The move that could Redesigning the balance of artificial intelligence comes from Meta, which according to media sources is evaluating the adoption of Google's TPU chips in its data centers. A negotiation that, if confirmed, would represent one of the most significant openings ever towards an alternative to Nvidia GPUs, which has been the cornerstone of global AI infrastructure until now. The rumors hit Nvidia shares yesterday, with a 7% drop in intraday trading and a close of 2,6%. Today, the group is attempting a rebound in the pre-market, while Alphabet remains in full swing.
Why TPUs Now Scare GPUs
If the Google-Meta axis were to consolidate, it would mark a turning point. For the first time, one of the world's largest consumers of computing power is evaluating the integration of Tensor Processing Units directly in their data centers, starting from 2027, and with an operational advance already in 2026 thanks to the rental of computing power from the Mountain View cloud.
The TPUs are the heart of Google's hardware strategyCustom-designed chips, extremely efficient at AI workloads, and until now, only available as a cloud service. The idea of selling them directly breaks with the group's traditional strategy and opens the door to a market worth hundreds of billions of dollars.
Gemini 3 and the new AI balance
The timing is not at all coincidental. Google just unveiled Gemini 3, the new linguistic model trained entirely with Tpu and, according to several analysts, superior to the systems that power ChatGPT. The Financial Times It has garnered assessments that the release could "reset the AI hierarchy," evoking the precedent of DeepSeek, the Chinese startup that shook up the industry earlier this year with a powerful, low-cost training model. The message to investors is unmistakable. In artificial intelligence, hardware is as important as models, and Google is putting pressure on a market accustomed to Nvidia dominance.
Nvidia sinks, Alphabet runs
The market reaction was immediate. On Monday Nvidia was overwhelmed by the leak, burning up up to $150 billion in market capitalization in the most turbulent phases and oscillating between 169 and 178 dollars before rising again in the final phase. Alphabet, on the other hand, continued to run, supported by the possibility that TPUs may finally gain space in the hardware intended for training generative models. In the background is the scale of the Meta investments: between 70 and 72 billion dollars of capex in 2025, a volume so large that every technological choice becomes a message for the entire market.
Nvidia is trying to today recover in the pre-market, but the balance remains precarious. Jensen Huang's group reiterates that it is "a generation ahead" of its competitors and claims the strength of an ecosystem that combines hardware, software, libraries, and networks. But the mere fact that one of its most important customers is evaluating an alternative was enough to to undermine certainties.
