Cerebras Systems

Most chipmakers cut a wafer into hundreds of small dies. Cerebras keeps the whole wafer as a single chip, the Wafer-Scale Engine, which is why its processor is 56 to 58 times larger than a GPU and, the company says, delivers AI inference up to 15x faster. When Andrew Feldman and four co-founders started it in 2015, it was not clear the thing was even manufacturable.

That wager is now a public one. Cerebras listed on the Nasdaq in May 2026 under the ticker CBRS, popped 68% on day one to a roughly $95 billion valuation, and pulled in about $5.55 billion, the biggest US tech IPO since Uber. Revenue hit $510 million last year on 76% growth, with the company swinging to an $88 million profit. The catch sits in plain sight in the filing: for years a single customer bought almost everything Cerebras made.

Cerebras designs the largest computer chip ever built and wraps it in systems and cloud services that run AI models fast. The pitch is speed: because everything sits on one piece of silicon, there is no slow chip-to-chip interconnect to fight, so training and inference run in a single latency budget where GPU clusters need many.

You can buy it three ways. Rent it by the token through the inference and training cloud with drop-in OpenAI-compatible APIs, take dedicated private capacity, or install a CS-3 supercomputer in your own data center for full control of models and data.

The problem Cerebras keeps pointing at is latency. On GPUs, a coding agent can take 20 to 30 seconds to generate, long enough to break a developer's concentration and force a context switch, and reasoning models that think out loud pay that tax on every step. Multi-step agents stall or time out; voice assistants feel robotic when the pause is a beat too long.

Cerebras reframes speed as a design parameter rather than a nice-to-have. Run a model at around 1,000 tokens per second and you can afford more reasoning, deeper retrieval, and tighter agent loops inside the same wall-clock budget, which is the argument behind its work with coding, search, and drug-discovery customers.

This is a B2B story aimed at three rooms. AI-native product teams who live or die on token speed, from coding agents to real-time voice. Large enterprises and the Global 1000 that want frontier models at production scale. And research and government labs doing physics, genomics, and drug discovery that a wafer-scale system can crunch in a fraction of GPU time.

The common buyer is technical: infrastructure and ML leads choosing where inference runs, and researchers who treat a hundreds-times speedup as the difference between months and years.

The lineup all traces back to the same piece of silicon. The Wafer-Scale Engine is the chip; the CS-3 is the system built around it; the Condor Galaxy network is what you get when you wire a lot of them together. The cloud and code tiers are how everyone who cannot install a supercomputer still gets to use one.

Cerebras makes money two ways. It sells and deploys wafer-scale systems to enterprises, labs, and cloud partners, the high-ticket contract side that has driven most of its revenue so far. And it rents inference by the token through a pay-as-you-go cloud, with a free tier to get developers hooked, a self-serve Developer tier starting at $10, and enterprise deals for guaranteed throughput and custom model weights.

The pricing page splits neatly: raw inference API access on one side, a Cerebras Code subscription for developers on the other. Everything above the developer tier is a sales conversation.

Cerebras is not trying to out-volume Nvidia; it is trying to be the fastest place to run inference, full stop. The moat is architectural. Wafer-scale keeps the whole model on one chip with enormous on-chip memory, so it sidesteps the interconnect bottleneck that GPU clusters have to engineer around.

The strategic bet for 2026 is inference at scale: build out data centers across North America and Europe, ride partnerships with OpenAI, AWS, Meta, and the developer ecosystem, and become the default when speed is the constraint. Whether that holds as Nvidia, Groq, and custom silicon all chase the same latency prize is the open question.

The customer quotes read like a who's who of AI: OpenAI calls Cerebras its dedicated low-latency inference option, Meta runs its Llama API on it, and Cognition, Notion, AlphaSense, GSK, and Mayo Clinic all show up with concrete wins. Cognition reports its SWE-1.6 coding agent runs up to 5x faster on Cerebras than on GPU, hitting around 950 tokens per second.

The research proof is older and harder to argue with. Argonne National Laboratory won a Gordon Bell Special Prize for COVID-19 variant work on a CS-2 cluster, and NETL clocked a CS-2 at nearly 500x faster than its Joule supercomputer on field-equation modeling.

The praise is consistent and mostly technical: the speed is real, the OpenAI-compatible API makes it easy to drop in, and researchers cite orders-of-magnitude gains on the right workloads. Independent benchmarks back the direction even where they land below the headline 15x.

The recurring worry is not the technology but the customer list. Analysts keep flagging revenue concentration: G42 was most of the business, and even as that fell in 2025, a small set of Abu Dhabi entities and now OpenAI still account for the bulk of sales. As one analyst put it, the concentration rotated, it did not go away. The other knock is that squeezing the full speedup out of the hardware means working in Cerebras's own tooling, not the familiar GPU stack.

Before going public Cerebras had raised around $2.7 billion privately, backed by an unusually operator-heavy cap table: Sam Altman, Ilya Sutskever, Andy Bechtolsheim, Adam D'Angelo, Intel CEO Lip-Bu Tan, and former AMD and Juniper leaders among them. G42 layered on more than $1.4 billion in commitments.

The May 2026 IPO priced at $185, opened at $350, and closed the first day up 68% for roughly a $95 billion valuation, raising about $5.55 billion. The market has since cooled the stock hard, with the cap sliding toward $60 billion within weeks, which is the price of going public on a story the market is still arguing about.

The technology story is strong and getting stronger: real speed advantages, marquee customers, and a data-center buildout aimed at making Cerebras the default for high-speed inference. If inference demand keeps climbing, wafer-scale is well positioned to catch it.

The risk is the same one the IPO prospectus flagged. A business where a couple of counterparties drive most of the revenue is exposed to any one of them changing course, and OpenAI in particular is a customer with its own shifting chip strategy. The next chapter is whether Cerebras can turn a handful of enormous contracts into a broad, durable book of business.

Cerebras is unusual in that most of its five co-founders are still in the building a decade later, with Andrew Feldman as CEO and Sean Lie, Jean-Philippe Fricker, and Michael James in senior architecture roles. The company frames itself as a place for people who want to work on genuinely new hardware, publish and open-source their research, and skip the bureaucracy, and its own pitch to engineers leans on job stability with startup vitality.

The work is deep-systems work: chip design, compilers for a novel dataflow language, ML frameworks, and the infrastructure to run some of the fastest AI machines on earth. Hiring runs across engineering, research, product, and operations, concentrated in Sunnyvale with teams in San Diego, Toronto, and Bangalore.

Disclosed pay clusters in engineering, where posted US base salaries run from roughly $120,000 to $300,000 depending on level, with a few new-grad roles landing in the mid-$100,000s. A handful of non-engineering and design roles disclose bands in a similar range.

These are base numbers only. Cerebras notes that actual compensation for most roles adds bonus and equity, and as a newly public company that equity is now liquid stock rather than paper, which materially changes the math for candidates.

The registered entity is Cerebras Systems Inc., operating from 1237 E. Arques Ave, Sunnyvale, CA 94085, per its own privacy and terms pages effective August 2024. The company publishes a CCPA disclosure and a subprocessors reference, and states it does not retain the inputs and outputs of its training, inference, and chatbot services.

One nuance worth knowing: for API use, ownership of prompts and outputs is governed by the terms of whichever third-party model you run, not by Cerebras alone.

The past year has been loud. Cerebras pulled off the biggest US tech IPO since Uber, signed a multi-year deal with OpenAI to deploy 750 megawatts of wafer-scale systems reported at more than $20 billion, and struck an AWS partnership pairing its CS-3 systems with AWS Trainium. The links below are the source trail.