Why AI credit spreads are suddenly flashing 2008-style warning signs

Most people think of the AI boom as a stock market story. Nvidia, the Magnificent Seven, soaring semiconductor indexes – it all feels like a replay of the dot-com era. But the most important AI stress signals right now aren’t coming from stocks at all. They’re coming from credit markets.

Behind the headlines about record AI profits and data center spending, the price of money itself is changing. Credit default swaps are blowing out, bond investors are demanding higher yields, and even the strongest tech giants are seeing their cash flows strained by the cost of the AI arms race.

From dot-com bubble to AI credit boom

In the late 1990s, the bubble was mostly about equity. You could add “.com” to your name and raise money in the stock market with barely any revenue and no clear path to profitability. Today’s AI leaders are very different. Amazon, Alphabet, Meta, Microsoft, Oracle, and others are profitable, global businesses with huge customer bases and strong balance sheets.

That’s why many people dismiss the idea of an AI “bubble.” The technology is real, the companies are mature, and AI is already being used across search, coding, productivity, and more. But bubbles don’t require fake technology. The internet was real. Fiber networks were real. E-commerce was real. The problem back then wasn’t the tech – it was the price paid, the timing, and the assumptions used to justify massive overbuilding.

The same distinction applies to AI. Artificial intelligence can be transformative while still generating terrible returns for many of the investors funding today’s buildout.

The AI race has become a financing race

For the first phase of the AI boom, it looked like Big Tech could fund everything out of its own cash flows. These companies generate enormous revenue and, historically, strong free cash flow. They also enjoy some of the best credit ratings in corporate America.

That picture is changing fast. By early July 2026, AI-related debt issuance reportedly hit around $270 billion – nearly double all of 2025 already. Roughly $194 billion of that came from the usual suspects: Amazon, Microsoft, Meta, Alphabet, and Oracle.

They’re not borrowing because their core businesses have collapsed. They’re borrowing because even their massive cash flows are being overwhelmed by the cost of staying competitive in AI. Every CEO has decided they must build now – more data centers, more GPUs, more power, more networking – because falling behind could be fatal.

When everyone reaches that conclusion at the same time, they all end up competing for the same scarce resources: chips, electricity, construction capacity, and capital. The AI race stops being just a technology race and becomes a financing race. And a lot of that spending will ultimately fund losing bets, not winning ones.

Usage vs. profitable usage: the core AI question

It’s clear that AI is being used. Millions of people interact with generative AI tools. Businesses are rolling out copilots, AI search, coding assistants, and automated customer support. The technology keeps improving and spreading.

But there’s a crucial difference between usage and profitable usage. Many AI features are currently free, bundled into existing products, or priced below their true cost. Unlike traditional software, where serving each additional user is almost free once the product is built, generative AI has a continuing cost every time someone runs an inference – every query, every image, every document processed burns compute.

The industry is betting that these costs will fall quickly and that AI will become so embedded in workflows that companies gain strong pricing power. That might happen. But cheaper models – including aggressive competition from Chinese players and open-source systems – are challenging the idea that only the most expensive, compute-hungry models will win. As we’ve covered in pieces like why Chinese AI is suddenly so good, you can increasingly get strong performance without eye-watering infrastructure costs.

If comparable results can be achieved with far less compute, today’s mega-scale AI infrastructure plans could turn out to be excessive. Adoption projections might be broadly right while revenue and pricing assumptions are wrong. That is exactly the kind of uncertainty credit markets are built to price.

What credit default swaps are suddenly signaling

One of the clearest signs of shifting sentiment is in credit default swaps (CDS) – a market that came to fame during the 2008 financial crisis and is now back in focus.

A CDS is like insurance on a company’s debt. The buyer pays an annual premium to a seller, and if a defined “credit event” (like default or restructuring) occurs, the contract pays out. The higher the CDS spread, the more investors are willing to pay to protect themselves – not necessarily because default is imminent, but because demand for protection is rising.

Recently, CDS spreads for major AI-linked names have jumped:

• Oracle’s CDS spread climbed to around 215 basis points, up from about 145 at the end of last year – a multi-year high.
• SpaceX’s CDS spread surged to roughly 185 basis points, rising by more than half since trading began just a month earlier.
• Nvidia, Meta, Amazon, Alphabet, and CoreWeave have all seen their CDS spreads hit new highs, even if at lower levels than Oracle and SpaceX.

These numbers don’t say “default is coming tomorrow.” What they say is that investors are no longer comfortable assuming everything will work out. A year ago, the main worry around a big AI-related bond sale was getting enough allocation. Now, in cases like SpaceX, dealers were quoting CDS prices before the bonds were even officially announced – because investors wanted a hedge before they had anything to hedge.

That’s a major behavioral shift: from “how much can I buy?” to “how do I protect myself?”

Bond markets are still open – but at a higher price

Credit stress rarely shows up first as a slammed-shut door. It appears as a rising price for walking through that door. Recent bond deals tied to AI show exactly that pattern.

Amazon’s bond sale: weaker enthusiasm than expected

Amazon recently sold about $25 billion of bonds across eight maturities from 3 to 40 years. Initial orders hit around $62 billion, but when the banks managing the sale cut the spreads (reducing the yield for investors), orders dropped to roughly $41 billion.

That still sounds like strong demand – about 1.6 times the deal size. But for Amazon, one of the world’s most trusted corporate borrowers, it was unusually weak. High-grade U.S. bond offerings this year have typically attracted orders around four times their size.

Amazon also had to offer a relatively high “new issue concession,” with some of the longest bonds reportedly paying an extra 18–21 basis points versus comparable existing debt. The company got the deal done, but only by paying more and accepting less exuberant demand. That’s a subtle but important warning sign.

BlackRock–Meta data center financing: a 7.53% wake-up call

Shortly after, a $12.5 billion bond deal emerged to finance a Texas data center project for Meta, arranged through BlackRock entities. The structure is more complex – project financing rather than direct corporate borrowing – but the message from the market was similar.

The bonds took nearly a week to place and ultimately priced at a yield of 7.53%, one of the highest yields for a blue-chip data center financing since the AI borrowing wave began. Again, the bonds sold. The stress is not about access; it’s about price. Investors are starting to demand much more compensation to fund AI infrastructure.

Google’s first-ever negative free cash flow

Perhaps the most symbolic data point so far is Alphabet (Google) posting negative free cash flow for the first time since going public in 2004. The figure was about –$5.9 billion.

Alphabet is not insolvent. Its advertising, YouTube, Android, and cloud businesses remain highly profitable. But free cash flow – what’s left after operating expenses and capital expenditures – turned negative because AI-related capex exploded.

Recent numbers highlight the scale:

• Quarterly capital expenditures jumped to around $45 billion, roughly double the level a year earlier.
• Alphabet raised its 2026 capex budget to between $195 billion and $205 billion and warned that spending will rise significantly again in 2027.

And crucially, spending twice as much doesn’t mean building twice as much capacity. Memory prices are climbing, power connections are scarce, transformers and turbines have long lead times, and suitable land is getting more expensive. With the same companies bidding against each other for the same inputs, costs are rising faster than expected – and not all of that translates into proportional increases in compute or future revenue.

For Alphabet, and others like it, that creates a double risk: will AI generate enough incremental revenue, and will each new dollar of capex buy less infrastructure than the last? That’s how an equity growth story turns into a credit and cash flow story.

Hidden leverage and off-balance-sheet AI bets

As AI spending soars, companies have a strong incentive to be creative about how that debt appears – or doesn’t appear – on their balance sheets. Project financing, special-purpose vehicles, and complex partnerships can all shift where the risk seems to sit.

The Texas data center project tied to Meta and BlackRock is a good example. BlackRock entities reportedly hold about 80% of the project’s equity and financed part of that stake with the $12.5 billion bond sale. Meta keeps roughly 20% of the exposure and guarantees lease payments linked to the project.

On paper, much of the debt belongs to the project itself rather than Meta. Economically, though, the project still depends heavily on Meta. That doesn’t automatically mean anything shady is happening – project finance is common and can be efficient. But as structures get more complex, it becomes harder for investors to see who is ultimately on the hook if things go wrong. In many cases, that opacity is intentional.

Banks financing these projects can also use credit derivatives to hedge or transfer risk. That means rising CDS activity may reveal growing anxiety that isn’t obvious from corporate balance sheets alone. The market is no longer just financing a handful of tech giants; it’s financing them plus a web of SPVs, infrastructure partnerships, data center developers, power projects, and chip suppliers – all anchored to the same AI revenue assumptions.

Semiconductor stocks and the late-cycle pattern

While credit markets are flashing yellow, equity markets are also starting to wobble, especially in semiconductors. In July, the MSCI World Semiconductor Index fell about 16%, putting it on track for its worst month since 2022, even though it remains up significantly for the year.

Semiconductor shares dropped across South Korea, Japan, Taiwan, and the U.S. One telling example: SK Hynix reported a six-fold increase in quarterly profit and margins above 80% – and its stock fell 19%.

Why would investors punish such strong results? Because expectations were even higher, and because SK Hynix also announced at least $31 billion in capital spending for the year, about 50% above its previous level. What looks like proof of extraordinary demand can also look like a classic late-cycle move: record profits driving record investment, which then creates excess capacity and eventually pressures prices and returns.

At the same time, reports of Chinese progress in advanced lithography raise fears of future competition and potential oversupply in global chipmaking. This is the part of the dot-com comparison that really matters: at the peak of a capital cycle, the best financial results often arrive just before investors start asking whether those results can possibly be sustained.

Stocks price the dream, credit prices the payments

Equity markets tell you what investors hope a company might become. Debt markets tell you what investors believe can be safely financed. Right now, those two stories are starting to diverge in AI.

Consider the signals together:

• Oracle’s CDS cost is at multi-year highs.
• SpaceX’s CDS spreads have surged shortly after launch.
• Amazon’s massive bond sale drew unusually weak demand and required higher concessions.
• A BlackRock-backed data center financing for Meta had to offer a 7.53% yield.
• Alphabet just posted its first-ever negative free cash flow as AI capex explodes.

None of these on their own prove an imminent crisis. But together they show that the market is actively recalculating risk. The AI boom is no longer just about who can spend the most and grow the fastest. It’s increasingly about who can turn that spending into durable, dependable cash flows – and who will be left holding the bag if AI revenues don’t live up to the projections.

For investors, that means paying attention not just to AI product launches and model benchmarks, but also to balance sheets, capex plans, and credit spreads. The most exciting AI technology won’t necessarily belong to the best investment if the financing behind it doesn’t add up.

If you’re exploring where AI value might actually accrue – to tools, platforms, or more efficient stacks – it’s worth also looking at how competition and cost structures are evolving. Our breakdown of why everyone is suddenly rethinking Google search is one example of how AI-driven shifts in business models can ripple through both revenue and risk.