US AI chipmaker Cerebras may file IPO as early as next week

Cerebras Systems is preparing to file for a US IPO as early as next week, according to sources.

The Sunnyvale, California-based company develops high-performance chips for AI workloads and competes with Nvidia and other AI chip makers.

The company had previously withdrawn its IPO filing in October 2025 after raising over US$1 billion at an US$8 billion valuation, following a US national security review of a minority investment by UAE-based G42.

In the upcoming IPO filing, G42 is not listed as an investor.

Cerebras is targeting a Q2 2026 listing, as US IPO activity in 2025 reached US$46.2 billion, the highest since 2021.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Cerebras’ IPO hopes hinge on breaking Nvidia’s ecosystem lock-in

Cerebras withdrew its IPO in October 2025 after a $1.1 billion Series G at an $8.1 billion valuation. CFIUS (Committee on Foreign Investment in the United States) cleared a G42 stake in late March 2025. G42, a UAE-based AI company, drove about 80% of sales and had roughly $1.4 billion in orders 1 2 3.
Its wafer-scale chip (a single processor from an entire silicon wafer) delivers about 3.5x FP8 (8-bit floating point) performance over Nvidia H100 systems on equal rack space and power 4. It also ran nearly 5x faster than Nvidia’s Blackwell GB200 on GPT-OSS‑120B, an open-source 120‑billion‑parameter GPT‑class model 5. Programming is simpler than multi‑GPU clusters, yet Nvidia’s CUDA, its proprietary GPU programming platform, keeps developers tied to its stack 3 4. Revenue rose from under $6 million in Q2 2023 to $70 million in Q2 2024, with 2023 at $79 million, so it must grow past G42 to support the $8.1 billion valuation 6 3 2.

AI infrastructure startups can benchmark Cerebras against Nvidia to get better GPU pricing or build hybrid systems

Cloud providers plus MSPs can pitch CS‑3, its current‑generation system for training plus inference, when speed matters. Cerebras logs 3,000 tokens per second at $0.75 per million tokens versus Baseten, an AI model‑serving platform, at 650 tokens per second on GB200 Blackwell at $0.50 per million tokens 5.
Startup founders building real‑time apps such as code generation, reasoning agents, or low‑latency search can use these numbers to bargain with GPU vendors, then run hybrids that send latency‑critical jobs to Cerebras and route batch runs to cheaper GPU clusters 2 5.