Targeting Nvidia! Cerebras (CBRS.US) unveils the "fastest AI accelerator" CS-4 system, claiming inference speed is 30 times that of GPU solutions

According to Zhitong Finance APP, AI chip company Cerebras Systems (CBRS.US) officially launched its next-generation rack-level AI system, Cerebras CS-4, at the SUPERNOVA event in San Francisco, proclaiming it as the “world’s fastest AI accelerator.” This system targets the AI inference sector and directly rivals Nvidia (NVDA.US) GPU server products. The CS-4 is expected to begin shipping this quarter (Q3 2026) and is currently undergoing sample testing with select customers.

Technical Specs: 750 PFLOPs of computing power with three wafer-scale processors

The CS-4 is the first product of Cerebras's next-generation Nexus rack-level platform architecture, powered by three new Wafer Scale Engine 3 Turbo (WSE-3T) wafer-scale processors. The WSE-3T is the largest AI processor to date, with each chip integrating four trillion transistors and 900,000 AI-optimized cores, covering 46,225 square millimeters of silicon and housing 44GB of built-in SRAM.

The CS-4 system delivers 750 PFLOPs of AI compute power, 129.6 PB/s of memory bandwidth, and 7.2 Tbps of I/O throughput per second. Compared to its predecessor CS-3, the CS-4 doubles the speed and increases throughput per watt by ten times, significantly improving data center economics. Compared to the previous generation WSE-3, the WSE-3T doubles the per-wafer AI compute power from 125 PFLOPS to 250 PFLOPS and doubles memory bandwidth from 21.6 PB/s to 43.2 PB/s. Power consumption per chip has jumped from 15 kilowatts to approximately 33 kilowatts.

In terms of architectural design, the CS-4 adopts a modular Nexus platform architecture, separating compute, power, and I/O into independent components. The power conversion module has been moved from the traditional 50 millimeters from the processor on GPU boards to about 0.5 millimeters, virtually eliminating board-level power loss. The pluggable “backpack” design integrates power conversion, liquid cooling, and control electronics, reducing deployment time from days to hours.

Performance Breakthrough: Over 4,400 tokens per second per user in GPT-OSS-120B test

Cerebras claims the CS-4 sets a new industry benchmark for inference speed. In direct comparison tests on the GPT-OSS-120B model, the CS-4 generates more than 4,400 tokens per second per user, up to 30x faster than GPU-based solutions. The system can support models with over 50 trillion parameters.

Cerebras CTO and co-founder Sean Lie said, “Being 30x faster isn’t just about making interaction feel smoother, it gives agent systems an order of magnitude more space to reason, validate, or call tools in the same real time window.” Cerebras CEO Andrew Feldman noted, “Historically, fast inference has meant using smaller, weaker models. The CS-4 achieves industry-leading speed on the largest frontier models, fundamentally changing this paradigm.”

In terms of energy efficiency, the CS-4 achieves up to 10x throughput per watt improvement over the CS-3, significantly improving data center economics. Cerebras CEO Andrew Feldman emphasized that fast tokens are more valuable than slow tokens, and the CS-4 can provide higher-quality tokens and greater total token output within a given power budget, “fundamentally changing the paradigm.”

The CS-4 also features a design that reduces component count by 50%, making deployment simpler and faster for customers. In the system architecture, the new Nexus platform separates computation, power supply, and I/O into modular components—power conversion has been shortened from about 50 millimeters on traditional GPU motherboards to the processor to about 0.5 millimeters, greatly reducing board-level power losses and supporting higher operating frequencies. The pluggable “backpack” design integrates power conversion, liquid cooling, and control electronics into a single unit, cutting deployment time from several days to a few hours.

Notably, Cerebras’s wafer-scale system uses static random access memory (SRAM) instead of the industry standard DRAM. SRAM is much faster than DRAM, but is larger and more expensive—Cerebras’s platter-sized wafers provide the needed physical space to use SRAM. At the same time, the monolithic chip design keeps data transfer distances much shorter than those of Nvidia or AMD’s multi-chip systems.

Ecosystem Partnerships: Collaborating with AMD Helios, Targeting 5x Efficiency Gain

The CS-4’s programmable I/O subsystem supports Ethernet-based RoCE v2 RDMA standards, allowing integration with heterogeneous systems. Cerebras has already partnered with AMD (AMD.US) on this architecture, pairing AMD’s Helios rack-level solutions with the WSE, resulting in five times the tokens per watt per second versus a pure Cerebras configuration. AWS Trainium is also listed as a partner.

SemiAnalysis founder and CEO Dylan Patel commented that CS-4’s improvements in deployability and networking capabilities enable it to scale to larger models, creating a “high-capacity token factory.”

Market Background: Shares Pull Back 35% From Highs Post-Listing, CS-4 Seen as Key Inflection Point

Cerebras went public on Nasdaq on May 14, 2026 at $185 per share, opening at $350 on its first day and closing at $311.07, up 68%. However, its share price has since steadily declined, closing around $220 on Tuesday, a drop of about 35% from its peak.

The company’s Q2 financial report was mixed: core revenue reached $209.9 million, up 103% year-on-year, but it posted a loss per share of $2.98 compared to a profit of $1.91 in the same period last year. The company’s guidance for tripling core revenue by 2027 and better-than-expected outlook for Q3 failed to fully convince investors. However, the company’s remaining performance obligations reached $25.4 billion, providing strong visibility into future revenues.

Industry Significance: AI Inference Market Landscape Faces Reshaping

The launch of CS-4 comes at a pivotal moment amid explosive growth in AI inference demand. As large models transition from training to massive deployment, inference efficiency is becoming a key competitive dimension for AI infrastructure. If Cerebras’ claimed 30x speed advantage is validated via independent benchmarking, it could change the procurement decisions of hyperscale data center operators.

The release of Cerebras CS-4 signals that competition in AI chips is rapidly shifting from model training to the inference market. Whether this system can truly challenge Nvidia’s dominance in AI inference will depend on independent benchmarking validating its performance claims—and whether CS-4 can be delivered at sufficient scale and speed to meet the growing demand for high-throughput token generation from hyperscale data centers.

Cerebras’s cloud inference business already saw a 287% year-on-year increase in Q2, reaching $127.7 million in core cloud revenue. The CS-4 launch marks a new phase in Cerebras’s business model shift from selling hardware to leasing inference capability.