Written by: Luke Sobieraj, Founder & COO, Digital Colliers
Broadcom is reportedly in talks to raise more than $60B in debt to finance AI chips, with the eventual beneficiaries being a small handful of frontier labs. Anthropic. OpenAI. A couple of hyperscalers. If your bank is anywhere in that syndication stack, directly or two hops removed through a data center REIT or a GPU-lease special purpose vehicle, you have a concentration problem that doesn't show up cleanly on any single credit memo.
The issue isn't the borrower on the paper. It's who the borrower's revenue depends on. And most regional banks I've watched underwrite this stuff can't answer that question in under a week.
The exposure isn't where the loan documents say it is
A regional bank syndicated into a Broadcom-style raise sees Broadcom on the tape. That's the named counterparty. But the cash flows servicing that debt depend on chip orders from a very short list of end customers, and those customers are themselves burning venture and strategic capital at a rate the public doesn't fully see.
Same shape shows up in adjacent deals. Data center construction loans where the anchor tenant is one lab. GPU-as-a-service financings where three customers make up 70% of contracted revenue. Power purchase agreements tied to a single hyperscaler's capex plan.
You can be diversified across six loans on paper and concentrated on two end-customers in reality. That's the cluster your risk committee needs a view on.
The internal data joins most banks are missing
To get a true cluster view, you need to join data that lives in at least four systems that don't talk to each other:
- The loan tape (borrower, facility, commitment, drawn balance)
- The KYC/CDD file (ultimate parent, related parties, beneficial ownership)
- The covenant and reporting package (customer concentration disclosures, revenue by top-5 customer, contracted backlog)
- Public and paid market intel (who's actually buying the chips, who's signing the offtake, who's on the tenant roster)
Most mid-market finance and risk teams still assemble this by hand in spreadsheets, which is why the same research shop that tracks close cycles finds month-end running eight to ten days at a typical mid-market outfit. If your close is that slow, your ad-hoc exposure query is slower. By the time the answer lands, rates have moved.
The operators getting ahead of this are building a per-borrower, per-project graph. Every facility gets tagged not just with the direct obligor but with the two or three end-customers whose demand actually services the debt. Then you can run the query the credit committee actually wants: if Anthropic's revenue growth slows by 30%, what's my drawn and undrawn exposure across the book.
Why this is a left-behind problem
The large money center banks already have some version of this. They've got the data engineering headcount, they've got the internal tooling, and they've been building cross-facility exposure views for a decade. A regional bank syndicated into the same paper does not have the same instrumentation.
That's the left-behind risk. Not that regionals can't participate in AI-infra credit. They can and they are. It's that they're participating with a view of exposure that assumes the borrower on the document is the risk, when the concentration actually sits one or two layers down. Around 95% of enterprise AI projects fail to reach production, and while frontier labs aren't average enterprises, the point stands: the end-customer demand curve for these chips is not as stable as the loan documents imply.
What a workable answer looks like in 2026
You don't need to boil the ocean. The teams shipping this in the next twelve months tend to do three things:
- Pick the ten to twenty facilities that carry AI-infra exposure and build the graph manually first. Borrower, project, end-customer, key covenant triggers.
- Automate the refresh from covenant reporting and public filings so the graph updates monthly, not annually at renewal.
- Wire the graph into stress scenarios the ALCO already runs, so the AI-infra cluster shows up as a named scenario alongside CRE and shared-national-credit reviews.
DORA has been in force in the EU since January 2025, and while it's an operational resilience regime rather than a credit one, the underlying assumption is the same: regulators expect you to know your concentrations across systems, not just within them. That expectation is going to arrive in credit supervision too. The banks that already have the graph built will answer the exam question in a meeting. The banks that don't will spend a quarter reconstructing it under a deadline.