Real Alpha
Try Pro free

Uncapped · 3 min · Sep 15, 2026

Faster inference can open new markets and increase throughput

Andrew Feldman argues faster inference can unlock applications that slow systems cannot support, and says disaggregating inference tasks boosts throughput. He reports 5x throughput with AMD and similar results with AWS.

Andrew Feldman

Andrew Feldman is cofounder and CEO of Cerebras Systems.

Co-founded and leads Cerebras Systems, whose wafer-scale engine is the largest processor in computing history.

  • Jack Altman, Host · General partner, Benchmark

    Co-founded Lattice and ran it until 2023, then started his fund, Alt Capital.

About this segment

Andrew Feldman argues faster inference can unlock applications that slow systems cannot support, and says disaggregating inference tasks boosts throughput. He reports 5x throughput with AMD and similar results with AWS.

From “Andrew Feldman on Building Cerebras and the Future of Chips | Ep. 57”, Sep 15, 2026. Watch the full episode

“And, you know, with AMD, we're seeing 5x additional throughput, I mean, five times as much throughput, while keeping the speed the same.”
Andrew Feldmanat 54:31 · Uncapped54:31

Takeaways

  1. Speed can create uses that slower performance cannot support.
  2. Cerebras splits inference work with partners such as AMD and AWS.
  3. The AMD collaboration delivers 5x throughput while keeping speed the same.

In this episode

Full episode
  1. 4:09Attacking Goliath requires a radically better productPro pick 5:39
  2. 9:37Failure analysis kept Cerebras going through 18 months of setbacks3:29
  3. 16:13Chip fabs are a rare, tightly coupled supply chain6:53
  4. 43:57Data center delivery depends on power gear and site readiness6:05

  5. 51:57Faster inference can open new markets and increase throughput3:26

  6. 55:12Nvidia's long fight as a public company shaped its grit2:25Up next
  7. 29:00Cerebras made architectural choices to adapt as AI changed4:15
  8. 12:55Cerebras is developing memory and optical links alongside compute3:07

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 52:04

    Constrained in some form or another whether it's memory or chips or data center space, whatever there's some constraint there and the demand is is overwhelming doesn't that argue then the only thing that matters is Like tokens per watt like basically we're gonna have a limited number of denominator and so however much whether it's tokens per watt or tokens per facility or tokens per chip or I don't know something like that but it feels like we need the most of the numerator for the limited denominator task per watt maybe and then so like if that's the case then you have cerebrus has a speed advantage a huge speed advantage but like how do you think about that dynamic if we fast forward in terms of what the constraint actually is?

  2. 52:51

    I think there are a couple things. First, speed, which is what we've chosen to focus on, has historically created markets, right? And if you think about it, right, there's no market for slow search, right? There's no market for dial-up. Your kids are, what, 14 and 12? If you want to punish them, right, don't take away their phone. ratchet it back to dial-up speed right this is a real punishment let him use it for a week at dial-up speed slowly right that why is we laugh and then we say it's okay for for ai to be slow i think think about it i mean when when the internet was slow netflix delivered dvds and envelopes yeah and when the internet got fast they became a movie studio all right that's not They didn't get better at their other thing.

  3. 53:46

    They became something entirely new. And I think what we're seeing with the launch of GPT-5-6 Sol, right, it's put out in limited availability last week, that people are thinking of whole new applications. You've got frontier intelligence instantly. And that opens up all sorts of new opportunities. So that's one thing you do with speed. The other thing you do is you try and think about how you can drive up throughput, how you can make more tokens.

  4. 54:13

    And one of the ways we're doing that is by partnering in something called disaggregation. And we're doing it with AMD, we're doing it with AWS, where you think about in the work of inference, is there a part that can be done by somebody else, and is there a part that can be done by you such that the result is higher throughput? And, you know, with AMD, we're seeing 5x additional throughput, I mean, five times as much throughput, while keeping the speed the same.

  5. 54:47

    And we're seeing similar numbers with AWS. We have an opportunity, because of our architecture, to do that with the entire GPU landscape. We could do it across the board. And so, you know, there are four major chip makers right now in our category. Obviously, NVIDIA, we'd love to partner with them. There's AMD, there's the Google TPU, and there's AWS with their training parts, and we're already working with two.

  6. 55:15

    So that's something we've thought a great deal about,