At @Cerebras, we’re aiming to double AI speed every year.
To achieve this, we build on key innovations including Nexus - our new rack-scale architecture.
The CS-4 is the first system built on it.
The CS-5 and the CS-6 will use it too.
Delivering fast AI starts with fast compute tightly coupled to SRAM.
Nothing else gets you there.
But you also need I/O and power.
In past systems, those were treated as an intertwined collection of parts.
Improve one and you had to re-engineer the whole system.
Nexus is different.
Power, compute, and I/O are decoupled.
Each is a self-contained module that can be developed independently.
Power sits at the front of the rack as modular power supplies.
When the next generation of supplies comes out, plug them in.
Compute sits at the back as three pluggable backpacks, each housing one Wafer-Scale Engine. The next generation of compute comes out, upgrades are dead simple.
I/O is its own silicon, connecting at the edge of the wafer.
Because each part is built as a separate module, an advance in one no longer requires a redesign of the others. Faster I/O can be plugged in as the industry advances.
Innovation in power, compute, or I/O can reach customers as soon as they are ready.
Nexus was built for scaling AI over generations and as the industry rapidly evolves.
It will support multiple generations of the fastest inference in the industry.
All on the same architecture.
Cerebras today announced financial results for Q3 2026, delivering record revenue, and raising full year outlook.
Q3 2026 highlights:
• Record Revenue: GAAP cloud revenue grew 265%; Core cloud revenue grew 272%
• 700 MW of data center capacity now under contract
• Commercial wins with 2 new hyperscalers
• Day-zero launch partner for @Kimi_Moonshot K3.1, running at up to 2,000 tps
• Raised Q4 guidance and full year outlook across the board
Today, @cerebras announced financial results for Q3 2026, delivering record revenue, and raising full year outlook.
Q3 2026 highlights:
• Record Revenue: GAAP cloud revenue grew 265%; Core cloud revenue grew 272%
• 700 MW of data center capacity now under contract
• Commercial wins with 2 new hyperscalers
• Day-zero launch partner for @Kimi_Moonshot K3.1, running at up to 2,000 tps
• Raised Q4 guidance and full year outlook across the board
Moonshot just dropped Kimi K3.1 Pro. We used it to speed run online shopping.
One prompt → monitor found → checkout → order confirmed.
11.5 seconds.
Watch @Kimi_Moonshot K3.1 Pro in action, accelerated by Cerebras.
Last week, a short seller, @CulperResearch, published a report arguing that @cerebras wafers wear out after about two years, and based on that premise the report suggested that the economics of our business are fundamentally challenged.
That number is wrong, and given that the basis of the entire report is built upon that assumption, all of the assertions that follow are also wrong.
We do agree that the question of how long our systems last is a fair question, so we want to address the topic plainly.
If you have a few minutes, I would encourage you to click on the link below, which will provide you with a better understanding of the pioneering work done by our engineering team over the past decade to ensure that the operating lifetime of our solutions is many, many years.
Link Here: https://t.co/mGKI3dSHiw
How we deploy multi-trillion parameter models at scale
Today, let's break down the math for how @cerebras:
1. Fits the full model into on-chip memory
2. Maps the model across multiple chips
3. Handles a high # of concurrent requests
The fastest inference in the world is coming to @OracleCloud.
@cerebras and @Oracle partner to enable blisteringly fast inference on OCI.
Cerebras brings fast AI.
Oracle brings deep expertise building and running AI infrastructure at scale.
Enterprises have trusted Oracle with their most important data and applications for decades.
Partnering with Cerebras gives Oracle customers more choice to match each AI workload to the compute that suits it best.
Coming soon, they’ll be able to build agents with true super fast AI in the cloud they already use.
Thank you to @MThiagarajan_ and the @Oracle team.
Excited to build this with you.
Full announcement here: https://t.co/wACT9YiG7F
Today’s guest is @achterbrain , co-founder and CTO of @CallosumAI .
Jascha’s bet is that the future of compute is heterogeneous: different chips working together on different parts of a problem.
Beyond the multichip future, we discuss:
- What neuroscience taught Jascha about designing computing systems
- How chipmakers discover what their hardware is best at
- Where thermodynamic, biological, and quantum computing could fit
- Hiring people who welcome having their ideas challenged
- Building an AI infrastructure company in London
Timestamps:
0:00 Introduction
2:55 What heterogeneous computing actually means
11:02 From PhD research to Callosum
18:36 Turning heterogeneous compute into a product
23:24 Sourcing compute and working with chipmakers
28:53 Matching workloads to the right hardware
34:24 What heterogeneous computing can do today
38:29 The next waves of chips
48:02 Building the team
51:33 Why build in London?
56:11 Helping enterprises upgrade their compute stacks
1:01:04 Fundraising and building momentum
1:03:40 Where specialization wins, and where it doesn’t
Listen to the show, and follow along for new episodes every Tuesday.
A very thoughtful explanation of the benefits of disaggregated inference. At @cerebras we deliver the fastest AI in the industry and have pioneered disaggregation.
There is some speculation about our partnership with Cerebras.
Cerebras is a close partner, and we have a deep engagement pushing on the frontiers of speed.
I joined AMD right after graduating from MIT. Two stints there and a decade building @cerebras later, I still lose sleep over how much memory and compute to put on a chip.
I’ve spent my entire career in computer architecture, and I’ve never seen the problems change this quickly. Models keep getting larger. The demands on the hardware keep growing. Decisions about where to put memory and how to connect it have become central to what people can build with AI.
We chose SRAM for its bandwidth. Now we’re working on wafer-scale stacked DRAM to give larger models more capacity while preserving fast inference. That brings us back to some very physical problems: how to power it, cool it, and manufacture it reliably.
This is the work I love. In part two of our AMA, I answered your questions about those decisions and where we’re going next.
0:30 What changes between CS4, CS5, and CS6?
2:19 HBM already stacks memory. What’s different here?
3:34 We chose SRAM for speed. Why add DRAM?
4:52 Drawing the stack is easy. Making it work isn’t.
7:23 The memory-versus-compute decision that keeps me up at night
Very interesting post by @mschoening of @NotionHQ .
In some domains AI speed even trumps intelligence, when the intelligence is good enough. Speed makes products better.
Is 500 tokens per second per user the new slow?
Great conversation with @jaltma and @ericvishria
Eric and Benchmark have been an excellent partner to @cerebras since the very beginning.
We get into the near-death years and much more.
Worth a listen.
Speed makes AI productive and fun.
@Alibaba_Qwen 3.8-27B is now live on @cerebras Inference Cloud.
Up to 16x faster than GPUs.
Thank you to the @Alibaba_Qwen and @cerebras teams for bringing this to life.
Build with it today: https://t.co/L6AWKTd6OB
What excites me most about Jalapeño is the design methodology behind it. Using AI to help build chips faster and push both throughput and latency forward could benefit the entire industry.
I spoke with @swyx on @latentspacepod about why I see this as the future of chip design.
When we started @cerebras, we knew we would have to solve difficult problems in yield, power, and cooling. What became clear was that making wafer-scale work would require rethinking almost the entire system.
Today, we answered questions from the community about those engineering decisions, how we run larger models, and what it takes to program the hardware.
2:15 Why hasn't NVIDIA gone wafer-scale?
4:10 How we build a working chip out of an imperfect wafer
8:07 What happens when a model won't fit on one wafer?
11:34 Why we chose SRAM over HBM
14:26 How much of a barrier is NVIDIA's CUDA moat?