Japan to catch up on gen AI with a new supercomputer

• Plans for the supercomputer were announced by the Ministry of Economy, Trade and Industry.
• The National Institute of Advanced Industrial Science and Technology will develop a supercomputer.
• The machine will have a computing capability roughly 2.5 times greater than its existing machine by 2024.
• AIST will make it available to domestic companies developing generative AI through a cloud service.

As a country known for its advanced technology, Japan needs to catch up in the generative artificial intelligence (AI) race. While countries like the US, China, and European Union (EU) are leading with their algorithms, Japan’s lag is attributed mainly to its shortcomings in deep learning and extensive software development.

For starters, the key to generative AI development is large language models, which underpin ChatGPT and Baidu’s Ernie Bot, capable of processing vast data sets to generate text and other content. But as Noriyuki Kojima, co-founder of Japanese LLM startup Kotoba Technology puts it, Japan is trailing behind the US, China, and the EU in developing those algorithms.

“Japan’s trailing position in the field of generative AI largely stems from its comparative shortcomings in deep learning and more extensive software development,” Kojima said in an interview with CNBC. To put into context, based on a report by Reuters in May, citing research from a consortium of state-run institutes, Chinese organizations, including tech giants Alibaba and Tencent, have launched at least 79 LLMs domestically over the past three years.

In the US, corporate powerhouses such as OpenAI, Microsoft, Google, and Meta play a significant role in propelling the country’s LLM advancements. But deep learning requires a “robust community of software engineers” to develop necessary infrastructure and applications, Kojima added. 

Japan, however, will face a deficit of 789,000 software engineers by 2030, according to the country’s Ministry of Economy, Trade and Industry (METI). Japan is also facing hardware challenges since LLMs need to be trained using AI supercomputers like IBM’s Vela and Microsoft’s Azure-hosted system. Unfortunately, Nikkei Asia reported that no private Japanese company possesses its own “world-class machine” with those capabilities.

How can Japan catch up on gen AI?

Kojima told CNBC how government-controlled supercomputers like Fugaku “hold the key” to Japan’s pursuit of LLMs. “Access to such large-scale supercomputers forms the backbone of LLM development, as it has traditionally been the most significant bottleneck in the process,” he explained.

The upside is that the METI has plans to introduce a new cutting-edge supercomputer through its affiliated laboratory to support the development of generative AI in Japan, according to Nikkei’s latest report. According to Nikkei, the plan is for the National Institute of Advanced Industrial Science and Technology (AIST) to develop a supercomputer with a computing capability roughly 2.5 times greater than its existing machine as early as 2024. 

Under METI’s supervision, the research arm will make it available to domestic companies developing generative AI through a cloud service. AIST, one of Japan’s most prominent research institutes, plans to establish a new research center for supercomputers and quantum technologies in July. METI will provide 32 billion yen (US$226 million) to finance the installation of facilities.

Nikkei also stated that AIST will operate the new supercomputer at the center for its studies on quantum technology and open it to Japanese companies. So far, AIST’s existing supercomputer is accessible from outside via a cloud service.

Separately, the Tokyo Institute of Technology and Tohoku University also have plans to work on the country’s LLMSs prowess. Based on an announcement in May, the institute and university plan to use the Fugaku supercomputer to develop LLMs based primarily on Japanese data in collaboration with the supercomputer’s developers, Fujitsu and Riken. 

The organizations also plan to publish their research results in the same year AIST intends to release its supercomputer, which is 2024. According to Fujitsu, the research results would help other Japanese researchers and engineers develop LLMs.

Above all the efforts mentioned above, the Japanese government also plans on investing 6.8 billion yen (US$48.2 million), about half the total cost, to build a new supercomputer in Hokkaido that will begin service as early as 2024, Nikkei reported. That particular supercomputer, built in Hokkaido, will specialize in LLM training to promote Japan’s development of generative AI.

Based on reports, the cloud service provider Sakura Internet is assembling the supercomputer in Ishikari City. It’ll contain more than 2,000 graphics processing units from Nvidia, which are necessary for LLMs. Interestingly, the Sakura Internet’s supercomputer would be able to develop GPT-3 within roughly three days and about a year to develop GPT-4.

---

---

---

---