Recently, as information about Google’s seventh-generation TPU (TPU v7) continues to circulate in Silicon Valley, this new generation of AI-accelerating ASICs is being viewed as a lower-cost, high-efficiency alternative to NVIDIA’s Blackwell, attracting attention from global capital and industry. The market debate around the core question of whether ASICs are challenging the GPU hegemony has intensified.
NVIDIA, as the dominant player in the GPU ecosystem, was the first to issue a counter-signal. Influenced by the controversy over the AI valuation bubble and competitive pressure from TPUs, NVIDIA’s stock price has corrected by a cumulative 12.18% since November. To stabilize market expectations, the company emphasized in a recent report: “Compared to ASICs that only support specific tasks or frameworks, NVIDIA’s solutions offer greater versatility, higher performance, and better substitutability.” Several overseas institutions also pointed out that although Blackwell is more expensive, its performance in large model training scenarios still has a significant leading advantage, with NVIDIA maintaining a market share of over 90% in the AI acceleration chip sector.
However, Google’s response was more direct: if the external adoption rate of TPUs expands, it is expected to cut into and “slice off about 10% of NVIDIA’s annual revenue share.” This statement has also been seen by institutions as Google’s first clear commercial expectation for a path to GPU replacement.
In fact, to reduce dependence on NVIDIA GPUs, more global tech companies are accelerating their shift towards self-developed or procured ASIC solutions. Musk recently announced the establishment of an internal AI chip R&D team and is trial-deploying self-developed solutions in supercomputing centers; meanwhile, several U.S. media outlets have reported that Meta is considering using Google TPUs as an additional computing power source for its data centers.
According to the latest forecasts from overseas semiconductor analysis firms IDC and TrendForce, as CSPs accelerate their ASIC layouts, global AI ASIC shipments are expected to double year-on-year between 2026 and 2027, becoming a new mainline for AI infrastructure investment.
In the domestic market, Alibaba’s self-developed PPU is leading in the ASIC direction. Multiple institutions have analyzed that this chip has already surpassed NVIDIA’s A800 in inter-chip interconnect bandwidth, memory capacity, and other metrics, with some parameters comparable to H20. This indicates that domestic AI inference chips are gradually approaching high-end forms.
Listed companies in the A-share market are also accelerating their ASIC layouts. Chipone’s AI ASIC business achieved year-on-year growth in the third quarter of this year and continues to engage in deep cooperation with leading internet companies; Aojie Technology focuses on smart wearables, edge SoCs, and customized ASICs, currently having ample orders, with institutions expecting a significant increase in ASIC revenue around 2026.
Brokerage firms believe that the path to self-controllable domestic AI chips is steadily advancing, with multiple technical routes, including Alibaba’s Hanguang, Huawei’s Ascend, and Cambricon, achieving continuous iteration, which is expected to gradually alleviate the structural dependence of the domestic market on overseas AI computing power in the coming years. Western Securities pointed out that CSPs promoting self-developed chips can reduce external supply constraints in algorithm and model iterations, thus achieving continuity and predictability in large-scale deployments, and the industrial value will continue to rise.
Overall, the competition for AI computing power has transitioned from a “GPU monopoly” to a new pattern of “GPU + ASIC” advancing in parallel. As Google TPUs, various CSP self-developed chips, and domestic AI chip systems accelerate their maturity, the global computing power landscape is at a critical window of reconstruction.
