CEOInterviews.AI
Start App
Dylan Patel
Founder, CEO, and Chief Analyst, SemiAnalysis

Cerebras Supernova: Dylan Patel (Semianalysis) on fast inference

📅 Sep 08, 2026 Cerebras 12 MIN 308 VIEWS 46 SEGMENTS · 2 SPEAKERS
Not all AI chips are chasing the same goal. Some are built for the cheapest possible tokens; others are built to generate tokens as fast as they can. At Cerebras Supernova, Dylan Patel breaks down the new inference landscape, including the class of hardware optimized purely for speed. He explains why Cerebras wafers excel at ultra-fast inference, how much performance is still hiding in software, why AI models end up shaped by the hardware they run on, and whether he’s still bullish on disaggregated inference. 0:00 Why expensive teams need faster inference 0:31 From Xbox repair to chip anal...

What Dylan Patel said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Dylan Patel discussed the chip industry, starting with his personal history and moving to the current landscape. He praised Etched for shipping hardware quickly, noting their success despite lacking initial chip expertise. He categorized hardware into 'yapper' (fast inference) and cheap token types, with Cerebras and Groq as examples. He argued that companies should use fast inference for expensive employees and highlighted the use of Claude Code by non-developers. He explained software optimizations like speculative decoding and predicted a 50-80% performance improvement on existing hardware. He discussed how model architectures, like sparse mixture-of-experts, are co-designed for Nvidia GPUs, potentially hurting alternatives like TPUs and Trainium. He was bullish on disaggregation partnerships, such as Nvidia-Groq and Cerebras-AMD, while noting risks from model architecture shifts.

Key takeaways

  1. Predicts 50-80% performance improvement still available on existing hardware and models.
  2. Categorizes hardware as 'yapper' for fast inference, citing Cerebras and Groq as examples.
  3. Believes companies should use fast inference for employees earning $500k or more.
  4. Praises Etched for shipping hardware in 'a couple of years' despite initial lack of expertise.

Numbers and commitments

FigureWhat it refers toTypeAt
500k or a million Annual salary threshold for using fast mode guidance 0:09
50 to 80% Potential performance improvement on existing hardware guidance 7:33
700 experts Number of experts in DeepSeek model metric 7:58

Chapters

  1. 0:00Origin story: Xbox and chips
  2. 0:58Disaggregation partnerships and risks
  3. 1:42Etched's rapid shipping and hardware bets
  4. 2:35Yapper vs cheap token hardware
  5. 3:49Fast inference for expensive employees
  6. 5:14Software optimizations and performance headroom
  7. 7:58Model-hardware co-design and sparse MoE

Questions asked in this interview

11
  1. 0:42Like, why did you care about chips in the first place?
  2. 1:21Do you think they stand a chance against the incumbents like Nvidia?
  3. 1:50How do you think Etched was able to have so much like monetarily and like pre-commitments before they launched anything?
  4. 3:30But are there other like non-coding use cases and industries that you're seeing like really start to invest in AI or speed and compute?
  5. 4:16Well that's what they should do, but are you seeing any industries actually like make these big bets and do it successfully?
  6. 4:57And two, like how well does this scale, especially in terms of like user cost?
  7. 6:22Are you going to reach like a ceiling?
  8. 7:31What do you think is like the ceiling?
  9. 7:45What do you think are some of the ways that we've either designed our models for our hardware or even ways that we've nerfed our models for our hardware?
  10. 9:08Does that make it hard for other architectures or other hardware to succeed?
  11. 10:53Are you still bullish on this partnership?
Dylan Patel 0:00 ↗
I think any company that like pays their people over a certain amount of money should just use ultra fast mode because their people are like—
Interviewer 0:08 ↗
Expensive.
Dylan Patel 0:09 ↗
They're expensive and it's like, you know, if your person's making, you know, 500k or a million or more, you should have them use fast mode.
Interviewer 0:17 ↗
And even in many cases, if they make less, you should have them use like a fast inference cuz their time's expensive and like yeah, iterating is super fast.
Dylan, we are so excited to have you here today at Cerebras Supernova.
Dylan Patel 0:35 ↗
Hi, my name is Dylan. I guess I'm well known for talking about chips, as boring as that is—
Interviewer 0:42 ↗
Or exciting. How did all of this start? Like, why did you care about chips in the first place?
Dylan Patel 0:45 ↗
My Xbox broke when I was eight. I had to fix it. Had to open it up and fix it. It's called the Red Ring of Death. And then I stayed on the forums. And by the time I was 12, I was like moderating a bunch of forums. Basically, that's where it all started. And I had a bunch of other passions. You know, I get really obsessed with things. Like, I'm obsessed with your boots. Like, holy shit those are sick. Sorry, I had to compliment. And so, you know, I get obsessed with things. There was a little bit where I was obsessed with other things besides chips, but then I came back to it. Or rather, it was just like in the background always as a hobby. And then I made it into a job in 2020, I guess, when I wasn't really sure what I wanted to do. You know, my professional career was like was good. It was just like soul sucking.
Interviewer 1:21 ↗
Yeah. There are more and more different players in the chip space. A lot of them are very early and haven't shipped anything. But this morning we saw Etched announced that they shipped their first batch of hardware for Jane Street. And so in terms of all of these new players coming up, like what are the different bets that you're seeing people take? Do you think they stand a chance against the incumbents like Nvidia?
Dylan Patel 1:42 ↗
I think it's super exciting that Etched was able to go from like nothing to shipping a rack—
Interviewer 1:47 ↗
Yes.
Dylan Patel 1:48 ↗
—in like just a couple of years. And so that's like super sick.
Interviewer 1:50 ↗
How do you think Etched was able to have so much like monetarily and like pre-commitments before they launched anything?
Dylan Patel 1:56 ↗
It's hard to say how many orders they actually have cuz like how real are those contracts? But ultimately, I think like a lot of people hated on them cuz they didn't have chip design expertise and it's like—
Interviewer 2:06 ↗
They're young.
Dylan Patel 2:07 ↗
Yeah. But then instead they just like yo—I mean they hired people that did, but like they YOLOed the crap out of it and that's why they were able to get to successfully shipping like really, really quickly. There are other companies who started in the same timeframe and they are not shipping a chip yet, which is interesting. But generally, I think there's like a couple different bets. One is yappers, and the other one is like—there's like two types of hardware. There's like, "I want to make cheap enough, cheap tokens," and then there's, "I want to be a yapper." And the reason I say yapper is basically like Cerebras is a yapper.
Interviewer 2:35 ↗
Why? What does that mean?
Dylan Patel 2:36 ↗
Like think about someone who just talks really fast. They're a yapper. So Cerebras wafers are really, really good at super fast inference.
Interviewer 2:43 ↗
Yes.
Dylan Patel 2:43 ↗
So I classify that as like yapper hardware.
Interviewer 2:45 ↗
Okay.
Dylan Patel 2:45 ↗
This is patented by the way. But like, you know, there's other hardware out there like Groq that Nvidia bought that is yapper hardware. And like on the other spectrum, there's different ranges. Like if you want to do like efficient inference at like low speed or if you want to do yappers. And so I think different people have different approaches and also different targets. Some people are trying to be the fastest token spot possible. Some are trying to be the cheapest. Some are trying to be like some optimal on that Pareto curve. But generally, most people see that Nvidia is really good at throughput and they're mostly not targeting low, you know, low interactivity throughput cuz you can't get that much better than Nvidia there. Ultimately, you can be better, but you can't get that much better on a refined basis, but you can definitely get a lot better on yapping.
Interviewer 3:30 ↗
That's true. So obviously with Etched you see them go out with Jane Street. OpenAI last week for Ultra Fast also went out with Jane Street as our marquee customer. So obviously high-frequency trading is a trend there. But are there other like non-coding use cases and industries that you're seeing like really start to invest in AI or speed and compute?

23 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Patel, D. (2026, September 8). Cerebras Supernova: Dylan Patel (Semianalysis) on fast inference [Interview transcript]. Cerebras. CEOInterviews.AI. https://ceointerviews.ai/interview/1324705/

MLA

Dylan Patel. "Cerebras Supernova: Dylan Patel (Semianalysis) on fast inference." Cerebras, 8 Sep. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/1324705/.

BibTeX
@misc{patel2026_1324705,
  author       = {Dylan Patel},
  title        = {Cerebras Supernova: Dylan Patel (Semianalysis) on fast inference},
  howpublished = {Interview transcript, Cerebras. CEOInterviews.AI},
  year         = {2026},
  month        = {sep},
  url          = {https://ceointerviews.ai/interview/1324705/},
  note         = {Speaker-attributed transcript with timestamps}
}