My view on the AI build out and the money financing it, is undecided. I sense it's a timing point. Which I never like.
Longer term, I’m sure AI will be as transformational as we all believe it to be. Shorter term, I feel there has to be a correction to wash this out as it cannot continue like this indefinitely.
‘shorter’ or ‘longer’ I know are unhelpful in actionable terms. But I'm setting a trigger that tells me when 'shorter term' may start. I feel that's the most prudent action to take now for my risk management purposes. As I'm not a pessimist on this theme, at all.
Show most people a map of their own industry, that they are an expert in that looked like this, and they would not invest in it. So why does the money keep flowing? Either not enough of us are seeing the whole map, or perhaps we have our blinkers on, given the actual and promise of returns.
Like most of us, I read a lot of investment newsletters and then move on. Let's be honest, they can be dry.
One from the Ares alternative credit team, titled In The Gaps, has stayed with me because of a single page. It is a diagram. Every financing raised for AI data centres and chips over twelve months, drawn as a web, with lines running from each deal to every company it depends on.
Short version:
• Ares counted well over 100 financings for AI data centres and chips in twelve months, issued across seven different debt markets. Underneath, the risk kept resolving to the same eight companies.
• Several of those companies sit on both sides of the same structure: paying the rent, guaranteeing the building value at end of lease, borrowing and acting as tenant at the same time.
• A $14bn data centre that property insurers will only cover to about 3% of its value has bonds secured on its rent, and those bonds are rated as low risk. Same building, and two quite different views of the risk.
• Every deal on the map, Ares says, rests on one assumption: that AI capital spending keeps going.
The map
Heavily simplified version recreated to highlight examples. If you follow the two highlighted lines, i) the same name pays the rent and guarantees the building ii) the same name sells the chips and guarantees the loan.
Each deal on its own reads well, of course. Put them on one page and you cannot help but feel uneasy. The lines keep arriving at the same eight circles. Ares traced more than 100 financings, across seven separate debt markets, and found that the risk kept resolving back to Meta, Oracle, Microsoft, Amazon, Google, Nvidia, OpenAI and Anthropic.
Both sides of the same deal
Meta pays the rent that services one set of bonds, and also guarantees the building's value at the end of the lease.
CoreWeave appears as a borrower three times and is also the anchor tenant whose leases support another issuer’s bonds.
Nvidia, Ares says, is on the map four ways at once: it issues its own notes, supplies the chips behind the structured deals, is reported as a tenant, and stands as guarantor behind other companies’ borrowing.
The companies are also wired to each other. Nvidia has reportedly committed funding to OpenAI, which buys compute from Oracle, which runs it on Nvidia’s chips. Microsoft, Amazon and Google hold stakes in the AI labs they also sell to.
In substance, I may be oversimplifying, but I see this coming down to one assumption: that AI capital spending keeps going. If a period arrives where AI revenue disappoints the spending built for it, and the guarantees get called at the same time, from the same companies, at the moment they are least able to pay, who steps in then? This is not 2008. I don't see a reliable lender of last resort for this industry in the way governments were for banks.
A building nobody will insure
The same newsletter has a second example. Ares describes a $14bn data centre site in Texas with Meta as the tenant. The property insurance is capped at roughly $450m, about 3% of the value. Insurers looked at a site that size and decided a total loss was too big a risk to cover.
Now look at the same site from the bond side. Debt investors have lent against the rent Meta pays for the building, at investment grade. If the building burns down, Meta moves its computing somewhere else. The lenders lose the only thing they were secured on, a rent stream from a building they have no claim to, and which nobody insured past 3% of its value.
Ares puts it in one line. “The same risk that the insurance industry finds too big to insure in the property market as a liability, it swallows whole as an asset when it arrives as an investment grade bond”.
Often it is the same industry on both sides.
Why the money keeps flowing
If the map is that clear, why does anyone buy?
Because investors aren't buying the map. They're buying individual transactions within it. A real estate fund sees a lease. An equipment lender sees collateral. Each one is doing its job well, judged against the criteria for that one piece.
I am somewhat bemused why the investing industry isn't talking enough about this. My guess is that as the chart still goes up and each deal passes the standard investment due diligence checks then it’s business as usual. I would like to see greater emphasis on the intersection between investment and operational due diligence. Both can assess their respective risks effectively, but it's the dependencies between them, and across apparently separate investments, that I think deserve more scrutiny, and more weight of this intersection placed into the investment decision making process.
Ares also looks at who is buying. US insurers hold around $9 trillion of assets, and record annuity sales mean they need to put a lot of it into ‘A rated bonds’. The bonds they buy are expected to fail about four times in every ten thousand each year. That is how they price them. But these bonds rest on one industry carrying on spending at the current rate, and a spending cycle turning is nowhere near a four in ten thousand event.
This is not only about AI
The pattern is older than data centres. Ares points to telecoms in 2000, where equipment makers financed their own customers and a technology story turned into a correlation story. The same shape sat under the financial crisis, where structures moved risk around rather than reducing it. Each time, things that looked separate shared one assumption, and when it broke, they moved together.
It applies at a smaller scale too. A thematic basket of ten companies in one sector shares one driver. A global tracker, a US tracker and a technology fund own the same handful of companies three times over. And while you'll be holding many instruments, the actual drivers or the number of independent bets is often one or two.
I don't have a settled view yet. What I do have is a growing awareness of this environment.
Each of these deals makes sense on its own. The rating, tenant, the rent. The issue I find only really shows up through relationships and interconnectedness. This is pulling me towards a view that a number of opportunities that appear more independent than they are or with acceptable levels of separation, feel like they actually share a stronger single assumption.
Having worked in the industry during 2008, it is hard not to see some similarities, or perhaps the better word is familiarities. Mortgages were packaged, rated and sold to investors who were required to buy highly rated paper. Each package looked diversified. Underneath, much of that apparent diversification depended on the same assumption: that house prices wouldn't fall together country wide (US).
The rating agencies had the information but saw it far too late. Criteria looked more at each package on its own terms. There was less focus on the interconnectedness.
It is not the same. I know that. In 2008 the borrowers at the bottom could not pay, and the packaging hid that. Here, the companies at the centre of the map can pay. They are some of the most profitable businesses in the world.
The risk here is when the decision to spend less, which can start a process that tests guarantees and leases built on the assumption that spending would continue.
If that happens, where the damage lands depends on the financing structures and the protections within them. The smaller companies in the middle of the map, including former bitcoin miners and chip rental businesses, look particularly exposed. But some of that risk could transfer back to the eight major companies through their guarantees and commitments.
What is familiar is the risk transfer: one assumption, many wrappers, each assessed on its own.
My View
Longer term: Comfortable as I expect AI to matter as much as most of us think it will.
Shorter term: I'll be watching AI capital expenditure guidance. If two or three of the major companies begin reducing their spending expectations over one or two quarters, that would be my first signal that the investment cycle may be changing.
That wouldn't automatically mean the financing structures were in trouble. I would want to understand why spending was slowing, and whether that slowdown was beginning to put pressure on the revenues, leases and guarantees supporting the wider industry.
That would be my trigger to reassess my exposure, with a view to reducing it if the evidence supported doing so.
Until then, my longer-term view remains positive. It's the timing of the potential correction that I cannot establish.
Source: Ares Management, In The Gaps, Alternative Credit Newsletter, Fall 2026. Figures on financings, insurance coverage, counterparties and funding are as reported there.
