Google is aggressively expanding its artificial intelligence chip product line, pushing competition with NVIDIA to a new stage. This move comes at a critical juncture when the demand for AI chips remains strong and the competitive landscape is accelerating its transformation.
Alphabet On Monday, it was announced the launch of new variants of the tensor processor (TPU) chips – TPU 8tx and TPU 8i Max, which are tailored for confidential computing and long-context window use cases, respectively. With this, Google's (TPU) chip series has expanded to four models, covering a wider range of market needs. Google's head of chip business, Amin Vahdat, stated: "The future belongs to specialization, and there is still tremendous room for value enhancement through specialization."
According to Bloomberg Industry Research, the sales of TPU, which includes related software and hardware, are expected to soar from $12 billion this year to over $135 billion by 2028. On the client side, Apple, Meta, Oracle, Anthropic, and Jump Trading have all already adopted TPU, and OpenAI also has plans to make purchases.
Product iteration has accelerated, from one new model every two years to multiple models released within a year.
The TPU 8tx, TPU 8i, and Max launched by Google this time are not entirely new chip architectures; rather, they are customized modifications based on the existing eighth-generation TPU architecture to meet specific requirements.
TPU To meet the security and compliance requirements of government agencies, as well as regulated industries such as finance and healthcare, 8tx adopts chip-level data encryption technology, ensuring that data remains encrypted throughout the processing stage without the need to be decrypted into plaintext in memory. This chip works in conjunction with CPU from Supermicro Semiconductor (AMD), representing an extension of their collaborative efforts.
TPU 8i Max is specifically designed for AI applications that require processing a large amount of contextual information. By encapsulating two TPU chips together, it enhances computing power and memory space, making it suitable for scenarios such as drug research and development, mathematical operations, and code writing. This new product will be launched next year.
It is worth noting that the iteration of Google's chips is clearly accelerating. In the past, the company typically released a new processor every two years; this year, they have launched two new processors simultaneously, and now they have added two more variant models in the same year.
Price advantages and a specialized approach attract a diverse range of customers.
According to Mandeep Singh, an industry research analyst at Bloomberg, one of the core competencies of TPU is that its price is about 30% lower than that of NVIDIA GPU, yet it performs similarly in terms of data throughput.
This advantage is gaining recognition from mainstream institutions. Quantitative trading firm Jump Trading has deployed thousands of TPU for the training of AI models. Its Chief Technology Officer, Alex Davies, stated in an interview that the company still primarily uses NVIDIA chips, but some employees highly recognize the performance of TPU in completing specific tasks.
"Adopting different approaches to certain issues is actually quite simple, and one can learn a lot from it," said Davies. He also pointed out that in scenarios where TPU performs exceptionally well, its higher cost efficiency can further enhance operational capabilities.
Oracle announced last month that it has reached a partnership with Google, allowing its cloud customers to use TPU for the first time starting next year. Karan Batta, Senior Vice President of Oracle Cloud Infrastructure, stated: "As the workloads of AI become increasingly complex, more options are needed." Previously, Oracle cloud customers could only choose from AI chips manufactured by NVIDIA and AMD.
Endorsed by flagship customers, Anthropic holds substantial procurement options.
At the high-end customer level, AI company Anthropic is already an important user of Google TPU. According to Bloomberg, based on its recent regulatory documents, Anthropic holds options to purchase up to $150 billion worth of TPU between 2028 and 2031. OpenAI also plans to make purchases of TPU.
Google itself, in its role as a AI software developer, also provides unique support for its chip business. The TPU research and development team can obtain Google's latest Gemini models in advance to optimize energy efficiency and identify defects in chip design. Vahdat indicates that AI is reducing the risk of major errors in semiconductor design; such errors, once they occur, can cost up to about $100 million and result in months of project delays.
NVIDIA still holds a dominant position, but the competitive landscape is beginning to open up.
Despite Google's increasingly strong offensive, NVIDIA's dominant position in the AI chip market remains unshakable in the short term. The demand for its GPU continues to exceed supply capacity, which in turn leaves room for growth for competitors such as Google.
The article points out that NVIDIA itself is also transitioning towards specialization. Its product line has extended into the CPU domain, and by acquiring the technology of Groq, it has launched chip solutions that focus more on the AI inference stage. This trend indicates that the entire AI chip industry is gradually evolving from "general-purpose" to "scenario-based and specialized."
The rise of Google TPU is a microcosm of this industry trend.
