Key Points Summary

This episode featured two AI infrastructure company CEOs. Andrew Feldman is the founder of Cerebras, a company specializing in inference chips, which just went through an IPO and holds a $25 billion backlog. He emphasized one thing repeatedly: the demand for AI compute has long been oversubscribed, with no "build it and they will come" scenario, as OpenAI, Anthropic, SpaceX, and Google have appetites far exceeding supply. The emergence of reasoning has once again driven up computational intensity, making it the battleground for fast machines. Robin Rombach is the founder of Black Forest Labs, working on generative image and video models (Flux series). He previously invented the latent diffusion algorithm, which is the foundation of all current image and video generative models. He recently partnered with Martin Scorsese to help the director visualize the images in his mind using AI; but what excites him more is the idea that the same multimodal model can be used to shoot a movie and deployed in a robot as its brain. The endpoint of generative video is not the screen; it's the physical world.

Insightful Highlights

Reasoning Is the Next Compute Black Hole

· "What's interesting is that this wave is different from the past; they are not betting on 'build it and they will come,' the demand has already booked up the capacity. We have a $25 billion backlog."

· "Reasoning is reasoning; reasoning consumes a massive amount of tokens, which is precisely the battleground for fast machines."

· "If Cerebras is 15 times faster, and you run for 24 hours, it's equivalent to weeks or even months of thought."

Open Source and Sovereignty: Enterprises Want Control

· "No one likes to be dependent. The lesson learned by the hyperscale vendors from the x86 era was being tied to Intel."

· "You don't need to have the fastest chip; you just need not to fully rely on someone else's chip."

· "If you want to run an open-source model now, it's either OpenAI's OSS 12B or a Chinese model. The U.S. needs more local open-source options."

AGI Has Arrived According to the Definition from Twenty Years Ago

· "Any definition of AGI that we proposed 20, 30, 40 years ago, we have far surpassed."

· "The Turing test? It was smashed a long time ago."

· "The problem is no longer that we don't know how to ask; AI can tell you in turn: hey, you stupid humans, you didn't consider this."

Generative Video is Not a Replacement for Human Creativity

· "These AI models are a medium, and we don't want to dictate how to use them, especially for people like Martin Scorsese."

· "Language is a somewhat lossy communication method; visual information is too rich. Turning the images in your mind into visible pictures is where technology shines the most."

· "The most interesting results almost always occur when a person is continuously iterating in the loop."

From Movies to Robots: Same Model

· "You can use the same multimodal model to shoot a movie and then deploy it as the brain on a robot."

· "Pretraining on videos 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 that you can instruct the robot with in-context prompt: 'Bring me that glass of orange juice.' We can't do that now, but that's the direction."

AI Infrastructure Craze: Data Centers Larger than Cities

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

Answer: When we talk about data centers, the amount of electricity they will consume in the coming years will exceed the total amount used on Earth in the past 50 years. A single building is as large as a football field and consumes more power than a medium-sized city. Data centers are being built all over the United States, in Canada, the Nordic countries, Paris and the entire France, the Middle East, and even in Kazakhstan, Tajikistan, and Georgia. Every country, every state wants to join in.

The ones footing the bill? OpenAI, Anthropic, SpaceX AI, Google — with insatiable appetites. Interestingly, this wave is different from many past technology trends: they are not betting on a "build it and they will come" approach; the demand has already booked out the capacity. We have $250 billion in backlogged orders. OpenAI wants more data centers, Microsoft wants more, AWS wants more. The demand is not waiting for customers to show up; the customers are already lining up.

Host: This has also given rise to a term called 'token maxing,' unlimited token minting. Some people doubt whether such high demand is actually creating real value?

Answer: Of course, a significant amount of value is being created. Of course, there is also a lot of trial and error. When AWS first came out, it was so satisfying to bypass your own IT department; every engineer would sign up with a credit card. Many things were indeed useful, while some made you think later, "Hmm, maybe I shouldn't have done that." But overall, it's still profitable, just that some directions turn out to be dead ends.

I still remember in 1988 when Costco opened in Palo Alto, everyone shopped at Costco like they did at Safeway, going through every aisle. That was a terrible way to shop because you ended up buying four unnecessary items, each at $22. Later, people learned the strategy: go to the back to get chicken, pick up 18 cupcakes for a kid's birthday party, and get out efficiently. AI token consumption is the same; initially, everyone used it openly, but now companies are strategizing: which tasks can be done with open-source models, and which must use cutting-edge models. We are starting to manage AI like running a business.

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

Host: Sam Altman has said on AllIn that the next step is reasoning, understanding intent, devising strategies, and cross-validating with agents on other threads. We have come a long way from 'guess the next word,' and now Cerebras is right at the center because reasoning is inference, requiring massive computational power.

Answer: Reasoning consumes a massive amount of tokens, which gives the fast machine a battlefield. Each step of reasoning internally consumes tokens, where you originally relied on spending a lot of time to come up with the right answer. The Cerebras speedup of nearly 15 times means that running 24 hours of reasoning is equivalent to the amount of thought for others for weeks or even months.

This morning, I tried a GLM-52 model on BitTensor with ZAI, gave it unlimited computational power, and asked it to tell me every hour about global trends that have not yet been identified. It began debating with itself: should it look on Hacker News and Reddit? Or do trends appear first on Instagram? I watched a reasoning model arguing with itself in the background, engaging in reasoning. Unlimited tokens equal unlimited reasoning, and with Cerebras' nearly 15x speedup, 24 hours is equivalent to others' weeks.

Host: Does Cerebras have its own Moore's Law? How long does it take to double internally?

Answer: All previous chips followed Moore's Law, doubling every 18 months. We disrupted that trend with this chip, charting a completely new path. My assessment is that in the next 18 months, it will far exceed a 2x improvement. The new architecture still has a lot of optimization potential. GPUs have been using a 20-year-old architecture and can only rely on shrinking process nodes to sustain themselves, but the new architecture has a lot to learn from and adjust.

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

Answer: Currently, the silicon is not sitting idle due to high demand. But you're right; OpenAI is developing its own chips, and so is Amazon. Nobody likes to be dependent. The lesson learned from the era of x86 for large-scale manufacturers is not to be tied to Intel; the lesson for GPU manufacturers is not to be tied to a few large-scale customers, so they have supported the new cloud. Developing your own chips is not about being the fastest but about not relying entirely on others and at least controlling a significant part of your destiny.

Open Source and Sovereignty: What Enterprises Want is Control

Host: Open source is having a moment. I used OpenClaude early on, then switched to Kimmy and found that my Claude's token was overwhelming, but with Kimmy, I couldn't tell the difference. Open-source models have started to delve into reasoning, and this year, the gap suddenly closed.

Q: You don't want to drive a Ferrari to the grocery store. Sometimes you drive a sports car, sometimes a minivan, and you don't mind if the kids spill Cheerios. It's the same for businesses: you entrust complex problems to cutting-edge models (OpenAI, Anthropic, Gemini), but a solid open-source capability is all you need for many everyday tasks. Think about how much time a company spends copying and pasting from Workday to Excel. You don't need gold-medal math skills; reliable open source is enough.

I recently turned over another card: regulated industries like finance and healthcare (HIPAA, FINRA) are afraid of data leaks and losing control of AI sovereignty. They want to keep models on-premises and use open-source versions to retain more control. OpenAI released a $12 billion open-source model a few months ago, which was decent. But now the U.S. is leaning towards open source, either with the $12 billion model or Chinese models, as there are too few local open-source options. NVIDIA has also seen this trend and is promoting its own open-source models, but Jensen is hesitant. His clients are Sam, Dario, Elon, and Sergey. Will going open source compete with his clients?

Cerebras is in a more neutral position. We run GLMs, Kimmy, Qwen series, as well as OpenAI's closed-source models. We also run models developed by GSK, UAE's G42, and MBZUAI. Sovereignty is a trend.

AGI Has Arrived, the paradigm won't die, people will

Host: When Fable 5 and o-56 were released, the government said "pause and reassess." There was tension between Anthropic and the administration, but it's starting to ease now. Do you think staggered releases are reasonable? Is the model really that dangerous?

A: I've never seen anything like this before. But thinking about it: when a model is sufficiently creative, the government saying "please release it in stages" actually makes sense. We regulate powerful drugs the same way. Of course, we don't encourage the FDA's seven-year pile of garbage paperwork, but saying "at least let the government conduct some red team testing to confirm our defenses can hold up," giving two to three weeks to patch obvious vulnerabilities, this is not an unreasonable request.

But now is the most polarized time. If it wasn't Trump doing this, if it were any other president, the reaction might be completely different. Polarization harms clear thinking. Both sides will do stupid things and smart things. The grassroots staff in the government are actually working very diligently; it's just that things are moving too fast.

Nikesh from Palo Alto Networks told me: they tested the model on their own software and found dozens of critical vulnerabilities within an hour, forcing them to stop everything and spend six weeks patching. When you realize this is a powerful tool, maybe show it to a small group of people first, maybe do a red team test.

Host: By any definition from 20 years ago, AGI has already arrived. What do you think?

Answer: Yes. The Turing Test? It's been aced ages ago. Any definition proposed 10, 15, 20, 30, 40, 50 years ago, we have far surpassed. The questions posed by science fiction writers have all been answered; they would say, "I'm fine, thank you." That's why the words of those who seem to be on the edge are worth listening to. Ilya talked about security eight years ago, and you said, "What?" Well, turns out he was right. Elon talked about reducing rocket costs to near zero, and you said, "What?" Well, he did it.

Host: Recursive learning, you ask it a question, learn from the answer, ask again, get a better response, cover more material, these iterative answers jump from "better" to "much better." The slope of the exponential curve is too steep.

Answer: Recursive gains are exponential, you get better, ask again, keep gaining, the slope is too steep. We are just beginning to see this. Keep pouring in computational power, will the answers keep getting better? Stop after running a token or a budget, but when does this exponential curve end? Or does it keep going forever in the top-right direction? This question is extremely interesting now.

The rate of human learning is blocked by generations; elephants and large mammals take 15-20 years per generation. To be swift, you have to be like a fruit fly, two generations in a day. AI is obtaining this learning speed across thousands of generations. When I was studying psychology, a professor said: paradigms don't die; people do. Followers of Freud, Skinner, Jung held the leadership for 20-40 years before the next generation questioned. AI has compressed the intergenerational gap to fruit fly speed.

Here's my bet: our children and everyone they know won't die from cancer. The economy will shake, cars are coming, and the days are numbered for those who used to shoe horses. But listing the gains and losses: infinite energy, infinite food, infinite knowledge, infinite education, infinite housing. For a thousand years, we've known that one-on-one tutoring is better than classroom teaching; Aristotle tutored Alexander, Socrates tutored his students, but we chose factory farming-style education. Now AI can give every child a tutor to learn in their own way.

Scorsese's AI Toolbox: Turning Mental Images into Reality

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

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

Answer: Intuitive intelligence and deep reasoning are two complementary forms of intelligence. We started from intuition, with images being the most natural entry point, as the computational load is not as heavy as with videos. However, we are now converging into multimodal models. Pre-training with videos implicitly teaches the model the laws of physical interaction, allowing us to obtain action prediction from the same model, in other words, robotic control.

Interviewer: Have you collaborated with Martin Scorsese? Did you sit next to him and have him use your tool?

Answer: Yes, I was in the same room with him, where he explored our model. As one of the core researchers, I sat beside him, and that felt surreal. At the same time, I am a big fan of his.

What he wanted was to visualize the scenes in his mind, a village in Eastern Europe, as he described it, and we looked at the output, iterated. In the end, what he said was: turning the mental imagery into visual expression, this mode of communication is far more efficient than language. Language is a somewhat lossy form of communication, whereas the signal in visual information is so rich. The amount of information in a single image or video clip is enormous; this is another form of communication channel.

We don't want to dictate how these models should be used, especially not to Martin Scorsese, saying, "This is how you should use it." AI models are a medium. The most interesting things almost always happen when humans are continuously iterating in feedback loops.

From Film to Robot: The Endpoint of Generative Models is Beyond the Screen

Interviewer: Startups now use Flux and your model to release videos. Previously, it cost $250,000 to make a launch video; now it can be done in one or two weeks. Gal Gadot just made a Bitcoin movie, where the actors perform on a sound stage without a green screen, and all backgrounds are generated using AI. With a budget of $30 million, they achieved the effects that previously required $150 million. Have you seen the model being used in production?

Answer: I have seen some examples. High-end film production is one of the most demanding use cases. I am glad to see people exploring it, but I also want to make it clear: the technology is still on a trajectory and is evolving rapidly. A few years ago, when we were doing our PhD, we could only generate 64×64 images. Now we are working with multi-input high-resolution videos, but it won't stop there.

What excites me the most is this: you can use the same multimodal model to shoot a movie and then deploy it as the brain of a robot. This is so interesting. Whether a computer can actually use it is still uncertain, but the technology is moving towards the physical world, with world models, action models—basically, they are all the same thing.

Host: Where does the training data come from? Do we make humans wear glasses and gloves to record a first-person view? Or is watching a thousand videos of people pouring drinks on YouTube enough?

Answer: The goal is to instruct the robot with an in-context prompt: "Bring me that glass of orange juice." We are not there yet. The current approach is this: the model already has a lot of visual understanding; it only needs a few hours of fine-tuning data to adapt to specific hardware. The direction is to minimize fine-tuning as much as possible, rely more on in-context prompts, but this is still a research problem.

Host: Open source is having a moment; companies want sovereignty. What should a powerhouse like Disney with a vast IP library do? Should they use your open-source model for training, or collaborate with you to train an exclusive model?

Answer: The most interesting use case is generating things that did not exist before; this is fundamentally the most intriguing aspect of this technology. Our public tools cannot be used to generate specific IP, which is reasonable. We do collaborate with some IP owners to develop models, some based on our open-source models, and some based on our stronger proprietary models.

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

Host: The most fascinating phenomenon now is fan films. In the past, there was fan fiction where fans wrote their own Star Wars stories, and then people started wearing Jedi costumes to shoot fan films. George Lucas said it's allowed as long as it's not for commercial use. Now, people are using AI to reinterpret untold Star Wars stories, and "Star Wars Stories Untold" videos have millions of views each. This is the future: letting consumers pay for licenses, allowing them to use characters to create their own stories.

Answer: It would be great if we could find a viable business model for IP and also open up this super creative custom gameplay. Whenever I read a book or watch a movie, I always wonder how things would develop if done differently. Now, we can finally visualize these ideas.

We have just passed 100 employees and are hiring in Germany and San Francisco: researchers for large-scale model training, individuals with experience in diffusion and flow matching training, engineers to develop custom solutions with clients, personnel for large-scale computing infrastructure operations, and those interested in bringing the technology to more people.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia