OpenAI Unveils Custom AI Chip with Broadcom
OpenAI is now powering its Codex-Spark coding model with a custom, "plate-sized" chip co-designed with Broadcom, sidestepping Nvidia for this key workload. The chip is part of a reported $10 billion development program, with early users describing the model's performance as "unusually fast". The partnership represents one of the largest investments in custom AI silicon and a significant move by a major model developer to control its hardware stack.
- This partnership follows a broader industry trend of major AI players developing custom silicon to optimize for specific workloads and reduce reliance on third-party vendors. Companies like Google with its TPU, Amazon with Trainium, Microsoft with Maia, and Meta with Artemis have all invested in in-house chip design. - Broadcom has established itself as a key enabler for companies developing custom AI ASICs, providing design and fabrication partnership. In 2024, Broadcom's AI-related revenue reached $8 billion. - While OpenAI is partnering with Broadcom for this custom chip, it also recently announced that its GPT-5.3-Codex-Spark model is the first to run on hardware from Cerebras, known for its large "Wafer Scale Engine" technology designed for low-latency AI tasks. - The move to custom silicon is often driven by the need for greater efficiency in inference workloads, which are projected to constitute up to 70% of all AI compute by 2027. Custom ASICs can be optimized for the predictable and repetitive nature of these tasks, improving power, latency, and cost. - This initiative is part of a larger strategy by OpenAI to diversify its hardware supply chain, which has also included using Google's TPUs to lower costs for inference. Concurrently, OpenAI has also deepened its partnership with Nvidia, announcing a plan to deploy 10 gigawatts of NVIDIA systems for its next-generation AI infrastructure. - The development of sophisticated custom AI chips represents a significant investment, with design and manufacturing costs for chips on advanced nodes potentially reaching into the hundreds of millions of dollars. However, for high-volume applications, the long-term cost savings on a per-unit basis can be substantial. - Nvidia, the dominant player with over 80% of the AI chip market share, is responding to the custom silicon trend by offering its NVLink Fusion technology. This allows hyperscalers to build semi-custom systems that integrate their own chips with Nvidia's GPUs and CPUs. - The new Codex-Spark model, powered by specialized hardware, has demonstrated significant performance improvements on specific benchmarks, achieving 77.3% accuracy on Terminal-Bench 2.0, an increase from the 64% achieved by a previous version.
