Inference speed and architecture
Cerebras runs Moonshot AI’s Kimi K2.6 trillion-parameter open-source model at 981 output tokens per second. This performance exceeds the next-best GPU cloud provider by 6.7x and beats the median by 23x, according to Artificial Analysis. The WSE-3 chip carries 44 GB of on-chip SRAM to keep model weights on-chip. The Cerebras WSE-3 architecture eliminates the memory bottlenecks that slow down traditional GPU clouds because the 44 GB of on-chip SRAM keeps model weights on-chip to avoid stalling during HBM memory transfers. A 10,000-token agentic coding task takes 5.6 seconds on Cerebras, while the official Kimi endpoint requires 163.7 seconds. SambaNova uses a CGRA architecture where compute and memory units map to the AI model’s graph. Artificial Analysis reports that SambaNova delivers high token rates. SambaNova claims its SN50 chip provides 5x the inference throughput of competitors.
The Cerebras WSE-3 architecture overcomes the limitations of traditional memory. Because the model weights stay on-chip, the system avoids the latency of moving data to external memory. This structural advantage allows Cerebras to win hyperscale deals for large frontier models. The 29x latency reduction on 10,000-token tasks for agentic workflows changes which product architectures work for enterprise users. Cerebras timed this benchmark announcement to arrive within days of its Nasdaq IPO. The performance gap is substantial, as Cerebras is 6.7x faster than the next-best GPU cloud provider.
Cluster specifications
The Cerebras CS-3 cluster uses 32 WSE-3 GPUs and 20 cabinets to manage workloads. The cluster consists of sixteen WSE cabinets holding two CS-3 systems each, plus four NET cabinets. These 32 GPUs run over an on-wafer fabric with 214 Pb/s bandwidth. The cluster holds 1.408 TB of on-chip SRAM and lacks DRAM. One engine delivers 125 petaFLOPS of sparse compute. The total provisioned heat load reaches 945 kW, with 875 kW requiring water cooling and 70 kW requiring air cooling.
| Feature | Cerebras CS-3 (32-node) |
|---|---|
| Total On-chip SRAM | 1.408 TB |
| External DRAM | 0 GB |
| Provisioned Heat Load | 945 kW |
| Network Bandwidth | 214 Pb/s |
| Sparse Engine Rating | 125 petaFLOPS |
Will hyperscalers adopt this high-density cooling requirement for massive sparse model deployments? You should check if your facility handles this thermal density. SambaNova relies on an array of compute and memory units. The company is currently teaming with Intel on a heterogeneous system that employs both SambaNova chips and GPUs. Because Intel is not a competitive GPU supplier, this arrangement helps SambaNova scale without needing to replace existing GPU deployments. SambaNova is also in talks to raise between $800 million and $1 billion. This follows a $350 million Series E in February 2026, which implied a $2 billion valuation. SambaNova must show performance data and customer traction to justify an $11 billion valuation.
Market competition and valuation
Cerebras holds an $8B valuation while SambaNova targets a $5B valuation for its SN40L. SambaNova closed a $1B round following a $350 million round earlier this year. The company is also discussing a heterogeneous system with Intel. This move comes as the market seeks alternatives to Nvidia’s dominance. Nvidia’s acqui-hire on December 24, 2025, was a turning point in the inference race. Instead of attempting to out-benchmark specialized hardware, Nvidia chose to absorb the talent and ideas that made deterministic inference possible.
Cerebras provides a 29x latency reduction on 10,000-token tasks, which shifts the viability of certain product architectures for AI practitioners. For founders and infrastructure investors, the Cerebras benchmark challenges the GPU-centric stack. SambaNova must show performance data and customer traction to maintain its market position. The company’s SN50 chip claims 5x competitor inference throughput. Cerebras uses its wafer-scale technology to handle models that exceed the capacity of standard hardware.
The ability to run trillion-parameter models like Kimi K2.6 depends on the hardware’s ability to bypass traditional memory limits. Cerebras achieves this with its architecture, which creates a performance gap compared to GPU clouds that rely on software optimization. SambaNova faces pressure to publish competing third-party verifications to avoid ceding the positioning narrative to Cerebras. The market will increasingly price and provision inference based on latency guarantees, variance bounds, and cost per token. The Cerebras WSE-3 is the superior choice for hyperscale deals requiring massive throughput for trillion-parameter models.




