Nvidia's dominant position in the AI chip market faces growing competition, but challengers are pursuing different goals rather than a full-spectrum replacement. Nvidia's strongest hold is in training, where its share exceeded 90% in 2025, according to Silicon Analysts. Its inference share sat between 60% and 75% over the same period, with the gap between the two ranges being roughly 15 to 30 percentage points. Some companies, like Groq, design chips only for inference; Groq was "established in 2016 for inference" and markets itself as "the only custom-built inference chip." Google's latest Tensor Processing Unit, Ironwood, is its first TPU "designed specifically for inference," and it is only accessible through Google Cloud. A key distinction among challengers is whether they sell chips or use them internally. OpenAI's partnership with Broadcom, announced in October 2025, involves co-developing custom AI chips that will draw up to 10 gigawatts of power.