Cerebras Systems announced a new server system called the CS-4 that uses three of its large, specialized chips to process AI chatbot queries faster than competing hardware. The company, which competes directly with Nvidia in the inference segment of artificial intelligence, is betting that its massive chip design will give it an edge in the rapidly growing market for running AI models rather than training them.

What Makes Cerebras Chips Different From Nvidia GPUs?

Cerebras chips gain their speed advantage through sheer size. The company's processors are large enough to fit entire AI model computations on a single piece of silicon, which eliminates the energy waste and slowdown that comes from moving data between multiple smaller chips. This is fundamentally different from Nvidia's approach, which uses general-purpose graphics processing units (GPUs) that can run almost any type of AI workload but require more data movement.

The new CS-4 system is powered by Cerebras' WSE-3 Turbo chip, fabricated using TSMC's 5-nanometer manufacturing process. The server rack includes new networking components designed to speed data movement between chips, and the company engineered the entire system to require 50% fewer components than previous generations. This simplification matters because it means data centers can be built faster and with less complexity.

"We're going to get four times as fast between now and the end of the year, end of 2027, and we're going to get 20 times more throughput," said Andrew Feldman, CEO of Cerebras.

Andrew Feldman, CEO at Cerebras Systems

Feldman's ambitious roadmap reflects the company's confidence in its specialized approach. Cerebras expects to deliver 600 megawatts' worth of computing power by the end of 2027, a massive expansion that signals the company's commitment to scaling its inference business.

How Does Inference Hardware Compete in the AI Market?

The inference market, where AI models generate answers to user queries, has become increasingly lucrative. Unlike training, which builds AI models from scratch, inference is the repetitive, high-volume task of running those models in production. This shift has created an opportunity for specialized chip makers to challenge Nvidia's dominance.

Other startups are pursuing similar strategies. Etched, a San Jose-based company, has built a chip called Sohu that is hardwired specifically for transformer models, the architecture underlying most large language models (LLMs). Etched claims its eight-chip server can process about 500,000 tokens per second on Llama 70B, a popular open-source AI model, which the company presents as roughly 10 to 20 times faster than Nvidia's H100 GPU on the same task. However, those figures remain unverified company claims.

The competitive landscape includes several players targeting the inference layer:

• Cerebras: Focuses on large chips that fit entire computations on single silicon, reducing data movement overhead and energy consumption.
• Etched: Specializes in transformer-only ASICs (application-specific integrated circuits) that sacrifice flexibility for extreme speed on a single workload.
• Groq: Another inference-focused competitor building custom silicon for AI model deployment.

What Are the Trade-offs of Specialized Inference Chips?

While specialized chips like Sohu promise raw speed, they come with significant limitations. Etched's chip cannot run mixture-of-experts models like DeepSeek V4 or Qwen3, which are among the most widely used open-source models today. It also cannot handle multimodal tasks, image generation, or training workloads. Customers must abandon standard tools like vLLM and switch to Etched's proprietary software stack.

Cerebras' approach is less restrictive than Etched's but still specialized. The company is betting that the speed gains from its large-chip architecture will outweigh the loss of flexibility compared to Nvidia's general-purpose GPUs. Nvidia's real competitive advantage has never been raw speed alone; it is CUDA, the software layer that millions of developers already know, combined with the ability to run whatever new AI model ships next.

Steps to Understanding Inference Chip Competition

• Recognize the market shift: More money is now flowing into inference (running models) than training (building models), creating a new opportunity for specialized hardware makers to compete with Nvidia.
• Understand the speed-versus-flexibility trade-off: Specialized chips like Sohu and Cerebras' WSE-3 Turbo sacrifice the ability to run all AI model types in exchange for extreme speed on specific workloads.
• Consider the software lock-in: Customers who adopt specialized chips often must use proprietary software stacks, which creates switching costs and reduces portability compared to Nvidia's CUDA ecosystem.

Cerebras' new CS-4 system will be available in the third quarter of 2026, and the company plans another generation of chips and servers in 2027. The company reported an adjusted loss of $6.9 million on sales of $180.1 million in its most recent quarter, indicating it is still in growth mode despite the competitive pressure.

The emergence of multiple specialized inference competitors suggests that Nvidia's dominance in AI hardware is being challenged, but not necessarily threatened. Nvidia's moat remains its software ecosystem and flexibility. For customers with specific, high-volume inference workloads, however, specialized chips from Cerebras, Etched, and others may offer compelling alternatives that deliver better performance per dollar spent.