Cheaper, and losing: why price does not decide the cloud market¶
A plain-language companion to the working paper "Cheaper and Losing: A Layered Model of Cloud Dependence and the Limits of Subsidy" (v0.3, July 2026). The paper carries the model, seven proved propositions, and the Monte-Carlo stress tests; this text carries the ideas. There is also a live version you can play.
The short version. European cloud providers are several times cheaper than the US hyperscalers for raw compute and win independent price-performance benchmarks. Their share of their own market has roughly halved since 2017; three US firms hold about 70 percent. If price mattered, this could not happen. The model explains why it happens anyway, why the obvious remedies barely move it, and which two levers actually work.
Altitude¶
A buyer's needs are a mix of two layers. The commodity basement: virtual machines, storage, bandwidth, where Europe is cheap and present. And the differentiated upper floors: managed databases, serverless platforms, AI services, where network effects rule and Europe is thin. Call the differentiated share of a firm's needs its altitude. A logistics company lives low; an AI startup lives high.
Each firm picks the provider that serves its mix, so the market sorts on a threshold: below it, firms go European and pocket the price edge; above it, they go American and pay for the ecosystem. And the ecosystem's value grows with the number of firms already on it, which closes a loop: every firm that goes American makes going American slightly more attractive for the next one. Markets with that loop can tip: past a critical strength of the network effect, two resting points exist, and history, never price, picks between them.
Why the market keeps sliding¶
The measured firm population is two-humped: infrastructure-heavy firms in one hump, platform-native firms in the other, few in between. And firms age upward: as a company matures it adopts more managed services, more analytics, more AI, so its altitude drifts up, and almost nobody drifts back down. Put the two together and the aggregate US share climbs on its own, passing through today's level on its way toward a corner. The AI wave is the latest and steepest push, and the paper shows the ascent has a ratchet: once a firm has re-architected upward, easing conditions do not bring it back.
Why the obvious levers fail¶
Subsidising the commodity layer saturates. Making European compute even cheaper moves the sorting threshold into the trough between the two humps, where almost no firms live. In the paper's sweep, the responsiveness of the US share to the subsidy collapses by an order of magnitude: money keeps flowing, the dot barely moves. Only a subsidy scaling with the entire commodity base could restructure the market, which no budget contemplates.
Out-featuring the incumbents fights a tipped market. At the calibrated network strength, the differentiated layer sits in its high basin; a symmetric feature-for-feature European challenger is playing against the basin, and capability funding alone does not clear the network constraint.
The two levers that work¶
Un-bundling. Interoperability's real payoff is to let a firm split its stack: keep the American differentiated core, return the commodity and data spend to Europe. In the calibration the US needs-share falls from 0.75 to 0.65 without Europe winning a single differentiated workload. The lever is real, and its target is the bundle.
The egress lever. Model the cost of coupling a workload to a second provider, and the realistic European position appears: a loosely coupled secondary, holding backups, archives, and the chattier workloads as coupling costs fall. That coupling cost is exactly what the EU Data Act's switching and egress rules push down. The beachhead this grows is worth more than its size: it is sovereign reversibility, the preserved option to leave, and the supplier base that any demand-side policy presupposes.
The honest caveats: this is a stylized model with a calibrated firm distribution; the up-stack drift is asserted from adoption patterns, and the results are reported as signs and robust rankings across hundreds of random parameter draws, never as forecasts. The interactive version recomputes the verified model on every slider you move, so you can try to break the conclusions yourself.