Credit repriced the AI buildout in August. The Treasury market hasn't.

In the middle of August, two markets quoted two different prices for the same companies. The market that sells insurance against a borrower failing to pay — credit default swaps, or CDS — widened sharply on the big AI names inside two weeks, while Nvidia and Oracle shares just sat near their highs ^[1]. Oracle's swaps touched a record 261 basis points, which Morgan Stanley reads as better than one-in-five odds of default within five years ^[2]. Two prices, one company. The gap couldn't hold, and it didn't — but note which price moved. By mid-September, equities had walked toward the credit column rather than the reverse: Oracle shares were down 28% on the year ^[3]. What the credit market was actually repricing matters more than the repricing itself. The verdict isn't on AI. It's on capacity that doesn't exist yet, and it's selective. Construction loans on Oracle's Project Jupiter trade at 89 to 91 cents on the dollar ^[4]. Firmus Technologies withdrew its Australian listing after cutting the offer price from A$11 to as low as A$5.50 ^[5]. The company's own explanation is the driest line in the whole repricing.

The board determined that the terms on which the offer could proceed would not appropriately reflect the strength of the company's business and long-term growth outlook. — Firmus Grid Pty Ltd

The two data-center real estate trusts floated this year are both down around 16% ^[6]. Paper tied to unbuilt megawatts now trades as risk. But the same market kept writing the other kind of paper. Amazon's $53.8 billion raise in March — the largest corporate bond sale on record — drew something like $126 billion of orders ^[7], and hyperscalers were still lining up fresh deals as of October 10 ^[8]. JPMorgan's head of global syndicate put the new logic plainly.

Notwithstanding just how volatile markets are right now, for the right companies – the truly great companies and credit stories – there is still a tremendous amount of capital available. — John Servidea

The market didn't stop lending to AI. It started distinguishing between a balance sheet and a blueprint. That distinction has already reordered the global risk hierarchy, which is the least-noticed line in the whole buildout. TSMC's bonds now trade at lower yields than Meta's, and SK Hynix's 2031 notes pay just nine basis points over Amazon ^[9]. Emerging-market corporate debt, historically 30 to 50 basis points wide of US benchmarks, has converged to a few ^[9]. The world's bond desks have decided that wealthy-Asian chipmakers are safer credit than the American companies borrowing to buy their chips. The money leaving the physical layer had to land somewhere, and you can see where. Palantir raised its 2026 growth projection to 82% and was upgraded to Buy at 81 times sales ^[10][11] — a multiple that towers over the 20-to-30 range typical of software, with growth expected to halve next year as OpenAI and Anthropic move into its market ^[12]. Even the accounts tied to the president rotated out of infrastructure suppliers and into monetization names and dividend stocks in August ^[13]. Google's enterprise backlog is the hinge: revenue rose 24% on enterprise AI demand and the backlog is up tenfold in three years ^[14]. But that backlog is booked, not built. It is take-or-pay contracts against capacity that does not exist yet, the same character now being repriced elsewhere. Which is the backdrop for the product line the labs shipped this fall. OpenAI launched Dots, an agent for routine corporate work ^[15], and after OpenAI's Decisions API came Microsoft's Decision-1, a model for speeding routine operational decisions ^[16]. The category has already turned into a price war: OpenAI cut its Decisions API in direct response to Jev, a classification model from a startup run by a former OpenAI employee, which charges $0.042 per million tokens against OpenAI's $0.10 ^[17]. The most advanced companies on earth are now racing to the bottom on the price of a routine decision. The labs are building on the side of their own trade the market is paying up for. The clock behind that move is the same clock behind everything else here. The buildout was financed on take-or-pay contracts — commitments to pay for compute whether or not it's used — that are still mostly under construction; in 2027 and 2028 those bookings flip into actual bills, the reset wall ^[18]. Roughly half of Oracle's $664 billion backlog rests on one counterparty, OpenAI, whose own commitments are financed ^[19]. Bain's framing of the race has become the industry's actual business plan.

The question is whether the applications arrive in time to pay for it — Bain & Company, Inc.

The underwriting of all this is circular in a way that reads like a bond-desk joke. SoftBank is selling a record £1 trillion of retail bonds to refinance the bridge loan behind its OpenAI stake — the same stake that secures a $10 billion margin loan. Its shares fell 2.6% on the announcement ^[20].

We are considering various options to refinance the bridge loan, but nothing has been decided, including the amount for each. — SoftBank Group

Amazon's record raise included up to $50 billion for OpenAI ^[8]. Broadcom is borrowing through a special-purpose vehicle that buys its own hardware and leases it to its own customers ^[21]. And the people inside the labs have spent the year quietly selling their own firms' equity ^[22]. One column hasn't moved at all. Washington still prices the buildout as the growth engine that pays for everything else. The US projects $10.3 trillion of AI spending through 2032, and Fed governors have cited AI demand as the thing that knocked inflation progress off course ^[23]. Treasury Secretary Bessent, facing more than $40 trillion of federal debt, says the country will have to grow its way out of it ^[23].

AI-related demand, along with higher energy prices, has helped knock inflation progress "off course." — Mike Sullivan

The BIS warned in June that this could end in a global crisis ^[24] — before the first spread in this piece moved. Deutsche Bank's George Saravelos finds the Treasury market still positioned long the exact vision a bust would vindicate ^[25]. His clients are the giveaway: they read AI as the driver of higher yields, which is to say the long end of the bond market is priced for the boom, not for the bust the swaps desk started pricing in August. Michael Burry has a one-word diagnosis for this market ^[26].

the stock market is in its first stage of grief: denial. — Michael Burry

Denial now has an address, and it isn't the swaps desk or the stock market. The booked backlogs flip to bills in 2027 and 2028. The credit column started that clock in August. The Treasury column hasn't priced it at all.