2026-11-13 14:00

Lead: The real battleground is the CUDA ecosystem, not individual chip pricing. Google launches two specialized TPUs, challenging NVIDIA’s general-purpose chip strategy.

Google unveils two new specialized TPUs, directly challenging NVIDIA’s general-purpose chip roadmap.

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

TPUs are Google’s custom AI accelerators, developed for over a decade—originally to reduce computational costs for internal search operations. This time is different: the 8tx is targeted at government agencies and regulated industries, emphasizing on-chip data encryption; the 8i Max integrates two chips into a single package, focusing on ultra-long context window support.

Google’s chip chief Amin Vahdat told Bloomberg: “The future belongs to specialization.”

This statement is aimed squarely at NVIDIA. For two decades, NVIDIA built its GPU business around the promise of “one chip to do it all.” Now, general-purpose chip companies are moving toward specialization, while specialized chip vendors are building general-purpose platforms—two divergent paths colliding in the middle.

From every two years to two per year: Why Google suddenly accelerated

Previously, Google typically launched one new AI processor every two years. In 2026, this rhythm was disrupted. Vahdat stated that four distinct TPU products will be optimized for specific use cases. His rationale was direct: “We can build a general-purpose chip aiming to serve the entire market—or we can launch specialized variants. The latter actually creates more value for our customers.”

Beneath the expanded product line lies a growing client base. Google’s TPU has already attracted Apple, Meta, and Oracle. In October, Oracle announced that starting next year, its cloud customers will gain access to Google’s TPU for the first time—previously restricted to NVIDIA and AMD options. Oracle Cloud Infrastructure Senior Vice President Karan Batta said: “As AI workloads grow increasingly complex, customers demand more choices.”

A larger opportunity comes from Anthropic. According to a recent regulatory filing, the company holds an option to purchase up to $150 billion worth of TPUs between 2028 and 2031. Bloomberg previously reported that OpenAI also plans significant procurement.

Google’s earnings report included a firm statement: All currently available TPUs have already been pre-booked by customers.

Why now? Two external factors converged. On one hand, concerns about a potential slowdown in AI infrastructure investment are weighing on tech stocks. On the other, the Democratic victory in the U.S. midterm elections has introduced new uncertainties for data center projects. By expanding its product line at this juncture, Google sends a clear signal to clients: regardless of construction pace shifts, the chip layer is already secured.

Who are the 8tx and 8i Max for?

These are not entirely new designs, but reconfigured versions of existing chips tailored to two distinct needs.

The TPU 8tx targets clients handling sensitive, confidential data—government institutions and highly regulated sectors. It pairs with AMD CPUs, enabling data encryption at the silicon level. Vahdat explained precisely: traditionally, data must be decrypted in memory before processing; the 8tx bypasses this step. This is part of Google’s collaboration with AMD.

The TPU 8i Max targets applications requiring ultra-long context windows. By integrating two TPUs into a single package, both compute power and memory capacity are amplified. Vahdat noted such capabilities are crucial for drug discovery and mathematical computation, as well as programming-related tasks.

Two customer segments, two distinct value propositions. Government and regulated industries buy “compliance,” willing to pay a premium for security. Model developers buy “context length,” willing to pay for the ability to process vast document inputs. One chip cannot satisfy both demands—so Google splits them.

30% cheaper, but less profitable than NVIDIA

Bloomberg Industry Research analyst Mandeep Singh provided a set of figures: compared to NVIDIA GPUs, TPUs cost approximately 30% less, while maintaining competitive throughput. He also noted that Google’s TPU business does not enjoy the same profit margins as NVIDIA.

Where does the 30% cost advantage come from? Two sources. First, TPUs are domain-specific accelerators, eliminating large portions of general-purpose computing units. Second, Google itself is the largest user—the R&D burden can be amortized across its own massive computational needs.

But lower cost doesn’t guarantee high sales. NVIDIA’s true moat isn’t the chip—it’s CUDA: decades of accumulated codebases, tools, and developer expertise. As Counterpoint Research VP Neil Shah put it, the time, code, and project experience invested in CUDA are sunk costs that cannot be migrated.

A client already heavily committed illustrates the point. Jump Trading’s Chief Technology Officer Alex Davies said the quant trading firm owns thousands of TPUs, primarily used to train models supporting trading research—but the company still relies predominantly on NVIDIA chips for model training, with only a small subset of engineers showing strong preference for TPUs: “They express their enthusiasm quite clearly.”

The insight here is stark: a client with a major agreement and thousands of TPUs still uses NVIDIA as its primary training platform. TPU captures “specific tasks,” not the “main battlefield.”

Google recognizes its weakness. It’s expanding supply chains—partnering with Broadcom on design and manufacturing coordination—and recently signed a deal with Broadcom’s competitor Marvell. Reports also suggest MediaTek is providing support for TPU operations. Meanwhile, it’s catching up on software. Vahdat highlighted a unique advantage: the TPU development team gets early access to Google’s latest Gemini models, enabling better chip efficiency tuning and earlier detection of software vulnerabilities. Chip design flaws could lead to losses in the tens of billions—AI is reducing the likelihood of such failures.

Competition is intensifying from both sides. A startup called Tenspresso, founded by two former TPU engineers, is developing a first-generation chip targeting the 8i Max market segment—design completion expected at least a year away. OpenAI and Anthropic are also advancing in-house chip initiatives, with teams including senior engineers who previously worked on TPU development.

Bloomberg Industry Research forecasts that by 2028, Google’s TPU-related chip, software, and hardware revenue will exceed $135 billion, up from an estimated $12 billion this year—this figure excludes revenue generated through customers using TPUs via Google Cloud.

Within the same study, a counterpoint emerged: opposition to data center construction could reduce projected tech industry spending on related infrastructure—originally expected to reach $1.4 trillion by 2027—by 5% to 7%. This includes expenditures on chips, networking equipment, land, and other materials required for data center deployment.

Together, these two figures encapsulate Google’s current reality: demand is rising, but resistance is growing too. And in the same year, it chose to expand its product portfolio, embedding the word “specialization” into its strategic catalog.

More than a decade ago, Google built TPUs to reduce reliance on NVIDIA’s cards. More than a decade later, it sells TPUs to Apple, Meta, Oracle—and now competes with NVIDIA for the same customers. In the final reckoning, the battle won’t be about which chip is faster, but which one lets developers change fewer lines of code.

NVIDIA’s brand still stands, but someone is now standing at the door.

Source: Cai Neng Quan