Big Tech Is Financing AI With $1.65T in Hidden Debt. Your App Doesn't Care.
The Financial Times reported last week that the five biggest US tech companies have piled up roughly $1.65 trillion in AI-related obligations, much of it kept off the balance sheet through opaque leasing and financing deals. Meta, Microsoft, Google, Amazon, Oracle. The kind of numbers that used to describe national economies.
I read the piece twice. Then I looked at our infrastructure bill for the month.
The gap between how AI is being funded and how most people actually use AI is getting wider every quarter.
What's actually happening
The hyperscalers are locking in GPU capacity, data center power, and long-term compute leases the way developers used to lock in office towers. Except a lot of it doesn't show up as debt on the standard financial statements. It's structured as commitments, off-balance-sheet vehicles, and "AI factories" financed by third parties. Regulators are starting to notice. Analysts are drawing comparisons to the pre-2008 shadow banking boom. You don't have to believe it's a bubble to notice the shape.
The trickle-down problem
Two things follow.
One, the cost of compute for the biggest players is climbing faster than their revenue from AI products. That gap gets closed somehow, and history says it gets closed by charging users more, at some point, for something.
Two, the small end of the market is diverging. Open-weight models, fine-tuned specialists, and inference optimization mean you can now do 80% of what a general purpose GPT does at roughly 1% of the cost. At DontCode we fine-tuned our own model specifically for building apps. Ours is a narrower model, not a smarter one. It handles the one task we throw at it 100,000 times a day faster and cheaper than a frontier model would.
The industry is bifurcating into "planet-scale bets" and "small, focused, cheap." Guess which end most founders live on.
If you're building right now
The boring take: don't marry your economics to the largest models unless you have to. Your competitors won't.
The more interesting take: build on infrastructure you don't own the debt for. That's basically what we do at DontCode. Someone else absorbs the compute risk, the deployment risk, the auth pool, the payments rails, the database. You get a working app.
When the hyperscalers eventually pass the bill along (rate hikes, tier changes, capacity throttling), you want to be sitting on top of an abstraction, not one bad quarter away from your app going dark.
The quiet advantage
There's a whole class of solo builders and small teams in Korea and elsewhere shipping real businesses right now who genuinely don't care what any of the big five just financed. Their apps run. Their KakaoPay charges go through. Their morning digests send. Their credits are metered, small, and predictable.
If that sounds better than reading FT footnotes about lease-adjusted EBITDA, you can try DontCode. More writing on this kind of thing over on the blog.