Part IA Loose Thread (The One That Unravelled Everything)

I didn’t set out to write about debt markets.

In December 2025, I was staring at a power‑consumption estimate for a simple AI agent task—”read this email thread and draft a reply”—and the number didn’t make sense. Not morally, not environmentally. Thermodynamically. The energy required to have an AI handle something I could do in ninety seconds was orders of magnitude higher than just… doing it myself. A laptop that’s already on, a few keystrokes, a click. Versus an entire orchestration pipeline of model calls, tool invocations, and container spin‑ups burning through a data center somewhere in Virginia.

That question—why is so much invisible infrastructure required for something so small?—turned into a two‑month investigation that followed the watts up through the financing stack: from per‑query energy, to trillion‑dollar build‑out plans, to GPU‑backed special‑purpose vehicles rated A+ by the same agencies that blessed mortgage CDOs, to private‑credit funds sold to retirees as “income products.”

Then, just as I was finishing the draft, the punchline arrived on its own.

In February 2026, Blue Owl Capital—one of the most aggressive private‑credit players in data‑center finance and a sponsor of Meta’s $27–30 billion AI data‑center SPV—gated a retail fund, sold over $1.4 billion in loans to raise cash, and helped trigger a CNBC headline about “private credit worries” rocking U.S. markets.¹²³

If 2008 had a canary in the coal mine with Bear Stearns, the AI credit cycle might have something more fitting: an owl in the coalmine. And unlike the canary, this one has a $201 billion balance sheet behind it.⁴

This piece traces how we got here. It starts with the boring math—watts per email—and ends at a question nobody building AI data centers wants to answer honestly.

Part IIThe Boring Math (That Isn’t Boring at All)

When you send a short email or check tomorrow’s weather, the physical world beneath your screen barely stirs. The marginal energy to push a small text message through existing networks is on the order of a millionth of a kilowatt‑hour—micro‑watt‑hours in data‑center terms. Even generously folding in your laptop’s idle draw, you’re talking milli‑kilowatt‑hours per email. For decades, the digital economy’s selling point was that these costs were so tiny they felt free.⁵⁶

A typical search query: roughly 0.3 Wh. Loading a lean weather page: same ballpark or lower. Thousands of such queries for the cost of a single kWh on your electricity bill.⁷⁸

Route that same question through a large language model, and the numbers jump. Public benchmarks put a single chatbot request at around 0.34 Wh per query, with some analyses stretching to 2.9 Wh depending on model size and infrastructure. On a per‑task basis, asking ChatGPT “What’s the weather tomorrow?” consumes 10–1,000× the energy of just loading a forecast page.⁷⁹¹⁰

Now hand the entire task to an agent—not just the answer, but the doing—and the gap blows wide open.

A chatbot call is like asking one smart person a question. An agent is like commissioning a small consultancy engagement: plan, research, back‑and‑forth, write‑up, follow‑through. All happening behind the scenes every time you say “handle this for me.”

Technically, that means 3–10 LLM calls per task: once to plan, again to choose tools, again to interpret results, again to refine. Recent research on dynamic reasoning agents measured this concretely—for an 8B‑parameter model, stepping from a single‑shot answer to an agentic chain increased energy from 0.32 Wh to 22–42 Wh. For a 70B model: 2.55 Wh to 158–348 Wh per request.¹¹

Add tool calls, serverless function spin‑ups, API hops, database queries. Each one modest. Stacked together: 1–100 Wh per completed agent task, or roughly 100–10,000× more energy than the manual version.^9115121314

Here’s what that looks like in a table, because the ratios deserve to be seen, not just read:

That’s the invisible physics beneath every “let the AI handle it” moment. An SUV‑to‑a‑block‑walk ratio, repeated billions of times per day, at scale. And the industry’s response to this overhead is not to question it—it’s to build more data centers.

Part IIIThe Productivity Gap (And Why the Trajectory Matters)

If the extra watts were buying something transformative, you could argue thermodynamics is just the price of progress. So what do we actually know about AI’s economic upside — and how fast is that upside arriving?

At the micro level, the current picture is rough. Somewhere between 75–95% of AI projects fail to deliver measurable ROI or never reach stable production. The successful minority averages roughly 1.7× ROI — but that sits atop a graveyard of pilots and experiments that burned capital and compute for nothing.^{15,16,17151617}

At the macro level, the gains are real but bounded by where adoption actually stands today. Wharton, OECD, McKinsey, and Goldman cluster around AI adding 0.2–1.0 percentage points to annual labor‑productivity growth when fully diffused — a condition that assumes years of deep integration across sectors, not the current patchwork of pilots and point solutions. Over decades, that compounds into low‑single‑digit to low‑teens percent higher GDP levels. Meaningful. But the keyword is decades.^18,19,20,2118¹⁹²⁰²¹

None of this settles the question. AI’s productivity contribution in 2026 is not its contribution in 2033. Models are improving. Integration is deepening. The optimists aren’t wrong to point at the trajectory rather than the snapshot — and this piece isn’t arguing that AI delivers nothing.

The argument is narrower and more specific: the infrastructure being built right now is priced for a productivity acceleration that hasn’t arrived yet, financed with debt that comes due on a schedule the technology may not match. The question isn’t whether AI eventually justifies its footprint. It’s whether it justifies this much leverage, this fast, against assets with this short a useful life.

That’s a timing problem as much as a technology one — and timing problems have a way of becoming credit problems.

Part IVThe Debt Machine (Follow the Money, Not the Marketing)

The answer lives in the financing stack. Follow that, and the picture changes fast.

By early 2026, Moody’s estimated that global data centers need roughly $3 trillion in investment through 2030, with AI as the primary driver. Some industry projections push the AI‑specific number north of $5 trillion when power and network upgrades are counted. That money isn’t coming from petty cash. It’s coming from debt.²⁴²⁵

On the safest end: bond markets. In 2025 alone, tech and AI‑related issuers accounted for roughly 16–17% of global non‑financial corporate bond issuance, up from 11.6% the year before. Meta and Google together sold on the order of $55 billion in bonds flagged for AI expansion. Oracle stacked a $38 billion term‑loan package on top of $18 billion in bonds.⁴²⁶²⁷

Further down: private credit. By late 2025, private‑credit lending to B‑rated and below borrowers had overtaken syndicated loans for four consecutive years. At least $200 billion in private‑credit exposure to AI‑related companies was already in place, with that figure expected to swell as mid‑market operators, GPU clouds, and software firms tap non‑bank lenders at 8–15% rates.⁴²⁵²⁸

Put roughly $1–1.5 trillion of AI‑linked debt at a blended 6–8% and the sector is on the hook for $60–120 billion per year in interest—before tariffs, insurance, or operating costs. In a world where AI’s aggregate benefits are measured in tenths of a percentage point of GDP growth, that interest bill starts to look less like smart leverage and more like ongoing drag.²⁵²⁷²⁹

Part VHardware’s Hidden Half‑Life (The SPV Time Bomb)

A growing share of this build‑out is being funnelled through special‑purpose vehicles—SPVs that hold long‑term tenant leases, GPU contracts, and sometimes tokenized “titles” to GPU racks. Those SPVs then issue notes or asset‑backed securities that rating agencies score and large asset managers buy.³⁰³¹

On paper, the collateral looks solid. In practice, a meaningful chunk of it is short‑cycle silicon. High‑end AI GPUs turn over on roughly 2–3 year product cycles. Google’s TPUs, Amazon’s Trainium and Inferentia chips are being marketed as delivering 2–4× better performance per dollar—accelerating the economic decay of older GPU fleets with each generation.^30323334

The debts these SPVs issue run 10–20+ years.

That’s it. That’s the structural problem.You’ve got paper maturing in 2049 backed partly by hardware that will be economically obsolete three or four times over before the first principal payment is due. The loans aren’t “backed by GPUs” in any durable sense—they’re backed by a narrative about sustained demand for whatever replaces today’s GPUs.

Part VIThe Real Estate Illusion (NIMBY Meets the Grid)

The strongest counterargument is real estate. Even if GPU economics weaken, the land and buildings under these campuses will appreciate. Lenders have a way out.

Recent developments make that assumption less reliable than it looks.

By 2025–26, local opposition to new data centers had become organized and explicit. At least 188 community groups in 17 states mobilized against data‑center projects, contributing to an estimated $64–98 billion in facilities blocked, delayed, or cancelled. Complaints are consistent: 24/7 noise, visual impact, heavy truck traffic, rising electricity bills, large tax abatements for operators, and intense water use in drought‑stressed regions.^35363738

Regulators have started responding. Counties in Northern Virginia that once allowed data centers “by right” have tightened zoning. Grid operators warn that power demand from U.S. data centers is expected to jump 28% in 2026 alone. And here’s the physical footprint in one snapshot—because the water and power numbers deserve to be stated once, clearly, and not repeated:

• Electricity by 2028: ~300 TWh/year in the U.S. alone—equivalent to powering 28+ million households. AI servers went from ~2 TWh in 2017 to a projected 300 TWh in 2028. That’s a 150× increase in a decade.²²²³
• Water by 2028: Up to 720 billion gallons annually for cooling—enough to cover the indoor water needs of 18–19 million people.²²²³⁵⁷
• Global electricity: Total data‑center and AI demand could exceed 1,000 TWh by mid‑decade, more than doubling from 2022.⁵⁸

None of this means every existing site loses value. But “data‑center land always appreciates” is no longer a safe modelling assumption. For SPVs and ABS that quietly bake in real‑estate upside as a backstop, the combination of local opposition, water and grid constraints, and environmental politics materially weakens one of the key lines of defence.^35363741

Part VIIWhose Retirement Is Holding This Paper?

From a distance, GPU‑backed SPVs sound like specialist instruments. In reality, their risk is already widely distributed into the portfolios of ordinary savers.

Rating agencies have built detailed methodologies for this exact purpose. S&P describes four approaches to rating data‑center deals; Moody’s and Fitch published frameworks allowing A–AAA ratings under the right structural conditions. Those ratings make the paper eligible for core bond mandates—the ones your pension fund uses.⁴²⁴³⁵⁰

In the Meta–Blue Owl Hyperion deal: a bankruptcy‑remote SPV issued roughly $27–30 billion of A+‑rated notes maturing in 2049. PIMCO anchored with around $18 billion. BlackRock bought more than $3 billion, distributing the bonds across income strategies and ETFs—the same ETFs sitting in millions of retirement accounts.⁴⁴⁴⁵

In parallel, private‑credit BDCs have become popular yield products for retail investors. When redemption pressure spikes—as we just saw in February 2026—managers respond by selling loans and gating withdrawals. That protects the fund’s balance sheet but locks investors in.⁴¹²

The risk of AI infrastructure overbuild and hardware obsolescence isn’t confined to a handful of specialists. It’s embedded, in tranched and rated form, in the bond funds, private‑credit products, and retirement portfolios of millions of people who experience it as “income” or “core fixed income.” When these structures get tested, those are the people who discover what “AI infrastructure risk” really means.

Part VIIIWe’ve Seen This Movie Before (And It Didn’t End Well)

None of this is new in financial history. The telecom bust of 2000–2002. Enron’s SPV shell game. The AAA‑rated mortgage tranches of 2008. Same pattern, different assets.

Telecom operators borrowed and spent $300+ billion building fiber and switching capacity, convinced internet traffic would justify the leverage. When demand and pricing fell short, at least 23 major companies went bankrupt, including WorldCom—then the largest U.S. bankruptcy. The fiber didn’t vanish. It became stranded overcapacity earning far less than needed to service the debts written against it.⁴⁶⁴⁷

Enron added the structural twist: SPVs and off‑balance‑sheet entities hiding risky assets while preserving investment‑grade ratings. Rating agencies and investors knew there was complexity—but treated the structures as robust until confidence snapped.⁴⁸

2008 scaled the pattern further. AAA‑rated tranches of mortgage‑backed securities where models recognized default correlations in theory but still assigned near‑zero risk to senior slices. When U.S. house prices fell nationally—a scenario many models treated as low‑probability—the loss assumptions failed.⁴⁹⁵⁰

The relevance isn’t that GPUs are subprime mortgages. It’s that we’ve once again created highly rated structures on top of assets whose long‑term cashflows are more fragile and correlated than the ratings imply—and the people holding that paper include retirees and bond‑ETF shareholders who never signed up to speculate on three‑year GPU cycles.

Part IXThe Leverage on Top of the Leverage

Part XThe Slow‑Burn Risks Nobody’s Pricing

Beyond credit and hardware, several second‑order forces could amplify stress.

Insurance. Large AI data centers concentrate billions in equipment and revenue in single facilities—often in regions exposed to floods, wildfires, heatwaves, or grid failures. Insurers are already flagging these sites as challenging risks. If premiums spike, coverage limits tighten, or capacity is withdrawn after a major event, operators face higher fixed costs on top of already substantial debt service.⁶⁰⁶¹

Grid stability. As data‑center power demand surges—that 28% jump in 2026—utilities are exploring new tariffs, demand charges, and curtailment obligations. If operators are forced to reduce load during peak events or face higher rates not in the original models, the predictability of their economics erodes. That matters for the SPVs built on assumptions of steady, high utilization.³⁸³⁹⁴⁰

Narrative reversal. AI is now at the centre of climate and water debates. If political consensus shifts from “essential infrastructure we must subsidize” to “extractive industry that must be contained,” future tax breaks and fast‑track permitting may become harder to obtain.³⁵⁵⁹

Software efficiency. Better architectures, quantization, sparsity, and compiler optimizations could reduce the FLOPs required per unit of useful work—undercutting the aggressive demand forecasts that today’s build‑out assumes.⁵⁸

Each introduces uncertainty about whether the future cashflows from current sites will justify the leverage piled onto them.

Part XIThe Bull Case (Honestly Presented, Not Strawmanned)

It’s important to give the defence its due. Sophisticated bulls aren’t blind to risk—they weigh it differently, and their arguments are serious.

They correctly note that SPVs derive revenue from more than GPU usage: power markups, land leases, and multi‑year tenant commitments generate cashflows you don’t see if you only model GPU‑hour pricing. Long‑term, take‑or‑pay contracts with counterparties like Meta or Oracle guarantee minimum revenues and impose termination fees—making “just walk away” scenarios less plausible. Power Purchase Agreements lock in energy costs for 10–15 years, insulating operators from rate volatility.⁴²⁴⁴

Restructuring is often easier here than in past crises. Instead of millions of homeowners, a typical SPV has a small number of institutional creditors and one or two anchor tenants. Maturities can be extended, coupons trimmed, equity re‑cut over a limited set of phone calls. Data‑center vacancy rates sit below 1–3% in key markets, with multi‑year waitlists. Sponsors like Blue Owl, with $201 billion under management, can inject capital to cure covenant breaches—exactly the manoeuvres we’ve just watched them execute.¹

From this angle, the right historical analogy is not Lehman in 2008 but the telecom hangover of 2000–2002: ugly for equity, painful for some bondholders, resolved through bankruptcies, asset sales, and consolidation without bringing down the financial system.

Taken seriously, these arguments don’t make the problem disappear. They change its shape. Instead of imminent implosion, the more plausible base case is a multi‑year period of stress and restructuring: downgrades, haircuts, extended maturities, and 10–20% cumulative losses in some bond and private‑credit products tied to AI infrastructure. Not the end of the world—but a very different picture from a frictionless AI boom that cleanly converts compute into productivity and shareholder value.⁴⁵⁶

Part XIIThe Owl in the Coalmine (The Signal, Not the Verdict)

That brings us back to February 2026.

Blue Owl—sponsor of Meta’s record data‑centre SPV, one of the largest private‑credit platforms in existence—halted redemptions in a retail‑facing middle‑market lending fund and sold roughly $1.4 billion of loans at 99.7 cents on the dollar. The stress was in corporate lending, not explicitly in AI infrastructure. Yet the market’s reaction told the story: Blue Owl’s stock fell 6–10%, peers like Blackstone, Ares, Apollo, and KKR traded down 3–5%, and CNBC led with a headline about U.S. markets being “rocked by private credit worries.”¹²³

This is not evidence that the Meta SPV or other AI deals are failing today. No covenant breaches, no downgrades, no forced sales in those structures—yet.

What it does show is that the funding layer we’ve been tracing—retail‑exposed private‑credit vehicles holding illiquid loans—is already under visible strain. When redemption pressure hit, the response was loan sales and gates, not quiet absorption. That is precisely the mechanism through which stress in one part of a sponsor’s business migrates to others—including AI infrastructure—if conditions tighten.

The owl in the coalmine is a timing signal, not a verdict. It tells us that private‑credit stress is arriving earlier than the late‑decade window many expected. That retail investors are more tightly bound into that stress than marketing materials admit. Whether the next wave lands on AI infrastructure in 2027–2029 will depend on demand, pricing, regulation, and hardware economics.

But the combination we have today—energy‑intensive agents handling trivial tasks, a multi‑trillion‑dollar AI build‑out financed with leverage and structured products, rising physical and political constraints, and now visible cracks in the private‑credit plumbing—is enough to justify a simple, uncomfortable question:

How much of this boom is genuine progress, and how much is thermodynamic and financial overreach waiting for a turn in the cycle?

I started this piece because of a number on a power‑consumption spreadsheet. I ended it watching a private‑credit fund gate retail investors while its parent sponsors the largest AI infrastructure deal in history. The spreadsheet and the headline turned out to be the same story.

The owl isn’t dead yet. It’s just stopped singing.