Cerebras CS-4 Delivers A 30x Uplift In AI This Year, But Next-Gen Racks Already In Works With CS-5 Pumping Out 10K TPS Per User In 2027 & CS-6 Bringing 3D Wafer-Scale SRAM

Cerebras is rolling out its CS-4 AI rack-scale solution this year but is also working on its next-gen CS-5 and CS6 solutions.

Cerebras Takes Wafer Scale Engine To The Next-Level With CS-4, CS-5, and CS-6 AI Racks

Last week, Cerebras took the curtains off its CS-4 rack-scale solution, which is powered by the WSE-3T chip. The WSE-3T is a boosted version of the WSE-3 (Wafer Scale Engine), offering much higher capabilities.

At Hot Chips 2026, Cerebras is providing a deeper dive into its rack-scale solutions while also giving us a look at its next-gen solutions.

So starting off with the Cerebras CS-4, it offers up to a 2x uplift in token generation speed and up to a 10x uplift in throughput per watt versus Cerebras's CS-3 solution. Being much larger than any GPU or compute accelerator due to its wafer-scale design, the current-gen WS-3 is already much ahead of the aforementioned solutions in Token Generation speed. With CS-4, that lead is taken up to 30x across various models.

One of the advantages of the wafer-scale design is that Cerebras has a lot of room to expand. The chip is massive in size versus today's leading AI accelerator, NVIDIA Rubin. A single NVIDIA Rubin chip delivers up to 22 TB/s memory bandwidth, and AMD's MI455X delivers up to 23.3 TB/s of raw memory bandwidth. Cerebras CS-4 with a single WSE-3T chip offers 43,200 TB/s bandwidth, which is 2000x more bandwidth than Rubin. But we also need to understand that the bandwidth figures for NVIDIA and AMD chips are based on HBM4 solutions, while Cerebras is measuring the raw BW offered by the SRAM onboard the wafer.

Moving forward, Cerebras is integrating the WSE-3T chips on CS-4 within its nexus rack-scale platform. Compared to CS-3, the new solution is designed with modularity in mind, offering a simpler build, faster deployment, and independent integration of power, compute, and IO.

Nexus Goes Modular With WSE Backpacks & A Wire-Free Package

Each Nexus rack is attached to pluggable backpacks, which contain one WSE-3T chip, each. Each of these backpacks is an innovation on its own, featuring a wafer package attached to twice the cooling and power, a wafer IO module with twice the bandwidth, 2x faster latency and room for additional modular upgrades, and the whole pack leverages an efficient manufacturing process which uses 50% fewer components and is 60% automated versus CS-3.

Power distribution is important for AI factories. Each loss can lead to a severe lack of compute output. On Rubin, it is claimed that the 50mm distance between the converters and silicon leads to major power losses along paths within the PCB. As such, the system requires more copper layers to reduce resistive losses, which add to the costs and complexity of the system.

Cerebras bypasses these losses through its 54.5VDC Busbar which involves no PCB between the power and the chip. The chip sits directly on the DC/DC Power convertors, and then there's also the shorter distance between the array of AC/DC power converters, which reduces extra resistive power losses and parasitic inductances. The result is a 100x improvement versus a traditional GPU setup since the distances are cut down to just 0.5mm (vs 50mm).

For IO, Cerebras makes use of a next-gen interface module which extends the fabric from wafer edges and is both modular and programmable for the future. The IO interface is also low latency and high bandwidth, made possible through new direct wafer link interfaces and a standard RoCE protocol network.

Each backpack also includes integrated water conditioning. The pack houses a flow regulation actuator which guarantees proper wafer flow rate, a leak detection module, valved dry quick disconnects, and an energy meter, making it easy to install while monitoring the system for potential leaks. The cooling itself goes in the rear.

On the front, Cerebras houses the power, which includes AC/DC PSUs with up to 277VAC input and 54.5VDC output. There are up to 30x PSU modules per backpack, so 90 in total, and all of these are air-cooled with dedicated fan modules for cooling smaller devices in the backpack.

Another area in which Cerebras shows its rack prowess is the fabric. CS-4 features a 53.5 PB/s fabric on the wafer itself and has no need for cables. Meanwhile, NVIDIA's Rubin NVL72 racks feature 5000 cables, offering up to 260 TB/s of NVLINK fabric speeds. This means CS-4 offers 200x higher fabric bandwidth than GPU interconnects.

The bandwidth and latency advantages don't stop at the fabric. The WSE-3T chip offers 2.4 Tb/s of aggregate bandwidth at 3us latency, while the network latency between the wafers is 1.7x higher.

We then move to a generalized comparison between the CS-3 and CS-4 rack-scale solutions. CS-3 was capable of 125 PFLOPs on a single WSE-3 chip while CS-4 offers 750 PFLOPs of AI compute with three WSE-3T chips (250 PFLOPs per chip). It has 132 GB of SRAM (44 GB per chip) versus 44 GB on the previous rack, with much higher bandwidth and less than half the bandwidth.

With the speeds and feeds done, Cerebras showcases what to expect in terms of AI compute while also highlighting the capabilities of its current CS-3 rack, which it claims already runs the largest frontier model (GPT-5.6 SOL @ 10T parameters).

Lastly, for the CS-4, Cerebras has already announced that the rack is in early access and general availability is scheduled for Q3 2026.

CS-5 Tackles The AI Wall With Boosted Capabilities While CS-6 Goes 3D

But there's more: Cerebras is also announcing its next-gen CS-5 and CS-6 rack-scale solutions for the first time.

According to Cerebras, CS-5 will be launching in 2027 and is "Designed to set another standard of speed and efficiency". This solution will scale from 30B to multi-trillion-parameter models.

For Gemma 4 31B and gpt-oss 120B, the company estimates up to 10,000 tokens per second per user, and in frontier models such as DeepSeek, Kimi, GPT 5.6 SOL, CS-5 is expected to hit up to 5000 tokens per second per user with up to 3 million tokens per second per MW.

There's also CS-6, which is expected to take full advantage of 3D packaging solutions with a yield-resilient architecture, a vertical power delivery solution, & a fully integrated cooling methodology.

The Wafer Scale Engine for CS-6 will integrate Wafer-Scale SCRAM on top of the WSE chip through 3D integration, while being an order of magnitude smaller in footprint with the fastest AI inference speeds on the market. This is a very forward-looking design for now, but it looks like Cerebras has the stage set for future AI models and is scaling its wafer-scale engines to meet the accelerated AI growth big time.

Follow Wccftech on Google to get more of our news coverage in your feeds.