.png){kind=logo}
Cerebras focuses on ultra-fast AI inference rather than trying to replace NVIDIA across every workload.
OpenAI, AMD, AWS, and other partnerships give Cerebras a path toward large-scale deployment.
Cerebras must convert strong benchmarks, contracts, and capacity into durable revenue and profits.
Andrew Feldman has placed Cerebras at the center of a highly competitive AI hardware sector. The company does not need to replace every NVIDIA GPU to make that fight matter. Feldman has a narrower target: the part of AI work that turns a model’s answer into tokens at high speed. Cerebras calls that task inference. Its case rests on a simple idea. As AI agents handle more tasks, fast responses can matter as much as raw model power.
This bet now has real scale behind it. OpenAI agreed to add 750 megawatts of Cerebras compute through 2028. Cerebras later said the deal has a value above USD 20 billion. OpenAI also gave Cerebras a USD 1 billion working-capital loan. The deal gives Feldman a major customer and a large test for the company’s wafer-scale system.
Cerebras takes a different path from NVIDIA. NVIDIA built a broad AI platform around graphics processors, CUDA software, network systems, libraries, and large clusters. Cerebras built a wafer-scale processor with a huge amount of compute and memory on one piece of silicon. The design aims to cut the delay that users feel when an AI model produces one token after another.
Cerebras has formed ties with AMD and AWS. The goal does not require a full break from GPU systems. AMD can handle the prefill stage, which processes the prompt and prepares the model state. Cerebras can handle the decode stage, which creates output tokens. AWS plans a similar setup with Trainium and Cerebras, with an expected path to Amazon Bedrock in 2027.
Cerebras unveiled CS-4 in August and claimed up to 30 times the speed of GPU systems for some inference workloads. That figure needs context. It does not mean CS-4 beats NVIDIA at every AI task. Cerebras targets cases where token speed and response time have high value.
A benchmark can show a large gap on a selected workload, but production systems add software, data, network, power, and cost limits. Cerebras must show that its speed can create a clear business benefit after those factors enter the picture. More than 600 megawatts of data-center capacity now sit live or under contract, the company says.
Also Read - The Rise of Reasoning AI: How Enterprises are Moving Beyond Generative AI
Cerebras posted USD 210 million in core revenue for the second quarter of 2026, more than twice the prior-year level. Cloud revenue rose 281 percent from a year earlier. Core gross margin reached 41 percent. The company also raised its 2026 core revenue forecast to USD 880 million to USD 890 million.
Cerebras reported USD 25.4 billion in future performance obligations. That number does not equal current revenue. It represents contract value that can turn into revenue as Cerebras meets its commitments. Cerebras also raised about USD 5.55 billion in its NASDAQ debut at USD 185 per share. The stock opened at USD 350, then closed at USD 166.43 on October 2.
That market shift shows the next hurdle. Cerebras must turn large contracts and strong benchmarks into durable profit, not just high demand.
OpenAI remains the clearest proof point for Cerebras, yet it also creates a major customer concentration risk. Recent market pressure showed how fast sentiment can shift when reports raise questions about the exact role Cerebras may play in OpenAI’s fastest inference work.
OpenAI CEO Sam Altman called Cerebras a close partner on October 5 and said the firms have deep engagement on AI speed. The statement helped lift Cerebras shares after the sharp drop. Cerebras needs more customers with large production workloads to reduce its exposure to one buyer.
Why this Matters
Cerebras’ rise matters as AI demand pushes the industry beyond traditional GPU systems. Andrew Feldman’s strategy could reshape how companies handle AI inference, response speed, and computing costs. The outcome could also affect NVIDIA’s market position, cloud infrastructure, chip design, and the way major AI companies build future systems.
Feldman does not need a world where NVIDIA disappears. A more practical path would place Cerebras beside GPUs within large AI systems. AMD and AWS already point toward that model. Gimlet Labs has also agreed to take about 100 megawatts of Cerebras systems, with deployment expected over one to two years and cloud access in 2027.
Cerebras has also announced a 165-megawatt AI data center in Mikkeli, Finland, backed by a seven-year capacity deal. The company says the project could add EUR 1.0 billion to EUR 1.7 billion in regional investment.
NVIDIA still holds major advantages in software, hardware, network systems, and installed capacity. Cerebras has a different tool for a different part of the AI workload. Feldman’s challenge does not rest on a claim that one chip can win every task. It rests on a sharper proposition: AI may value response speed so much that specialized inference hardware earns a permanent place beside the GPU.
1. Who is Andrew Feldman?
Andrew Feldman is the CEO and co-founder of Cerebras, an AI computing company focused on specialized hardware.
2. How does Cerebras challenge NVIDIA?
Cerebras targets AI inference with wafer-scale processors designed to deliver very high token-generation speed.
3. Does Cerebras aim to replace NVIDIA GPUs?
Not necessarily. Its strategy increasingly focuses on working alongside GPUs and other processors for different parts of AI workloads.
4. Why is the OpenAI deal important?
OpenAI agreed to add 750 megawatts of Cerebras compute through 2028, giving Cerebras a major customer and a large-scale test of its technology.
5. What is the biggest challenge for Cerebras?
Cerebras must turn technical performance, major contracts, and data-center capacity into sustained revenue, strong margins, and a broader customer base.
