Broadcom & OpenAI build LLM accelerator for gigawatt-scale AI
Gigawatt-scale AI accelerators: Broadcom partners with OpenAI on custom silicon...
Broadcom and OpenAI have announced a joint venture to build a new LLM (large language model) accelerator chip, targeting deployment at gigawatt-scale data centers. The custom silicon will be optimized for transformer-based models, promising up to 10x better energy efficiency compared to current GPUs. This partnership combines Broadcom's expertise in networking and custom chip design with OpenAI's advanced model architectures and training insights.
The accelerator is expected to handle inference and training workloads for models with over a trillion parameters, while consuming significantly less power per operation. The move positions both companies to compete directly with NVIDIA's dominance in AI hardware. The first prototypes are slated for 2027, with production for hyperscale clients like Microsoft and Google. This could reshape the AI infrastructure landscape, making large-scale deployments more cost-effective and environmentally sustainable.
- Broadcom and OpenAI co-develop a custom LLM accelerator for gigawatt-scale deployments
- Targets 10x better energy efficiency than current GPUs for transformer models
- First prototypes expected by 2027, targeting hyperscale data centers
Why It Matters
Custom AI hardware could lower costs and energy use, democratizing access to massive LLM deployments.