Conversation with Cerebras CEO: Holding $25 Billion in Backlog Orders, AI Computing Demand Is Already Fully Booked; We Are Not "Building Facilities and Waiting for Customers"

Collated & Compiled by: Odaily TechFlow

Guests: Andrew Feldman, CEO & Co-founder of Cerebras; Robin Rombach, CEO & Co-founder of Black Forest Labs

Host: Host of the All-In Podcast

Podcast Source: All-In Podcast

Original Title: Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit — Cerebras & Black Forest Labs CEOs

Release Date: July 10, 2026

Key Takeaways

This episode features two CEOs from AI infrastructure companies. Andrew Feldman is the founder of Cerebras, a company specializing in inference chips, which recently completed its IPO with a backlog of $25 billion in orders. He repeatedly emphasized one thing: demand for AI compute power has already been booked solid; there is no "build it and they will come" scenario. The appetites of OpenAI, Anthropic, SpaceX, and Google far exceed the supply. Furthermore, the emergence of reasoning has caused a surge in computational intensity, which is precisely the battleground for fast machines. Robin Rombach is the founder of Black Forest Labs, which develops generative image and video models (the Flux series). He previously invented the latent diffusion algorithm, the foundation for all current image and video generation models. He recently collaborated with Martin Scorsese, allowing the director to visualize scenes from his mind using AI. However, he is even more excited about the potential for the same multimodal model used to make movies to be deployed as the 'brain' for robots. The endpoint of generative video is not the screen, but the physical world.

Highlights and Insights

Reasoning is the Next Compute Black Hole

• "Interestingly, this wave is different from the past. They aren't betting on 'build it and they will come'; demand has already booked production capacity. We have a $25 billion order backlog."
• "Reasoning consumes a massive number of tokens, and this is exactly the battlefield for fast machines."
• "If Cerebras is 15 times faster, running for 24 hours is equivalent to weeks or even months of thinking time."

Open Source and Sovereignty: Enterprises Crave Control

• "Nobody likes being dependent. The lesson hyperscalers learned from the x86 era was being locked into Intel."
• "You don't need to make the fastest chip; you just need to not be completely dependent on another company's chip."
• "If you want to run open-source models today, you either use OpenAI's OSS 12B or Chinese models. The US needs more domestic open-source options."

AGI Has Arrived, By Definitions from Twenty Years Ago

• "We have far surpassed any definition of AGI we formulated 20, 30, or 40 years ago."
• "The Turing Test? It was crushed long ago."
• "The issue isn't that we don't know how to ask questions anymore. AI can now tell you, 'Hey, you dumb humans, you haven't considered this.'"

Generative Video Isn't About Replacing Human Creation

• "These AI models are a medium. We don't want to dictate how they should be used, especially for someone like Martin Scorsese."
• "Language is a somewhat lossy communication method. Visual information signals are incredibly rich. Turning mental images into visible pictures is where this technology is most powerful."
• "The most interesting results almost always emerge when humans are in the loop, continuously iterating."

From Movies to Robots: One Model for All

• "You can make a movie with a single multimodal model, and then deploy that same model as the brain for a robot."
• "Pretraining on video implicitly teaches the model the laws of physical interaction. Then, from the same model, you extract action predictions, which equates to robot control."
• "The goal is to be able to command a robot via an in-context prompt: 'Bring me that glass of orange juice.' We can't do it yet, but that's the direction."

The AI Infrastructure Boom: Data Centers Bigger Than Cities

Host: We've never seen this scale of construction. Not since the Great Wall and the Pyramids has humanity poured so much capital, time, and intellect into building something. You're in the middle of it. Your clients are building data centers; you're a key component. What is Cerebras doing in 2026? What's happening with those massive projects in Texas?

The data centers we're talking about will, in a few years, consume more electricity than the entire world has used in the last 50 years. A single building can be the size of a football field, drawing more power than a medium-sized city. They are being built all over the US, in Canada, Northern Europe, Paris and across France, the Middle East, and even large-scale data centers in Kazakhstan, Tajikistan, and Georgia. Every country, every state wants a piece of it.

Who is paying? OpenAI, Anthropic, SpaceX AI, Google. Their appetites are terrifyingly huge. Interestingly, unlike many past tech booms, this one is different: they aren't betting on "build it and they will come." Demand has already reserved all the production capacity. We have a $25 billion order backlog. OpenAI wants more data centers, Microsoft wants more, AWS wants more. The demand is not waiting for customers to walk in; customers are already lining up.

Host: This has also spawned a term called 'token maxing' – generating tokens infinitely. Some people question whether all this massive demand is actually creating real value?

Of course, immense value is being generated. And of course, there is a lot of random experimentation. I compare it to the early days of AWS. It was so liberating to bypass your own IT department; every engineer would just pull out a credit card and sign up. Much of it was useful. Some of it, in hindsight, was "yeah, probably shouldn't have done that." But overall, it was a net positive, even if some directions were dead ends.

I remember when Costco opened in Palo Alto in 1988. People shopped there like it was a Safeway, walking down every single aisle. That was a terrible way to shop because you'd buy four things you didn't need for $22 each. Later, people learned the strategy: go to the back for the rotisserie chicken, grab 18 cupcakes for the kid's birthday party, and get out efficiently. AI token consumption is the same. At first, everyone used it liberally. Now, companies are becoming strategic: figuring out which tasks are good enough for open-source models and which require frontier models. We're starting to manage AI like a business.

Reasoning Replaces Training: Why Fast Machines are the Stars of This Wave

Host: Sam Altman said on All-In that the next step is reasoning – understanding intent, formulating strategies, cross-validating with other agent threads. We've come a long way from 'predicting the next word,' and Cerebras is right at the center because reasoning is inference, requiring massive computation.

Reasoning consumes massive amounts of tokens, creating a battlefield for fast machines. Each step in the reasoning process devours internal tokens. Conventionally, you trade huge amounts of time for better answers. Since Cerebras is 15 times faster, 24 hours of reasoning on our hardware is equivalent to weeks or months of thinking time on others.

I tested a model this morning – ZAI's GLM-52 on BitTensor. I gave it unlimited compute power and asked it to tell me, every hour, what global trends were going unrecognized. It began to debate with itself: Should I look on Hacker News and Reddit? Do trends appear first on Instagram? I watched a reasoning model arguing with itself in the background, performing reasoning. Unlimited tokens equal unlimited reasoning. With Cerebras being 15 times faster, 24 hours equals weeks for others.

Host: Does Cerebras have its own Moore's Law? Do you internally discuss when performance will double?

All previous chips followed Moore's Law, doubling every 18 months. We broke that trajectory with this chip and charted a completely new path. My assessment is that over the next 18 months, we will see far more than a 2x improvement. There's still huge room for optimization in the new architecture. GPUs are 20-year-old architectures, only sustained by shrinking process nodes. But a new architecture still has a vast amount left to learn and tune.

Host: With a $25 billion order backlog, you also need to keep pace with OpenAI, who might be a potential future competitor. How do you run the company?

Right now, no silicon sits idle; the demand is too large. But you're right, OpenAI is making its own chips, and Amazon is too. Nobody likes being dependent. The lesson hyperscalers learned from the x86 era was being locked into Intel. The lesson GPU vendors learned was being dependent on a few hyperscale customers, which is why they fund new clouds. The point of making your own chip isn't necessarily to be the fastest; it's to avoid being completely dependent on someone else, to at least control a significant part of your own destiny.

Open Source and Sovereignty: Enterprises Want Control

Host: Open source is having its moment. I used OpenClaude early on, later Kimmy. I found my token usage for Claude was exploding, but I couldn't tell the difference with Kimmy. Open-source models are starting to do reasoning, and the gap suddenly closed this year.

You don't drive your Ferrari to the grocery store. Sometimes you drive the sports car, sometimes you drive the minivan where you don't mind if your kids spill Cheerios. It's the same for businesses: hard problems go to frontier models (OpenAI, Anthropic, Gemini), but a huge number of everyday tasks just require solid open-source capability. Think about how much time a company spends copying from Workday into another Excel cell? That doesn't require gold-medal math; a reliable open-source model is sufficient.

Recently, another factor has emerged: regulated industries like finance and healthcare (HIPAA, FINRA) fear data leaks and their intelligence sovereignty being controlled by others. They want to run models on-premises, using open-source versions to gain more control. OpenAI released OSS 12B a few months ago; it's okay. But if you want to run open source in the US today, your options are basically OSS 12B or Chinese models. There are far too few domestic open-source choices. NVIDIA sees this window too and is pushing its own open-source models, but Jensen is hesitating. His customers are Sam, Dario, Elon, Sergey. Would launching open-source models compete with his clients?

Cerebras takes a more neutral position. We run GLM, Kimmy, the Qwen series, and also OpenAI's closed-source models. We also run models developed by GSK themselves, and proprietary models from UAE's G42 and MBZUAI. Sovereignty is a trend.

AGI is Here, Paradigms Don't Die, People Do

Host: When Fable 5 and o-56 were released, the government said, 'Pause before releasing.' There was tension between Anthropic and the administration, which is now easing. Do you think staged releases are reasonable? Are the models really dangerous enough?

I've never seen anything like this before. But thinking back: when a model becomes powerful enough in creative thinking, for the government to say, "Please release it in stages," I think that's actually not unreasonable. We handle powerful medicines this way. I'm not endorsing the FDA's seven years of paperwork, but asking for "at least let the government do some red-teaming to ensure our defenses can hold," giving a few weeks to patch obvious holes, doesn't seem like an unreasonable demand.

However, we are in the most polarized time ever. If this weren't being done by Trump, under any other president, the reaction might be completely different. Polarization harms clear thinking. Both sides will do stupid things, and both sides will do smart things. The career people within the government are actually working very seriously on this; it's just that this technology is moving so fast.

Nikesh from Palo Alto Networks told me: they tested the model against their own software and found dozens of critical vulnerabilities within an hour. They had to drop everything and spend six weeks patching. You realize this is a powerful tool; maybe it's wise to let a small group see it first, or do red-teaming first.

Host: By any definition from 20 years ago, AGI is already here. Do you think so?

Yes. The Turing Test? It was crushed long ago. We have far surpassed any definition proposed 10, 15, 20, 30, 40, or 50 years ago. We've answered all the questions science fiction writers posed, and they would say, "I have no more questions, sorry." This is why what people on the fringe say is worth listening to. When Ilya talked about safety eight years ago, you said, "What?" But he turned out to be right. When Elon talked about reducing rocket costs to near zero, you said, "What?" But he did it.

Host: Recursive learning. You ask a question, learn the result, ask again, get a better answer covering more material. These iterative outputs jump from 'slightly better' to 'dramatically better.' The slope of the exponential curve is too steep.

The recursive gains are exponential. You get better, do it again, gain more. The slope is incredibly steep. We are just starting to see this. Will answers just keep getting better with more compute? It stops when tokens or budget run out, but where does this exponential curve end? Does it just keep going up and to the right? This question is incredibly interesting right now.

Human learning speed is constrained by generations. Elephants and large mammals take 15–20 years per generation. To learn fast, you need to be like a fruit fly, with two generations a day. AI is achieving learning speeds that span thousands of generations. When I studied psychology, a professor said something memorable: Paradigms don't die, people do. The disciples of Freud, Skinner, and Jung held leadership positions for 20–40 years before the next generation challenged them. AI has compressed the intergenerational interval to fruit fly speed.

Here's my bet: our children and everyone they know won't die from cancer. The economy will experience shocks. When the car came, the people shoeing horses had a tough time. But list the gains versus the losses: infinite energy, infinite food, infinite knowledge, infinite education, infinite housing. We've known for a thousand years that one-on-one tutoring is better than classroom learning. Aristotle tutored Alexander. Socrates tutored his students. But we