Conversation with Cerebras CEO: $25 Billion Backlog, AI Computing Capacity Already Fully Booked – We’re Not "Building and Waiting for Customers"

Compiled & Translated: TechFlow

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

Host: All-In Podcast Host

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 featured two CEOs from AI infrastructure companies. Andrew Feldman is the founder of Cerebras, a company specializing in inference chips that recently completed its IPO and holds $25 billion in backlog orders. He repeatedly emphasized one thing: demand for AI compute power is already 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 emergence of reasoning has once again skyrocketed compute intensity, which is exactly the battleground for fast machines. Robin Rombach is the founder of Black Forest Labs, which creates generative image and video models (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 more excited about the direction where the same multimodal model can be used to make movies and also deployed as the brain for robots. The endpoint of generative video isn't the screen; it's the physical world.

Highlight Reel

Reasoning is the Next Compute Black Hole

• "Interestingly, this wave is different from the past. They aren't betting on 'if we build it, they will come'. Demand has already booked up capacity. We have $25 billion in backlog orders."
• "Inference is reasoning, and reasoning consumes a massive number of tokens. This is precisely the battleground for fast machines."
• "If Cerebras is 15x faster and you run it for 24 hours, it's equivalent to running weeks or even months of thinking."

Open Source & Sovereignty: Enterprises Want Control

• "Nobody likes dependency. 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 entirely dependent on someone else's chip."
• "If you want to run open-source models now, it's either OpenAI's OSS 12B or Chinese models. America needs more native open-source options."

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

• "By any definition of AGI we proposed 20 years ago, 30 years ago, 40 years ago, we have already far surpassed them."
• "The Turing Test? We blew past it a long time ago."
• "The issue is no longer 'we don't know how to ask'. The AI can turn around and tell you: 'Hey, you dumb humans, you didn't consider 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. The signal in visual information is incredibly rich. Turning the picture in your mind into a visible image – that's the most powerful aspect of this technology."
• "The most interesting results almost always come from the process where humans are iterating in the loop."

From Movies to Robots: One Model Suite

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

The AI Infrastructure Boom: Data Centers Bigger Than Cities

Host: We've never seen this scale of construction. Since the Great Wall and the Pyramids, humanity hasn't poured this much capital, time, and smart people into building something. You're actually doing it. Your clients are building data centers, and you're a key part. In 2026, what is Cerebras doing? 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 planet used in the last 50 years. A single building is the size of a football field, drawing more power than a medium-sized city. They're being built all across the US, in Canada, in Northern Europe, in Paris and all over France, in the Middle East, and even large data centers are going up in Kazakhstan, Tajikistan, and Georgia. Every country, every state wants a piece of it.

Who is paying for it? OpenAI, Anthropic, SpaceX AI, Google. Their appetites are terrifying. Interestingly, unlike many past tech booms, they aren't betting on 'if we build it, they will come'. Demand has already booked up capacity. We have $25 billion in backlog orders. 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 has also spawned a term called 'token maxing', burning tokens endlessly. Some question whether all this massive demand is actually creating real value?

Of course, there's massive value being created. And of course, there's a massive amount of experimentation. I compare it to when AWS first came out. It was so liberating to bypass your own IT department; every engineer would just sign up with a credit card. A lot of it was genuinely useful, and some things, in hindsight, you'd think, 'Wow, we shouldn't have done that.' But overall, it was profitable; it's just that some directions are dead ends.

I remember when Costco opened in Palo Alto in 1988. People shopped there like it was a Safeway, going down every single aisle. It 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, get the chicken, grab 18 cupcakes for the kid's birthday party, and get out efficiently. AI token consumption is the same. Initially, everyone used it liberally. Now companies are becoming strategic. They're figuring out which tasks are fine with an open-source model and which require a frontier model. We're starting to manage AI like a business.

Reasoning Replaces Training: Why Fast Machines Are the Stars Now

Host: Sam Altman said on All-In that the next step is reasoning – understanding intent, formulating strategy, cross-referencing with agents on other threads. We've come a long way from 'predicting the next word'. Cerebras is now right in the middle of this because reasoning *is* inference, and it requires immense computation.

Reasoning consumes a massive number of tokens, giving fast machines their battleground. "Reasoning" swallows tokens internally at every step. You used to trade a lot of time for a better answer. Cerebras being 15x faster means running inference for 24 hours is equivalent to weeks or months of thinking for others.

I tried a GLM-52 model from ZAI on BitTensor this morning. I gave it infinite compute power and told it to tell me every hour what global trend hasn't been identified yet. It started debating with itself: "Should I look on Hacker News and Reddit? Or do trends appear first on Instagram?" I watched a reasoning model self-debating in the background; it was doing reasoning. Infinite tokens equal infinite reasoning. With Cerebras being 15x faster, 24 hours is equivalent to weeks for others.

Host: Does Cerebras have its own Moore's Law? How often do you internally discuss doubling performance?

All previous chips have followed Moore's Law, doubling every 18 months. We broke that trend with this chip and created a completely new trajectory. My assessment is that over the next 18 months, the improvement will be far more than 2x. There's still plenty of room to optimize the new architecture. GPUs are a 20-year-old architecture, relying solely on shrinking process nodes to keep up. But a new architecture has a vast amount left to learn and tune.

Host: With a $25 billion backlog, you also have to keep up with OpenAI, which could be a potential future competitor. How do you run the company?

Right now, silicon doesn't sit idle. The demand is too great. But you're right, OpenAI is making its own chips, and Amazon is too. Nobody likes dependency. The hyperscalers learned from the x86 era the lesson of being locked into Intel; the GPU companies learned the lesson of being locked into a few hyperscale clients, which is why they funded new cloud providers. The point of making your own chip isn't to be the fastest; it's to not be entirely dependent on others, to at least control a significant part of your own destiny.

Open Source & Sovereignty: Enterprises Want Control

Host: Open source is having a moment. I used OpenClaude early on, then later Kimmy, and found my Claude tokens were burning, 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 where you don't mind if the kids spill Cheerios. It's the same for enterprises. Hard problems get sent to frontier models (OpenAI, Anthropic, Gemini), but a huge number of daily problems just need solid open-source capability. Think about how much time a company spends copying data from Workday into an Excel cell? You don't need gold-medal math for that; a reliable open-source model is enough.

Recently, another card has been turned over. Regulated industries like finance and healthcare (HIPAA, FINRA) are afraid of data leaks and having their AI sovereignty controlled by others. They want to deploy models on-premises and use open-source versions to grab a bit more control. OpenAI released OSS 12B a few months ago; it's okay. But if America wants to run open-source now, it's either OSS 12B or Chinese models. The number of US-native open-source options is too scarce. NVIDIA sees this window too, promoting their own open-source models, but Jensen is also hesitant. His customers are Sam, Dario, Elon, Sergey. Would promoting open-source compete with his clients?

Cerebras occupies a more neutral position. We run GLM, we run Kimmy, we run the Qwen series, and we also run 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, 'Hold on, release in stages.' Tensions were high between Anthropic and the administration, though it's easing now. Do you think staged releases are reasonable? Are the models truly dangerous enough?

I've never seen anything like this before. But looking back: when a model becomes powerful enough in creative thinking that the government asks you to release it in stages, I actually think there's some merit to it. We manage powerful drugs this way. I don't endorse the FDA's seven years of paperwork, but saying, 'At least let the government do some red-teaming to confirm our defenses are adequate,' giving two or three weeks to patch obvious vulnerabilities, doesn't seem like an unreasonable request.

But this is the most polarized time ever. If this weren't done by Trump, under any other president, the reaction might have been completely different. Polarization hurts clear thinking. Both sides will do stupid things, and both sides will do smart things. The people at the working level in government are actually working seriously on this; it's just that things are moving too fast.

Nikesh from Palo Alto Networks once told me: they tested a model against their own software and found dozens of critical zero-day vulnerabilities within an hour. They had to stop everything they were doing for six weeks to patch them. You realize this is a powerful tool. Maybe you should show it to a small group first, maybe do some red-teaming.

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

Yes. The Turing Test? We blew past it a long time ago. By any definition proposed 10, 15, 20, 30, 40, 50 years ago, we have far surpassed them. We have answered the questions sci-fi writers posed, to the point where they'd say, 'I have no more questions, sorry.' This is why listening to those on the fringe is worthwhile. Ilya talked about safety eight years ago, and you thought, 'What?' He was right. Elon said we could reduce rocket costs to near zero, and you thought, 'What?' He did it.

Host: Recursive learning. You ask it a question, it learns the result, you ask again, the answer is better, covering more material. These outputs from the cycle jump from 'a little better' to 'a lot better'. The slope of the exponential curve is just too steep.

Recursive gains are exponential. You get better, you do it again, you get more gains. The slope is incredibly steep. We are just starting to see this. If we keep feeding it compute power, will the answers keep getting better? It stops when it runs out of tokens or budget, but when does this exponential curve end? Does it just go up and to the right forever? This question is incredibly fascinating right now.

Human learning speed is constrained by generations. Elephants and large mammals take 15-20 years for a generation. To go fast, 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 something that stuck: 'Paradigms don't die; people do.' The disciples of Freud, Skinner, and Jung would hold leadership positions for 20-40 years before the next generation could challenge them. AI has compressed inter-generational time to the speed of a fruit fly.

My bet is this: our children and everyone they know won't die from cancer. There will be economic upheavals. When cars came, people who