OpenAI multi-vendor chip strategy

This reveals that OpenAI is building a multi supplier compute stack so no single chip vendor can set its speed, costs, or product roadmap. Nvidia remains the backbone for giant training and general purpose clusters, while Broadcom is helping OpenAI build its own inference chip, and Cerebras is filling the ultra low latency tier where fast reply speed matters more than standard GPU flexibility. Potential Arm involvement fits the same pattern of owning more of the stack over time.1 2 3 4 5 6

• Nvidia is still the base layer. OpenAI and Nvidia signed a plan for at least 10GW of Nvidia systems, with first deployment set for the second half of 2026 on Vera Rubin. Nvidia also said it could invest up to $100B over time as each gigawatt is built out, showing that OpenAI still needs hyperscale GPU supply even while diversifying.1 2
• Broadcom is about cost and control. OpenAI and Broadcom said they will deploy 10GW of OpenAI designed accelerators by the end of 2029. Jalapeño is built specifically for LLM inference, not adapted from older graphics workloads, which matters because ChatGPT style serving is mostly about answering huge volumes of prompts cheaply and reliably.3 4
• Cerebras is the specialty tier, not the replacement. OpenAI and Cerebras agreed to deploy 750MW of wafer scale inference capacity through 2028. Cerebras and its filings describe this as a latency focused serving layer for OpenAI workloads, which matches a split architecture where GPUs handle the broad fleet and custom or wafer scale systems handle the fastest response paths.5 6

The next step is a more vertically integrated OpenAI, where model design, chip design, and data center layout are tuned together. If Jalapeño and later generations work at scale, OpenAI can shift more inference off merchant GPUs, protect margins, and turn hardware choice into a product advantage, especially in coding, voice, and agent workloads where latency directly shapes user behavior.3 4 5 6