--- title: "Is the \"Dual Supplier\" strategy taking shape? It is reported that Google is partnering with MediaTek to develop the eighth generation TPU" type: "News" locale: "en" url: "https://longbridge.com/en/news/284222478.md" description: "Google collaborates with MediaTek to develop the eighth-generation Tensor Processing Unit (TPU), splitting training and inference tasks into separate chips. The TPU 8t is designed by Google, focusing on training, while the TPU 8i is designed by Broadcom, focusing on inference. Both chips are planned to be supplied later in 2026, with performance significantly improved over the seventh-generation TPU, increasing by 2.8 times and 80% respectively. This move responds to the trend of differentiation in AI workloads, with a design philosophy centered around scalability, reliability, and efficiency" datetime: "2026-04-27T13:07:38.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/284222478.md) - [en](https://longbridge.com/en/news/284222478.md) - [zh-HK](https://longbridge.com/zh-HK/news/284222478.md) --- # Is the "Dual Supplier" strategy taking shape? It is reported that Google is partnering with MediaTek to develop the eighth generation TPU According to media reports citing informed sources, MediaTek may participate in the design of some upcoming chips for Google (GOOGL.US). It is reported that MediaTek may be responsible for the I/O chip and backend design services (i.e., packaging). Google will be responsible for the computational core design of the training-focused TPU 8t. Broadcom (AVGO.US) will be responsible for the design of the inference-focused TPU 8i. The report also states that TSMC will use its 3-nanometer technology to manufacture the TPU 8t and provide its CoWoS advanced packaging technology. At the Google Cloud Next 2026 conference held in Las Vegas last Wednesday, Google Cloud announced two new products of the eighth-generation Tensor Processing Unit (TPU) — TPU 8t and TPU 8i. This is the first time Google has split training and inference tasks into separate chips, marking a significant shift in its AI hardware strategy. Both chips are scheduled to be officially supplied later in 2026. Compared to the seventh-generation Ironwood TPU released last November, the TPU 8t offers a 2.8 times performance improvement at the same price, while the TPU 8i sees an 80% performance increase; both chips have more than doubled performance per watt compared to the previous generation, with TPU 8t reaching 124% and TPU 8i reaching 117%. The division of the eighth-generation TPU into two is a direct response from Google to the increasingly differentiated trends in AI workloads. Pre-training, post-training, and real-time inference have significantly diverged in computational characteristics: training tasks pursue extreme throughput and scalability, while inference tasks are more sensitive to latency and concurrency. A single chip is difficult to optimize efficiency for both types of scenarios simultaneously. Google pointed out in its technical blog that the design philosophy of the eighth-generation TPU revolves around three pillars: scalability, reliability, and efficiency. Both chips share the core genes of Google's AI software stack but have been specifically optimized for different bottlenecks. It is worth mentioning that this is not the first collaboration between Google and MediaTek. In December last year, reports indicated that as demand for Google's AI chip TPU surged, Google had doubled the order volume for the next-generation "TPU v7e" chips assigned to MediaTek to ensure the supply of the next generation of products. This move indicates that Google is eager to reduce risks through a diversified supplier strategy, and MediaTek's position in Google's AI chip landscape is rapidly rising, expected to gain substantial benefits from this additional order. Google's TPU, with its excellent energy efficiency ratio, not only meets the enormous computational demands internally but also attracts a large number of external customers. In response to the surging market interest, Google has decided to expand its supply chain system, no longer relying solely on Broadcom, but instead bringing in MediaTek as a new key partner to address the supply-demand imbalance. Reports indicate that Google's choice of a "dual supplier" strategy (Broadcom + MediaTek) is backed by deep tactical considerations. Currently, the core bottleneck limiting AI chip shipments lies in TSMC's CoWoS advanced packaging capacity. By diversifying orders, Google can leverage the long-standing close cooperation between MediaTek and TSMC to more efficiently secure scarce CoWoS quotas and advanced process capacity ### Related Stocks - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) - [GGLS.US](https://longbridge.com/en/quote/GGLS.US.md) - [GGLL.US](https://longbridge.com/en/quote/GGLL.US.md) ## Related News & Research - [Forget one chip to rule them all: With TPU 8, Google has an AI arms race to win](https://longbridge.com/en/news/283660813.md) - [Google Expands TPU Push As Demand Surges, Targets AI Inference Market](https://longbridge.com/en/news/283391187.md) - [Alphabet's (GOOGL) $40B Anthropic Gambit: Why Google Is Doubling Down on Its Biggest Al Rival](https://longbridge.com/en/news/284085625.md) - [Google Conference 'Further Proves that Demand Is Not Just Anthropic and OpenAI'](https://longbridge.com/en/news/284009280.md) - [Sorry, Microsoft: Google Just Put The 'Golden Handcuffs' On Geothermal Leader Fervo](https://longbridge.com/en/news/283533404.md)