Conversation with Cerebras CEO: Holding $25 Billion in Backlog Orders, AI Computing Demand Has Long Been Fully Booked – We Didn't Build It and Wait for Customers

Organized & Compiled: 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

Air Date: July 10, 2026

Key Takeaways

This episode features the CEOs of two AI infrastructure companies. Andrew Feldman, founder of Cerebras, a company specializing in inference chips that just completed its IPO with a $25 billion backlog of orders, repeatedly emphasized one thing: demand for AI compute power has long been fully booked; there's no "build it and they will come" scenario. The appetites of OpenAI, Anthropic, SpaceX, and Google far outstrip supply. Furthermore, the advent of reasoning has dramatically increased computational intensity, which is precisely the battlefield for fast machines. Robin Rombach, founder of Black Forest Labs, creates generative image and video models (the Flux series). He previously invented the latent diffusion algorithm, foundational to all current image and video generation models. He recently collaborated with Martin Scorsese, enabling the director to visualize images from his mind using AI. However, his more exciting direction is that the same multimodal model can both make films and be deployed as the 'brain' for robots. The endpoint of generative video isn't on the screen; it's in the physical world.

Key Insights Summary

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 pre-booked our entire capacity. We have a $25 billion order backlog."
• "Inference is reasoning. Reasoning consumes a massive number of tokens, and this is precisely the battleground 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 & Sovereignty: Enterprises Want Control

• "No one likes being dependent. The lesson hyperscalers learned from the x86 era was being locked in by Intel."
• "You don't need to make the fastest chip; you just need to avoid being completely dependent on someone else's chip."
• "If you want to run an open-source model today, your options are either OpenAI's OSS 12B or a Chinese model. America needs more domestic open-source choices."

AGI, by Definitions from Twenty Years Ago, is Already Here

• "We have already far surpassed any definition of AGI proposed 20, 30, or 40 years ago."
• "The Turing Test? That was crushed long ago."
• "The problem is no longer that we don't know how to ask. AI can now tell you back: 'Hey, you dumb humans, you didn't consider this.'"

Generative Video is Not About Replacing Human Creation

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

From Film to Robots: The Same Model

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

AI Infrastructure Frenzy: Data Centers Bigger Than Cities

Host: We've never seen a construction scale like this. Not since the Great Wall or the pyramids has humanity invested so much capital, time, and intelligence into building something. You're actually doing it; your customers are building data centers, and you're a critical part of it. In 2026, what is Cerebras doing? And what's the deal with those massive projects in Texas?

The data centers we're talking about will, in the coming years, consume more electricity than the Earth 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 across the US, in Canada, Northern Europe, Paris and all over France, the Middle East, and even large data centers are being built in Kazakhstan, Tajikistan, and Georgia. Every country, every state wants in.

Who's paying for it? OpenAI, Anthropic, SpaceX AI, Google – their appetites are terrifyingly large. Interestingly, unlike many past tech booms, they aren't betting on "build it and they will come"; demand has already pre-booked the capacity. We have a $25 billion order backlog. OpenAI wants more data centers, Microsoft wants more, AWS wants more. The demand isn't waiting for customers; the customers are already lining up.

Host: This also spawned a term called "token maxing," endlessly burning tokens. Some question whether all this massive demand is actually creating real value?

Of course, massive value is being generated. And of course, there's a huge amount of trial and error. I compare it to the early days of AWS. It was so liberating to bypass your own IT department, every engineer signed up with a credit card. A lot of it was really useful; some of it, in hindsight, you thought, "Man, we shouldn't have done that." But on the whole, it was profitable, just with some blind alleys.

I remember when Costco opened in Palo Alto in 1988. People shopped at Costco like it was Safeway, walking down every single aisle. It was a terrible way to shop because you'd end up buying four things you didn't need, each for $22. Later, people learned the strategy: go to the back for the chicken, grab 18 cupcakes for the kid's birthday party, and get out efficiently. AI token consumption is the same. At first, people used it liberally, but now enterprises are getting strategic: defining which tasks are fine with an open-source model and which require frontier models. We are starting to manage AI like a business.

Reasoning Replaces Training: Why Fast Machines Star in This Wave

Host: Sam Altman said on AllIn 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 now Cerebras is right at the center because reasoning is inference, which requires immense computation.

Reasoning consumes a massive number of tokens, creating the battlefield for fast machines. Reasoning involves consuming internal tokens at every step; you traditionally trade a huge amount of time for a better answer. A Cerebras being 15x faster means running inference for 24 hours is equivalent to weeks or even months of thinking time for others.

I tried a GLM-52 model from ZAI on BitTensor this morning. I gave it unlimited compute and asked it to tell me, every hour, what trends worldwide were still unidentified. It started debating with itself: Should it look on Hacker News and Reddit? Or do trends appear first on Instagram? I watched a reasoning model debating in the background; it was doing reasoning. Unlimited tokens equate to unlimited reasoning. With Cerebras being 15x faster, 24 hours is like weeks for others.

Host: Does Cerebras have its own Moore's Law? What's the internal discussion on the doubling period?

All previous chips followed Moore's Law, roughly doubling every 18 months. Our chip broke that trend and is on an entirely new trajectory. My assessment is that, in the next 18 months, the improvement will be far more than 2x. The new architecture still has significant room for optimization. GPUs are based on a 20-year-old architecture, surviving only by shrinking process nodes, but a new architecture 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 future competitor. How do you operate the company?

Right now, silicon won't sit idle; demand is too large. But you're right, OpenAI is also making its own chips, as is Amazon. No one likes being dependent. The lesson hyperscalers learned from the x86 era was being locked in by Intel; the lesson learned from the GPU era was being dependent on a few large customers, which is why they funded new cloud ventures. Making your own chip isn't about being the fastest; it's about not being completely dependent on others, at least taking control of an important part of your destiny.

Open Source & Sovereignty: Enterprises Want Control

Host: Open source is having a moment. I used OpenClaude early on, then Kimmy, and found my Claude tokens bleeding, 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 want to drive a Ferrari to the supermarket. Sometimes you take the sports car, sometimes you take the minivan, and you don't mind if the kids spill Cheerios. It's the same for enterprises: hard problems go to frontier models (OpenAI, Anthropic, Gemini), but a huge volume of daily tasks just need solid open-source capabilities. Think about how much time a company spends cutting and pasting from Workday into another Excel cell? That doesn't require gold-medal math; a reliable open-source solution is sufficient.

A new card was recently played: regulated industries like finance and healthcare (HIPAA, FINRA) are worried about data leaks and having their AI sovereignty controlled by others. They want to deploy models on-premises, using open-source versions to gain more control. OpenAI released OSS 12B a few months ago, which is okay. But today in the US, if you want to run open source, your options are either OSS 12B or Chinese models. There are too few domestic open-source choices. NVIDIA also sees this window and is pushing its own open-source models, but Jensen is hesitant because his customers are Sam, Dario, Elon, Sergey – would releasing open source compete with them?

Cerebras is in a relatively 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 definitely 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." Anthropic had tense relations with 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 is powerful enough in creative thinking, and the government says "please release it in stages," I actually don't think that's unreasonable. We manage potent drugs this way, though I certainly don't encourage the seven-year paperwork mess from the FDA. But saying "at least let the government do some red teaming, confirm our defenses can hold" – giving a few weeks to patch obvious vulnerabilities – isn't an unreasonable request.

But this is the most polarized time ever. If this wasn't done by Trump, under any other president, the reaction might have been entirely different. Polarization hurts 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, but things are moving too 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 stop everything they were doing and spend six weeks patching. You realize this is a powerful tool. Maybe it's wise to show it to a small group first, maybe do red teaming first.

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

Yes. The Turing Test? That 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. They would say, "I have no more questions. I'm sorry." That's why what seemingly fringe people say is worth listening to. Ilya talked about safety eight years ago, and you thought "What?" He turned out to be right. Elon talked about reducing rocket costs to near zero, and you thought "What?" He did it.

Host: Recursive learning – you ask it a question, learn the result, ask again, get a better answer covering more material. The output from these cycles jumps 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 too steep. We are just beginning to see this. If you keep pouring in compute, will the answers keep getting better? It stops when tokens or budget run out, but when will this exponential curve end? Or will it keep going up and to the right? This question is incredibly fascinating right now.

The speed of human learning is limited by generations. Elephants and large mammals take 15-20 years per generation. To go faster, you need to be like a fruit fly, with two generations a day. AI is achieving this learning speed across thousands of generations. When I studied psychology, a professor said: "Paradigms don't die, people do." Students of Freud, Skinner, and Jung held leadership positions for 20-40 years before the next generation could question them. AI has compressed the inter-generational gap to fruit fly speed.

My bet is this: our children and everyone they know will not die of cancer. There will be economic disruption. When cars came, the farriers sharpening horseshoes had a tough time. But look at the balance sheet of gains vs. 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, yet we chose factory-farming style teaching