Google (GOOGL.US) TPU expands to four models in a year, pointing to Nvidia! The agency expects related sales to exceed US$135 billion in 2028

The Zhitong Finance App learned that Google, a subsidiary of Alphabet (GOOGL.US), is expanding its AI chip product portfolio to challenge the dominant position of Nvidia (NVDA.US) and seize share in this market worth hundreds of billions of dollars. Google will launch two new variants of its Tensor Processing Unit (TPU). TPUs are custom semiconductors that help develop and run AI software. The latest chips will meet the needs of confidential computation and help models access a wider range of information when processing queries. More broadly, they aim to serve an industry that is shifting to the use of AI agents and facing security challenges.

This launch marks an acceleration in product development for Google's chip division. The department has released a new AI processor every two years; two TPUs were launched at the same time this year, and the latest variant to be announced on Monday will further expand the product line.

Product line expansion: from one model in two years to four models in a year

Amin Vahdat, the head of the Google chip, said in an interview that the four TPU products will be optimized for different needs. “This really follows Google's core philosophy, that the future lies in specialization, and the benefits of specialization are still very significant.”

Google's release of a new chip comes at a time when the AI chip industry is in turmoil. Concerns about a possible slowdown in infrastructure spending have left tech stocks in turmoil, and the Democratic Party's victory in the US midterm elections has raised doubts about data center projects. However, demand for Nvidia's AI processors is still greater than its supply capacity, which provides Google and other competitors with room to grow.

According to Vahdat, the new TPU 8tx and 8i Max variants are not new silicon designs, but have been reconfigured to take on the tasks customers have always requested.

8tx handles secure and confidential work for clients such as government agencies and regulated industry companies. The system works with AMD (AMD.US) chips and is part of a partnership between the two companies.

8i Max is used to run AI applications that require a long context window — that is, the amount of data the AI system can consider in a single request. According to Vahdat, this is essential for tasks such as drug development and mathematics, and is also important for programming applications.

From personal use to export sales: TPU sales are expected to exceed 100 billion US dollars

TPU competes in a market dominated by Nvidia graphics processing units (GPUs). Google developed this technology more than a decade ago with Broadcom (AVGO.US), which helped design and coordinate production. Google's initial idea was to handle growing computing demands more efficiently. Google then realized the potential of TPU for AI work and has now begun selling these chips outside of its cloud computing business.

Google's new role as a chip seller is bringing significant benefits. The agency predicts that TPU sales will exceed $135 billion by 2028, including related software and hardware. This is higher than the estimated $12 billion this year, and does not include sales generated by customers using these chips through Google Cloud.

TPU has attracted some well-known customers, including Apple (AAPL.US), Meta (META.US), and Oracle (ORCL.US). Vahdat said that providing a dedicated version of the chip should help Google make further progress. These new chips will be available next year.

“We can build a single chip that tries to cover the entire market, but by creating a dedicated version of the chip, we can actually deliver more value.” he said.

Competition is heating up: Nvidia is also turning to specialization

Even Nvidia, which has always promoted GPUs as suitable for all AI tasks, has begun offering more dedicated chip systems, including launching its own central processing unit (CPU) — a category pioneered by Intel (INTC.US). These processors don't carry the heavy data computation required to create AI models, but they are useful for managing AI agents.

Nvidia is also using technology acquired from startup Groq to provide chips that focus more on the AI inference stage — the stage where a trained AI model performs a task.

Google's 8Tx works with AMD's CPU, which is a competitor to Nvidia and Intel. This technology allows data to be encrypted at the chip level, making it more secure. Normally, information must be unlocked in memory as the computer processes it, Vahdat said. The 8i Max packages two TPU chips together to provide stronger computing power and memory.

Analyst Mandeep Singh said that part of TPU's appeal lies in its cost performance ratio. They're around 30% cheaper than Nvidia GPUs, but they're no less impressive in terms of throughput — that is, the amount of data they can process in a given time.

The customer landscape is expanding, and TPU export sales are gradually becoming a climate

This path has attracted a variety of customers, including in the financial sector. Google announced in last month's earnings report that Jump Trading is a TPU customer.

Alex Davies, chief technology officer of Jump Trading, said in an interview that the company spent several years testing these chips and signed a larger contract with Google about a year ago. This trading company has thousands of TPUs to train AI models such as trading research. Davies said Jump still mainly uses Nvidia for training, but “a small group” of employees really prefer to use TPU to complete their tasks.

He said that TPU performed well in some of the tasks. Davies also said that using different types of chips to meet the challenge is beneficial. In fields where TPU performance is comparable to or superior to Nvidia chips, TPU also has higher energy efficiency advantages, thereby reducing costs.

Last month, Oracle said it had signed an agreement to allow its cloud customers to use TPU for the first time starting next year. Oracle Cloud customers were previously only able to choose AI chips from Nvidia and AMD.

“We're partnering with Google because as AI workloads get more demanding, customers need more choices,” Karan Batta, senior vice president responsible for Oracle Cloud Infrastructure, said in a statement. “This partnership expands our ability to help customers build and run AI on the infrastructure that best fits their needs.”

Claude developer Anthropic PBC is another TPU customer. According to a recent regulatory document, the company has the option to purchase TPU chips worth up to $150 billion between 2028 and 2031. According to reports, OpenAI is also planning to buy TPU.

Google's own status as an AI software developer has also helped chip research and development. TPU developers rely on early use of Google's latest Gemini model to improve energy efficiency and find bugs. This is critical in semiconductor design—flaws can cost billions of dollars and waste months of work. According to Vahdat, AI is reducing the probability of these risks occurring.

Challenges Remain: Competition, Slowing Spending, and Supply Bottlenecks

But Google is also facing other challenges, including new entrants to the market. TensPresso, a startup co-founded by two former TPU engineers, is focusing its first chip on long context window inference — this is exactly what the 8i Max is targeting. However, according to media reports last week, the product is at least a year away from being designed.

OpenAI and Anthropic are also advancing efforts to develop their own chips — there is no shortage of TPU veterans in the team — which could increase future competition.

There is also a risk that AI infrastructure spending will slow down. Analyst Singh estimates that opposition to data centers may reduce the expected spending of 1.4 trillion US dollars by technology companies by about 5% to 7% by 2027. This includes expenses on chips, network equipment, land, and other data center construction materials.

According to Singh, Google's TPU's profit margins are not as lucrative as Nvidia.

Like other companies in the chip and AI infrastructure sector, Google is facing a shortage of key components such as memory and manufacturing capacity, which makes it difficult to meet demand. The company said in its financial report that every TPU that can be used by Google has been contracted to customers.

Vahdat's team is working to diversify supply sources. In recent months, Google and Broadcom's competitor Marvell Technology (MRVL.US) have signed an agreement that includes some TPU work. There are also reports that MediaTek is also providing technical support for Google's TPU business.

Vahdat declined to comment on the two companies' roles in the TPU business.

Analyst William Tong said that Google is likely finalizing a longer-term agreement to “determine supply and prevent further bottlenecks.”