对话Cerebras CEO ：手握250亿积压订单，AI算力需求早就订满，我们不是「建了等客来」

Arranged & 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

Release Date: July 10, 2026

Key Takeaways

This episode features the CEOs of two AI infrastructure companies. Andrew Feldman is the founder of Cerebras, a company specializing in inference chips that just went public and holds $25 billion in backlog orders. He repeatedly emphasized one thing: demand for AI compute is already fully booked; there's no "build it and they will come" situation. The appetites of OpenAI, Anthropic, SpaceX, and Google far outstrip supply. The emergence of reasoning has caused compute intensity to skyrocket again, making this the perfect battleground for fast machines. Robin Rombach is the founder of Black Forest Labs, which builds generative image and video models (Flux series). He previously invented the latent diffusion algorithm, the foundation of all current image and video generation models. He recently collaborated with Martin Scorsese, allowing the director to visualize scenes from his mind using AI. But even more exciting for him is that the same multimodal model can be used to make movies and also deployed as the brain of a robot. The endpoint of generative video isn't the silver screen; it's the physical world.

Highlights of Key Insights

Reasoning is the Next Compute Black Hole

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

Open Source & Sovereignty: Enterprises Want Control

• "Nobody likes being dependent. The lesson the 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 not be completely reliant 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 domestic open-source options."

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

• "Any definition of AGI we proposed 20 years ago, 30 years ago, or 40 years ago, we have far exceeded."
• "The Turing test? We crushed it a long time ago."
• "The problem is no longer 'we don't know how to ask.' AI can tell you back: 'Hey, you dumb humans, you didn't consider this.'"

Generative Video is Not About Replacing Human Creativity

• "These AI models are a medium. We don't want to dictate how to use them, especially for someone like Martin Scorsese."
• "Language is a somewhat lossy communication method. Visual information signals are incredibly rich. Turning the images in your mind into visible pictures—that's where the power of this technology lies."
• "The most interesting results almost always emerge when humans are iterating in the loop."

From Movies to Robots: The Same Model

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

The AI Infrastructure Frenzy: 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 brainpower into building something. You are actually doing it; your clients are building data centers, and you are a key part of it. In 2026, what is Cerebras doing? What's the situation with those massive projects in Texas?

In the next few years, the data centers we're talking about will consume more electricity than the earth has consumed 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 across the US, in Canada, in Northern Europe, in Paris and all of France, in the Middle East, and even large data centers are going up in Kazakhstan, Tajikistan, and Georgia. Every country, every state wants in on this.

Who's paying for it? OpenAI, Anthropic, SpaceX AI, Google. Their appetite is terrifying. Interestingly, unlike many previous tech booms, they aren't betting on "if you build it they will come"; demand has already consumed the capacity. We have $25 billion in backlog orders. OpenAI wants more data centers, Microsoft wants more, AWS wants more. Demand isn't waiting for customers; customers are already lining up.

Host: This has also spawned a term, 'token maxing'—unlimited generation of tokens. Some question whether this massive demand is actually creating real value?

Of course, a huge amount of value is being created. And of course, there's also a huge 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 sign up with a credit card. A lot of it was useful. In hindsight, some things were like, "Ah, we shouldn't have done that." But overall, it was profitable, just with some wrong directions.

I remember when Costco opened in Palo Alto in 1988. People shopped there like Safeway, walking down every aisle. It was a terrible way to shop because you'd buy four things you didn't need for $22 each. Eventually, people learned the strategy: go to the back, get the chicken, get the 18-pack of cupcakes for the kid's birthday party, get in and out. AI token consumption is the same. At first, everyone used it liberally. Now, businesses are learning strategies: which tasks are fine with open-source models, and which require frontier models. We are starting to manage AI like we run a business.

Reasoning Replaces Training: Why Fast Machines Are the Main Character

Host: Sam Altman mentioned 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.' Now Cerebras is right at the center because reasoning is inference, and it's computationally massive.

Reasoning consumes a massive number of tokens, giving fast machines a battlefield. Each step of reasoning internally swallows tokens. You used to trade a lot of time for good answers. Cerebras being 15x faster means that running inference for 24 hours is equivalent to weeks or even months of thinking for others.

I tried a ZAI GLM-52 model on BitTensor this morning. I gave it unlimited compute and told it to tell me every hour about trends in the world that weren't yet identified. It started debating with itself: Should we look on Hacker News and Reddit? Or do trends appear first on Instagram? I was watching a reasoning model debate itself in the background. It was reasoning. Unlimited tokens equal 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 about the doubling period?

All previous chips followed Moore's Law, doubling every 18 months. We broke that line with this chip and created a completely new trajectory. My assessment is that over the next 18 months, it will be far more than 2x. There's huge room for optimization in the new architecture. The GPU is a 20-year-old architecture, barely holding on by shrinking process nodes. But the new architecture has a lot left to learn and tune.

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

Right now, silicon doesn't sit idle. The demand is too large. But you're right, OpenAI is making its own chips, Amazon is too. Nobody likes being dependent. The lesson the hyperscalers learned from the x86 era was being locked in by Intel; the lesson the GPU vendors learned was being locked in by a few hyperscale customers, so they funded new clouds. The point of making your own chip isn't primarily to be the fastest; it's to not be completely reliant on others, to have at least some control over your own destiny.

Open Source & Sovereignty: Enterprises Want Control

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

You don't want to drive a Ferrari to the supermarket. Sometimes you drive a sports car, sometimes a minivan where you don't worry about the kids spilling Cheerios. It's the same for enterprises: tough problems go to frontier models (OpenAI, Anthropic, Gemini), but the vast majority of daily tasks just need solid open-source capabilities. Think about how much time a company spends copying and pasting between Workday and an Excel cell? That doesn't require gold-medal math; a reliable open-source model is good enough.

Recently, another factor emerged: regulated industries like finance and healthcare (HIPAA, FINRA) are afraid of data leaks and losing AI sovereignty to external entities. They want to run models locally, using open-source versions to grab more control. OpenAI released OSS 12B a few months ago. It's okay. But if the US wants to run open-source now, it's either OSS 12B or Chinese models. There are too few domestic open-source options. NVIDIA also sees this window and is pushing its own open-source models. But Jensen is also hesitating. His customers are Sam, Dario, Elon, Sergey. Would pushing open-source models compete with his customers?

Cerebras occupies a relatively neutral position. We run GLM, we run Kimmy, we run 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.' Relations between Anthropic and the administration were tense but are now easing. Do you think staged releases are reasonable? Are the models actually dangerous enough?

I've never seen anything like this before. But thinking back: when a model becomes powerful enough in creative thinking, and the government says 'please release it in stages,' I actually think that's not unreasonable. We do this with potent medicines. I don't encourage the FDA's seven years of garbage paperwork, but saying 'at least let the government do some red teaming to confirm our defenses can hold,' giving two or three weeks to patch obvious vulnerabilities, that's not an unreasonable request.

But this is the most polarized time. If this wasn't done by Trump, with 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 staff in the government are actually working hard on this; it's just that the pace is 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 and spent six weeks patching. You realize this is a powerful tool. Maybe show a small group first, maybe do red teaming first.

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

Yes. The Turing test? We crushed it a long time ago. Any definitions proposed 10, 15, 20, 30, 40, 50 years ago, we have far exceeded them. We've answered all the questions science fiction writers asked. They would say, 'I have no more questions, I'm sorry.' This is why what people on the fringe say is worth listening to. Ilya talked about safety eight years ago, you said 'What?' and he was right. Elon talked about reducing rocket costs to near zero, you said 'What?' and he did it.

Host: Recursive learning: you ask a question, learn the result, ask again, get a better answer covering more material. The output from these loops jumps from 'a bit better' to 'much better'. The slope of the exponential curve is too steep.

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

Human learning speed is capped by generations. Elephants and large mammals take 15-20 years per 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 was studying psychology, a professor said: Paradigms don't die, people do. The disciples of Freud, Skinner, Jung hold leadership positions for 20-40 years before the next generation questions them. AI has compressed the generational interval to fruit-fly speed.

Here's my bet: Our children and everyone they know will not die from cancer. The economy will have shocks. The car came, and the horseshoers had a hard time. But list the gains versus 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 teaching. Aristotle tutored Alexander, Socrates tutored his students. But we chose factory-farming education. Now AI can give every child a tutor that teaches them in their own way.