According to reports, Japanese telecom giant SoftBank Group will invest 150 billion yen ($960 million) in the next two years to upgrade its computing facilities. The plan includes the purchase of large quantities of Nvidia GPUs. In addition to using GPUs to drive its internal large language models (LLM), SoftBank will also provide access to enterprise customers in need of supercomputing services.
The investment will dwarf the 20 billion yen ($130 million) SoftBank invested in computing infrastructure last year, the largest such investment by a Japanese company, according to Nikkei.
The development and application of GenAI shows no signs of slowing down, so SoftBank hopes to develop a Japanese version of GenAI to meet user needs.
SoftBank established a new department in March last year to develop LLM and provide GenAI services. The unit was subsequently established last August as SB Intuitions, a subsidiary of SoftBank.
SBIntuitions completed the deployment of its computing platform in October last year and is developing its first domestic LLM, which will be launched this year with 350 billion parameters. That would make it more powerful than OpenAI’s GPT-3, which has 175 billion parameters, but far less powerful than GPT-4, which reportedly has 100 trillion parameters.
What is concerning is that OpenAI said earlier this month that it opened a new office in Tokyo and began to provide early services of the Japanese version of GPT-4 to Japanese companies. SoftBank is therefore under pressure to accelerate the delivery of viable alternatives.
SoftBank’s AI computing platform runs on Nvidia’s hardware, specifically the Nvidia DGX SuperPODAI supercomputer built with more than 2,000 Nvidia TensorCore GPUs. Given the good cooperation between the two companies, SoftBank has reason to choose Nvidia in its AI supercomputing expansion plan.
In addition, French telecommunications group Iliad demonstrated a SuperPOD in September last year, which the company claimed was Europe’s most powerful AI supercomputer. But Nvidia, which is working hard to seize GenAI hardware market share, has launched a more powerful AI supercomputer.
In May last year, Nvidia combined GraceGPU and HopperGPU to create the GraceHopper chip, a chip specifically designed for GenAI. Nvidia’s upcoming DGXGH200 supercomputer will use 256 GraceHoppers and plans to sell them to companies looking to build their own LLMs.
Although there is no news on the price of GraceHopper, it will definitely be eye-poppingly expensive. However, given the current competitive GenAI market, companies like SoftBank may decide that investing more in AI will be worthwhile.
