SoftBank just posted record profits driven by OpenAI's soaring valuation. So why is Masayoshi Son selling $15 billion in assets, expanding margin loans, and converting paper gains into cash? The answer reveals uncomfortable truths about AI's peak.
Meta's chief AI scientist Yann LeCun reports to a 28-year-old after twelve years building FAIR. Now he's leaving to raise billions for the exact research Meta couldn't afford. The death of corporate AI labs and the VC arbitrage replacing them.
Gamma hit $50M ARR with 52 people while AI peers burn billions. Now at $2.1B valuation, the profitable presentation tool faces its real test: can a purpose-built AI product beat Microsoft and Google's free bundled features?
SoftBank's Nvidia Exit Reveals the Paper Billionaire's Paradox at AI's Peak
SoftBank just posted record profits driven by OpenAI's soaring valuation. So why is Masayoshi Son selling $15 billion in assets, expanding margin loans, and converting paper gains into cash? The answer reveals uncomfortable truths about AI's peak.
Masayoshi Son just reported SoftBank's best quarter ever. Net profit: $16.2 billion. More than doubled from last year. The Vision Fund booked massive gains as OpenAI's paper valuation rocketed from $300 billion to $500 billion across seven months. SoftBank's stock tripled since April. Son's now Japan's richest man.
He's also selling everything he can get his hands on.
October brought a full exit from Nvidia: all 32.1 million shares, $5.8 billion. Same month, $9.2 billion in T-Mobile stock got liquidated. The Arm margin loan grew from $13.5 billion to $20 billion. Add an $8.5 billion bridge loan for the OpenAI commitment. This isn't riding high. This is someone converting paper wealth into actual cash before the window closes.
The timing stings. Nvidia hit $5 trillion in October, first company ever. The infrastructure backbone of AI. SoftBank grabbed roughly 5% in 2016. Sold it in 2019. That stake today? Worth over $210 billion. Son called it "the fish that got away." Now he's watching it swim off again, possibly at peak valuation.
Peak of what, though?
Key Takeaways
• SoftBank sold $5.8B in Nvidia and $9.2B in T-Mobile despite posting $16.2B profit, expanding debt to fund OpenAI
• Circular financing pattern emerged: Nvidia invests $100B in OpenAI, which buys Nvidia chips, inflating both valuations artificially
• Stargate's $500B commitment produced few actual data centers; Altman signed $550B in separate computing deals instead
• OpenAI hit $500B valuation on $13B revenue while losing $5-7.8B annually; secondary sale let insiders cash out
The Circular Logic of AI Capital
Standard narrative says SoftBank's rotating from one AI winner to another. Sell the mature chipmaker, fund the transformative application layer. Portfolio rebalancing, strategic stuff.
Look at the actual flows. Nvidia announces up to $100 billion in OpenAI investment. OpenAI commits to buying Nvidia chips with that money. OpenAI's valuation rises (these announcements help). SoftBank's paper gains increase. Those gains justify pouring $30 billion into OpenAI. OpenAI spends much of it on Nvidia chips. Microsoft owns a piece of OpenAI but pays billions to CoreWeave for compute. Nvidia holds equity in CoreWeave. OpenAI just grabbed 10% of AMD while promising to buy their chips.
Capital moves in circles. Each deal validates the previous one.
Bloomberg called it "an increasingly complex and interconnected web of business transactions." Diplomatic. Dot-com veterans might use different terminology. Circular financing. The late-1990s game where equipment vendors and customers inflated each other's numbers without generating real external demand. Telecom companies bought fiber-optic gear with vendor financing, vendor reported it as revenue growth. Beautiful system until someone asked where the actual end customers were.
The scale now makes 2000 look quaint. Over $1 trillion headed toward AI infrastructure over the next few years, if you believe the announcements. Who's buying what they produce? OpenAI projects $13 billion in 2025 revenue while burning $5 billion to $7.8 billion annually. That's a $500 billion valuation on a 39x revenue multiple for a company that doesn't cover costs. The recent $6.6 billion secondary share sale wasn't a funding round. Existing shareholders cashed out. OpenAI itself got zero new capital.
The Infrastructure That Wasn't
Stargate. Remember that one? Trump's White House announcement in January. $500 billion data center joint venture, OpenAI, SoftBank, Oracle. American AI dominance, 100,000 jobs. Son stood there promising to "begin deploying $100 billion immediately."
Ten months in, here's the score. Abilene, Texas site: operational. Five more locations announced in September, bringing claimed investment past $400 billion. But the Wall Street Journal reported in November that "progress proved slow, leading OpenAI CEO Sam Altman to turn elsewhere for his computing needs." Since spring, OpenAI signed multiyear computing deals totaling $550 billion with Oracle and Microsoft. Separate from Stargate.
So what did Stargate's $500 billion commitment actually produce? Some data centers that were getting built anyway under different names? The pattern repeats. Trump announced Foxconn's Wisconsin factory in 2017. The promise: $10 billion investment, 13,000 jobs. What actually happened: scaled-down facility, about 1,000 employees, most of the original plan scrapped.
Big number at press conference, modest reality later, everyone moves on before the accounting happens.
Forced Liquidation at the Top
Here's the tell. SoftBank just posted $16.2 billion quarterly profit. Yet CFO Yoshimitsu Goto said they need to "divest our existing portfolio so that capital can be utilized for our financing." Not rebalance. Not optimize. Divest for financing.
The Paper Billionaire's Paradox in motion. SoftBank's books show huge gains because OpenAI's mark-to-market valuation keeps climbing. Those gains boost reported profits, make the Vision Fund look brilliant. But paper gains don't cut checks. Can't buy Ampere Computing for $6.5 billion with an unrealized gain. Can't acquire ABB's robotics arm for $5.4 billion with accounting entries.
So SoftBank liquidates assets that actually generate cash. Nvidia pays dividends, has real revenue growth. T-Mobile throws off billions in free cash flow. These get sold to fund OpenAI, which burns billions annually with no disclosed path to profitability. Like selling your index funds to buy more lottery tickets because the tickets you're holding appreciated on paper.
The Arm margin loan expansion signals stress too. Borrowing $6.5 billion more against your chip designer to fund AI bets isn't confidence. It's needing cash now and not wanting to sell more winners.
The Warning Signs Are Flashing
SoftBank's not alone sensing something off. Bank of England Governor Andrew Bailey warned about AI bubble risks in November. "Very positive productivity contributions" from tech could get offset by uncertainty about future earnings, he noted. Apollo Global Management's chief economist Torsten Slok declared the AI bubble bigger than dot-com, with the top 10 S&P 500 companies more overvalued than the 1990s. Even Sam Altman admitted in August that "investors as a whole are overexcited about AI."
The CEO whose company just hit $500 billion while losing billions thinks the market's overheated. Worth considering.
Nvidia at $5 trillion assumes AI buildout continues indefinitely. Palantir reportedly trades at 700x earnings. Companies with minimal revenue command billions based purely on AI association. These multiples don't price current fundamentals. They price a future where AI transforms everything and generates returns justifying today's spending.
What if that future takes longer? Or looks different? DeepSeek emerged in China showing effective AI doesn't necessarily require massive Western-style capital expenditures. More efficiency breakthroughs make the infrastructure overbuild harder to justify.
The Exit Before the Exit
SoftBank's behavior suggests Son sees this. He's not fleeing AI, he's converting solid positions into speculative ones at valuations that may never repeat. Nvidia will probably remain valuable. Whether it stays at $5 trillion or reverts to something more normal like $3 trillion or $2 trillion is the question. A 40% decline still leaves it a massive success. It destroys anyone who bought at the peak with borrowed money, though.
OpenAI at $500 billion needs to become more profitable than most Fortune 500 companies. Possible? Sure. Likely at that exact valuation? History suggests caution. When secondary sales dominate over primary funding, insiders are cashing out. They've seen this movie.
Son's famous for big bets that either pay spectacularly or implode. His $20 million Alibaba investment became $150 billion, funding the Vision Fund and cementing his legend. He also dumped $18 billion into WeWork before it collapsed, called it "a stain on my life," watched the Vision Fund post an $18 billion loss in 2022 when his pre-rate-hike investments cratered.
This Nvidia exit echoes that pattern. Son tends to ride momentum past the peak, then suffers. This time he's exiting earlier. Wisdom or just age? Either way, when the world's most famous AI bull liquidates the infrastructure layer to pile into the loss-making application layer, worth asking what he knows that isn't in the press releases.
⏱️
Miss one day. Miss everything.
AI waits for no one. We'll keep you caught up.
Free daily. Quit anytime.
The dot-com crash left real companies behind. Google, Amazon, Microsoft survived and thrived. Thousands of others didn't. Investors who bought Pets.com or Webvan at peak never recovered. AI will likely follow the same script. Survivors and casualties. Real value and vaporized capital.
SoftBank's asset sales suggest Son's betting he knows which category OpenAI belongs in. The $5.8 billion question is whether he's right.
Why This Matters
For AI Infrastructure Investors: The largest AI bull just exited the dominant hardware play to fund application-layer companies burning billions. If Nvidia at $5 trillion doesn't interest SoftBank, either they're wrong or the risk/reward shifted dramatically. Hardware has proven revenue and margins. Applications mostly burn capital on promises of future monetization.
For Late-Stage Startup Employees: Secondary sales at $500 billion valuations with minimal primary capital suggest insiders are de-risking. Founders and early employees cashing out at these numbers implies "this might be the peak." If you're holding illiquid options in AI companies, that's a data point.
For Market Timing: Circular financing patterns and forced liquidations despite record paper profits mirror late-cycle behavior from previous bubbles. Not timing advice, but certainly a flag worth noting when your capital's at stake. AI boom likely has years left. Whether today's valuations have years left is a different question.
❓ Frequently Asked Questions
Q: What is circular financing and why is it a red flag?
A: Circular financing happens when companies invest in each other or their customers, creating artificial revenue growth. Nvidia invests $100B in OpenAI, which then spends it on Nvidia chips. Both report the money as business growth, but no external demand actually exists. This pattern inflated telecom valuations before the 2000 crash, then collapsed when real customer revenue never materialized.
Q: How much money did SoftBank lose by selling Nvidia in 2019?
A: SoftBank bought roughly 5% of Nvidia in 2016 for about $4 billion and sold in early 2019. If held until November 2025, that stake would be worth over $210 billion. The opportunity cost exceeds $200 billion, making it one of the most expensive timing mistakes in investment history. Son called it "the fish that got away."
Q: Why is OpenAI losing billions if ChatGPT has millions of users?
A: OpenAI projects $13 billion in 2025 revenue but loses $5-7.8 billion annually because training and running AI models costs enormous amounts. Each ChatGPT query costs roughly 10x more to process than a Google search. The compute infrastructure, Nvidia chips, and energy bills dwarf subscription revenue. Profitability requires either massive price increases or radical cost reduction.
Q: What's the difference between OpenAI's secondary sale and a funding round?
A: In a funding round, new investors give money to the company, which uses it for operations. In a secondary sale, investors buy existing shares from employees and early investors. OpenAI's recent $6.6 billion secondary put zero new capital into the company. Existing shareholders cashed out while OpenAI itself got nothing. It signals insiders want liquidity now.
Q: How much of OpenAI does SoftBank own after its $30 billion commitment?
A: SoftBank's stake will grow from 4% to 11% after completing its $30 billion investment through the Vision Fund 2. At OpenAI's current $500 billion valuation, that 11% stake is worth $55 billion on paper. However, OpenAI remains unprofitable and private, meaning SoftBank can't easily convert that paper gain into actual cash without finding buyers.
Tech journalist. Lives in Marin County, north of San Francisco. Got his start writing for his high school newspaper. When not covering tech trends, he's swimming laps, gaming on PS4, or vibe coding through the night.
Meta's chief AI scientist Yann LeCun reports to a 28-year-old after twelve years building FAIR. Now he's leaving to raise billions for the exact research Meta couldn't afford. The death of corporate AI labs and the VC arbitrage replacing them.
OpenAI projects $20 billion in revenue while lobbying Washington for tax credits designed for struggling manufacturers. The company has committed $1.4 trillion to AI infrastructure it cannot finance privately. Is this confidence or desperation?
Moonshot trained an AI model for $4.6 million that beats OpenAI's GPT-5 on reasoning tests. While OpenAI seeks trillion-dollar infrastructure, Chinese labs prove the math no longer works. The twist: even they worry the tech works too well.
Microsoft declares it's building "humanist superintelligence" to keep AI safe. Reality check: They're 2 years behind OpenAI, whose models they'll use until 2032. The safety pitch? Product differentiation for enterprise clients who fear runaway AI.