AI Is Becoming a Bond-Market Problem

The conventional chain from expensive energy to sticky inflation to higher Treasury yields is still operating. What changed this week is the borrower list. The Wall Street Journal reported that Broadcom, Oracle and SpaceX are each pursuing blockbuster financings — tens of billions of dollars apiece — specifically to buy AI chips. At the same time, the U.S. 10-year yield touched about 5.35 percent, its highest level since 2002. The market is no longer only pricing oil and deficits. It is discovering that AI infrastructure has become a capital-markets peer.

The deals remain early-stage talks, not closed offerings. Their architecture still matters more than their press-release totals.

Three structures, one capital call

Broadcom has been arranging more than $50 billion in financing for OpenAI’s custom chips, the companies’ joint program, with private-credit names including Apollo and Blackstone in early discussions. SpaceX is seeking roughly $40 billion — about $30 billion of investment-grade debt plus $10 billion of bank financing — to buy Nvidia chips, with Apollo again expected to lead and a close sketched for 2027. Oracle is talking with Apollo, Goldman Sachs and others about financing a large chip purchase for roughly a one-gigawatt data-center commitment; the structure under discussion would have investors fund a separate entity that owns the hardware and leases it back, keeping the debt off Oracle’s corporate balance sheet.

Those are three different wrappers around the same fact: cash flow from cloud and AI services arrives after the hardware invoice. Amazon’s reported GPU vehicle and Broadcom’s Anthropic convertible facility already showed how compute is being turned into a financeable claim. This week’s cluster shows the pattern scaling into sovereign-adjacent ticket sizes.

The killer ratio

MarketWatch, citing FactSet, put SpaceX’s projected 2026 revenue near $44.5 billion. A $40 billion chip financing would therefore approach a full year of company revenue — borrowed to purchase one supplier’s hardware. That is not a normal equipment lease. It is a balance-sheet event large enough to reprice how lenders think about SpaceX after its mid-year investment-grade debut and $25 billion bond sale.

A loop is forming in plain sight: investors fund debt; debt buys AI infrastructure; infrastructure pays Nvidia and Broadcom; capacity expands; the next round of borrowing gets easier to justify. Private credit sits in the middle because the public bond market alone is no longer the only pipe large enough — or patient enough — for these tickets.

Governments vs. AI for capital

PIMCO’s Christian Stracke has already argued that hyperscaler build budgets compete for scarce capital, labor, power and equipment, lifting real rates even before inflation fully settles. Reuters and other desks have framed the same contest from the sovereign side: AI corporate borrowing competing with governments for investor dollars. Treasury’s own response to a 24-year high in the 10-year has been to lean on bills and buybacks rather than feed more duration into a weak book — a reminder that Washington is already managing scarce demand for long paper.

None of that proves AI debt “caused” 5.35 percent. Fed policy, inflation compensation, oil and Treasury supply still dominate the attribution. The residual claim is narrower and more useful: when Broadcom, Oracle and SpaceX show up as $40–50 billion borrowers in the same week the sovereign curve marks a multi-decade high, the crowding channel stops being theoretical. The competition is no longer only government versus government. It is governments versus AI infrastructure for the same pool of capital — and the financing architecture is being built to keep that competition running.