By 2028, A Top AI Giant Will Be Downgraded Over Data Centers

My call: by the end of 2028 at least one of the AI gods, NVIDIA or a US hyperscaler, will be politely told by a ratings agency to take a step down from Olympus. The reason will not be antitrust, politics, or some rogue chatbot. It will be something far less glamorous: AI data-center debt.

The equity market still behaves like these companies are un-downgradable. Infinite cash flows, infinite narrative, infinite GPUs. But credit analysts do not live on vibes. They live on capex, coverage ratios, and the uncomfortable math of a multi-year AI build-out that suddenly looks less like a one-off splurge and more like a permanent lifestyle choice.

The Bet, Cleanly

Here is the forecast in scorable English.

By December 31, 2028, at least one of the following issuers: NVIDIA, Microsoft, Amazon, Alphabet, or Meta, will be downgraded on its long-term issuer or senior unsecured rating by Moody’s, S&P, or Fitch. In the agency’s own rationale, AI data-center capex or AI data-center related debt will be cited as a primary or co-primary factor.

One notch is enough. One agency is enough. The AI debt has to be named in the write-up, not buried in a footnote about “miscellaneous business risk.”

Driver 1: AI Capex Stopped Being A Phase

Somewhere around 2025, AI capex stopped looking like a spike and started looking like a base rate. NVIDIA calls itself a “data center scale AI infrastructure company” and is not being cute. Hyperscalers are re-laying the physical internet so every software team can call an LLM without their CFO having a heart attack.

That is expensive. It is also long duration. These are not laptops. These are multi-billion-dollar campuses with their own power politics, literally.

Ratings agencies tolerated the first wave as strategic investment. Now the run rate matters more than the press release. If AI build-out keeps growing faster than actual, provable AI cash flows, the math starts to look like this:

• Free cash flow after capex thins out or goes negative for a year or two.
• New data centers get funded less by retained earnings and more by term debt or debt-like commitments.
• Leverage and coverage creep toward the top of the current ratings ranges.

At that point a ratings committee has a choice: stretch the methodology again, or admit that the AI arms race is now a credit risk, not just a cute strategy slide.

Driver 2: The Disclosures Already Smell Smoke

NVIDIA’s own filings now point at credit ratings as a live risk, tied to its capital intensity and infrastructure role. That is not Reddit drama, it is the company telling you the agencies are watching.

Meanwhile, the cloud world is getting wobblier at the edges. Neocloud names that chased AI adjacency without mega-cap cash flows are already catching short-seller attention for hidden AI commitments and high data-center costs. Utilities like Consolidated Edison talk openly about AI-driven grid investment that must somehow be financed and recovered. AI sellers are minting money. AI’s plumbing is quietly stacking obligations.

Then you have iPower, a small operator that has experimented with Bitcoin, Ethereum, and yield-bearing AI tokens like sUSDai as part of an AI infrastructure financing strategy. When financing the future of compute starts to rhyme with DeFi summer, that is not a vote of confidence in cheap, friendly bank debt.

The majors are not staking sUSDai, to be clear. But the frontier always shows up in miniature first. The more we see weird financing at the edges, the more attention the core gets from people who sign off on AAA.

Driver 3: Duration Mismatch Meets Human Analysts

The basic tension is boring and lethal: data-center debt is long term, AI demand is assumed long term.

Assumed is doing too much work. Monetization is still a moving target. Copilots, AI search, summarization tools, advertising uplift, new SaaS layers, fine. The question is not whether they make money. The question is whether the margin stack and volume justify permanently higher capex and power contracts across multiple economic cycles.

Ratings agencies have been here before, with fiber, 3G, 4G, 5G. The pattern is familiar. At first everything is strategic, no one wants to be the skunk that “missed” the future. Then one day the coverage ratios and off-balance-sheet commitments are too loud to ignore. Entire sectors get reframed from unstoppable platforms to capex addicts.

AI data centers are being pitched as the new grid. That is exactly how you tempt a credit analyst to question whether you are a tech platform with optionality or a utility with a chronic spending habit.

Who Blinks First?

The safest path for the agencies would be to grumble in research notes and keep the halos shiny. Just move outlooks to Negative if things get strange and blame everything on macro.

I do not think they all take that path. The pressure is coming from three directions at once:

Methodology hardening. As AI moves from story to line item, agencies will be forced to spell out how they treat AI capex, long-term capacity deals, and power contracts. That codification will expose who is running hot on leverage and who is not.

Fixed income patience. If credit spreads for the AI leaders start to widen versus non-AI tech peers, it gives the agencies cover. “We are just catching up to the market” is a time-honored way to downgrade a national champion.

Contagion optics. The first time a big utility or AI-adjacent name gets cut with explicit mention of AI loads and grid capex, analysts will be asked why the hyperscale buyers of that capacity are still treated as if capital is optional.

The Countercase, Honestly

Could I be wrong? Absolutely.

If AI revenue ramps even faster than spend, if debt stays cheap enough, and if the agencies manage to convince themselves that every data hall is self-funding within a couple of years, the majors can sail through 2028 with pristine ratings and maybe a few cranky footnotes.

They also have nuclear options that normal companies do not. Massive equity cushions. The ability to dial back buybacks, slow hiring, stretch vendor terms, and pull cost levers if credit ever looks tight. A single quarter of discipline from a hyperscaler can clean up two years of overenthusiasm.

There is also politics. Downgrading the poster children of American AI in the middle of an industrial policy arms race is not career risk free. Some committees will prefer to mutter about “heightened capital intensity” without touching the rating grid.

I still think at least one of them crosses a line. All it takes is one issuer that leans a bit harder into debt, one set of AI commitments that turns out larger and stickier than disclosed, or one macro wobble that hits high-flying AI revenue before the capex ramp can be trimmed.

Stakes: What Breaks If I Am Right

The first downgrade will not shut off the lights in a data center. It will do something more annoying: reset the cost of capital story for AI.

Higher spreads, more scrutiny of every “strategic” AI partnership, tighter leash on utility-like capex. The debate will stop being “Is AI overhyped?” and start being “How much leverage can the AI stack safely carry?”

If I am right, equity investors will finally have to price AI growth as something that has a balance sheet, not just a multiple. If I am wrong, congratulations: you live in a world where AI data centers are so cash generative and so blessed by ratings agencies that trillion-dollar companies can industrialize the planet’s power grid without losing a single notch.

In that scenario, the joke is on me, and on every credit analyst, because we will have discovered the first capital project in history that scales forever and never needs to be paid for.