In 2000, Masayoshi Son spent six minutes with a Chinese English teacher named Jack Ma and wrote a US$20 million check. The company had no revenue and no finished business plan.

Son said later he had decided before Ma finished talking, because the man across the table had “shining eyes.”1

The company was Alibaba. SoftBank’s US$20 million bought a 34 percent stake, and Son held it through two decades of swings rather than cash out early. By the time SoftBank sold its last shares, in its 2024 financial year, that final tranche, the last slice of the stake it sold off, alone returned US$8.5 billion, about 425 times what Son had paid for it, and the full run is reckoned at close to US$150 billion.

He ran the identical instinct closer to Southeast Asia. In March 2019, SoftBank’s Vision Fund, the giant investment vehicle Son built to write cheques at this scale, put US$1.46 billion into Grab, the Singapore-based ride-hailing and payments company, at a US$14 billion valuation.

Vision Fund rode Grab’s rise, then sold the position down as it retreated from venture investing, the same way it has since exited Nvidia entirely and sold off billions of dollars of T-Mobile stock.

The same appetite is now aimed at something twenty times the size of Alibaba’s ceiling. SoftBank has committed close to US$64.6 billion to OpenAI, ChatGPT’s maker, enough to own about 13 percent of the company by October. This month it went further into debt to keep that commitment moving.

On September 13, SoftBank shares fell as much as 13 percent in a single session, the steepest one-day drop since July.

The trigger was OpenAI’s own chief executive, Sam Altman, who ruled out a public listing this year, the sale of shares to outside investors that would let Son turn a paper gain into cash. Altman called it “an ill-advised moment,” pointing to AI-safety concerns that Anthropic’s chief executive had just urged the whole industry to slow down for.

Two days before that, SoftBank had closed a two-year loan of US$11.87 billion from a syndicate of roughly twenty banks, more than the US$10 billion it had originally sought. Separately, it has been arranging a bond sale of US$10 billion to US$20 billion, and offering shares of Arm, the British chip designer SoftBank controls, as collateral, the security a lender can seize if a loan goes unpaid, for a further loan rather than sell a single Arm share outright.

The instinct has cost SoftBank before it has paid out. Jack Ma quit SoftBank’s own board in 2020, after an earlier wave of Vision Fund losses became too large to ignore.

In the year to March 2023 alone, the fund lost US$32 billion, most of it the price of a co-working company called WeWork and start-up valuations that never recovered.

None of this means Son is wrong about OpenAI. It means the reader should notice what he is not doing: Vision Fund booked a US$46 billion paper gain on the position in its 2025 financial year, driven mainly by OpenAI’s rising valuation.

He has still not sold a single OpenAI share to book any of it.

He has still not sold a single OpenAI share to book any of it.

Expect Son to close the pending bond sale, sized at US$10 billion to US$20 billion, within the next six months, and to keep offering up Arm shares as loan collateral rather than sell them outright to raise the same cash. The load-bearing assumption has held for twenty-six years: whatever he is holding is worth more leveraged than sold, and OpenAI’s delayed listing is a postponement of the payout, not proof the bet was wrong.

The load-bearing assumption has held for twenty-six years: whatever he is holding is worth more leveraged than sold.

If he is right, nobody will remember that the stock fell 13 percent in a single afternoon in September 2026.

If he is wrong, that afternoon is where the historians will say the pattern finally caught up with him.

Footnotes

  1. Son and Ma have each described the meeting differently across interviews over two decades; Ma said in a 2008 speech it lasted six minutes, other retellings put it closer to ten, and neither man has produced a transcript. ↩