Google invades NVIDIA’s territory: TPU expansion strategy targets new cloud service providers
```
Google is moving its self-developed AI chip TPU from internal infrastructure to the external market, and for the first time is targeting emerging cloud service providers (Neocloud) that have long relied on Nvidia GPUs, attempting to compete for Nvidia’s share in the AI computing power market through lower costs and a more flexible business model.
According to The Information’s report on July 13, Google has recently been actively promoting TPU to several new cloud service providers, not only emphasizing advantages in cost and stability, but also offering financing support for data center projects and promising to lease back some TPUs to reduce clients' financial pressures for building computing power infrastructure.
Facing Google's offensive, Nvidia quickly launched a defense. On one hand, it discussed new financial incentives for some customers; on the other, it introduced a GPU lease-back guarantee program, hoping to further solidify the ecosystem of new cloud service providers. This competition over AI chips and cloud infrastructure is extending from product performance to financing, leasing, and the broader ecosystem.
Google Markets TPU to New Cloud Service Providers for the First Time
Previously, TPUs were almost exclusively deployed in Google’s internal data centers, and only offered externally via Google Cloud computing services. Now, Google has begun to directly sell TPUs to new cloud service providers, a market that was nearly completely dominated by Nvidia GPUs.
Google emphasizes to potential customers that TPUs have higher operational efficiency, more stable hardware platforms, and simpler network architectures. In contrast, Nvidia’s different generations of GPU platforms have substantially varying hardware architectures, which increases the complexity for clients deploying and upgrading. The Information previously reported that the Grace Blackwell platform encountered reliability issues in the early stages of large-scale deployment.
Aside from the product itself, Google's bigger bargaining chip is its business model.
According to sources cited in the report, Google has discussed providing financing support for some new cloud service providers to help build TPU data centers, aiding in raising funds needed to purchase chips and equipment. At the same time, Google proposed leasing back these TPUs for its own AI model training, search, advertising business, and Google Cloud clients, thus helping clients lock in demand ahead of time and reduce investment risk.
Google spokesperson Ted Ladd stated that Google maintains a long-term partnership with Nvidia, GPUs remain the most popular chips for Google Cloud, and the company hopes to offer customers more choices. Many customers already use both GPUs and TPUs simultaneously for different workloads.
Nvidia Responds Quickly, Launches Lease-back Guarantee
One of the new cloud service providers Google recently approached is Nscale. This company, founded just two years ago, has a close relationship with Nvidia, not only having obtained priority GPU supply from Nvidia, but by the end of last year, Nvidia was its largest preferred shareholder. Both Microsoft and Google plan to lease Nscale’s European data center GPU resources within the next year, and the company is also establishing a large AI data center in West Virginia, USA.
However, Nscale has indicated that all signed projects and ongoing negotiations currently revolve around GPUs, and there are no plans to pursue TPU deals.
The report states that upon learning of Google’s contacts with Nscale, Nvidia quickly discussed new financial incentive arrangements with them. According to a person close to the deal, these arrangements aim to reduce Nscale’s motivation to shift to TPUs. However, Nscale denied that Nvidia ever offered any incentives on the condition of abandoning TPU cooperation.
Meanwhile, Nvidia has also started offering GPU lease-back guarantee programs to new cloud service providers.
According to the program, if partners are unable to lease Vera Rubin GPUs to other customers in the future, Nvidia will lease them back at an agreed price. However, according to sources involved in the negotiations, this guarantee price is about half of the expected market rental rate; if clients eventually rent out GPUs at a higher price, Nvidia will share half of the excess income over the guarantee price.
Although the terms are not lenient, industry insiders believe such income guarantees can help new cloud service providers obtain bank financing more easily. Especially in areas where major cloud vendors have not yet landed, there remains some attraction.
Google Hopes to Build a TPU Ecosystem
Promoting external sales of TPUs doesn’t simply mean increasing chip sales.
The report notes that some Google executives believe TPUs could reach about 10% of Nvidia’s AI server chip business scale in the future, and several major AI customers, including Meta, have been looking for cheaper alternatives.
A larger customer base can also increase TPU production scale and improve overall manufacturing economics. Meanwhile, if new cloud service providers or their financier bear part of the chip procurement costs, Google can also reduce its own capital expenditure pressure, which is becoming an important metric for investors watching tech giants' AI investments.
Additionally, Google hopes to gain more say in advanced process capacity allocation at TSMC. Currently, both TPUs and Nvidia GPUs rely on TSMC for manufacturing, and Nvidia has recently become one of TSMC’s largest customers for advanced packaging and AI chips. Expanded TPU demand would help Google secure more manufacturing resources.
Currently, TPU has attracted enterprises including Anthropic, Meta, Apple, Citadel Securities, and Safe Superintelligence, among others. Google has recently teamed up with Blackstone to launch a new cloud service provider, planning to lease TPU computing power to AI developers starting next year, further driving TPUs to evolve from Google’s internal infrastructure to an independent computing power platform for the whole AI industry.
Risk Warning and DisclaimerThe market entails risks, and investments should be made cautiously. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investments made accordingly are at your own risk. ```