Cerebras CEO Interview: With $25 Billion in Backlogged Orders, AI Computing Demand is Already Fully Booked

Original Title: Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit---Cerebras & Black Forest Labs CEOs
Original Source: All-In Podcast
Original Compilation: Deep Tide TechFlow

Key Takeaways

In this episode, we invited the CEOs of two AI infrastructure companies. Andrew Feldman is the founder of Cerebras, a company specializing in inference chips, which has just completed its IPO and holds $25 billion in backlogged orders. He repeatedly emphasizes one thing: the demand for AI computing power is already fully booked, and there is no situation of "waiting for customers after building it"; the appetite of OpenAI, Anthropic, SpaceX, and Google far exceeds supply. The emergence of reasoning has caused computational intensity to soar again, making it a battlefield 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, which is the foundation of all current image and video generation models. He recently collaborated with Martin Scorsese to visualize the director's ideas using AI; however, he is more excited about the direction where the same multimodal model can make movies and be deployed as a brain in robots. The endpoint of generative video is not on the screen but in the physical world.

Highlights of Insights

Reasoning is the Next Computing Power Black Hole

• "Interestingly, this wave is different from the past; they are not betting on 'once built, people will come.' The demand has already booked the production capacity. We have $25 billion in backlogged orders."

• "Reasoning consumes massive tokens, which is precisely the battlefield for fast machines."

• "If Cerebras is 15 times faster, running for 24 hours is equivalent to weeks or even months of thinking."

Open Source and Sovereignty: Companies Want Control

• "No one likes to be dependent. The lesson learned by large-scale vendors from the x86 era is being bound by Intel."

• "You don't need to make the fastest chip; you just need to not be completely reliant on others' chips."

• "If you want to run open-source models now, it’s either OpenAI's OSS 12B or Chinese models; the U.S. needs more local open-source options."

AGI Has Already Arrived by Definitions from Twenty Years Ago

• "Any AGI definition we proposed 20, 30, or 40 years ago, we have already far surpassed."

• "Turing Test? It has long been surpassed."

• "The problem is no longer that we don't know how to ask; AI can tell you: 'Hey, you foolish humans, you haven't considered this.'"

Generative Video is Not a Replacement for 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 detrimental means of communication; visual information signals are too rich. Turning the images in your mind into visible images is where technology is most powerful."

• "The most interesting results almost always occur 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 and then deploy it as the brain of a robot."

• "Pre-trained video implicitly teaches the model the laws of physical interaction, and then you get action predictions from the same model, which is robot control."

• "The goal is for you to be able to instruct the robot with in-context prompts: 'Bring that glass of orange juice over here.' We can't do that yet, but that's the direction."

AI Infrastructure Boom: Data Centers Larger than Cities

Host: We have never seen such a scale of construction. Since the Great Wall and the Pyramids, humanity has not invested so much capital, time, and smart people into building something. You are actually doing this, and your clients are building data centers; you are a key part. What is Cerebras doing in 2026? What’s the situation with those massive projects in Texas?

Answer: The data centers we are talking about will consume more electricity in the coming years than the total amount consumed on Earth in the past 50 years. A single building is as large as a football field, with power connections exceeding that of a medium-sized city. Construction is happening all over the U.S., Canada, Northern Europe, Paris, and all of France, the Middle East, and even Kazakhstan, Tajikistan, and Georgia are building large data centers. Every country and every state wants to get involved.

Who is paying? OpenAI, Anthropic, SpaceX AI, Google, with appetites that are frighteningly large. Interestingly, this wave is different from many past tech booms: they are not betting on 'once built, people will come'; the demand has already booked the production capacity. We have $25 billion in backlogged orders. OpenAI wants more data centers, Microsoft wants more, AWS wants more. The demand is not waiting for customers to come; the customers are already lined up.

Host: This has also given rise to a term called 'token maxing,' endlessly churning tokens. Some question whether such a large demand is creating real value.

Answer: Of course, a lot of value is being generated. There is also a lot of trial and error. I compare it to when AWS first came out; it was so refreshing to bypass my IT department, with every engineer signing up with a credit card. Many things are indeed useful, while some later make you think, 'Hey, I shouldn't have done that.' But overall, it’s still profitable; just some directions have gone astray.

I remember when Costco opened in Palo Alto in 1988, people browsed Costco like they did Safeway, walking through every aisle. It was a terrible way to shop because you ended up buying four unnecessary items, each costing $22. Later, people learned strategies: go to the back for chicken, grab 18 cupcakes for a kid's birthday party, and be efficient. AI token consumption is similar; at first, everyone used it freely, but now companies are starting to strategize: which tasks can be done with open-source models and which must use cutting-edge models. We are beginning to manage AI like a business.

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

Host: Sam Altman mentioned on AllIn that the next step is reasoning, understanding intent, formulating strategies, and cross-validating with other agents. We have come a long way from 'guess the next word,' and now Cerebras is right at the center because reasoning is inference, which requires massive computation.

Answer: Reasoning consumes massive tokens, which creates a battlefield for fast machines. Each step of reasoning internally consumes tokens; you originally relied on spending a lot of time to get good answers. Cerebras being 15 times faster means that running reasoning for 24 hours is equivalent to weeks or even months of thinking for others.

This morning, I tried a GLM-52 model on BitTensor, giving it unlimited computing power and asking it to tell me about trends that have not yet been identified worldwide every hour. It started debating with itself: should it look for trends on Hacker News and Reddit, or are trends appearing first on Instagram? I watched a reasoning model debating itself in the background; it was reasoning. Unlimited tokens equal unlimited reasoning; with Cerebras being 15 times faster, 24 hours is equivalent to weeks for others.

Host: Does Cerebras have its own Moore's Law? How long do you discuss before doubling?

Answer: All previous chips followed Moore's Law, doubling every 18 months. We broke that line with this chip, creating a completely new trajectory. My judgment is that in the next 18 months, it will far exceed double. The new architecture still has a lot of optimization space. GPUs are based on a 20-year-old architecture, only able to sustain themselves by shrinking process nodes, but the new architecture has a lot to learn and adjust.

Host: With $25 billion in backlogged orders, you also have to keep up with OpenAI's pace, as they may be potential competitors in the future. How do you operate the company?

Answer: Right now, silicon chips will not be idle; the demand is too great. But you are right, OpenAI is also making its own chips, and Amazon is too. No one likes to be dependent. The lesson learned by large-scale vendors from the x86 era is being bound by Intel; GPU vendors learned the lesson of being bound by a few large customers, so they funded new clouds. Making your own chips is not about being the fastest but about not being completely reliant on others, at least controlling a significant part of your own destiny.

Open Source and Sovereignty: Companies Want Control

Host: Open source is reaching a moment. I used OpenClaude early on, then Kimmy, and found that 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 has suddenly closed this year.

Answer: You don’t want to drive a Ferrari to the supermarket. Sometimes you drive a sports car, and sometimes you drive a minivan; you don’t mind if your kid spills Cheerios. Companies are the same: hard problems are given to cutting-edge models (OpenAI, Anthropic, Gemini), but behind them, many everyday issues only need solid open-source capabilities. Think about how much time a company spends copying from Workday to another cell in Excel? This doesn’t require gold medal math; solid open-source is enough.

Summary:

Recently, another card has been flipped: regulated industries like finance and healthcare (HIPAA, FINRA) are worried about data leaks and the potential for others to control their intelligent sovereignty. They want to run models locally and use open-source versions to gain more control. A few months ago, OpenAI released OSS 12B, which is decent. However, the U.S. is now pushing for open-source, either OSS 12B or Chinese models, with very few local open-source options available. NVIDIA has also noticed this window and is promoting its own open-source models, but Jensen is hesitating; his clients are Sam, Dario, Elon, and Sergey, and he wonders if promoting open-source will compete with his clients' businesses.

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

AGI Has Arrived, Paradigms Won't Die, People Will

Host: When Fable 5 and o-56 were released, the government said, "Pause before releasing more." Anthropic's relationship with the administration has been tense, but it is starting to ease. Do you think a phased release is reasonable? Are models really that dangerous?

Answer: I've never seen anything like this before. But looking back, when a model is powerful enough in creative thinking, and the government says, "Please release it in phases," I think that's actually reasonable. We manage potent drugs this way; of course, we don't encourage the FDA's seven years of bureaucratic paperwork, but saying, "At least let the government conduct some red team testing to confirm our defenses can hold up," and giving two or three weeks to patch obvious vulnerabilities is not an unreasonable request.

But now is the time of the most severe polarization. If this were not done by Trump, any other president's reaction might have been completely different. Polarization has harmed clear thinking. Both sides will do foolish things and smart things. The grassroots personnel in the government are actually working seriously; it's just that 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, forcing them to stop everything and spend six weeks patching. You realize this is a powerful tool; maybe let a small group see it first, or conduct red team testing first.

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

Answer: Yes. The Turing Test? It has long been surpassed. Any definitions proposed 10, 15, 20, 30, 40, or 50 years ago, we have far exceeded. The questions posed by science fiction writers have all been answered; they would say, "I have no more questions, sorry." This is why what seems to be on the fringe is worth listening to; Ilya talked about safety eight years ago, and you said, "What?" It turns out he was right. Elon talked about reducing rocket costs to near zero, and you said, "What?" He did it.

Host: Recursive learning, you ask it a question, learn the result, ask again, and the answer is better, covering more material. The answers produced in these cycles jump directly from "a little better" to "much better." The slope of the exponential curve is too steep.

Answer: Recursive gains are exponential; you get better, then come back for more, continuing to gain, and the slope is too steep. We are just starting to see this. If we keep investing in computing power, will the answers keep getting better? Once we run out of tokens or budget, we stop, but when does this exponential curve reach its peak? Does it keep going up forever? This question is fascinating right now.

The speed of human learning is constrained by generations; elephants and large mammals take 15-20 years for a generation. To learn quickly, you have to be like a fruit fly, which has two generations in a day. AI is achieving this learning speed across thousands of generations. When I studied psychology, a professor said: paradigms won't die, people will. Freud, Skinner, and Jung's disciples held leadership positions for 20-40 years before the next generation questioned them. AI is compressing the intergenerational gap to the speed of fruit flies.

I bet on this: our children and everyone they know will not die from cancer. The economy will have shocks; the arrival of cars made life difficult for those who shod horses. But if we list the gains and losses: infinite energy, infinite food, infinite knowledge, infinite education, infinite housing. We have 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-style education. Now AI can give every child a tutor who teaches in their own way.

Scorsese's AI Toolbox: Turning Imagery into Reality

Host: Robin Rombach is the co-founder and CEO of Black Forest Labs, headquartered in Freiburg and San Francisco. You previously worked on Stable Diffusion and invented the latent diffusion algorithm. What is the business of Black Forest Labs? What is the goal?

Answer: My partners and I founded this company two years ago. We previously worked on Stable Diffusion and earlier invented latent diffusion, which is the foundational algorithm behind all current image generation, video generation, and even physical AI models. The principle is to compress natural data (images, videos, audio) into an efficient representation space and then train transformers on it, similar to the principles of JPEG and MP3, but implemented using neural network algorithms. We developed it during our PhD studies in Munich.

Now we are tackling multimodal visual models, pre-training on image and audio data simultaneously, entering a new paradigm: combining action prediction so that the same model can handle images, videos, audio, and predict actions, ultimately deployable on real-world robots.

Host: From images to videos to audio to robots, if a model can generate video, it means it understands the world.

Answer: Intuitive intelligence and deep reasoning are two complementary forms of intelligence. We start from the intuitive side; images are the most natural entry point, and the computational load is not as high as video. But now it is converging into multimodal models. Pre-training on video implicitly teaches the model the laws of physical interaction, obtaining action predictions from the same model, which is robot control.

Host: Do you collaborate with Martin Scorsese? Are you sitting next to him while he uses your tools?

Answer: Yes, I sat in the same room with him; he explored our model, and as one of the core researchers, I sat next to him. It felt incredible. At the same time, I am a big fan of his.

What he wants is to visualize the scenes in his mind, describing a village in Eastern Europe, we look at the output as he iterates. In the end, he said: turning the images in your head into visual expressions is a communication efficiency far superior to language. Language is a somewhat lossy form of communication; visual information carries a rich signal; the amount of information in a picture or a video is enormous; this is another communication channel.

We don't want to dictate how to use these models, especially not to Martin Scorsese, saying, "You should use it this way." AI models are a medium. The most interesting things almost always emerge when people are iterating in the loop.

From Movies to Robots: The Endpoint of Generative Models Is Not on the Screen

Host: Startups are now using Flux and your models to create launch videos; previously, it cost $250,000 to make a launch video, but now it can be done in a week or two. Gal Gadot just made a Bitcoin movie where actors performed on a sound stage without green screens, with all backgrounds created using generative AI, achieving effects that originally would have cost $150 million for a $30 million budget. Have you seen this in production?

Answer: I've seen some. High-end film production is one of the most demanding use cases. I'm glad people are exploring it, but I also want to clarify: the technology is still on a trajectory and is rapidly iterating. A few years ago, when we were doing our PhDs, we could only generate 64×64 pixel images; now we can create high-resolution videos with multiple inputs, but it won't stop there.

What excites me the most is this: you can use the same multimodal model to make a movie and then deploy it as the brain of a robot. This is fascinating. Whether computer use can be applied is still uncertain, but the technology is moving toward the physical world; world models, action models, are essentially the same thing.

Host: Where does the training data come from? Do you have humans wear glasses and gloves to record first-person perspectives? Or is it enough to watch a thousand videos of people pouring drinks on YouTube?

Answer: The goal is to use in-context prompts to instruct the robot: "Bring me that glass of orange juice." We can't do that yet. Currently, the approach is: the model has already been equipped with a large amount of visual understanding, and it only needs a few hours of fine-tuning data to adapt to specific hardware. The direction is to minimize fine-tuning and rely as much as possible on in-context instructions, but this is still a research question.

Host: Open source is having its moment; companies want sovereignty. How should IP giants like Disney handle this? Should they train their own models using your open-source models or collaborate with you to train exclusive models?

Answer: The most interesting use cases lie in generating things that have never existed before; that is fundamentally the most interesting aspect of this technology. Our public tools cannot generate specific IP, which is reasonable. We do collaborate with some IP holders to develop models, some based on our open-source models and some based on our stronger proprietary models.

The most interesting angle is that technology is becoming faster and more interactive. You can imagine various interactive content creation tools available on Disney+.

Host: The most interesting phenomenon now is fan films. Previously, there was fan fiction writing personal Star Wars stories, and then people started making fan films in Jedi costumes. George Lucas said it was allowed as long as it wasn't for commercial use. Now people are using AI to reinterpret untold Star Wars stories, with each video in Star Wars Stories Untold getting millions of views. This is the future: allowing consumers to pay for authorization to create their own stories using characters.

Answer: If a viable business model can be found for IP holders that also opens up this super creative customization, that would be fantastic. When I read a book or watch a movie, I always think, "What if it developed this way?" Now we can finally visualize those thoughts.