OpenAI Explores Custom Chips to Reduce NVIDIA Dependence
OpenAI is reportedly exploring alternatives to NVIDIA's chips amid concerns over high compute costs. The company is considering a partnership with manufacturer Foxconn and developing its own custom silicon to better tailor hardware for its AI workloads and gain supply chain control.
- NVIDIA currently holds a dominant position in the AI accelerator market, with an estimated market share ranging from 80% to over 90% for GPUs used in AI training. - The cost of training state-of-the-art AI models is a significant driver for seeking alternatives; training for GPT-4 was estimated to cost between $78 million and over $100 million, with Google's Gemini Ultra estimated at $191 million. Some projections suggest future models could cost over $500 million for a single training run. - The announced collaboration with Foxconn focuses on co-designing and manufacturing next-generation AI data center infrastructure in the U.S., including racks, networking, cooling, and power systems, to strengthen the domestic supply chain. Foxconn plans to invest $2 billion to $3 billion annually in AI development. - This move follows a broader industry trend where major technology companies are developing their own custom AI silicon to optimize for specific workloads and reduce costs. Notable examples include Google's Tensor Processing Units (TPUs), Amazon's Trainium and Inferentia chips, Microsoft's Maia AI Accelerator, and Meta's Training and Inference Accelerator (MTIA). - Beyond the Foxconn partnership for infrastructure, OpenAI is also reportedly working with Broadcom to design a custom AI processor, with plans to use Taiwan Semiconductor Manufacturing Company's (TSMC) advanced 3-nanometer fabrication technology. - Developing a custom Application-Specific Integrated Circuit (ASIC) is a resource-intensive process with high non-recurring engineering costs that can range from hundreds of thousands to millions of dollars and development timelines that can span months or even years.
