Cerebras Reveals $25 Billion Backlog and Says It's Breaking Moore's Law by More Than 2x Every 18 Months
All-In Podcast interview with Cerebras CEO Andrew Feldman and Black Forest Labs CEO Robin Rombach, published this week
Cerebras Systems founder and CEO Andrew Feldman disclosed that the inference chipmaker is now sitting on a $25 billion order backlog, with hyperscalers and frontier labs booking capacity years before Cerebras finishes building the chips. "The demand is way outstripping our ability to build data centers and to fill them with hardware," Feldman said, describing a dynamic where OpenAI, Anthropic, Google, Microsoft and AWS are not speculatively building infrastructure but chasing demand that is already booked. The comment reframes the current AI infrastructure cycle less as a bet on future usage and more as a scramble to catch up with existing customer commitments, a distinction that matters for anyone modeling capex durability across the hyperscaler and neocloud landscape.
Breaking Moore's Law: Cerebras Claims Its Own Performance Curve
The most concrete technical disclosure in the conversation was Feldman's claim that Cerebras has moved past the traditional Moore's Law cadence of doubling performance every 18 months. "We crushed it with this chip and we've carved out a whole new trajectory," Feldman said, adding that his internal expectation is that Cerebras will be "way over 2x" on that same 18-month timeframe going forward. He attributed the gap to architectural maturity: GPUs, built on a 20-year-old design, are increasingly dependent on shrinking to the next fab node for gains, while Cerebras' wafer-scale architecture is young enough that there is still substantial headroom to optimize independent of process node improvements. This is a claim investors should treat as a company assertion rather than an audited benchmark, but it is a specific, falsifiable one that the market can track against future product launches.
Why Inference Speed Is Becoming the Bottleneck, Not Just Training
Feldman connected Cerebras' speed advantage directly to the economics of reasoning models, which consume large volumes of tokens internally before producing an answer. "It's exactly the fact that this reasoning consumes a huge amount of tokens internally that allows a blisteringly fast machine like ours" to matter, he said. The argument is that as reasoning chains lengthen — Feldman referenced runs stretching 24 to 48 hours — a 15x speed advantage compounds into weeks or months of effective "thinking" time compressed into a single day. This is the clearest articulation yet of why Cerebras believes its hardware advantage widens rather than narrows as the industry shifts from single-pass inference to multi-step agentic reasoning.
Token Maxing, Enterprise Discipline, and the Real Signal Inside AI Spend
Feldman pushed back gently on the idea that current AI usage is undisciplined speculation, drawing a direct parallel to the early days of AWS adoption inside enterprises. "For sure there's experimentation, but it doesn't mean that the net value isn't enormous. It means some of it is going to go nowhere," he said, comparing the current phase to shoppers wandering every aisle of a newly opened Costco before learning to shop strategically. He argued that enterprises are now moving past unconstrained "token maxing" toward disciplined allocation — using cheaper open-source models for lower-stakes tasks and reserving frontier models for hard problems — which he sees as evidence of a maturing, not overheating, market.
Open Source and Sovereignty Emerge as a Structural Business Line for Cerebras
Feldman was direct about a shift in customer demand toward open-source and sovereign deployment options, driven by data leakage concerns and regulatory exposure in sectors like finance and healthcare. "In the US we need more domestic open-source models. We need to give the world a choice. If they want to run open source right now, it's OSS 120B or Chinese models," he said, referencing OpenAI's open-weight release as a step in the right direction but an insufficient one. Cerebras' own positioning benefits from this trend: the company said it currently runs GLM, Kimi, and the Qwen model family alongside OpenAI's closed models and custom models built by clients including GSK and UAE-based G42 and MBZUAI. Feldman noted that Nvidia has largely avoided pushing its own open-source models aggressively because doing so would put it in direct competition with the same labs — OpenAI, Anthropic, xAI — that buy its chips, a dynamic that leaves room for a more neutral infrastructure player like Cerebras to serve that demand.
On Anthropic's Government Disclosure Episode, Feldman Sides With Caution
Asked about the controversy surrounding Anthropic's coordination with the government on a model rollout, Feldman was measured but supportive of the precautionary approach, independent of politics. "At a time that a model is sufficiently creative in its thinking that it poses a meaningful threat, for the government to say we'd like you to roll it out in steps — this doesn't seem unreasonable to me," he said, comparing it to phased safety review processes used for pharmaceuticals. He cited a conversation with Palo Alto Networks' Nikesh Arora, who reportedly told him the model "killed" existing security software defenses, forcing a six-week patch cycle. Feldman's broader point — that political polarization is degrading the industry's ability to reason clearly about legitimate safety tradeoffs — is a notable data point for investors trying to gauge regulatory risk heading into the next generation of frontier releases.
Feldman Declares AGI Already Achieved, Reframes the Debate Toward Deployment
In one of the more quotable moments, Feldman stated plainly that artificial general intelligence has already arrived, at least by any standard that would have been used to define it a decade or two ago. "We've hit it. We just haven't exactly deployed it fully," he said, arguing that older benchmarks like the Turing test have been "blown away." This is a notable escalation from a chip executive rather than a lab CEO, and it signals that infrastructure providers — who have visibility into actual model capability through their customers' workloads — are now comfortable making AGI claims publicly, shifting the debate from "have we arrived" to "how fast can it be distributed and organized around."
Black Forest Labs: Robin Rombach on the Convergence of Image, Video, and Robotics Models
The conversation's second half turned to Black Forest Labs co-founder and CEO Robin Rombach, whose company is now over 100 employees and has raised a fresh funding round following its work on the open-source Flux model. Rombach, who previously helped build Stable Diffusion and invented the latent diffusion algorithm underlying most modern generative image and video systems, laid out a roadmap that extends well beyond content generation into robotics. "We are now like entering a new paradigm which is combining that with something that's called action prediction, such that you can actually use the same model to make images, to make videos, to make audio, and to predict actions — which means you can ultimately deploy it on a robot in the real world," he said. This is the most consequential disclosure from Black Forest Labs in the interview: the company is explicitly positioning its multimodal architecture as a bridge into physical AI and robotics, not just media generation, a strategic direction that broadens its addressable market well beyond creative tooling.
The Scorsese Partnership: What It Actually Involves
Rombach confirmed and detailed a collaboration with director Martin Scorsese, describing sessions where Scorsese used Black Forest Labs' models to visualize a scene — reportedly a village in Eastern Europe — for a potential future project. "Getting the mental picture of something out of your head and communicating it in a visual way, by making these images or a series of images, is something that just makes it easier to communicate and convey an idea of what is actually in your head," Rombach said, relaying Scorsese's own framing. Importantly, Rombach was careful to temper expectations that generative video is close to producing finished, director-quality feature films. "I'm not sure if that is the ultimate goal," he said of full AI-generated movies, arguing instead that the near-term value is concentrated in pre-production ideation, storyboarding, and iterative human-in-the-loop workflows rather than replacing production entirely. That is a meaningfully more conservative framing than some of the more promotional narratives circulating around generative video, and it is a useful check for investors trying to size the near-term commercial opportunity versus the long-term one.
IP Licensing Strategy: Black Forest Labs Positions Itself as Neutral Infrastructure
On the question of how major intellectual property holders like Disney should approach generative tools, Rombach said Black Forest Labs already blocks generation of certain copyrighted characters on its public-facing tools and separately develops custom models directly with IP holders, built on either open-source or proprietary architectures depending on the client's needs. He did not name specific studio partners beyond confirming the Scorsese collaboration, but the framing suggests the company is trying to build a licensing and custom-model business alongside its open-source consumer tools, a dual-track strategy similar to what Cerebras described with its sovereign and enterprise deployments. Rombach also pointed to the rise of AI-assisted fan films, such as unofficial Star Wars content drawing millions of YouTube views, as a preview of how studios might eventually monetize fan creativity through licensing rather than blanket restriction.
Where the Technology Still Falls Short
Rombach was candid about current limitations, noting that high-end film production remains one of the most demanding use cases and that fully autonomous, prompt-driven robot control is not yet achievable. "You would want to go to a place where you could prompt a robot in context, as you can do with a language model — we're not there yet," he said, explaining that current deployments require several hours of fine-tuning data per robot and task rather than the kind of zero-shot generalization the industry is aiming for. That gap between ambition and current capability is a useful counterweight to the more aggressive robotics narratives being pushed elsewhere in the market, and it suggests actual commercial deployment of these multimodal world models into physical robotics is still an earlier-stage bet than the image and video generation business, which is already producing revenue through licensing and enterprise customization work.