"Not being able to invest in Anthropic before its IPO was a mistake."
Masayoshi Son said it plainly at a Tokyo event on November 18. He did not say he misread the company. He said SoftBank Group (9984.T) had a chance in 2023 and passed, because it had already decided to concentrate on OpenAI.
One horse
"In the same horse race, it is hard to bet on more than one horse."
That was the explanation. At the time, he thought he should not have a foot in two boats.
In horse racing the horse is not the point, the odds are. Pick one right and you take everything; spread the bets and a win only returns your stake. That worked when one winner ate the whole track.
AI is not like that.
There is more than one model company. The cloud providers build their own. The chip companies keep moving down the stack. Who eats whom has no answer today. Betting one horse here is not courage. It is mistaking odds for probability. Probability can be calculated. Odds can only be guessed.
And a race ends. AI does not. Lose today and you can change lanes tomorrow; win today and a new architecture can unseat you next year. Betting several horses is not diversification. It is admitting you do not know who wins, the honest read in 2023.
The $64.6 billion conversion
Since September 2024, SoftBank has put money into OpenAI through the Vision Fund: about $64.6 billion committed, for roughly 13%.
Back out the implied valuation: about $496.9 billion. Run the same sum against Anthropic's $1.5 trillion and 13% is $195 billion.
A threefold gap. In AI, where winners may number only a few, that is the distance between surviving and not.
This is not SoftBank's loss to book: it holds no Anthropic stock. Which company you pick is a gap of three times, not tens of percent, and it comes from the asset, not the effort.
$64.6 billion spent today at $1.5 trillion buys 4.3% of Anthropic. Same money, three times the stake on the other side. The choice of horse sets the weight of the position.
Why he chose it anyway
On results alone, the answer is simple: wrong horse. His explanation is that he could not bet both. The first is an error of judgment, fixable; the second is a method.
SoftBank's capital is not unlimited. To hold a heavy position in AI you concentrate, and not only money: board attention, leverage over cloud and chips, the willingness to add when things go wrong. Diversify and you are a small holder everywhere, with no capacity and no voice.
Last year he sold Nvidia (NVDA) stock. His word for it was "through tears." He sold it to raise money for OpenAI.
That fact matters. The method has a price, and he has paid it. Selling a rising asset to load another is the standard concentrated move, and the hardest. Win and it is foresight. Lose and it is stubbornness.
Now look at Anthropic itself. Purchase and lease commitments: $900.9 billion. SoftBank's $64.6 billion is equity; Anthropic's $900.9 billion is procurement.
But they point at one object. The barrier is no longer whether you can fund a position. It is whether you dare sign. Sign and capacity is locked. Do not sign and no amount of money gets you past the queue. OpenAI took SoftBank's money and signed longer contracts with it. Son was betting on the ability to sign, which is about credit, not technology.
$900.9 billion is 13.9 times $64.6 billion. One company's purchase book exceeds one investor's entire commitment by an order of magnitude. In a model company's hands, money is a commitment. In an investor's hands, it is a stake.
What he actually admitted
He said he had met Dario Amodei, thought him a remarkable person, and wished he had invested then. Then he added: "But Anthropic is a great company."
The first is regret, the second manners. The word with information in it is mistake.
There are three kinds: wrong industry, wrong company, wrong moment. He claimed the third, the lightest. Misread an industry and you change tracks; misread a company and you change targets; miss a moment and you only change pace. Three errors, three fixes, and he picked the one that requires no self-revision.
A public admission of error is rarely an apology. It re-prices the next bet. He conceded the outcome, not the method.
The audience was wider than reporters. SoftBank's LPs were listening, OpenAI's management, and Anthropic too. To the LPs it is an accounting, to OpenAI loyalty, to the market disclosure that SoftBank has one horse.
The next fork
So will he bet one horse again? His answer was that it is hard to bet several. Hard, not wrong.
The 2023 call looks right in 2023 terms: OpenAI had a product, revenue, Microsoft (MSFT); Anthropic had a model and a few people. It looks wrong now. Same man, same reasoning, opposite answers. The odds structure was rewritten by compute, by orders, by that $900.9 billion.
Judging 2023 with today's numbers is the rear-view mirror. There was no $81 billion run rate, no $900.9 billion of commitments, no $1.5 trillion price. What was visible was capability, team and funding cadence. The word mistake can only come from him.
So next time, he may not change. Concentration is not temperament; it is capital structure. The larger the assets, the harder it is to hold a heavy stake everywhere, and without one you get no voice and no capacity. Position sets the view. That logic expires only if the next concentration is wrong.
When an investor says publicly that he missed a $1.5 trillion company, he has usually found the next target or is building the case for a bigger bet.
Good for him either way. Not necessarily good for the people listening.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
