Written by: Luke Sobieraj, Founder & COO, Digital Colliers
Qualcomm just signed a multiyear deal to supply custom silicon for AWS AI inference workloads. The stock jumped roughly 5% on the news. If you're underwriting loans to companies building on AWS infrastructure, you just inherited a new concentration risk that doesn't show up in your current credit model.
The Hidden Layer in Infrastructure Debt
When you finance an AI inference startup, you're betting on their ability to generate revenue from customers who want fast, cheap predictions. You model their customer concentration, their burn rate, their gross margins. What you probably don't model is whether AWS can actually deliver the compute capacity their business plan assumes.
That compute capacity now depends on Qualcomm shipping multiple generations of custom chips on schedule and at the promised performance levels. If Qualcomm hits a fabrication snag, or if geopolitical events disrupt their supply chain, your borrower's AWS capacity might shrink or get more expensive. Their unit economics break. Your loan goes sideways.
This isn't theoretical. More than 80% of AI projects fail already, roughly twice the failure rate of conventional IT projects. You're lending into a sector where the base success rate is around 20%. Now add chip-supply concentration risk on top.
The Concentration Data Model You Need
Credit committees used to track customer concentration and vendor concentration. Now you need a third layer: silicon concentration. Here's what the data model looks like.
Start with borrower revenue by hyperscaler. If 70% of your borrower's inference traffic runs on AWS, and AWS inference depends on Qualcomm silicon, you have a path from Qualcomm execution risk straight into your borrower's P&L.
Next, map the hyperscaler silicon roadmap by vendor. AWS uses Nvidia for training, but they're building out custom inference silicon with partners like Qualcomm. Azure uses AMD and Nvidia. Google Cloud uses its own TPUs plus Nvidia. Your borrower's hyperscaler choice determines their indirect chip-vendor exposure.
Then add vendor risk flags. Qualcomm isn't just a US company designing chips. They rely on TSMC for fabrication, which introduces Taiwan geopolitical risk. They rely on global logistics networks for delivery. They rely on complex EUV lithography equipment from ASML. Each dependency is a potential chokepoint.
What Good Looks Like in Practice
The operators who are ahead on this build a simple dashboard. One column for each borrower in the AI infrastructure portfolio. Rows for hyperscaler dependency percentages, silicon vendor behind each hyperscaler, and red-yellow-green flags for vendor execution and geopolitical risk.
When a borrower's AWS dependency hits 60% or more, the credit team adds a quarterly check-in on AWS silicon roadmap health. If Qualcomm or any other AWS silicon partner shows delivery delays or performance misses, the credit team models the revenue impact on the borrower and adjusts covenants or reserves accordingly.
It's not about predicting chip-supply disruptions with perfect accuracy. It's about knowing which of your loans would break if a disruption happened, and pricing that risk into your credit decision before you write the check.
Credit Committee Checklist
Add three items to your next infrastructure loan review. First, quantify borrower revenue by hyperscaler and map each hyperscaler's silicon vendors. Second, flag any concentration above 50% on a single hyperscaler as a new covenant trigger for quarterly silicon-roadmap monitoring. Third, build a scenario model for a six-month delay in the borrower's primary hyperscaler's chip supply and calculate the revenue impact.
88% of AI proof-of-concepts never reach widescale deployment. For every 33 AI POCs a company launches, only four graduate to production. You're already lending into a sector where most bets fail. The last thing you want is to lose money on a borrower who had a real shot but couldn't get the compute they needed because their hyperscaler's chip partner missed a deadline.
Chip-supply concentration is now a credit risk. Your data model should reflect that.