On August 18, 2026, Cerebras Systems announced the CS-4, the fourth generation of its wafer-scale AI rack system and the fastest production AI inference hardware in the industry. The headline figure: up to 30 times more tokens per second per user than GPU-based systems. The enterprise implication is significant, not because faster is always better in isolation, but because inference latency is the binding constraint on how much reasoning an agentic system can fit into a given response window.
Cerebras is a public company (NASDAQ: CBRS) and a known quantity in the inference infrastructure market. OpenAI already uses Cerebras to power the ultrafast preview of GPT-5.6 Sol, which delivers up to 14x real-time speed for interactive workloads. CS-4 is the hardware generation that makes that kind of speed available at hyperscale.
What Launched: CS-4 and the WSE-3 Turbo
The CS-4 rack holds three of Cerebras’s new Wafer Scale Engine 3 Turbo (WSE-3T) processors. The WSE-3T is not a new chip in the traditional sense: it uses the same 5nm TSMC silicon, the same 900,000 cores, and the same 44 GB of on-wafer SRAM as the WSE-3 that launched two years ago. What changed is how hard Cerebras pushes it.
By improving the power delivery architecture, moving conversion electronics 100 times closer to the processor (from approximately 50 millimeters on a conventional GPU board to 0.5 millimeters), Cerebras can now feed twice as much power to each wafer. That enables a clock speed roughly doubling to around 2.8 GHz, which in turn doubles both AI compute (250 PFLOPS per wafer) and memory bandwidth (43.2 petabytes per second per wafer).
For inference, memory bandwidth is the dominant performance lever. The model weights must flow from storage to compute cores faster than the cores can process them, or the cores sit idle waiting. Doubling memory bandwidth translates into roughly doubled decode speed. Three WSE-3T processors per rack compounds that gain further.
| Metric | CS-3 (1 wafer) | CS-4 (3 wafers) |
|---|---|---|
| AI compute | 125 PFLOPS | 750 PFLOPS |
| Memory bandwidth | 21.6 PB/s | 129.6 PB/s |
| On-chip fabric bandwidth | 26.7 PB/s | 160.5 PB/s |
| System I/O bandwidth | 1.2 Tbps | 7.2 Tbps |
| I/O latency | 5 microseconds | 2 microseconds |
| Inference speed vs GPU | up to 15x | up to 30x |
Sources: Cerebras official announcement, Business Insider / Globe Newswire
In head-to-head benchmarking on GPT-OSS-120B, CS-4 delivers more than 4,400 tokens per second per user. The fastest GPU-based inference service today manages roughly 350 tokens per second per user on comparable models, according to Artificial Analysis, which conducted third-party evaluation alongside Cerebras’s internal testing.
The Agentic AI Angle
Cerebras CTO and co-founder Sean Lie framed the enterprise relevance directly in the CS-4 announcement: “Being 30 times faster doesn’t just make a response feel fast. It gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use in the same wall-clock time.”
That statement is the key insight for enterprise AI teams. A single agentic task rarely completes in one model call. A well-designed agent uses multiple steps: plan, research, verify, draft, check, refine. Each step is a round-trip to the model. If each round-trip takes 500 milliseconds on GPU infrastructure, a 10-step agentic loop takes 5 seconds before any output reaches the user. On CS-4 infrastructure, the same loop runs in under 200 milliseconds total.
This is not a marginal improvement in user experience. It is the difference between agents that feel instantaneous and agents that feel slow enough to interrupt the workflow. For enterprise deployments where employees interact with agents in real time, the wall-clock time of a full agentic loop directly determines whether staff will use the system or route around it.
The implication extends to autonomous background agents as well. When a workflow agent runs dozens of planning and verification calls without any human waiting in real time, faster inference increases the number of complete task loops per hour, directly affecting throughput. Teams sizing infrastructure for agentic workloads can either run more concurrent users on the same compute budget or achieve the same concurrency at meaningfully lower cost per completed task.
Nexus: A Redesigned Rack Architecture
The physical form factor of CS-4 is as significant as the processor upgrade. Cerebras introduced the “Nexus” rack platform alongside CS-4, a modular architecture that separates compute, power delivery, and I/O into independent modules.
In previous Cerebras generations, the entire system was a tightly coupled chassis. Adding a new WSE meant replacing the chassis. With Nexus, the rack holds separate “backpacks,” each containing one WSE-3T, that plug independently into power and networking shelves at the front of the rack. A CS-4 rack holds three backpacks.
The practical benefits:
Deployment speed: Cerebras says CS-4 deploys up to three times faster than CS-3 and requires 50 percent fewer components. The modular design allows racks to ship pre-assembled with power shelves, then operators plug in compute backpacks on-site in hours rather than days.
Independent upgrades: When Cerebras releases a WSE-4 or later processor, operators can upgrade just the compute backpacks without replacing the rack infrastructure, a meaningful consideration for operators who lock in rack deployments for multi-year contracts.
Scalability: The new I/O module supports direct wafer-to-wafer links with 2-microsecond latency (down from 5 microseconds on CS-3) and standards-based RoCE v2 RDMA over Ethernet. Both enable large CS-4 clusters to serve models with more than 50 trillion parameters, expanding the addressable model tier well beyond what previous Cerebras systems could run efficiently.
Disaggregated Inference: AMD and AWS Partnerships
A structural challenge for wafer-scale inference is that each WSE holds only 44 GB of SRAM on-wafer. While that is far more SRAM than any GPU, frontier models with trillions of parameters require many wafers in pipeline parallel configuration. For long-context inference especially, keeping key-value cache data on-wafer adds pressure.
Cerebras’s answer is disaggregated inference: use CS-4 exclusively for the decode phase, where its memory bandwidth advantage is decisive, and offload the compute-intensive prompt prefill phase to partner hardware. CS-4 now formally supports AMD Helios and AWS Trainium as prefill partners, with more announced as coming.
The Register’s analysis of the CS-4 architecture notes that Cerebras is positioning CS-4 similarly to how Nvidia uses Groq LPUs in its LPX configurations, as specialized decode accelerators paired with other hardware for prefill. The combination lets operators run heterogeneous inference pipelines tuned for both workload phases.
SemiAnalysis’s deep-dive estimates that running a 1.6-trillion-parameter model like DeepSeek V4 Pro at 256K context with reasonable concurrency (256 requests) requires approximately 40 CS-4 systems, representing over $20 million in capital expenditure and roughly 1 megawatt of power. That positions CS-4 as infrastructure for large AI providers and enterprises running dedicated inference at scale, not a general-purpose GPU replacement for smaller deployments.
Enterprise Inference Economics in 2026
CS-4 enters a market that has shifted dramatically over the past 12 months. As covered in the IBM-Together AI partnership, enterprise AI teams now spend more on inference than on training, a reversal that has made inference economics central to AI budgets. Faster inference changes the cost-per-task calculation: if a task that required 10 seconds on GPU infrastructure completes in under 1 second on CS-4, operators can either serve 10x the concurrent users on the same hardware or achieve the same throughput at one-tenth the infrastructure footprint.
The caveat is that CS-4 is not competing with GPU cloud spot instances for small-scale or variable workloads. It is purpose-built for operators and enterprises that need guaranteed high-speed inference at volume. The first CS-4 shipments go to select customers this quarter, and Next Platform’s analysis suggests that per-wafer economics are broadly similar to CS-3, meaning operators get roughly double the performance at comparable hardware cost.
For enterprise AI architects evaluating infrastructure for 2026 and 2027 agentic deployments, the CS-4 raises the baseline expectation for what “fast inference” means. A workload that runs well on GPU infrastructure does not necessarily require CS-4. A workload where inference latency is the bottleneck on agent loop throughput, or where the interactive quality of real-time AI conversations determines user adoption, should now include CS-4 in its evaluation set.
Availability and What Comes Next
First CS-4 shipments begin in Q3 2026. Cerebras has committed to a public roadmap of roughly 2x performance improvement per year through 2027. Next Platform expects a WSE-4 in a future generation that will address the SRAM capacity constraint by increasing on-wafer memory alongside the compute core count.
The Nexus rack architecture is explicitly designed to carry forward across multiple processor generations, a signal that Cerebras is building toward a longer hardware lifecycle for CS-4 customers and a cleaner upgrade path from WSE-3T to WSE-4.
For enterprise teams building agentic AI systems today, the practical question is not whether CS-4 changes the physics of inference (it does) but whether your specific workload’s latency requirements and volume justify purpose-built wafer-scale infrastructure versus provisioning GPU capacity from cloud providers. If your agents are bottlenecked on inference round-trip time and your usage volumes are large, CS-4 has moved that conversation from theoretical to procurement-ready.
If you are designing AI workflows for your enterprise and want to think through the infrastructure architecture required, reach out to Enera.