The race to build specialized chips for AI inference is accelerating, with two major players unveiling dramatically faster hardware and securing massive funding rounds. Cerebras announced its next-generation WSE-3T chip and CS-4 rack systems on August 19, 2026, doubling per-chip performance while cramming three times as many accelerators into a single rack. Meanwhile, Etched raised $700 million at a $21 billion valuation after its first customer, quantitative trading firm Jane Street, tested and deployed its custom inference hardware. Together, these announcements signal a fundamental shift in how the AI industry approaches the inference problem, the computational bottleneck that occurs after a model is trained and begins answering user questions.

Why Is Inference Becoming the Industry's Biggest Bottleneck?

Inference has emerged as the critical constraint in AI deployment. Fewer than 1% of people worldwide have access to frontier AI models, and the cost and volume of running them has become prohibitively expensive. Unlike training, which happens once per model, inference happens billions of times per day as users interact with chatbots, search engines, and other AI applications. General-purpose graphics processing units (GPUs), originally designed for training AI models, are proving inefficient for this task. Both Cerebras and Etched are betting that custom silicon built specifically for inference can deliver dramatic improvements in speed and power efficiency.

The core insight driving both companies is straightforward: if you sacrifice flexibility and design a chip for one specific job, you can optimize it ruthlessly. Cerebras' approach focuses on maximizing memory bandwidth, the speed at which data flows through the chip. Etched takes a different path, splitting inference into two distinct stages and building separate hardware for each.

How Are These Companies Approaching the Inference Problem Differently?

Cerebras and Etched represent two distinct philosophies for solving inference, each with its own advantages:

• Cerebras' Strategy: The company doubled the clock speed of its WSE-3T chip from 1.4 GHz to an estimated 2.8 GHz by improving power delivery efficiency, allowing twice as much power to flow through the same silicon without overheating. This enables the chip to generate tokens, the individual words or word fragments that make up AI responses, much faster. The new CS-4 rack system houses three of these dinner-plate-sized accelerators, tripling the compute density compared to previous generations.
• Etched's Strategy: Rather than building one chip for all inference tasks, Etched splits the problem into prefill and decode stages. Prefill is the computationally intensive phase where the model reads and understands the user's prompt. Decode is the token-generation phase where the model produces the answer one token at a time. Etched built separate hardware optimized for each stage, with the decode chip featuring custom memory and interconnect technology that lets many chips share a single fast, low-latency memory pool.
• Interconnect Innovation: Both companies are rethinking how chips communicate with each other. Cerebras reduced interconnect latency from five microseconds to just two microseconds by eliminating unnecessary switches and having chips talk directly to one another. This matters because inference at scale requires many chips working together, and slow communication between them becomes a bottleneck.

What Makes These New Chips Faster Than Nvidia and AMD GPUs?

Cerebras' WSE-3T boasts 43.2 petabytes per second of memory bandwidth, roughly 1,000 times faster than Nvidia's or AMD's best GPUs. To put this in perspective, memory bandwidth determines how quickly a chip can access the weights and parameters that define an AI model. For inference, where the same model weights are accessed repeatedly, this bandwidth advantage is transformative. The chip can keep more of the model in its on-board memory, reducing the need to fetch data from slower external storage.

Etched's first chip worked on its first manufacturing attempt, an uncommon achievement for a semiconductor startup. The company manufactures on TSMC's N4P process, the same cutting-edge manufacturing node that Cerebras uses. Etched's founders, Gavin Uberti, Chris Zhu, and Robert Wachen, dropped out of Harvard in 2022 to start the company as Thiel Fellows, betting that the transformer architecture underlying GPT, Claude, Gemini, and nearly every commercial AI system would remain dominant. Four years later, that bet has paid off.

How Are These Chips Being Deployed in Real-World Systems?

Cerebras is no longer trying to run entire inference workloads on its own accelerators. Instead, it has partnered with Amazon Web Services (AWS) and AMD to offload the compute-intensive prompt-processing phase onto their Trainium XPUs and Instinct GPUs. Cerebras' chips now function primarily as decode accelerators, similar to how Nvidia is using Groq's LPUs, or language processing units, in its LPX rack systems. This disaggregated approach allows each type of hardware to do what it does best.

Etched's first customer deployment is particularly noteworthy. Jane Street, a major quantitative trading firm, tested Etched's hardware before committing and found that the chip delivered the precision needed for its most demanding workloads. Jane Street's decision to deploy Etched's first production rack signals that proprietary trading capital is moving directly into inference hardware. The company's order book had already passed $1 billion as of June 2026, before this latest funding round.

What Does This Mean for the Broader AI Hardware Market?

The inference chip market has attracted more funding than any other hardware category in 2026, but the strategies diverge sharply. Some companies are building their own chips, while others have retreated from hardware entirely. Groq, for example, signed a $20 billion technology licensing deal with Nvidia and shifted from selling its own chips to offering inference as a cloud service. This split between selling custom silicon and selling services on someone else's hardware is now an open question in the industry.

The pressure from hyperscalers building in-house chips is intensifying. Google split its eighth-generation TPU, or tensor processing unit, into separate training and inference parts, and OpenAI completed its first in-house inference chip with Broadcom in just nine months. When the largest buyers of compute start designing their own silicon, the addressable customer pool for specialist startups narrows significantly. That Etched found its first customer in finance rather than in a hyperscaler like Google or OpenAI is worth noting in this context.

Cerebras' decision to double performance this generation rather than increase on-chip memory capacity is curious. The company's chips have 44 gigabytes of static random-access memory (SRAM), unchanged since the WSE-2 launched five years ago. In a disaggregated inference environment where prefill is handled by GPUs, one might expect Cerebras to prioritize SRAM capacity over compute. However, given that these disaggregated compute architectures are relatively new, Cerebras may have been too far along in production to pivot its design.

The power consumption of these new systems is substantial. Cerebras estimates that each CS-4 backpack will consume around 46 kilowatts, with a total system power of between 120 and 140 kilowatts. While this is significant, it looks conservative compared to the 240 to 250 kilowatt rack systems coming from AMD and Nvidia later in 2026. The efficiency gains from custom silicon are allowing these companies to pack more compute into systems that consume less power per unit of performance.

The competitive landscape is reshaping rapidly. Cerebras went public on Nasdaq and raised $5.5 billion with its stock up 108% on day one. SambaNova raised $1 billion at an $11 billion valuation, and d-Matrix took $275 million at a $2 billion valuation. MatX, founded by former Google TPU designers, raised $500 million for its LLM-specific chip, taking an approach similar to Etched's. These funding rounds reflect investor confidence that specialized inference hardware will become essential infrastructure for AI deployment.