The Banks Said No. The Savers Are Next.

Goldman Sachs, JPMorgan, and Mizuho were asked to lend SoftBank $10 billion against its stake in OpenAI. They looked at the collateral — a private company's shares, no public price, no easy exit — and said no. ^[1][2] That refusal is the whole story in miniature. SoftBank cut its ask to $6 billion, and S&P lowered its credit outlook to negative. Then it turned to a different lender: Japanese households. This month SoftBank is preparing a record 1-trillion-yen retail bond sale, the largest by any issuer in the country's history, seven-year paper at 4.3% to 4.9% — a spread of barely 1.4 points over government debt. ^[3][4] The party that could price the risk said no, so the risk moved to a party that can't. The same transfer is happening everywhere, in different flavors. Alibaba raised $10.2 billion this week through the largest share placement in Hong Kong exchange history, pricing 710 million shares at an 8.4% discount, while its quarterly profit is expected to fall 50% as it spends on AI. ^[5][6] The cost of the buildout is being paid by diluting the shareholders who already own the company. Meta kept $27 billion in data-center debt off its balance sheet by putting its Louisiana facility in a joint venture 80% owned by Blue Owl Capital, with PIMCO and BlackRock buying the bonds. Consumer advocates say the structure was altered the very day regulators approved the power plan, and that four-year renewable leases contradict the 15-year terms originally approved. ^[7][8] The power-cost risk, meanwhile, flows to Louisiana ratepayers. OpenAI is rushing toward an IPO its own CFO has said the company isn't ready for, because public markets offer more scale than private ones. ^[9][10] The company carries roughly $900 billion in infrastructure commitments against a projected $200 billion in 2030 revenue. ^[11] Apollo's chief economist Torsten Slok puts a number on the gap: the buildout may need $2 trillion in debt, and traditional bond markets can absorb only about $1 trillion through 2030. The other trillion has to come from somewhere — private credit, asset-backed structures, and ultimately retail and public investors. ^[12] Andromeda Capital describes the setup as one where credit investors accept record-tight spreads and capped upside while the risk migrates toward private-equity vehicles, life insurers, and circular financing. ^[13] None of this is because AI has no revenue. Cloud revenue is surging — Google Cloud up 82%, Azure 43%, AWS 37% ^[14] — and Anthropic just posted its first operating profit, $559 million on $11.6 billion in quarterly revenue. ^[15] OpenAI's sales are running ahead of forecast, 15% above 2025 projections. ^[16] The problem is not the absence of revenue. It is the gap between revenue and the capital required to produce it. Morgan Stanley's math: a fully optimized data center using the latest Nvidia chips costs $25 billion a year to run and generates $23 billion in output. ^[17] The buildout can't pay for itself, and the shortfall is being underwritten by people who cannot price what the banks already refused. Consider the person at the end of the chain. A Japanese saver buys a seven-year bond at 4.3% from a company with a negative credit outlook and $40 billion in bridge loans due in March 2027. ^[2][3] For holding risk the banks would not touch, they get a point and a half over what the government pays. That is the spread the AI boom is now offering the people it has come to depend on.