Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit with CEOs of Cerebras & Black Forest Labs

Why does it matter? Because the AI race has already lapped its own definition of AGI — and the real bottleneck is concrete, not code

Two founders with direct lines into the actual buildout — Cerebras CEO Andrew Feldman and Black Forest Labs CEO Robin Rombach — offer a picture that diverges sharply from the headline debate. Supply is the actual bottleneck, not demand. AGI crossed its finish line without a press release. And the architecture generating film and the architecture controlling robots are converging into one system.

• Cerebras holds a $25 billion backlog — demand was booked before the chips were even finished • AGI is already here by every prior definition; the race now is toward superintelligence • 15x faster inference turns 24 hours of compute into weeks of equivalent reasoning • The architecture powering video generation today will power robotics tomorrow — it's already one system

The biggest AI companies aren't chasing future demand — they're already years behind orders already on the books

Supply is the binding constraint on AI progress — not model quality, not market adoption. "They're trying to capture yesterday's demand," Feldman says. "The demand is way outstripping our ability to build data centers and fill them with hardware." Cerebras alone holds a $25 billion backlog. Customers placed chip orders before production finished. OpenAI, Anthropic, Google, Microsoft, AWS — none are building speculatively. "The demand is booked."

The physical scale defies easy intuition: buildings the size of football fields drawing more power than midsize cities, going up across the U.S., Europe, the Middle East, Kazakhstan, Armenia. Collectively, these facilities will consume more energy in coming years than Earth used over the previous 50 years combined. The variable choking AI progress is infrastructure, not intelligence.

AGI has arrived by every definition we once used — and we've run out of the questions we used to ask

Nobody announced it. Feldman doesn't hedge: "AGI — I think I suspect you'll agree with me that we've hit it. We just haven't exactly deployed it fully." By every prior benchmark: the Turing test — "blew it away." The milestones science fiction once imagined — "we've blown past it." If those writers looked at current systems today, Feldman says, "they'd be like, 'I'm out of questions.'"

The debate about when AGI would arrive is now closed. What remains is the harder problem: what superintelligence demands of regulation, infrastructure, and human organizational design. We built toward a finish line carefully defined over decades. We crossed it. The questions that actually matter now are ones we haven't yet learned to formulate.

Unlimited tokens means unlimited reasoning — and 15x faster inference compresses weeks of thinking into a single day

Unlimited tokens, unlimited reasoning — Feldman confirms it. The part worth internalizing: "if by using Cerebras we were 15 times faster and then you ran it for 24 hours, you got weeks or months worth of thinking." Not quicker answers. A different class of answers.

"If you run these for 25 or 48 hours, you get amazing things now." The mechanism: reasoning models consume enormous tokens internally during deliberation. That burn rate is exactly what a fast chip was designed to serve. "It is exactly the fact that this reasoning consumes a huge amount of tokens internally that allows a blisteringly fast machine like ours."

Speed here isn't a cost variable. It's the lever determining reasoning depth at any fixed time budget.

The model generating film and the model running inside a robot are the same thing — creative AI and physical AI have already merged

The same AI model that generates film can serve as a robot's brain. Most investors still price image generation and robotics as separate categories — Rombach's framework says they're already one.

"Pre-training on videos gives implicit understanding of the physics of interactions with the real world — and then you can get stuff like action prediction, like robotics, out of the same model." The underlying logic: "In order to make a video of the world, you have to understand the world." That understanding — objects, motion, cause and effect — transfers directly to action prediction.

Black Forest Labs is now entering "a new paradigm" combining image, video, audio, and action prediction in a single architecture. The creative content upside and the robotics upside are one bet, placed once.

Open source has closed the gap for most enterprise work — and sovereignty fears are cementing a permanent two-tier market

Open source has already closed the capability gap for most of what enterprises actually do. "You don't want to take your Ferrari to the grocery store," Feldman says — frontier models for genuinely hard problems, open source for everything else. "The cutting and pasting economy is real. This doesn't need gold medal math. What this needs is rock solid open-source capabilities."

Two forces are cementing a two-tier split: cost discipline (routing tasks by complexity rather than defaulting to the most powerful model) and sovereignty (regulated industries in finance, healthcare, HIPAA, FINRA demanding on-premise control). "If they want to run open source right now, it's OSS 12B or Chinese models." That's a strategic gap the U.S. hasn't addressed. "Sovereignty is a trend," Feldman says — and Europe's cautious reaction to recent open-source releases confirms it's become geopolitical.

A $30M Bitcoin film would have cost $150M and never been greenlit — the threshold for what gets made has permanently shifted

$30 million. That's what Gal Gadot's Bitcoin film cost — shot on a sound stage with no green screens, all the scenery generated by AI behind the actors. Without it, building those sets runs $150 million. "The film would have never been greenlit." She told Jason at Yuri Milner's Breakthrough Prize.

The story isn't cost reduction. It's expansion. AI isn't cutting costs on films that would have existed anyway — it's pulling previously impossible projects across the economic threshold. The universe of viable films has grown, not just gotten cheaper.

Rombach confirms production-level deployments are happening. The trajectory moves one way: from 64x64 pixel images when he was a PhD student a few years ago to multi-minute high-resolution video today.

Scorsese uses AI image generation to get the picture out of his head — because language is too lossy a medium for a filmmaker's vision

The highest near-term return on AI visual tools isn't autonomous filmmaking. It's pre-production. Rombach sat in a room with Scorsese multiple times — "insane," he says — watching him explore imagery of an Eastern European village for a project in development. What Scorsese articulated: "getting the mental picture of something out of your head and communicating it in a visual way... just makes it easier to communicate what is actually in your head."

Rombach's frame: "Language ultimately is like a lossy communication medium. Visual information is so rich. There's so much signal in it."

The ROI is immediate — no autonomous filmmaking problem needs to be solved first. The gap between imagination and what a crew can execute closes right now, in pre-production.

Cerebras claims to have broken Moore's Law — and a 20-year-old GPU architecture can't match its improvement trajectory

Every chip before Cerebras followed Moore's Law: roughly 2x performance every 18 months. Feldman says they've already beaten that curve — and the next 18 months will bring "way over 2x."

The structural argument is clean. GPUs are a 20-year-old architecture; squeezing more performance out means moving to smaller fab nodes, incremental gains on a diminishing curve. "In a newer architecture, you have a huge amount of room still to learn about the work being presented and make optimizations that give you huge gains."

Nvidia's current inference dominance is real. Whether it's structurally permanent — especially for the token-intensive, long-duration reasoning workloads where Cerebras specifically competes — is the more interesting question.

Whoever controls inference capacity is about to control both the intelligence layer and the physical world

Feldman's $25 billion backlog and Rombach's unified model thesis point at the same pressure point: compute access is no longer just an AI abstraction. If the same architecture generating content also runs robots — and if demand for both already outstrips the world's ability to build infrastructure — then inference capacity is physical world capacity. The sovereignty conversation around open-source models is about to get significantly more consequential than its current framing suggests.

The race for superintelligence runs through a data center. Whoever builds it first is in a different business than they currently realize.

---

Topics: AI infrastructure, inference chips, Cerebras, AGI, open source AI, Black Forest Labs, generative video, robotics, AI regulation, film production, model sovereignty, Moore's Law, reasoning models, Martin Scorsese