Google to Supply Oracle With Its Own AI Chip, TPU, as Seoul Metro Data Center Siting Disputes Intensify

Cloud developments this week were highlighted by Google’s expanded external supply of TPU, disputes over data center siting in the Seoul metropolitan area, and a space-based computing demonstration. Google has decided to supply its AI semiconductors, TPU, to Oracle Cloud for the first time outside its own platform, marking the first case in which Google is supplying TPU to an external public cloud infrastructure. The external platform is Oracle Cloud, while Google’s own cloud platform is GCP.

Google and Oracle have agreed to provide Google-designed AI semiconductors, Tensor Processing Units (TPU), to Oracle Cloud customers. Through this partnership, Google is pursuing collaboration among hyperscalers and expansion of its ecosystem. As a result, Google has widened the scope of TPU supply beyond GCP through its cooperation with Oracle.

At the same time, in South Korea, resident opposition and disputes over data center construction near residential areas in key parts of the Seoul metropolitan area, including Gwacheon and Incheon, are intensifying. As siting disputes surrounding data center construction in the Seoul metropolitan area grow, related issues were also treated as a major trend this week.

In space, a prototype satellite for Google’s space data center project operated data-center-class TPU in Earth orbit. The satellite succeeded in producing the first AI response ever generated, accelerating efforts to demonstrate space-based computing infrastructure. Image: AI-generated.

Oracle said it plans to introduce the latest eighth-generation TPU into OCI by 2027. The rollout will include 8t TPU for training and 8i TPU for inference. Oracle also plans to support power, cooling, and networking infrastructure for large-scale operations. Oracle’s adoption of Google TPU shows that this partnership is an extension of the broader trend in which Big Tech rivals in the cloud market are expanding mutual cooperation to secure AI computing resources and ecosystems.

Oracle has also adopted OpenAI’s Jalapeno chip and plans to integrate ChatGPT and Codex into its own applications. Google also plans to offer OpenAI models, which are competing models, on its own cloud. As a result, flexible coalition building across hardware and platforms is accelerating.

Meanwhile, resident backlash is spreading around major residential areas in the Seoul metropolitan area over data center construction. In the Juam district of Gwacheon, Gyeonggi Province, a 9.5 megawatt (MW) data center is being pursued about 20 to 100 meters (m) from apartment complexes. Residents in the Juam district of Gwacheon are strongly opposed, citing concerns over noise, fire, and deterioration of the residential environment.

In Dohwa-dong, Michuhol-gu, Incheon, a data center is being pursued near apartment complexes with more than 5,200 households and an elementary school. Undergrounding work for a high-voltage power line is also being pursued in the area. Residents of Dohwa-dong are continuing sit-ins and rallies to block the data center and the undergrounding work for the high-voltage power line.

The absence of a legal benchmark on the minimum separation distance between residential facilities and data centers is seen as a factor prolonging the dispute. Under current law, data centers are classified as broadcasting and telecommunications facilities, and whether they can be built is judged based on whether the zoning requirements are met. As a result, responsibility for resolving complaints and persuading residents about safety falls on individual operators and local governments.

Operators say the environmental impact is minimal because they use 22.9-kV distribution lines and install noise-reduction facilities. However, the dispute continues, with suspicions also raised of split development intended to avoid power grid impact assessments.

In this regard, there are calls for the government to urgently establish clear siting criteria to mediate the conflict.

Meanwhile, Google is pursuing 'Project Suncatcher,' a long-term research effort for a space data center. A prototype satellite for 'Project Suncatcher' succeeded in generating the first AI response using data-center-class TPU in Earth orbit.

A satellite equipped with the open-source AI model Gemma 4 created a space-themed haiku in response to commands from ground researchers and transmitted the result to a ground station in Svalbard, Norway. Through this, the satellite completed real high-performance inference using data-center-class hardware in space, which was presented as the first proof beyond mere survival. The satellite was jointly developed with Planet, a satellite imagery company, and launched last month on a SpaceX rocket. Power was supplied through solar panels, and TPU operated periodically.

Google has been exploring ways to overcome the power and cooling limits of terrestrial data centers, while pursuing a plan to use solar power directly in Earth orbit. This time, it carried out an in-orbit operational verification based on that concept. Google plans to launch two additional satellites in 2027, and those added satellites will be equipped with ultra-high-speed inter-satellite optical communication links. The purpose of the additional launch is to demonstrate distributed computing.

Source: IT DAILY · Kwon Young-seok
Original: https://www.itdaily.kr/news/articleView.html?idxno=242618

References

• https://www.itdaily.kr/news/articleView.html?idxno=242618