Could an AI product thrill customers while the company selling it quietly loses money on every sale?

This explores whether AI's value can flow mostly to buyers and users while the companies building and serving the models struggle to profit. The corpus has no direct data on AI company margins, so it can't say whether developers are losing money today. It does show how such a split could happen. The pieces are on the cost side, the pricing side and the adoption side.

Start with prices. Epoch AI found that the cost of reaching a given benchmark score drops about 66% per quarter right after a capability first becomes state of the art, then slows to about 32% per quarter within two years Do AI performance prices fall faster when capabilities debut?. Their explanation is that a new capability briefly carries monopoly pricing, and competitors quickly bid it away. That is great for customers. For developers it means the window to charge a premium for a new capability is short, and it closes fastest at the moment the capability is newest. Now set that against costs. Epoch's model of a one-gigawatt data center puts annualized cost at about $8.5B, and servers make up roughly 60% of it. Energy is only about $0.6B What actually drives AI data center costs?. Most of the bill is paying back hardware bought up front, and that bill doesn't shrink when prices fall. Put the two findings together and you get a squeeze: what you can charge collapses within quarters, while the cost of the hardware is paid off over years. The corpus doesn't directly show whether that squeeze produces losses, but it lays out exactly how it could.

The same split shows up one level down, among companies building AI apps for their own use. a16z's survey of CIOs found enterprises moving away from building their own AI apps and toward buying them. The reasons they gave were the constant work of keeping up with new models and the high cost of maintaining internal tools Are enterprises abandoning custom AI apps for third-party solutions?. These companies still value AI. They just found that building it didn't pay off. Benedict Evans adds that making tools easier to build doesn't solve the harder problems: workers often can't see which of their own tasks could be automated, and adoption depends on decisions across many departments Does easier tool-building actually solve enterprise adoption problems?.

The less obvious point is that "customers value it" is shakier than it sounds. Stack Overflow's 2025 survey found 80% of developers using AI tools while trust in their accuracy fell from 40% to 29% Why do developers keep using AI tools they don't trust?. People use these tools heavily without being convinced they're good. Another note describes "cognitive surrender": users accept AI output without checking it because checking is costly and fluent answers feel reliable When do users stop checking whether AI output is actually backed?. On the work side, AI agents that win benchmark contests still fail on long, real job workflows Why do agent benchmarks not predict real economic value?. Historical cases also suggest that capability without economic value, trust and standards tends to stall Why do capable AI agents still fail in real deployments?. So the split may have two layers. Developers may struggle to capture value, and some of the value customers think they're getting may never have been checked.

The takeaway: yes, the corpus makes this pattern plausible. Falling prices push value to customers faster than developers can pay down their hardware. But the question it raises is whether everyone is measuring value correctly in the first place.