Google releases two TPUs at once, and NVIDIA's "general-purpose" brand edge begins to slip.

Google has launched two new professional TPUs, challenging NVIDIA's general-purpose chip roadmap.

Recently, Google released two new TPU chips, the TPU 8tx and the TPU 8i Max.

TPUs are Google's self-designed AI-specific chips that have been developed for over a decade, initially built only to cut computing power costs for its own search business. This time is different: the 8tx is sold to government agencies and regulated industries, featuring chip-level data encryption; the 8i Max packages two chips in the same enclosure, focusing on ultra-long context processing.

Amin Vahdat, head of Google's chip business, made a seemingly ordinary statement in an interview with Bloomberg: "The future belongs to specialization."

This statement is directly targeted at NVIDIA. NVIDIA has been manufacturing GPUs for 20 years, with its core selling point always being "one chip that can handle all tasks". Now, companies that make general-purpose chips are beginning to launch specialized versions, while companies that make specialized chips are starting to build general-purpose platforms — the two development paths are colliding midway.

From one new product every two years to two new products a year, why is Google suddenly accelerating its pace?

In the past, Google usually launched a new AI processor every two years. This rhythm was disrupted in 2026. Vahdat said that the four TPU products will be optimized for different needs respectively, and he gave a straightforward reason: "We can build a general-purpose chip that tries to meet the needs of the entire market, or we can launch different specialized versions. Through the latter approach, we can actually create more value for our customers."

Behind the expansion of the product line is a growing list of customers. Google TPUs have already attracted Apple, Meta and Oracle. In October, Oracle announced that starting from next year, it will allow its cloud customers to use Google TPUs for the first time, while these customers could only choose NVIDIA and AMD products before. Karan Batta, Senior Vice President of Oracle Cloud Infrastructure, stated that: "As AI workloads become increasingly complex, customers need more choices."

An even larger order comes from Anthropic. According to a recent regulatory filing, the company has an option to purchase up to $150 billion worth of TPUs between 2028 and 2031. Bloomberg previously also reported that OpenAI also plans to purchase TPUs.

Google wrote a very assertive statement in its financial report: all currently available TPUs have already been booked by customers.

Why right now? Two external conditions are overlapping. On one hand, market concerns about a possible slowdown in AI infrastructure investment are weighing on tech stocks; on the other hand, the Democratic Party's victory in the US midterm elections has brought new uncertainties to data center projects. By choosing to expand its product line at this point in time, Google is sending a signal to customers: no matter how the construction pace changes, we will have the chips ready at the underlying level first.

Who are the 8tx and 8i Max targeted at respectively?

These two are not completely newly designed chips, but reconfigured based on existing chips to meet two specific types of demands.

The TPU 8tx is targeted at customers that need to process secure and confidential data — government agencies, as well as strictly regulated industries. It works in tandem with AMD CPUs to encrypt data at the chip level. Vahdat explained in detail: Traditional computers usually need to decrypt data in the memory before processing it; the 8tx can bypass this step. This is part of Google's cooperation with AMD.

The TPU 8i Max is targeted at applications that need to run ultra-long context windows. It integrates two TPUs in the same package, with both computing power and memory capacity improved at the same time. Vahdat said this capability is very important for drug R&D and mathematical computing, and also critical for programming-related applications.

Two customer groups, two sets of reasons to pay. Governments and regulated industries purchase "compliance" and are willing to pay a premium for security; model companies purchase "context" and are willing to pay a premium for the number of documents that can be processed at one time. A single chip cannot meet both types of demands, so Google has developed them separately.

30% cheaper, but less profitable than NVIDIA

Mandeep Singh, an analyst from Bloomberg Intelligence, released a set of figures: compared with NVIDIA GPUs, TPUs cost about 30% less while remaining competitive in throughput. He also pointed out that compared with NVIDIA, the profit margin of Google's TPU business is not that high.

Where does the 30% cost reduction come from? There are two sources. TPUs are specialized chips that remove a large number of general-purpose computing units; Google itself is the largest user, so R&D costs can be amortized against its own computing power demands.

But lower price does not equal easy sales. NVIDIA's real moat is not its chips, but CUDA — the decades of accumulated code, tools and experience from developers. Neil Shah, Vice President of Counterpoint Research, once said that the time, code and project experience developers have invested in CUDA are all sunk costs that cannot be migrated.

A customer that has already placed heavy bets on TPU can best illustrate this point. Alex Davies, Chief Technology Officer of Jump Trading, said that this quantitative trading firm has thousands of TPUs, which are mainly used to train models to assist trading research; but the company still mainly uses NVIDIA chips for model training at present, and only a small number of employees are particularly fond of TPUs, "they are quite positive about it".

The key information in this statement is: A customer that has signed a larger cooperation agreement and owns thousands of TPUs still uses NVIDIA chips for its main model training. TPUs have seized "specific tasks" rather than the "main battlefield".

Google itself is well aware of its own shortcomings. On one hand, it is expanding its supply sources — it cooperates with Broadcom for chip design and production coordination, signed a cooperation agreement with Broadcom's competitor Marvell in recent months, and there are also reports that MediaTek is also providing support for the TPU business; on the other hand, it is making up for its shortcomings in software. Vahdat mentioned a unique advantage that no other peer has: the TPU development team can get early access to Google's latest Gemini model, which is used to improve chip energy efficiency and detect software vulnerabilities. Chip design flaws can cause billions of dollars in losses, and AI is reducing the probability of such problems occurring.

Competition is also squeezing in from both sides. A startup called Tenspresso, founded by two former TPU engineers, is developing its first chip that directly targets the 8i Max's market, and the product will take at least another year to finish design. OpenAI and Anthropic are also advancing their self-developed chip projects, with many senior engineers who have participated in TPU development on their teams.

Bloomberg Intelligence estimates that by 2028, Google's TPU-related sales of chips, software and hardware will exceed $135 billion, and the figure for this year is expected to be $12 billion; this number does not include the revenue generated by customers using TPUs through Google Cloud.

The same research also mentions a reverse figure, that is, the opposition to data center construction may reduce the $1.4 trillion in related spending that tech companies originally expected to invest in 2027 by 5% to 7%. These expenditures cover chips, network equipment, land and other materials required for data center construction.

The two figures together represent Google's current full situation: market demand is rising, and resistance is also growing. And it has chosen to expand its product line in the same year, putting the concept of "specialization" at the core of its product portfolio.

More than a decade ago, Google developed TPUs to reduce its purchases of NVIDIA GPUs. More than a decade later, it sells TPUs to Apple, Meta and Oracle, and even competes with NVIDIA for the same customers. At the end of this battle, the competition is not about whose chips are faster, but about who can make developers modify fewer lines of code.

NVIDIA's sign is still hanging at the door, but competitors have already arrived.