The financial press covers hyperscaler capex in aggregate. $50 billion here, $80 billion there. What it rarely quantifies is the cost of that capital sitting idle because the infrastructure to deploy it does not exist.
THE ANATOMY OF A STRANDED INVESTMENT
A hyperscaler following the standard deployment model orders GPUs 12–18 months in advance of their intended data center commissioning date. The GPU arrives. The data center is not ready. The GPU sits in a warehouse or in a partially built facility with no power connection.
Capital cost at 8% annual rate = $2.4M per year idle
Depreciation over 5-year life = $6M per year
Revenue foregone at $2/GPU/hour = $17.5M per year
Total annual cost of idle cluster: ~$26M
Multiply this across the 7 GW of stalled US data center capacity, and the numbers become very large very quickly. This is not hypothetical — it is what is happening now to projects that announced 2026 commissioning dates without confirmed transformer orders.
THE DEPRECIATION TRAP
Nvidia's GPU upgrade cadence runs on roughly an annual cycle. The H100 was followed by the H200, then the B100 and B200. Each generation delivers substantially better performance per watt than the last.
A hyperscaler holding idle H100s while waiting for transformer delivery is not just losing rental revenue. It is watching its hardware depreciate against a next-generation product that its competitors may already be deploying. Every quarter of delay is a quarter of compounded waste.
WHO IS MOST EXPOSED
Exposure to this problem is not uniform across the hyperscalers:
Medium risk: Companies relying on colocation providers to handle power
High risk: Companies with announced projects lacking confirmed power strategy
Very high risk: The 25% of projects with no disclosed power sourcing
THE COLOCATION BUFFER
Some hyperscalers have partially insulated themselves from this problem by leasing capacity from colocation providers who placed their own infrastructure orders years earlier. But colocation capacity in the right locations with sufficient power density for AI workloads is itself now scarce and expensive.
The premium for pre-powered colocation capacity in US AI markets has increased significantly. Data center REITs with existing power contracts are trading at premiums that reflect this scarcity. The transformer shortage is being capitalised into real estate valuations.
THE EUROPEAN ARBITRAGE
For hyperscalers with the organisational flexibility to deploy internationally, European data center markets — particularly France — offer a genuine arbitrage. Faster grid connection timelines, available industrial sites with legacy HV infrastructure, and lower electricity costs can translate directly into earlier compute deployment and better capital efficiency.
The challenge is organisational rather than technical: most hyperscaler procurement processes are optimised for US domestic deployment. Adapting them for European sourcing requires effort that the magnitude of the problem is increasingly justifying.
THE INVESTOR PERSPECTIVE
For infrastructure funds and data center investors, the transformer shortage creates a clear valuation signal: assets with confirmed power capacity are worth significantly more than assets with announced power capacity. The gap between these two categories — which was negligible in 2021 — is now a fundamental credit and valuation distinction.
Projects in the 7 GW stalled pipeline are not just delayed. Some will never be built at their announced locations. The developers that survive this cycle will be those that either placed equipment orders early enough or are flexible enough to relocate to where the power exists.