SoftBank wants to invest in AI infra from balance sheet; use cases from Vision Fund

SoftBank founder and chief Masayoshi Son

While SoftBank has been saying for the past few quarters that it is back to an offensive mode of investing amid a broader pivot into the era of artificial intelligence (AI), its Vision Fund unit’s fresh deployment dropped 60 percent to $620 million in the June quarter.

For a moment, it might seem as if the Japanese tech investor is not walking the talk. But, that’s not true — the conglomerate actually made a mammoth $1.9 billion of investments in the quarter at a group level.

As such, the lion’s share of deployment is happening from the conglomerate’s own balance sheet at present and not the Vision Fund unit.

According to sources close to the developments, this is because SoftBank has settled on an approach to invest in AI infrastructure plays from its own balance sheet while it will back AI use-cases from the Vision Fund unit that is structured like a venture capital firm.

Two recent deals underline this strategy: Last month, SoftBank acquired a UK-based AI chip startup called Graphcore, whereas the Vision Fund unit is investing $10-20 million in search engine startup Perplexity.

While Graphcore competes with NVIDIA, the San Francisco-based Perplexity has seen its valuation jump to $3 billion as it garnered a large user base who want AI-generated answers for their search queries, not just links to pages thrown up in the traditional search engine format.

Meanwhile, SoftBank also bought out the Vision Fund’s holdings in chip company Arm last year at a valuation of $64 billion, around a month before it went public at a valuation of about $55 billion.

What explains this strategy?

"One reason could be that AI infrastructure needs a higher holding period as it takes more time to mature. Also, ideally you would like to hold it for longer because if infrastructure investments work out, they can create significant value such as Arm or NVIDIA. If it is from the balance sheet, then there is no tenure on the fund. That's why use cases are more conducive for a conventional fund structure," said a general partner of a venture fund based in India.

In an earnings call last week, SoftBank Group chief financial officer Yoshimitsu Goto highlighted at least three broad components in its AI infrastructure play — AI chips, AI data centres and solar energy to power the large energy demands of running AI data centres.

"AI infrastructure like data centres, chips and solar energy to power AI have more reliable cash flows than AI use cases. These are similar to other infra plays, being a secure stream of predictable cash flows. It could be the reason that SoftBank wants to invest in AI infrastructure from its balance sheet. In contrast, AI use case startups are still nascent and likely to be more risky bets with lower cash flow expectations in the immediate future and a lot of execution and equity risk," said Siddarth Pai, managing partner at 3One4 Capital.

Think of a mall and brand outlets in it. The mall owner has a decent yield that is quite steady, whereas the brands may be high margin but more risky. Brands can be churned out, but the infrastructure remains. That's the same difference between an infrastructure play and a use case play. Also, the biggest expense for use case companies will be compute they need for training and inference.

“The expenses of an AI startup will be the revenue of an AI infra company — a great strategy to play the entire field. One company's burn is another's bounty," Pai added.

This strategy has also been visible in the Vision Fund’s push to its portfolio companies to adopt  generative AI. An investor presentation for the Vision Fund unit gave examples of how its portfolio companies like Klarna, Meesho, Revolut, Grab and ByteDance have improved their business metrics or built new use cases for their consumers with the help of the technology.