Softbank reportedly taps up Nvidia for $960m AI push

According to a report by Nikkei, Softbank will spend JPY150 billion ($960 million) over the next two years on upgrading its computing facilities. The plan will involve procuring graphics processing units (GPUs) from Nvidia. Softbank will not only use them to power its in-house large language models (LLMs), but will also offer access to them for any corporate clients in need of a supercomputer.

The investment will dwarf the JPY20 billion ($130 million) Softbank spent on computing infrastructure last year, and according to Nikkei, is believed to be the largest such investment of any Japanese company.

The GenAI hype train is showing no signs of slowing, and Softbank wants to capitalise on the demand for AI that can converse in Japanese.

Last March, the telco launched a new division tasked with developing LLMs and offering GenAI services. That division was subsequently launched as a fully-fledged subsidiary, SB Intuitions, last August.

By last October, SB Intuitions had finished deploying its computing platform and was busy working on its first homegrown LLM. It is due to be completed this year and will be powered by 350 billion parameters. That would make it significantly more capable than OpenAI's GPT-3, which has 175 billion, but significantly fewer than the rumoured 100 trillion parameters powering GPT-4.

This is worth bearing in mind because earlier this month, OpenAI revealed it has opened a new office in Tokyo, and has begun providing businesses with early access to a version of GPT-4 that speaks Japanese. The pressure is on Softbank to offer a viable alternative.

Softbank's AI computing platform runs on Nvidia hardware – specifically an Nvidia DGX SuperPOD AI supercomputer powered by more than 2,000 Nvidia Tensor Core GPUs. Given their track record of working together, it stands to reason that Softbank would tap up Nvidia for its AI supercomputing expansion plans.

Incidentally, France-based telco group Iliad last September treated itself to a SuperPOD, prompting it to lay claim to Europe's most powerful AI supercomputer. It's a serious bit of kit, but Nvidia – which hopes to grab as much of the GenAI processing market as it can – has already unveiled its successor.

Last May, the US-based chip maker combined its Grace general processing unit (GPU) with its Hopper GPU to create Grace Hopper – a new product line designed specifically to excel at GenAI. Nvidia's upcoming DGX GH200 supercomputer will use 256 of these Grace Hoppers, and will be pitched to companies that want to build their own LLMs.

There's no word on pricing yet, but it is bound to be eye-wateringly expensive. However, given the GenAI land-grab that's currently underway, companies like Softbank may consider it money well spent.