Skip navigation
When Cerebras started, data centers were built for roughly 5 kW per rack.
0:00 / 11:39

When Cerebras started, data centers were built for roughly 5 kW per rack.

Sarah Chieng

@MilksandMatcha35K followers

Visit on X
137 likesDownloadOriginal post

24,973 views

When Cerebras started, data centers were built for roughly 5 kW per rack. A single one of our chips needed more than that. Today, AI racks from NVIDIA and AMD also draw 100–200 kW. The power and liquid cooling infrastructure we needed early on has become industry standard. I sat down with our cofounder and CTO Sean Lie to talk behind-the-scenes what it actually takes to deploy Cerebras hardware, and how we bring up larger models. 0:25 Does a dinner-plate-sized chip need a special data center? 1:48 Why liquid cooling makes denser racks possible 3:24 From delivered racks to production tokens 4:17 Things fail constantly. How do you keep serving? 5:39 What changes when you want to run a bigger model? 7:38 Can our CTO recognize our own data centers?

12 Comments

Comments from the original post on X (Twitter)

  • William's profile picture

    @718Maven

    the same CTO that sold all his shares?

    9 likes

  • Alexander Bittan's profile picture

    @alexanderbittan

    Just bought 15,000 shares! This company is under estimated big time!

    4 likes

  • Alexander Bittan's profile picture

    @alexanderbittan

    Excited to watch this !

    2 likes

  • Sarah Chieng's profile picture

    @MilksandMatcha

    thanks Alex!

    2 likes

  • Walter Hobbs's profile picture

    @WHobbs84

    Still down 69%

    2 likes

  • ar://web3's profile picture

    @VentureWeb3

    Stop pls,这样的视频发得越多,反而让FUD群体把你们的行为当作一场表演,拿出财报数据砸他们脸上。

    1 like

  • don marcos's profile picture

    @donmarcos_watts

    The underrated part is what that does to the building. 100 MW of IT at 5 kW/rack was ~20,000 racks. At 150 kW it's ~670. Same power bill, a fraction of the floor. The bottleneck moved from square footage to switchgear, chillers and the grid connection.

    1 like

  • Max Chen's profile picture

    @0xMaxChen

    One of best PR teams in the industry. shipping videos more than another other large 10b+ listco I believe

  • Soumil Bhandari's profile picture

    @SoumilBhandari

    Curious about the 5:39 part. With weights living in on-wafer SRAM (44 GB on WSE-3), precision seems to directly set how many wafers a model needs: FP8 vs BF16 roughly halves it. What precision do you serve the big models at today?

  • Derrick Dao's profile picture

    @derrick_dao

    5 kW to 100-200 kW a rack is a 20-40x jump in density, and the grid never moved with it. That's why the real shortage stopped being silicon and became substations — copper, transformers, and the megawatts to feed them. The chips were the easy part of this buildout.

  • Willie's profile picture

    @Dreaming09092

    a "good" CTO, pitty that he hate your company by dumping all his shares👎

  • terminüll's profile picture

    @terminulll

    5 kw used to be the entire rack budget. now it's a rounding error.