Interview with Cerebras CEO: Holding 25 Billion in Backlog Orders, AI Compute Demand Already Fully Booked, We Are Not "Build It and Wait for Customers"
TechFlow Selected TechFlow Selected
Interview with Cerebras CEO: Holding 25 Billion in Backlog Orders, AI Compute Demand Already Fully Booked, We Are Not "Build It and Wait for Customers"
AGI has already arrived according to the definition from twenty years ago.
Navigating Web3 tides with focused insightsOrganized & Compiled: 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
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 recently completed its IPO with a $25 billion backlog. He repeatedly emphasizes one thing: demand for AI compute is already fully booked; there is no "build it and they will come" scenario. The appetites of OpenAI, Anthropic, SpaceX, and Google far exceed supply. The emergence of reasoning has sent computational intensity soaring again, which is exactly the battlefield for fast machines. Robin Rombach is the founder of Black Forest Labs, creating generative image and video models (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, enabling the director to visualize scenes from his mind using AI; but the direction he is more excited about is that the same multimodal model used to make movies can be deployed as a brain on robots. The endgame of generative video is not on the screen, but in the physical world.
Highlights
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 out capacity. We have a $25 billion backlog."
- "Reasoning is reasoning, reasoning consumes massive amounts of tokens, which is exactly the battlefield for fast machines."
- "If Cerebras is 15x faster, running for 24 hours is equivalent to weeks or even months of thinking."
Open Source and Sovereignty: Enterprises Want Control
- "No one likes being dependent. Hyperscalers learned from the x86 era that being tied to Intel was a lesson."
- "You don't need to make the fastest chip; you just need to not be completely dependent on someone else's chips."
- "If you want to run open source models now, it's either OpenAI's OSS 12B or Chinese models. The US needs more local open source options."
AGI Has Arrived by Definitions from 20 Years Ago
- "Any definition of AGI we proposed 20, 30, or 40 years ago, we have far surpassed."
- "Turing Test? Blown away long ago."
- "The problem is no longer that we don't know how to ask; AI can tell you: Hey, you dumb humans, you didn't consider this."
Generative Video Is Not 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 too rich. Turning mental images into visible images is where the technology is most powerful."
- "The most interesting results almost always appear when humans are in the loop iterating constantly."
From Movies to Robots: The Same Model
- "You can use the same multimodal model to make a movie, and then deploy it as a brain on a robot."
- "Pre-trained video implicitly teaches the model physical interaction laws, then you get action prediction from the same model, which is robot control."
- "The goal is to be able to instruct the robot with an in-context prompt: 'Bring that glass of orange juice over.' We can't do that yet, but that's the direction."
AI Infrastructure Boom: Data Centers Larger Than Cities
Host: We have never seen construction scales like this. Since the Great Wall and the Pyramids, humanity has never invested so much capital, time, and brainpower into building one thing. You are actually doing this; your clients are building data centers, and you are a key link. What is Cerebras doing in 2026? What is the situation with those huge projects over in Texas?
The data centers we are talking about will consume more electricity in the next few years than the total of the past 50 years on Earth. A single building is the size of a football field, connected to more power than a medium-sized city. They are being built across the US, in Canada, in Northern Europe, in Paris and throughout France, in the Middle East, and even large data centers are being built in Kazakhstan, Tajikistan, and Georgia. Every country, every state wants to get involved.
Who is paying? OpenAI, Anthropic, SpaceX AI, Google, appetites are terrifyingly large. Interestingly, this wave is different from many past tech booms: they are not betting on "build it and they will come." Demand has already booked out capacity. We have a $25 billion backlog. OpenAI wants more data centers, Microsoft wants more, AWS wants more. Demand is not waiting for customers to come; customers are already lining up.
Host: This has also spawned a term called "token maxing," infinite token brushing. Some question whether such huge demand is actually creating real value?
Of course, a lot of value is being generated. Of course, there is also a lot of blind testing. I compare it to when AWS first came out; bypassing your own IT department was so cool, every engineer registered with a credit card. A lot was indeed useful, some in hindsight you think "hey, shouldn't have done that." But overall it was profitable, just some directions went empty.
I still remember when Costco opened in Palo Alto in 1988; everyone walked through Costco like it was Safeway, walking down every aisle. That was a terrible way to shop because you bought four things you didn't need, each $22. Later everyone learned strategy: go to the back for chicken, get 18 cupcakes for the kid's birthday party, clean and efficient. AI token consumption is the same; at first everyone used it openly, now enterprises are starting to talk strategy: which tasks are fine with open source models, which must use frontier models. We are starting to manage AI like running a business.
Reasoning Replaces Training: Why Fast Machines Are the Star of This Wave
Host: Sam Altman said on All-In that the next step is reasoning, understanding intent, formulating strategies, and cross-validating with agents in other threads. We have come a long way from "guessing the next word," and now Cerebras happens to be right in the center because reasoning is inference, computationally intensive.
Reasoning consumes massive amounts of tokens, which gives fast machines a battlefield. Every step of reasoning swallows tokens internally; you originally traded a lot of time for good answers. Cerebras being 15x faster means running 24 hours of reasoning is equivalent to weeks or even months of thinking for others.
I tried a GLM-52 model from ZAI on BitTensor this morning, gave it infinite compute, and asked it to tell me trends worldwide that haven't been identified yet every hour. It started debating itself: should it look on Hacker News and Reddit? Or do trends appear first on Instagram? I watched a reasoning model debating itself in the background; it was doing reasoning. Infinite tokens equal infinite reasoning; using Cerebras 15x faster, 24 hours equals weeks for others.
Host: Does Cerebras have its own Moore's Law? How often do you discuss doubling internally?
All previous chips stepped on Moore's Law, doubling every 18 months. We used this chip to break that line and ran a completely new trajectory. My judgment is that in the next 18 months, it will be far more than 2x. There is still plenty of optimization space for the new architecture. GPU is a 20-year-old architecture; it can only rely on shrinking process nodes to hold on, but there is still a lot to learn and tune in the new architecture.
Host: With a $25 billion backlog in hand, you still have to keep up with OpenAI's pace; they might be potential competitors in the future. How do you operate the company?
Silicon won't sit idle now; demand is too huge. But you are right, OpenAI is also making its own chips, Amazon is too. No one likes being dependent. Hyperscalers learned from the x86 era that being tied to Intel was a lesson; GPU vendors learned the lesson of being tied to a few hyperscale customers, so they funded new clouds. Making your own chips is not about being the fastest; it's about not being completely dependent on others, at least controlling an important part of your own destiny.
Open Source and Sovereignty: Enterprises Want Control
Host: Open source is having a moment. I used OpenClaude early on, later used Kimmy, and found my Claude tokens were exploding, but with Kimmy I couldn't tell the difference. Open source models are starting to do reasoning; the gap suddenly closed this year.
You don't want to drive a Ferrari to the supermarket. Sometimes you drive a sports car, sometimes a minivan; you don't heartache if kids spill Cheerios. Enterprises are the same: hard problems go to frontier models (OpenAI, Anthropic, Gemini), but behind the scenes大量 daily problems just need solid open source capabilities. Think about how much time a company spends copying and pasting in Workday to another cell in Excel? This doesn't need gold-medal math; steady open source is enough.
Recently another card was flipped: regulated industries like finance and healthcare (HIPAA, FINRA) fear data leaks, fear intelligence sovereignty being held by others, want to put models locally, use open source versions to grab a bit 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 local open source options. NVIDIA also sees this window, pushing its own open source models, but Jensen is hesitating; his clients are Sam, Dario, Elon, Sergey; pushing open source might compete with clients?
Cerebras stands in a relatively neutral position; we run GLM, run Kimmy, run Qwen series, and also run OpenAI's closed source models. We also run models developed by GSK itself, run UAE G42 and MBZUAI's own models. Sovereignty is a trend.
AGI Has Arrived, Paradigms Don't Die, People Do
Host: When Fable 5 and o-56 were released, the government said "pause before releasing." Anthropic and the administration had tense relations, now starting to ease. Do you think phased release is reasonable? Are models really dangerous enough?
I haven't seen anything like this before. But thinking back: when a model is powerful enough in creative thinking, the government says "please release in phases," I think this is actually fine. We manage potent drugs this way; of course, we don't encourage that pile of seven-year garbage paperwork from the FDA, but saying "at least let the government do some red team testing, confirm our defenses can hold," give two or three weeks to patch obvious vulnerabilities, this is not an unreasonable request.
But now polarization is at its most severe. If this wasn't done by Trump, any other president, the reaction might be completely different. Polarization hurts clear thinking. Both sides will do stupid things, and also smart things. Grassroots personnel in the government are actually working very seriously; it's just that this is moving too fast.
Nikesh from Palo Alto Networks told me: they tested models against their own software and found dozens of critical vulnerabilities within an hour; they had to stop everything at hand and spend six weeks patching. You realize this is a powerful tool; maybe show it to a small group first, maybe do red team testing first.
Host: By any definition from 20 years ago, AGI has arrived. Do you agree?
Yes. Turing Test? Blown away long ago. Any definition proposed 10, 15, 20, 30, 40, 50 years ago, we have far surpassed. Questions raised by sci-fi writers we have answered; they would say "I have no more questions, sorry." This is why words spoken by those who look marginal are worth listening to; Ilya talked about safety eight years ago, you said "what?" Turns out he was right. Elon talked about rocket costs dropping to near zero, you said "what?" Turns out he did it.
Host: Recursive learning, you ask it a question, learn the result, ask again, the answer is better, covers more material; answers produced from these loops jump from "a bit better" directly to "much better." The slope of the exponential curve is too steep.
Recursive gains are exponential; you get better, do it again, continue gaining, the slope is too steep. We are just starting to see this. Keep throwing compute at it, will answers keep getting better? Stop when tokens or budget run out, but when does this exponential curve end? Or does it go up and right forever? This question is extremely interesting now.
Human learning speed is stuck by generations; elephants and large mammals take 15-20 years per generation. Want to be fast, be like fruit flies, two generations a day. AI is getting this learning speed across thousands of generations. When I studied psychology, a professor said one sentence: paradigms don't die, people do. Freud, Skinner, Jung's disciples occupied leadership positions for 20-40 years before the next generation questioned. AI compresses the generational interval to fruit fly speed.
What I bet on is this: our children and everyone they know will not die from cancer. There will be economic shocks; cars came, days were hard for people shaving horse hooves. But list the earnings 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-farming style teaching. Now AI can give every child a tutor that learns according to their own way.
Scorsese's AI Toolkit: Turning Mental Images into Reality
Host: Robin Rombach is the Co-founder and CEO of Black Forest Labs, headquartered in Freiburg in the Black Forest region and San Francisco. You previously worked on Stable Diffusion and invented the latent diffusion algorithm. What is Black Forest Labs' business? What is the goal?
My partners and I founded this company two years ago. Previously worked on Stable Diffusion, earlier invented latent diffusion, which is the foundational algorithm behind all current image generation, video generation, and even physical AI models. The principle is compressing natural data (images, video, audio) into an efficient representation space, then training transformers on top, like the principles of JPEG and MP3, but implemented with neural network algorithms. We did this during our PhD period in Munich.
Now we are tackling multimodal visual models, pre-training simultaneously on image and audio data, entering a new paradigm: combining action prediction, letting the same model do images, do video, do audio, and also predict actions, ultimately deployable on robots in the real world.
Host: From images to video to audio to robots; if models can generate video, it means they understand the world.
Intuitive intelligence and deep reasoning are two complementary forms of intelligence. We started from the intuitive side; images are the most natural entry point, computational load is not as large as video. But now it is converging into multimodal models. Pre-trained video implicitly teaches the model physical interaction laws; from the same model you get action prediction, which is robot control.
Host: You have a collaboration with Martin Scorsese? You sit next to him letting him use your tools?
Yes, I sat in the same room with him; he explored our models, sitting next to him as one of the core researchers; the feeling was too crazy. At the same time, I am still a big fan of his.
He wants to visualize scenes from his mind; some village in Eastern Europe, he describes, we look at the output, he iterates. Finally what he said was: turning pictures in the brain into visual expression; this communication efficiency is far higher than language. Language is a somewhat lossy communication method; visual information signals are too rich; the amount of information in a picture or a video is huge; this is another communication channel.
We don't want to dictate how to use these models, especially not tell Martin Scorsese "you should use it this way." AI models are a medium. The most interesting things almost always come out when humans are in the loop iterating constantly.
From Movies to Robots: The Endgame of Generative Models Is Not on Screen
Host: Startups are now using Flux and your models to make launch videos; previously spent $250,000 making a launch video, now can finish in one or two weeks. Gal Gadot just made a Bitcoin movie; actors performed on a sound stage without green screens; all backgrounds made with generative AI; $30M budget achieved effects that originally required $150M. Do you see this in production use?
Seen some. High-end film production is one of the most demanding use cases. I am glad someone is exploring, but I also want to make clear: technology is still on trajectory, iterating quickly. A few years ago when we did PhDs we could only generate 64x64 pixel images; now doing multi-input high-resolution video, but it won't stop here.
What excites me most is this: you can take the same multimodal model to make a movie, and then deploy it as a brain on a robot. This is too interesting. Whether computer use can actually be used is uncertain, but technology is moving towards the physical world; world models, action models, basically all the same thing.
Host: Where does training data come from? Have humans wear glasses and gloves to record first-person? Or is watching a thousand videos of people pouring drinks on YouTube enough?
The goal is to instruct the robot with an in-context prompt: "Bring that glass of orange juice over." Can't do it yet. Current approach is: the model is already loaded with lots of visual understanding; only needs a few hours of fine-tuning data to adapt to specific hardware. Direction is to fine-tune as little as possible, rely as much as possible on in-context instructions, but this is still a research question.
Host: Open source is having a moment; enterprises want sovereignty. What should IP giants like Disney do; take your open source models and train themselves, or collaborate with you to train exclusive models?
The most interesting use case lies in generating things that didn't exist before; this is essentially the most interesting place for this technology. We cannot generate specific IPs on our public tools; this is reasonable. We也确实 collaborate with some IP owners to develop models; some based on our open source models, some based on our stronger proprietary models.
The most interesting angle is: technology is becoming faster, more interactive. You can imagine various interactive content creation tools hanging on Disney+.
Host: The most interesting phenomenon now is fan films. Previously there was fan fiction writing their own Star Wars stories; later someone wore Jedi costumes to shoot fan films. George Lucas said as long as not commercial use it was allowed. Now people use AI to reinterpret Star Wars stories not told; Star Wars Stories Untold each video has millions of views. This is the future: let consumers pay for licensing, let them use characters to create their own stories.
If a business model feasible for IP parties can be found, and this kind of super creative custom play can be opened up, that would be great. I always think "what if it developed this way" when reading a book or watching a movie; now finally people can visualize these thoughts.
We just passed 100 people; hiring in Germany and San Francisco: researchers for large-scale model training, people with diffusion and flow matching training experience, engineers developing custom solutions with clients, people for large-scale compute infrastructure operations, and people interested in getting technology into more hands.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
深潮TechFlow
