AI's 'Capital Wall' Is a Price Tag

In February, a $4 billion data-center financing fell apart because lenders declined CoreWeave's B+ credit rating, leaving the asset manager Blue Owl holding a $500 million bridge ^[1]. A month later, Amazon raised $53.8 billion in a single morning — the largest corporate bond sale on record — against roughly $126 billion of demand ^[2]. The only real difference between the two deals was the name on the borrower. J.P. Morgan's John Servidea put that difference 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

Behind the word is a number. Bank of America now calculates the hyperscalers' cost of capital — what it costs them, blended across debt and equity, to raise a dollar — at 10%, double the 4% of 2023 ^[3]. That is the price of admission, and it is the price doing the deciding. Amazon is the strongest credit on the ladder, and even it felt it. In February the company guided to $200 billion of 2026 spending, $53 billion above what analysts expected, and its chief executive vouched for the returns ^[4].

With such strong demand for our existing offerings and seminal opportunities like AI, chips, robotics, and low earth orbit satellites, we expect to invest about $200 billion in capital expenditures across Amazon in 2026, and anticipate strong long-term return on invested capital. — Andy Jassy

The stock fell as much as 11.5% that day ^[4]. By July the pushback had a number attached in the bond market: a $25 billion Amazon sale drew the weakest hyperscaler demand since October 2025, forcing 18 to 21 basis points of extra yield onto its longest bonds, and the company told underwriters it would issue no more debt in 2026 after borrowing $92 billion for the year ^[5]. Bank of America described the result.

Investors are pushing back. — Bank of America

Two rungs down, Alphabet posted its first negative free cash flow since going public — minus $5.9 billion — even as it raised 2026 capital-spending guidance to $195 billion to $205 billion, closing the gap with $84.75 billion of new equity and more than $114 billion of debt assumed since early 2025 ^[6].

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

Below the giants, the price climbs fast. Anthropic, with no public balance sheet to lean on, has had to offer Blackstone more than 10% for data-center money, up from 8.5% four months earlier ^[7]. Oracle — a mega-cap, not a startup — now trades near junk, the rating tier below investment grade where borrowers pay up for default risk, pushed there specifically by its AI data-center spending ^[8]. CoreWeave, the B+ borrower from February, can still borrow — it placed $4.2 billion of convertible bonds on October 13 — but on terms that show the strain: its yields have run near 15%, and its projected third-quarter interest expense of $860 million to $940 million runs about four times its projected adjusted operating income of $200 million to $260 million ^[9][10]. And the Australian data-center IPO cleared no price at all — the listing failed to find buyers ^[7]. The February deal did not die for lack of demand in the abstract; it died on a rating. Lenders declined the B+, and the money never came. The bar they are applying has a voice.

I would expect a significant deceleration in capex growth in 2027, and again further deceleration in 2028. — James Russo

The repricing has already rearranged who counts as safe. With more than $600 billion of hyperscaler borrowing now outstanding, TSMC's bonds yield less than Meta's, SK Hynix trades nine basis points over Amazon, and emerging-market corporate spreads have converged with US benchmarks from a historical 30-to-50-basis-point gap ^[11]. Around $100 billion of investment-grade debt now trades wider than the junk curve ^[11]. The European Central Bank has its doubts about the ratings themselves, noting that agencies' assessments of big tech may rest on revenue and leverage assumptions that will not survive contact ^[12].

The way rating agencies approach this sector may be based on assumptions on future revenue growth and leverage which may not stand the test of time, heightening the vulnerability to mispricing of credit risk. — European Central Bank

The money that didn't stay in AI didn't vanish; it walked back to the old economy, and it has been walking for a year. The Dow set records last November while the Nasdaq fell ^[13]. MUFG is now Japan's most valuable company, the Topix bank index is up more than 40% this year, and HSBC and BOC Hong Kong have each gained more than 25% ^[14]. Healthcare has become the formal AI hedge, with FactSet tracking a negative correlation between the semiconductor ETF and the healthcare ETF ^[15]. Israel Englander and Ken Griffin loaded the SPDR Gold ETF as the year opened, even as roughly 90% of surveyed investors said they intended to hold or add AI ^[16]. None of it stopped the spending, which is the thing a wall is supposed to do. The headline numbers keep climbing: Nvidia's finance chief puts top-five hyperscaler capex at $800 billion this year, rising to $1.3 trillion in 2027 ^[17], and Goldman Sachs sees $1.2 trillion next year ^[18]. What the price has done instead is sort — which is why Amazon can raise $53.8 billion in one morning while CoreWeave's $4.2 billion at near-15% yields is the best a B+ can get, and the Australian listing cleared nothing. Some of the slowdown is physical rather than financial: satellite data suggests 40% of US data centers may slip their 2026 completion dates on shortages of pipe fitters and electricians, tariff-bottled Chinese transformers, and utilities that cannot expand power fast enough ^[19]. But the borrowing isn't free to everyone else either. The ECB calculates US tech firms now account for nearly 10% of gross new euro-zone bond issuance on their way to a possible $1 trillion of AI investment by 2028, and warns it may crowd out other borrowers and push up costs for governments and companies alike ^[12].

U.S. big tech companies could push up borrowing costs for all sectors as they accumulate debt and account for a growing share of bond markets, with a potential spillover to the sovereign and supranational segment of the bond market. — European Central Bank

The bullish forecasts carry their own condition: the spending persists as long as hyperscaler credit spreads don't widen ^[20]. Apollo says they already have, by about 60 basis points since October 2025 ^[21]. CoreWeave's books show what the bottom rung is carrying: $64.6 billion in lease liabilities that extend beyond its existing customer contracts ^[9]. The company has said its long-term stability depends on what those contract renewals price after 2026 ^[9]. That is the bet, and it now carries a price.