Cerebras says demand is already ahead of supply
Andrew Feldman, CEO and co-founder of Cerebras, said on the All-In Podcast that the current AI infrastructure wave is not a case of building first and waiting for customers later. He said demand has already reserved capacity, with Cerebras carrying a $25 billion backlog. Feldman named OpenAI, Anthropic, SpaceX, Google, Microsoft and Amazon Web Services as part of the demand picture for more data centers and more compute.
He described the scale of the buildout in stark terms, saying data centers coming over the next few years will consume more electricity than the planet used over the past 50 years, with single buildings as large as football fields and power draw above that of a mid-sized city. He said projects are underway across the US, in Canada, the Nordics, Paris and France, the Middle East, as well as Kazakhstan, Tajikistan and Georgia.
Asked whether this level of demand is creating real value or simply encouraging waste, Feldman said both are happening. He compared the current phase to the early days of AWS, when engineers rushed in and experimented broadly before companies began to develop clearer operating discipline around what should be run where. He said businesses are now getting more deliberate about which jobs can run on open-source models and which require frontier systems.
Reasoning is the next compute-heavy workload
The hosts pointed to Sam Altman’s earlier remarks on All-In that reasoning is the next step for AI systems. Feldman said reasoning is inference, and inference at that level burns through huge amounts of tokens. That, he argued, is where faster machines matter most.
He said that if Cerebras is 15 times faster, then 24 hours of reasoning on its systems can amount to weeks or even months of “thinking” on slower infrastructure. As an example, Feldman said he tested ZAI’s GLM-52 model on BitTensor that morning, gave it effectively unlimited compute, and watched it debate where to look for under-recognized global trends, comparing Hacker News, Reddit and Instagram as possible sources.
On chip progress, Feldman said older generations of semiconductors generally tracked Moore’s Law, with performance doubling every 18 months. He argued that Cerebras has moved onto a different path and that gains over the next 18 months could be well above 2x. He added that GPUs are built on a 20-year-old architecture, while newer designs still have plenty of room for improvement.
Control and sovereignty are shaping enterprise demand
Feldman said it is not surprising that large customers are building their own chips. He mentioned OpenAI and Amazon and said the motivation is not simply to make the fastest hardware. Companies do not want to be fully dependent on someone else’s chip stack. In his telling, hyperscalers learned from the x86 era what it meant to be tied to Intel, and now want to keep more control over their own fate.
He applied the same logic to models. Frontier systems from OpenAI, Anthropic and Gemini may handle the hardest tasks, he said, but a large share of day-to-day enterprise work can run on solid open-source models. That matters even more in regulated sectors such as finance and healthcare, where concerns around data leakage and “intelligence sovereignty” push companies toward local deployment.
Feldman said OpenAI released OSS 12B a few months ago, but argued that US-based open-source options remain limited. In his view, companies that want open models today are often choosing between OpenAI’s OSS 12B and Chinese models. He said Cerebras sits in a relatively neutral position, running GLM, Kimmy, the Qwen family, OpenAI’s closed models, as well as models built by GSK, the UAE’s G42 and MBZUAI.
Feldman says AGI has already arrived by older definitions
When asked whether AGI has already arrived by standards people used decades ago, Feldman gave an unambiguous answer: yes. He said AI has gone well beyond definitions proposed 10, 15, 20, 30, 40 or even 50 years ago, and said the Turing test was surpassed long ago.
He tied that view to recursive improvement. As systems ask, learn and ask again, he said, output quality improves on an exponential curve. Humans, by contrast, are constrained by generational turnover. AI is not. Feldman framed that difference as a compressed learning cycle that can move far faster than biological systems.
On staged releases and government review, he said he had not seen many precedents, but does not object in principle to giving governments time for red-teaming and allowing companies a few weeks to patch obvious weaknesses. He cited a story from Palo Alto Networks CEO Nikesh, who told him that model-based testing of the company’s own software surfaced dozens of critical vulnerabilities within an hour, forcing a six-week patching effort.
Feldman also said he is betting that his children, and the people they know, will not die from cancer. He paired that with a broader vision that included unlimited energy, food, knowledge, education and housing, while also acknowledging economic disruption, much as earlier industrial shifts hurt workers tied to old industries.
Black Forest Labs is pushing multimodal models toward robotics
Robin Rombach, CEO and co-founder of Black Forest Labs, used the same conversation to outline his company’s direction. He said he and his co-founders started the company two years ago after earlier work on Stable Diffusion and, before that, latent diffusion. He described latent diffusion as a foundation for current image generation, video generation and even some physical AI models.
Black Forest Labs is now training multimodal visual systems across image and audio data, he said, while adding action prediction so that one model can handle image, video, audio and motion. The goal is to carry that same model into real-world robotic deployment.
Rombach said intuitive intelligence and deep reasoning are complementary forms of intelligence. His team started with images because the compute burden is lower than video, but he said the field is converging toward unified multimodal systems. In that setup, video pretraining teaches physical interaction patterns implicitly, and action prediction becomes a path into robotic control.
Martin Scorsese used the company’s tools for visual ideation
Rombach said he worked directly with Martin Scorsese in the same room while the filmmaker explored the company’s model. Scorsese, he said, wanted to visualize scenes from his imagination, including a village in Eastern Europe, by describing an idea, reviewing the output and iterating.
According to Rombach, Scorsese’s takeaway was that turning a mental image into a visual output can be far more efficient than language alone. Language, he said, is a somewhat lossy medium, while images and video carry much richer information.
Rombach stressed that Black Forest Labs does not want to dictate how creators should use the models, especially not someone like Scorsese. He described AI models as a medium and said the most interesting results usually appear when humans stay in the loop and keep iterating.
The endpoint for generative video is not the screen
The hosts said some startups are already using Flux and related models to produce launch videos in one or two weeks that might previously have cost $250,000, and mentioned an example involving a Bitcoin film with Gal Gadot, where performers worked on a sound stage without green screens and generative AI handled the backgrounds, producing visuals associated with a $150 million production on a $30 million budget.
Rombach said he has seen early production use, though he added that high-end filmmaking remains one of the toughest applications. The technology is still moving quickly, he said, and is far from finished.
What excites him most is not film alone. He said the same multimodal model used to make a movie could also be deployed as the brain of a robot. In that framing, world models and action models are part of the same technical direction.
On data collection for robotics, Rombach said the target is to instruct a robot with an in-context prompt such as “bring me that glass of orange juice,” though he added that the field is not there yet. For now, he said, models already come with substantial visual understanding and need only a few hours of fine-tuning data to adapt to specific hardware. The longer-term goal is less fine-tuning and more instruction through context.
IP partnerships and hiring plans
Rombach said the most compelling use cases involve generating things that did not exist before. He added that Black Forest Labs does not allow its public tools to generate specific IP, which he called a reasonable limitation, but said the company is working with some rights holders on custom models. Some of those projects are based on the company’s open models, while others use stronger proprietary systems.
The discussion also touched on fan films. The hosts pointed to AI-made reinterpretations of untold Star Wars stories and cited Star Wars Stories Untold as a project with videos reaching millions of views. Rombach said it would be compelling if the industry could find a business model that worked for IP owners while also opening a path for this kind of highly customized creativity.
He closed by saying Black Forest Labs has just passed 100 employees and is hiring in Germany and San Francisco for large-scale model training research, diffusion and flow matching training, custom engineering for clients, and large-scale compute infrastructure operations.
