Google Launches Two New TPUs for the 'Agentic Era,' Cementing Marvell Partnership and Nvidia Challenge

Google has spent months quietly turning its Tensor Processing Units into must-have silicon for AI developers. According to Bloomberg AI and the Los Angeles Times, demand has become so intense that even direct competitors such as Meta and Anthropic are lining up to buy the chips. That momentum emboldened Alphabet to accelerate a second act, specialized inference processors that run trained models rather than train them. The timing was deliberate, training-chip margins are already shrinking as cloud giants over-provision, while inference workloads explode with every new user-facing feature.

Multiple outlets, including The Information via The Decoder and finance-focused sites like Yahoo Finance and GuruFocus, reported Google was in late-stage talks with Marvell Technology to co-design two new data-center chips. Ars Technica AI now confirms the deal is done, Google has unveiled two new TPUs: one memory-processing unit meant to sit beside existing TPUs, the other a ground-up inference accelerator. Volume targets remain eye-popping, nearly 2 million units across the next two years. Marvell shares leapt to record highs when the talks surfaced, then spiked again on launch day, underscoring how much Wall Street sees Google as a serious foundry customer rather than a one-off experiment.

CNBC and BBC both captured Nvidia’s public bravado, Jensen Huang told investors his GPUs remain “a generation ahead.” Yet the same stories note that Nvidia stock slipped on the headlines, a rare wobble for the $3 trillion titan. Google’s pitch to hyperscalers is simple, Ironwood TPUs already deliver better performance-per-watt for large-scale inference, and the newly launched Marvell-designed parts will double down on power efficiency for AI agents and reasoning models. The subtext, if you’re running Gemini, you can do it faster and cheaper on Google silicon than on a rented H200 cluster.

The first chip, dubbed TPU v6e, is the inference monster. Google claims it cranks out 2.7x more tokens per watt than an Nvidia H200 on typical Gemini workloads. The second chip, TPU v6t, handles training but is optimized for post-training fine-tunes and reinforcement learning loops that feed back into agentic systems. Both are built on a 5-nanometer process at Samsung’s Austin fab, with Marvell supplying the SerDes, interconnect fabric, and custom SRAM blocks. Shipments begin in July to internal Google Cloud and to “select design partners” that include, yes, Meta and Anthropic.