Google's Custom Semiconductor Chips: Unveiling the Underestimated Power and Game-Changing Tech Behind Its Innovation

The story of Google's chip manufacturing dates back to 2015.

For Chinese peers, it serves as both a mirror and a source of pressure. It proves one thing: the end point of software giants may lie in chips, and the end point of chips lies in systematic patience spanning more than a decade.

01 A Hidden Giant

There is an easily overlooked fact: in the context of Chinese internet discourse, Google has long been labeled as a "software company". When it comes to chips, people reflexively think of Intel, AMD, NVIDIA, Qualcomm, TSMC, or the iconic leather jacket worn by Mark Zuckerberg and Jensen Huang, CEO of NVIDIA.

But if we take a longer-term perspective, the conclusion is unsettling: Among global internet companies, it is Google that has developed self-developed chips with the deepest, most systematic and most sustainable efforts.

No other company can match it.

This cognitive dislocation has a simple yet profound underlying reason — the United States is far too "affluent". Intel, AMD, NVIDIA, Qualcomm, Broadcom, Marvell, any of them picked at random is a giant spanning half the globe. In a country with such a rich chip foundation, Google's TPU has long been in the state of an "internal tool" and does not sell chips directly to the outside world, so it is naturally difficult to step into the spotlight. It is just a footnote in the grand narrative of Silicon Valley giants.

Until 2026, this footnote becomes the main text.

02 Ten Years of Unrecognized Efforts, One Stunning Breakthrough

The story of Google's chip development dates back to 2015.

That year, AlphaGo defeated Lee Sedol and shocked the whole world. Few people realized that behind every move of AlphaGo stood a dedicated chip that Google had just successfully developed in-house — the first-generation TPU (Tensor Processing Unit). It was not built for general-purpose computing, but was optimized frantically for only one task: matrix multiplication.

This is where Google's strategic vision is most sharp. When the entire industry was still discussing "whether CPU is sufficient and whether GPU represents the future", Google had already keenly judged from the massive data load of its own search engine, advertising system and YouTube recommendation that the computing power demand of AI would explode exponentially, and general-purpose chips would sooner or later become a cost black hole. It would be better to develop chips on its own.

In 2017, TPU v2 introduced the BF16 format and began to support training; in 2018, TPU v3 was launched with large-scale liquid cooling; in 2024, the sixth-generation Trillium was released, with 4.7 times performance improvement and 67% energy efficiency improvement over the previous generation, undertaking the heavy responsibility of training the Gemini large model; in 2025, the seventh-generation Ironwood pushed the single-chip FP8 peak computing power to 4614 TFLOPS; in April 2026, the eighth-generation TPU was even split into two product lines for the first time, TPU 8t (training) and TPU 8i (inference), completing the mature transition from "general large computing power" to "task classification".

What is most worth pondering for the industry in this evolution curve is that none of the generations was a "gimmick release", and every generation has been truly deployed on Google's own search, advertising, YouTube recommendation and Gemini training and inference. The chip industry is a capital black hole level industry. Without the support of huge self-use demand, it is very difficult for any self-developed chip plan to survive more than three generations of iteration — a point that Chinese peers understand most deeply.

03 The Real Watershed: Self-use or Commercial Sales

This is exactly the difference in competitive landscape that we most want to emphasize from the perspective of "co-opetition".

In the Chinese internet circle, Baidu's Kunlunxin is often mentioned in the same breath as Google's TPU. Objectively speaking, Kunlunxin is a representative of self-developed chips by Chinese internet companies, with top-tier technical accumulation and engineering capabilities. But if we make a strict comparison with Google, the gap between the two lies not in the chip itself, but in the commercialization radius: Google's TPU is not only for self-use, but also provides services to the outside world as the core computing power of Google Cloud, and now it is even sold directly to third-party customers; while most of the self-developed chips of Chinese internet companies including Kunlunxin are still in the self-use stage, and it is difficult to form external, large-scale third-party supply.

Of course, China is not without chip forces for external supply — independent chip companies such as Cambricon, Enflame, Biren and MetaX are catching up rapidly, constituting the vibrant side of China's AI computing power industry. But note the key difference here: China's chip supply comes from "scattered startups", while the U.S. chip supply is "directly led by giants". The former faces high risks at every step in financing, tape-out and ecological construction; the latter is backed by hundreds of billions of dollars in annual capital expenditure, trillion-parameter model consumption of its own, and a complete cloud computing distribution channel.

Several landmark events in 2026 have brought this landscape change to the surface completely:

• Anthropic's landmark order: In October 2025, Google promised to supply up to 1 million TPUs to Anthropic, worth tens of billions of dollars; in April 2026, Anthropic placed another TPU order at the scale of several gigawatts. The computing power lifeline of one of its biggest competitors is in the hands of Google — a scene that is almost unprecedented in commercial history.

• Meta leases Google's TPU: Meta, which is extremely short of computing power, was exposed to have signed a contract worth tens of billions of dollars to lease Google's TPU clusters.

• Ant Group's 200% performance improvement: In August 2026, it was disclosed at Google I/O Connect China that Ant Group migrated high-intensity computing workloads to Google TPU v7, achieving a 200% training performance improvement.

• Chain-style capital moves: In August 2026, Google reached an agreement with Marvell Technology, under which the latter will customize AI inference accelerators, storage controllers and other products for Google to access the TPU ecosystem; Google will invest up to 12.182 billion US dollars, holding about 6.7% of the shares.

So far, Google has completed the three-stage rocket closed loop from "self-use chips" to "chips + cloud services + investment binding". CFO Ruth Porat admitted at the earnings call that TPU revenue in 2026 only accounts for a small part, and most orders will be fulfilled in 2027. In other words, Google has just made the first move in this grand game.

04 Why No One Noticed It In Advance: The "Blind Spot" in the Competitive Landscape

Now let's answer the core question: Why has the industry long turned a blind eye to Google's decade-long chip development?

First, NVIDIA's "light" is too strong. In the narrative field of AI chips, NVIDIA almost monopolizes the right to define topics. Twenty years of accumulation of the CUDA ecosystem has made "GPU = AI computing power" the muscle memory of the industry. In contrast, the JAX/XLA ecosystem of TPU has higher threshold and lower mind share. This is not only a real short board, but also a cognitive shield.

Second, Google's business model is naturally "low-key". Intel sells chips, AMD sells chips, NVIDIA sells chips — while Google has long hidden TPUs inside the walls of its own data centers. A technology that is not sold externally is regarded as a "cost item" rather than an "income item" in the capital market, so the media naturally lacks the entry point for reports. It was not until the second quarter of 2026 that TPUs were delivered to third-party data centers for the first time, stepping into the public view as a "commodity".

Third, and the most underestimated point: systematic capability. Google's real moat is not a single chip, but the vertical integration of chip - model - compiler - network - data center - cloud. The self-developed optical circuit switch (OCS) supports dynamic reconfiguration of network topology in the Pod; the XLA compiler maximizes the efficiency of software-hardware collaboration; DeepMind can predict the future model form in advance and lock the architecture direction in advance. This capability of "full-stack self-development and endogenous iteration" can never be replicated by companies that simply buy NVIDIA GPUs. Some analysis points out that Google TPU can achieve several times the cost performance of NVIDIA's solution on its own optimized workloads.

05 Co-opetition Game: Google's Three Roles

From the intersection of industrial competition and cooperation, Google actually plays three roles at the same time:

For NVIDIA, it is a challenger. TPU shipments were about 2.5 million units in 2025, and the 2026 target was raised by 50%, with the total volume expected to exceed 3.7 million units; in the inference market, ASIC is moving rapidly from 15% share in 2024 to the predicted 80% share in 2030 as forecasted by IDC. Google is exactly the biggest promoter of this replacement curve.

For OpenAI, Meta, Anthropic and other companies, it is both a competitor and a symbiotic partner. It competes with everyone for model leadership with Gemini, while selling and leasing TPUs to them. Google's capital expenditure was about 31 billion US dollars in 2022, and it is expected to surge to about 180 billion US dollars in 2026, with self-developed chips as the core of investment.

For Chinese peers, it serves as both a mirror and a source of pressure. It proves one thing: the end point of software giants may lie in chips, and the end point of chips lies in systematic patience spanning more than a decade. The case of Ant Group migrating to TPU also reminds us: when chips become part of cloud services, the boundary of computing power competition has long crossed national borders.

06 Conclusion

Back to the sentence in the title: Google's chips should never be underestimated.

It is not the victory of a single chip, but the victory of a paradigm — demand defines chips, and system determines the outcome. In an era dazzled by the light of NVIDIA, Google spent ten years quietly building a computing power empire, and then turned it into a business in the right year.

For the industry, the cost of ignoring it may come faster than expected. For China's AI chip industry, while the rapid catch-up of companies like Cambricon and MetaX is commendable, a deeper issue remains: Can we have a platform-level player that has both top-tier self-use demand and open distribution capabilities?

This competition may have just begun.

This article is from the WeChat Official Account "Co-opetition AI", author: Co-opetition, authorized for release by 36Kr.