Google Ironwood TPU: 10x performance + four partners taking on Nvidia

According to a deep report by Bloomberg and an official announcement from Google, on April 22 Google officially expanded its in-house AI chip lineup: the inference-focused Ironwood (7th-generation TPU) is now fully available on Google Cloud, and it is also simultaneously launching next-generation design collaborations with four partners—Broadcom, MediaTek, Marvell, and Intel. The goal is to use a custom chip supply chain to directly challenge Nvidia’s dominant position in the AI compute market.

Ironwood: 7th-generation TPU, first designed specifically for inference

Ironwood is the 7th-generation product in Google’s TPU series and the first chip designed specifically for inference under the company’s strategy of “training and inference separation.” The specifications Google disclosed are as follows: peak performance per chip is 10x that of TPU v5p, with 192GB of HBM3E memory, memory bandwidth of 7.2 TB/s, and each individual superpod can scale to 9,216 liquid-cooled Ironwood units. Total FP8 compute reaches 42.5 exaflops.

Google’s official statement says Ironwood has been “fully opened for Google Cloud customers to use,” and this year shipment volume is expected to reach a “million-unit” level. Anthropic has committed to using up to 1 million Ironwood TPUs. Meta has signed a “multi-year contract worth billions of dollars” to use TPUs via Google Cloud.

Four-partner division of labor: training goes to Broadcom, inference goes to MediaTek

Google’s next-generation chip supply chain is clearly divided as follows:

Partner Code Role Highlights Broadcom Sunfish Training-focused Dedicated Continue the existing TPU collaboration relationship; leads large training nodes MediaTek (聯發科) Zebrafish Inference-focused Claimed to be 20–30% lower cost than the Broadcom solution Marvell In negotiations Memory Processing Unit (MPU) + additional inference TPU, optimized for HBM and inference Intel Not disclosed Involved in design, strengthening supply-chain diversification

This is the first “four-partner parallel, with training and inference clearly split” model seen in the AI industry. By dispersing IP risk and negotiating power, Google avoids the structural dependency of Nvidia as a single-supplier setup. The roadmap extends to TPU v8 by the end of 2027, to be produced using TSMC’s 2nm process.

Strategic significance: what Google is challenging isn’t a single chip, but the supply chain

Over the past three years, Nvidia has nearly monopolized the AI chip market, with CUDA software ecosystems and the H100/GB200 chips forming a double moat. Google’s Ironwood and four-partner strategy are not aimed at “surpassing at a single point in specs.” Instead, the goal is to replicate Nvidia’s logic in the industry—a “standardized platform + multi-customer procurement” model—so that TPUs are not just for Google’s own use, but a commercial compute option that major AI companies like Anthropic and Meta can share.

The significance of Anthropic’s commitment to 1 million TPUs is especially critical: this is the largest compute commitment from any single AI company outside Nvidia. It also complements Anthropic’s 5GW/100-billion AWS commitment with Amazon reached through AWS on 4/20—one side ties to AWS Trainium, the other ties to Google TPU. With Anthropic’s “dual custom-chip” strategy, Nvidia dependency is reduced. Meta, meanwhile, is the first to publicly include TPUs in its own AI training/inference workloads, sending another signal.

Market reaction and industry linkage

Before this disclosure, MediaTek had already been seen as a “beneficiary of Google’s custom chips.” The exposure of the Zebrafish codename is the first time MediaTek is directly listed as a design partner for Google’s inference chips. This extends the narrative line of “non-Nvidia chip alliances,” including the recent AMD × GlobalFoundries silicon photonics and Marvell × Google MPU initiatives.

Although Nvidia still has GB200 and its next-generation Rubin platform supporting the same period, the compute mix on the customer side is shifting from “all Nvidia” toward “Nvidia + TPU + AWS Trainium” operating on three tracks in parallel. This also means that for TSMC’s 2nm capacity, the four major customers—Google, Nvidia, Apple, and Amazon—are lining up, and negotiating power around foundry supply is continuing to rise.

This article, “Google Ironwood TPU: 10x performance + four-partner lineup to take on Nvidia,” first appeared on Lianxin ABMedia.