AI Contracts Are Pricing In Stability That Isn't There

If you're signing a multi-year AI vendor contract this quarter, the thing to worry about isn't the price. It's that you're buying stability nobody can actually sell you right now.

OpenAI is reportedly weighing drastic cuts to token prices ahead of a sharpening fight with Anthropic. A price cut reads like a gift to buyers. It isn't, not if you've locked terms. A vendor slashing prices in anticipation of a war is telling you the floor isn't a floor.

Look the other direction and the picture gets worse. Oracle grew its contract pipeline past a reported $638 billion and the stock slid anyway, because capex landed higher than guidance. That pipeline signals appetite, not that delivering on it pencils out. The cost overrun is the part procurement should sit with.

So the model layer is deflating while the infrastructure underneath it gets more expensive. That split squeezes margins across the whole stack, and the squeeze doesn't stay with the vendor.

Cheaper tokens sound like your win. They're the vendor telling you the floor you contracted against isn't a floor.

Here's the trap. Lock a long cloud AI deal and you may have fixed your price on infrastructure the vendor will eventually have to reprice to survive. You think you've protected yourself. What you've actually done is bet that today's vendor economics hold for three years, in a market where the people setting those economics are visibly bracing for a fight that breaks them.

I don't know which way the net pressure breaks for any given contract. Deflating models could offset inflating infrastructure, or the infrastructure bill could reach you through a renegotiation buried in the fine print. That uncertainty is the point. You can't price a contract against a number that's in freefall in both directions.

This is a governance problem before it's a budgeting one. Ungoverned tool sprawl is the obvious failure. The quieter one is signing for stability you were never offered, structuring a deal as if price discovery is finished when it's barely started. Build for repricing. Short terms, renegotiation triggers tied to published rate changes, exits that don't cost you a year of spend.