AMD and Broadcom press Nvidia in inference chips

Nvidia remains the leading supplier for training AI models, backed by CUDA and a developer ecosystem optimized around its GPUs. The next phase of demand is shifting toward inference, where Bloomberg Intelligence projects a 32% compound annual rate through 2032 and a market reaching $1.3 trillion, compared with $658 billion for AI training. Inference relies more on memory access and operating cost efficiency than raw compute, giving rivals room to compete.

AMD is targeting that opening with lower-cost GPUs, ROCm software improvements, chiplet designs built for high-bandwidth memory, and acquisitions focused on memory optimization and inference chips. Its customer wins include OpenAI, Meta Platforms, Anthropic, and Microsoft, including multiyear deals worth $100 billion each with OpenAI and Meta Platforms. Broadcom is pursuing custom ASICs, helping customers such as Alphabet, OpenAI, and Meta Platforms turn chip designs into scalable products.

Nvidia still reported second-quarter revenue of $96.2 billion, up 106% year over year, and guided to $108 billion for the current quarter. Its stakes in OpenAI, Anthropic, SpaceX, and Safe Superintelligence widen its exposure across AI labs, but concentrated customer sales and short-term compute commitments could amplify any pullback in orders.