I ran a large prompt using Fable and ChatGPT 5.6 Sol to examine the underlying conditions and financial structures behind the physical AI infrastructure buildout taking place right now. One of the top questions on investors' minds across the globe is whether AI, and the physical buildout around it, is a bubble ready to pop, or a long-term trend upward. Let's dig in, shall we?
(The full interactive report is here: AI Infrastructure Financing: State of the Market. The portfolio project card and Agent Research paper are on magro.dev.)
Historical context: sizing it against dotcom and '08
For historical context, I wanted a rough estimate of the TAM for this AI buildout, compared against both the dotcom bubble and the '08-'09 financial crisis. This is admittedly not an apples-to-apples comparison, but it gives us some idea of size and scope. Comparing a peak subprime stock, or a six-year fiber buildout, to a single year of AI debt is not a useful signal. On a multi-year basis, Street estimates for AI data-center / hyperscaler capex run roughly $2.0T to $2.5T for 2025 through 2027e (about $2.2T at the midpoint). Peak U.S. subprime was about $1.3T outstanding in 2007, still smaller than I expected given that the U.S. MBS market alone is currently between $11T and $12T. Then again, that $1.3T sat against a total MBS market of roughly $8.5T in 2007.
Estimates of the dotcom-era telecom and fiber buildout still land somewhere around $750B (with a wide range), spent over roughly six years and funded by more companies than I could name. That remains the closest physical-infrastructure analogue. The financing sitting under AI is a different number again: a base-path cumulative AI-related debt figure for 2025 to 2027e is also around $750B, with a wider scenario range if issuance accelerates. The dotcom comparison intuitively feels more apt for the buildout itself: we were building out the infrastructure for the internet, which has obviously been an enormous success. The bubble popped and the infrastructure still won. But the dotcom bubble was much more of an equity story than a fixed-income one. High-flying valuations sent the NASDAQ from a peak of 5,048.62 in March 2000 down to 1,114.11 by October 2002, a 78% drawdown. Credit spreads to risk-free assets widened, but the pain was overwhelmingly in equities. Subprime was the opposite: a credit and funding story. That distinction, equity risk versus funding risk, is the lens for everything that follows.
Risk one: funding. Balance sheet is the scarce resource
I spent many years trading fixed income: UST basis, liquidity premium, financing (repo) rates, off-the-runs versus on-the-runs, and futures calendar spreads dominated by inflexible real-money buyers expressing duration views through the futures market. Working in this space is like going through first principles of the shadow banking system in real time. Traders here see patterns over time and play the distortions that arise precisely because balance sheet is a precious resource to be maximized. One of my best mentors always framed it as looking for the best return on balance sheet.
Balance sheet, the finite resource an investment bank allocates across its desks (including the repo it offers institutions), is usually saved for higher-margin products with much higher carry: loans, and various credit and mortgage products. The Treasury desk was typically given just enough to safely facilitate the auction process and let its traders hold decent off-the-run positions while still capturing bid/offer. The rest went to customers with prized relationships.
In practice, this allowed obvious anomalies to persist (I'll get into that in a later post), but it's the funding risk in these products that really accelerated the collapse in 2008. CDOs were the product of the day. Much of the senior paper was genuinely AAA: stamped by the rating agencies and, at the top of the capital structure, it suffered far fewer defaults than the headlines would suggest. What took down Bear Stearns and Lehman Brothers wasn't the paper they sold and distributed to real-money end investors. It was what stayed behind on their own dealer balance sheets: the equity tranches (the truly awful bonds with the worst credit risk), plus the unsold super-senior slices still sitting in the warehouse, all of it funded overnight in repo and commercial paper. That's a great strategy... until your overnight lenders stop showing up. Going back to the carry point above: nothing beats a high yield, except a really high yield, and in the ZIRP years the temptation to keep that carry in-house was irresistible.
So the first risk I care about in the AI buildout is the same one: who holds the paper, and how is it funded? The question I was really pushing the agent to address was: how much of this AI debt is sitting on levered dealer balance sheets, financed short? The honest answer was "we don't know precisely, but nothing close to the subprime setup." The paper is largely in the hands of long-term investors in private credit, locked-up money that doesn't face an overnight roll, and so far they're doing great with it. That's the single most important structural difference from 2008.
Risk two: correlation
The second market risk I'm particularly attuned to is correlation. Even USTs and basis can blow out to nearly a full point on bond-future basis in times of stress. The participants who normally make these markets efficient get forced out, and by the time it gets to that level, everyone knows it. Correlation in the AI buildout shows up in the customer base: the end consumers of all this compute are largely the same handful of companies, Meta, Microsoft, Google, Amazon, and the labs they bankroll. Whatever hits one of them tends to hit all of them, and it hits the borrower and the collateral at the same time. The inputs, on the other hand (land, power, hardware, heating and cooling, memory, which we all got educated on during the DRAM crunch of the last few months, and the rest of the supply chain), are genuinely diversified across suppliers. But that cuts the wrong way here: because the verticals are chained in series rather than in parallel, it's almost an anti-diversification play. A bottleneck anywhere in the stack can shock the entire process, so diversification of suppliers buys you less than it looks like it should.
Risk three: GPU depreciation. Real, but not the main event
The third risk gets the most attention in the press, but in my opinion it's nowhere near as important as the funding question: GPU useful life and utilization. I'm not an AI hardware expert, but the accounting lives being used are five to six years (CoreWeave discloses six years for technology equipment; Amazon five to six for servers), and lenders are already haircutting that by taking the chips as collateral at initial advance rates around 60 cents on the dollar of book. That layered conservatism seems pretty fair given everything I understand from reading on the topic. These are great GPUs, H100s and the like, and while they may lose the frontier training race to newer generations, they remain perfectly good inference machines, and regardless of the state of the world we will need inference. If somewhere between now and then they become obsolete faster than the accounting assumes, it seems reasonable that investors will push new datacenter buildouts and newly issued bonds to be tied to the newer technology, and reprice the advance rates accordingly.
Putting the three together, the fact that the paper is in strong hands gives me some hope that we've learned from history and proceeded with sensible financial innovation. This debt market is built on bidirectional, complete information between sophisticated parties, as opposed to subprime, where people were buying houses they couldn't afford with no documentation, sitting in ARMs tied to the level of rates, and were decidedly not the types to be hedging their rate exposure with eurodollar futures. The scale story is more subtle than "AI is smaller." Physical AI buildout on a 2025 to 2027e capex path is already in the same league as, or larger than, peak subprime stock and the whole telecom/fiber episode. The financing under that buildout is closer to the fiber total on a cumulative base path ($750B), with 2025 alone still in the low hundreds of billions ($200B of AI-related debt, plus about $27B of data-center securitization). What keeps me from treating the headline size as the main risk is still who holds the paper and how it is funded, not whether the bars on the chart look big.
What actually worries me
I'm far more worried about open source and AI software development: harnessing models more effectively, optimizing the right model for the job, interoperability between memory, models, and sessions, and open source proving to be the better alternative for most use cases. Even then, we'll still need these datacenters to put in the robots!