Cerebras pivots from hardware to cloud

The refile mattered because Cerebras was no longer selling investors a single customer story, it was selling a transition from one off $2M hardware deals to a broader cloud inference business. By April 2026, the company could point to $510M of 2025 revenue, cloud revenue of $152M, a mix shift to 30% inference, and new demand from coding agents and OpenAI, which made the business look less like a bespoke government hardware vendor and more like an AI infrastructure platform with repeat usage.1 2 3

• The old problem was concentration. G42 accounted for about 83% of 2023 revenue, and concern over that UAE exposure helped stall the October 2025 IPO. By the April 2026 refiling, Cerebras could point to MBZUAI, OpenAI, and AWS as additional anchors, making the revenue base look wider and more durable.2 3
• The product story also got simpler. Instead of mainly explaining wafer scale hardware sold to labs like Argonne and Livermore, Cerebras could show a cloud API used by Perplexity, Notion, Windsurf, and Cognition, where customers pay per token for very fast inference rather than buying and operating exotic machines themselves.2 4
• That cleaner narrative also helped relative valuation. Cerebras was targeting roughly a $35B IPO, around 69x 2025 revenue, versus Groq at about a 40x multiple on projected 2025 revenue after Nvidia acquired its team and licensed its technology. The pitch was not just better chips, but being the last scaled independent pure play in the category.1 3 5

Going forward, the investor debate shifts from whether Cerebras can escape a single customer risk to whether it can keep turning low latency inference wins into a larger recurring cloud business. If coding agents and multi agent workflows keep growing, the companies that own the fastest token generation path will look more like strategic AI platforms than component vendors.1 2 3 5