Google (GOOGL.US) teamed up with MRVL.US (MRVL.US) to “reinvent the core”: memory processing unit+new TPU challenges Nvidia and makes another surprising move

The Zhitong Finance App learned that people familiar with the matter revealed that Google's parent company Alphabet (GOOGL.US) is negotiating cooperation with Mewell Technology (MRVL.US) to jointly develop two new chips aimed at improving the operation efficiency of artificial intelligence models. One chip is a memory processing unit, which is intended to work in collaboration with Google's self-developed TPU (Tensor Processing Unit); the other is a new TPU specially designed for AI inference tasks, that is, the stage where the model provides services to users rather than learning from data. Maywell will assume a design service role similar to how MediaTek participated in Google's latest Ironwood TPU project. Currently, the relevant negotiations have not yet resulted in a formal contract.

Google has been working to position TPU as an alternative to Nvidia (NVDA.US) GPUs. Currently, Nvidia's GPUs still dominate the AI chip market. As Google tries to prove that its huge investment in AI is turning into commercial returns, sales of TPU have become an important driving force for Google's cloud business growth.

According to people familiar with the matter, Google and Mwell Technology hope to complete the design of the memory processing unit as early as next year, and then move it to trial production.

Just a few days ago, Broadcom (AVGO.US) just finalized a TPU cooperation agreement extending until 2031. The negotiations also reflect that Google is shifting its focus to reasoning — the main computing power cost item in AI applications. Currently, the customized ASIC market is expected to grow by 45% in 2026 and is expected to reach US$118 billion by 2033.

Judging from the point of time, Google is not replacing Broadcom, but rather adding a third design partner to its supply chain. Currently, Google's custom chip supply chain includes Broadcom, which is responsible for the high-performance TPU variant, MediaTek, which is responsible for the cost-optimized “e” series TPU (20-30% lower cost), and TSMC, which is responsible for manufacturing. The core of this strategy is diversification, not substitution.

The value of Mywell

Maywell's data center business achieved a record revenue of US$6.1 billion in the fiscal year ending February 2026, with total revenue of US$8.2 billion, an increase of 42% over the previous year. Its custom chip business has annual revenue of about 1.5 billion US dollars, and has received design orders from 18 cloud providers to build chips for customers such as Amazon (Trainium processor), Microsoft (Maia AI accelerator), and Meta (new data processing unit). In addition, it has also undertaken the existing design work for Google's Axion ARM CPU.

At the end of March this year, Nvidia invested 2 billion US dollars in Mewell to integrate Mywell's custom chips and network products with Nvidia's interconnect architecture through the NVLink Fusion framework. The collaboration enabled Maywell to simultaneously stand at the intersection of two major ecosystems, GPU and ASIC. In December 2025, Maywell acquired Celestial AI for up to 5.5 billion US dollars and obtained photonic interconnection technology. Maywell CEO Matt Murphy said it would create “the most complete connectivity platform for AI and cloud customers.” Murphy aims to gain a 20% share of the custom AI chip market and expects revenue growth of around 30% year over year in the 2027 fiscal year.

Maywell's stock price has accumulated a cumulative increase of about 50% since the beginning of the year. In April alone, it rose 30% due to news of the Nvidia partnership and Google negotiations. Barclays analyst Tom O'Malley raised Maywell's rating to “Overweight” and raised the target price from $105 to $150.

GPU (Graphics Processing Unit, Graphics Processing Unit) is dominated by Nvidia. It was initially designed for graphics rendering, and has become the mainstream chip for AI training and inference due to its powerful parallel computing capabilities. It is characterized by strong versatility, mature software ecosystem (such as CUDA platform), and is widely used in computational power-intensive scenarios such as large model training and autonomous driving. Currently, GPUs account for more than 80% of the AI chip market, and Nvidia's H100 and B200 products are regarded as industry benchmarks.

TPU (Tensor Processing Unit) is a dedicated AI acceleration chip developed by Google. It uses an ASIC (Special Integrated Circuit) architecture and is deeply optimized for its TensorFlow framework and AI operations (especially matrix multiplication). Compared to GPUs, TPUs are more energy efficient and have lower latency in specific inference tasks, but their versatility is limited. Google provides TPU computing power to the outside world through cloud services, using it as a differentiating competitive point, with the aim of reducing dependence on Nvidia and providing cost-effective AI computing power options for cloud customers. Currently, TPU has been iterated to the sixth generation (Trillium) and is the core hardware asset of Google's AI strategy.