerous AI models.
“at a time that a model is sufficiently creative in its thinking that it poses a meaningful threat, for the government to say, we'd like you to roll it out in steps”
Listen at 21:40
Political polarization substantially harms clear thinking about AI policy.
“the polarization hurts a great deal. It hurts clear thinking”
Listen at 22:54
AI safety guardrails can increase latency and make systems feel slower.
“the very guardrails can add time and make it feel slower”
Listen at 24:38
A massive data breach will eventually occur.
“there'll be a massive breach”
Listen at 27:01
Existing historical definitions of AGI have already been surpassed.
“Any definition we would have previously put forward, we've blown past it”
Listen at 30:03
Recursive AI improvement can produce exponential gains.
“powerful recursive gains are exponential”
Listen at 32:16
World models will provide behavioral insights by observing human behavior.
“I think that's some of the things the world models are going to bring us as they begin to watch human behavior”
Listen at 34:07
AI systems learn at a pace equivalent to thousands of biological generations.
“we're getting sort of learning so quickly over the equivalent of thousands of generations”
Listen at 36:37
AI could help eliminate cancer deaths among future children and their loved ones.
“There's a shot that our children, none of them, nor their people they love, will die of cancer”
Listen at 38:48
Personalized AI teaching agents could adapt instruction to each child’s learning style.
“Imagine if we built agents that taught children for their way of learning”
Listen at 39:35
AI will create substantial abundance.
“I think it will create abundance for sure”
Listen at 40:15
Latent diffusion is a foundational algorithm behind deployed generative models.
“we invented an algorithm called latent diffusion, which is basically the fundamental algorithm behind all of generative models”
Listen at 41:55
Multimodal models with action prediction can ultimately control real-world robots.
“you can ultimately deploy it on a robot in the real world”
Listen at 43:24
Video pretraining gives models implicit understanding of real-world interaction physics.
“Pre training on videos gives like implicit understanding of the physics of interactions with the real world”
Listen at 44:36
Human-in-the-loop iteration produces the most interesting generative-model outputs.
“the real interesting use cases, they come when you have a human in the loop who iterates and uses it as a medium”
Listen at 50:10
Rapidly improving generative AI will unlock more high-end production uses.
“it's continued to improve and I think like, then it's going to unlock like even more of these like high end use cases”
Listen at 53:56
A single multimodal model can generate films and serve as a robot’s control system.
“you can use the same kind of AI model to make a movie and deploy that as a brain on a robot”
Listen at 54:28
The long-term goal is to control robots through contextual natural-language prompts.
“you would want to go to a place where you could like prompt a robot in context”
Listen at 57:02
Robot models can adapt to specific tasks with only a few hours of fine-tuning data.
“you need only a very little bit of a few hours of fine tuning data to adjust the model on that specific task”
Listen at 57:44
Public generative-AI tools can restrict generation of specified intellectual property.
“you cannot generate certain IP with these models”
Listen at 59:22
Interactive content-creation tools will emerge around generative AI.
“I can envision a whole bunch of very interesting interactive content creation tools”
Listen at 1:00:15
Generative models can let people visualize alternative ideas for existing stories.
“you can actually enable people to visualize these ideas”
Listen at 1:01:59
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
All-In with Chamath, Jason, Sacks & Friedberg
To stay updated on the intersection of venture capital, technology infrastructure, and global economic trends