Recording intelligenceAI-generated brief · check the source for context

Cerebras AI Build-Out in Texas

1:03:57 recording · EN · 3 speakers

Listen to the original

Listen to the episode
Executive Summary AI
  • Andrew Feldman describes a build-out without precedent: data centers that in the next several years will use more power than the previous 50 years on earth took, individual buildings the size of football fields drawing more power than midsize cities, going up across the US, Canada, the Nordics, France, the Middle East, Kazakhstan, Tajikistan, Georgia and Armenia.2:22
  • Cerebras carries a $25 billion backlog and buyers such as OpenAI, Anthropic, SpaceX AI and Google order chips before they are finished, so the industry is chasing demand already booked rather than betting that it will appear.3:50
  • On the pace of improvement, Feldman notes every processor before Cerebras followed Moore's law at roughly a doubling every 18 months, while he expects his newer architecture to be way over 2X in the next 18 months because early architectures still have room a 20-year-old GPU design does not.13:16
  • The second half turns to Robin Rombach, who started Black Forest Labs two years ago in Freiburg and San Francisco after inventing latent diffusion as a PhD student in Munich and building stable diffusion on it, and is now training multimodal models on images, video and audio that add action prediction so the same model can run on a robot.41:55
  • Rombach says production use is arriving, and Jason cites Gal Gadot's Bitcoin movie shot on a soundstage with generative scenery instead of green screens for a $30 million budget she said would otherwise have cost $150 million and never been greenlit.53:03

Brief overview

Demand for AI compute is booked years ahead, and reasoning at scale is where the value lands.

  1. Demand is booked, not speculativeFeldman's $25 billion backlog and pre-completion chip orders mean the constraint is supply of data centers and hardware, not customer appetite.
  2. Reasoning is inference, so speed compoundsReasoning burns huge internal token counts, so a 15x faster machine turns a 24-hour run into weeks or months of thinking.
  3. Open source buys control, not just costRegulated and sovereignty-minded buyers want on-prem models they own, and today the choice is OSS 120B or Chinese models.

Questions this recording answers

10 questions, each answered where it is said
How large is Cerebras's order backlog, and who else is chasing data center capacity?

Andrew Feldman says Cerebras has a $25 billion backlog and is not alone. OpenAI, Anthropic, Google, Microsoft and AWS all want more data centers. Buyers are not betting that customers will come; the demand is already booked, and they are trying to capture yesterday's demand faster than data centers can be built.

Answered around 3:50
How much power will the new AI data centers consume?

Feldman says data centers built over the next several years will use more power than the previous 50 years on earth took. Single buildings the size of football fields draw more power than midsize cities, and they are going up across the US, Canada, the Nordics, France, the Middle East, Kazakhstan, Tajikistan, Georgia and Armenia.

Answered around 2:16
Is heavy token spending creating real value or just waste?

Feldman sees massive value alongside massive experimentation, like early AWS when engineers signed up on credit cards. He compares it to shoppers learning Costco after 1988: enterprises now give highly productive people as many tokens as they need and move other work to cheaper or open-source models, running it like a business.

Answered around 4:44
Why does fast inference hardware matter for AI reasoning models?

Feldman explains that reasoning is inference and is computationally intensive, consuming huge numbers of tokens internally. Fast compute keeps it tractable instead of taking hours for a good answer. He adds that running models 15 times faster for 24 hours could yield weeks or months of thinking, and that speed makes safety guardrails less painful.

Answered around 11:37
Will Cerebras chips improve faster than Moore's law?

Feldman says earlier processors followed Moore's law, doubling about every 18 months, but Cerebras carved out a new trajectory and expects to be way over 2X in the next 18 months. A new architecture leaves room for big optimizations, while a 20-year-old architecture like the GPU relies mostly on smaller geometry and new fab nodes.

Answered around 13:10
Why are OpenAI and Amazon building their own chips?

Feldman says nobody likes being dependent. Hyperscalers learned that lesson depending on Intel in the x86 world, and GPU makers, dependent on a few hyperscalers, funded neoclouds to get more customers. Building chips is about controlling an important part of your destiny; you need not make the fastest chip, just avoid total dependence on others.

Answered around 15:20
Was a staged government rollout of a powerful new model reasonable?

Feldman says it does not seem unreasonable when a model is creative enough to pose a meaningful threat. Like powerful pharmaceuticals, though without FDA-style years of trials, the government can red-team it, check national infrastructure and ask for two or three weeks to patch obvious holes. He says political polarization hurts clear thinking about it.

Answered around 21:26
What is latent diffusion, the technique Black Forest Labs' founders invented?

Robin Rombach says he and his co-founders invented latent diffusion as PhD students in Munich. It compresses natural data such as images, video and audio into a much more efficient representation, the principle behind JPEG and MP3, and then trains a transformer on it. They built Stable Diffusion on top, and it underpins today's image and video models.

Answered around 41:55
How did Martin Scorsese react to Black Forest Labs' models?

Rombach says they explored the scenery of a village in Eastern Europe that Scorsese described, iterating on outputs. Scorsese said the value was getting the mental picture out of his head and communicating it visually through a series of images. Rombach calls language a lossy medium and sees human-in-the-loop iteration as the most interesting use.

Answered around 48:27
Is Black Forest Labs hiring, and what skills does it want?

The company just crossed 100 people and is hiring in Germany and San Francisco. It wants researchers experienced in large-scale, diffusion and flow matching training; engineers who build customized physical AI solutions or joint models with IP owners; engineers who run large-scale compute infrastructure and maximize MFU; and people who get the technology into users' hands.

Answered around 1:02:12
Key Quote

“And so, you know, we have a $25 billion backlog.”

— Andrew Feldman, Cerebras AI Build-Out in Texas, at 3:50▶ listen at 3:50
Key Quote

“If they want to run open source right now, it's OSS 120B or Chinese models.”

— Andrew Feldman, Cerebras AI Build-Out in Texas, at 18:52▶ listen at 18:52
Key Quote

“So not our children nor anyone they know dies of cancer.”

— Andrew Feldman, Cerebras AI Build-Out in Texas, at 38:25▶ listen at 38:25
Key Quote

“It was insane sitting in the same room with him multiple times”

— Robin Rombach, Cerebras AI Build-Out in Texas, at 47:59▶ listen at 47:59

Cite this

“Cerebras AI Build-Out in Texas.” https://content.libsyn.com/p/c/d/b/cdba2ab95745211c/PARIS_Versailles_Ch4.mp3. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/cerebras-ai-build-out-in-texas.html.

Report this page
What is wrong with this page?