Compute is now credit.
That sentence should bother anyone who builds agent systems for a living. I spend my time thinking about scheduling, memory hierarchies, and how a planning loop degrades when inference latency doubles. Those are architecture problems. But the reporting through August and September of 2026 has made something clear that I can no longer treat as somebody else’s department: the supply of compute my systems run on is being shaped less by silicon roadmaps and more by lending decisions.
Nvidia has picked up a nickname this year, repeated across the financial press. Central bank of AI. Federal Reserve of AI. The framing shows up because the company is doing more than selling chips. It is providing financial support and loans to AI infrastructure projects, effectively backstopping the buildout of the very demand that fills its order book. Critics are uneasy about what this does to the debt market. Analysts, meanwhile, project revenue reaching $1 trillion by 2029.
Why a central bank analogy actually fits
Analogies in tech writing are usually decoration. This one has structural teeth.
A central bank does a few specific things. It sets the price of the scarce resource everyone needs. It acts as lender of last resort when borrowers cannot fund themselves. And through both functions it decides, implicitly, which parts of the economy get to grow.
Substitute compute for currency and the mapping holds. Nvidia sets the effective price of the scarcest input in AI. When AI model companies cannot raise enough capital on their own, financial backing from the chip supplier fills the gap. And the allocation of that backing determines which labs, which clouds, and which architectures get to exist at scale.
The reporting points at the weak spot underneath: the poor financial shape of AI model developers and the real possibility that they cannot pay. That is a credit risk story wearing a technology costume.
What this means for people who design agents
Here is where my discipline stops being a bystander. Agent architecture is, at its core, a series of bets about what compute will cost and how reliably you can get it.
Consider the design choices that have become normal in the past two years:
- Multi-step planning loops that call a large model dozens of times per task
- Reflection and self-critique passes that multiply token spend for marginal accuracy gains
- Tool-use graphs where a single user request fans out into parallel model invocations
- Long-context memory strategies that pay for retrieval by stuffing the window
- Ensemble voting across model variants to stabilize outputs
Every one of these trades compute for reliability. They are reasonable trades when compute is cheap, abundant, and priced predictably. They become architectural debt the moment any of those three conditions changes. And when the entity setting the price is also the entity extending credit to your suppliers, the conditions are correlated in a way that no amount of clean abstraction layers will insulate you from.
I keep thinking about circularity. If chip revenue is partly funded by loans that enable chip purchases, then the demand signal my capacity planning depends on is partly self-generated. A planner that assumes exogenous supply growth is modeling the wrong system.
Designing for a credit cycle, not just a compute curve
The practical response is not panic. It is treating compute availability as a variable with a failure mode rather than a constant with a growth rate.
Some concrete implications for agent design:
- Make the model tier a runtime decision. Any system that hard-codes a specific frontier model into its control flow has taken on a price exposure it cannot hedge.
- Instrument compute per completed task, not per call. If you cannot state the cost of one successful agent run, you cannot reason about what happens when that cost doubles.
- Build graceful degradation paths. An agent that either runs at full reflection depth or fails entirely is brittle in the least interesting way.
- Prefer architectures where accuracy gains come from structure rather than volume. Better task decomposition, tighter tool interfaces, and cleaner state management are cheap in a way that additional sampling passes are not.
None of this is a prediction that the credit story ends badly. The $1 trillion revenue projection for 2029 suggests plenty of people expect the opposite. My point is narrower and, I think, harder to argue with: the assumption of cheap and endlessly expanding compute has quietly become a load-bearing element in how we build agent systems, and that assumption now rests on a debt market rather than on physics.
Thirty years to a $1 trillion valuation. Nine more months to double it. That kind of curve does not come from manufacturing improvements alone. It comes from financialization, and financialization has a rhythm that semiconductor engineering does not.
Build accordingly.
đź•’ Published: