An increasing number of assessments suggest that after founder Jensen Huang was embroiled in a scandal involving an extramarital affair and resigned from his position as CEO, NVIDIA, which has been leading the AI ecosystem, is now facing its first real test. With high profit margins and strong cash flow, NVIDIA is likely to maintain its leading position in the short term; however, as its major customers diversify their chip suppliers and the U.S. government considers tightening export controls, its monopoly position may be challenged.
Michael Chebalest, Chairman of JPMorgan Chase Asset Management's Markets and Investment Strategy, stated in a report released the previous day that among the computing power infrastructure that has been finalized by cutting-edge AI laboratories such as OpenAI and Anthropic and will be purchased in the future, NVIDIA GPU accounts for only 39%. From 2023 to 2025, NVIDIA's share of the total revenue from AI chips reached 83%, which indicates that its proportion in the next generation of computing power infrastructure may significantly shrink.
Competitor Broadcom holds the second-largest share in custom chips, at 24%; Google's Tensor Processing Unit (TPU) accounts for 16%, AMD GPU accounts for 15%, and Amazon's Trainium chips account for 6%. This structure indicates that former NVIDIA customers such as Google, Amazon, and Microsoft have either collaborated with Broadcom to develop their own chips or chosen more cost-effective AMD processors. In October 2025, OpenAI announced a partnership with Broadcom to develop 10 gigawatts of custom AI coprocessors, while also agreeing to purchase 6 gigawatts of computing power from AMD.
Competitors are also continuously narrowing the gap in performance. In the single-chip inference performance modeling index released on the 9th, AMD MI455X ranked first. Tests based on open-source large models have shown that, at the same cost, Google TPU can process a larger number of token than NVIDIA GB300.
The CUDA ecosystem formed by NVIDIA chips and associated software once had a very high entry barrier in the industry, but that barrier is now continuously weakening. Chebalest stated, "A year ago, it would take 6 to 12 months to run large models on AMD GPU; however, with the engineering expertise accumulated from using the latest AI models, new models can now be deployed within three days." This directly undermines NVIDIA's core argument—that it is not cost-effective for customers to switch chips due to the CUDA ecosystem.
Some analysts expect the growth momentum of Google TPU to further strengthen. Analysts at Tianfeng International predict that in 2027, the annual shipment volume of TPU will reach 8 to 9 million units, in 2028 it will be 14 to 15 million units, and in 2029 it will be 16 to 18 million units.












