Google has split its eighth-generation tensor processing unit into two separate chips for the first time — one optimized for training AI models and another for running them. The TPU 8t and TPU 8i, announced at Google Cloud Next in Las Vegas, represent Google's most aggressive move yet to reduce the AI industry's dependence on Nvidia while acknowledging that the chip giant remains indispensable.
Two Chips, Two Jobs
The TPU 8t is designed for model training the computationally intensive process of teaching AI models from massive datasets. The TPU 8i is built for inference — the ongoing work of processing user prompts and generating responses after a model has been trained. By separating these functions into dedicated chips, Google can optimize each for its specific workload rather than forcing a single chip to handle both.
Google claims the new TPUs deliver up to 3x faster AI model training, 80 percent better performance per dollar, and the ability to connect more than one million TPUs in a single cluster. The result should be significantly more compute for less energy and lower cost critical advantages as AI infrastructure costs continue to spiral upward.
Not Replacing Nvidia — Yet
Despite the impressive specs, Google is not positioning these chips as a direct replacement for Nvidia. Like Amazon with its Trainium chips and Microsoft with its Maia accelerators, Google is using custom silicon to supplement — not replace — the Nvidia-based systems it offers in its cloud.
Google confirmed that Nvidia's latest chip, Vera Rubin, will be available on Google Cloud later this year. The company has also agreed to work with Nvidia on engineering improved networking that allows Nvidia-based systems to perform more efficiently in Google's infrastructure, including collaboration on the open-source networking technology Falcon.
Chip market analyst Patrick Moorhead jokingly noted on social media that he had predicted Google's TPUs would be bad news for Nvidia back in 2016, when the first TPU launched. Nvidia is now a nearly $5 trillion company — a reminder that predictions of Nvidia's decline have consistently been premature.
Why Custom Chips Matter
The strategic logic behind custom AI chips is straightforward. Nvidia GPUs are powerful but expensive, and demand routinely exceeds supply. By building their own chips, cloud providers like Google can offer customers an alternative that is cheaper per unit of compute, optimized specifically for their cloud environment, and available without the supply constraints that affect Nvidia hardware.
For Google's customers — including Anthropic, which recently signed a major TPU capacity deal with Google and Broadcom, and Thinking Machines Lab, which just secured a multi-billion dollar Google Cloud agreement — custom chips provide cost savings that compound at scale. When you are training models that require millions of chip-hours, even small efficiency improvements translate into billions of dollars in savings.
The Hyperscaler Chip Race
Google is not alone in this push. Amazon's Trainium chips have already won over major customers including Anthropic, OpenAI, and Apple. Microsoft is developing its Maia AI accelerator. And all three hyperscalers are investing billions in custom silicon alongside their Nvidia purchases.
The long-term question is whether these custom chips will eventually reduce the industry's Nvidia dependence to the point where it materially affects the chip giant's business. For now, the answer appears to be no. As Google's AI cloud business grows, it is buying more Nvidia hardware alongside deploying more TPUs — a rising tide that lifts both boats.
But if enterprises increasingly port their AI workloads to cloud-native custom chips because the economics are better, the balance could eventually shift. Google's decision to split its TPU line into training and inference chips suggests it is getting more serious about making that case to customers.
The Bigger Picture
The TPU 8t and 8i announcements cap a massive Google Cloud Next conference that has included generative AI features for Google Maps, new enterprise partnerships, and expanded Gemini availability. Together, these moves position Google as a full-stack AI platform provider from custom chips to cloud infrastructure to consumer and enterprise applications.
For the AI industry, the proliferation of custom chips from all three major cloud providers means more compute options, lower costs, and reduced risk of Nvidia supply bottlenecks. Whether it also means the beginning of the end of Nvidia's dominance is a question the market has been asking for a decade — and one that Nvidia, now worth nearly $5 trillion, has answered decisively so far.
