Rapidly Rising Data Center Costs May Complicate CoreWeave's Push to 8 Gigawatts of AI Power by 2030

CoreWeave’s backlog has grown nearly 300% year over year to $99.4 billion, but the company is funding a capital-intensive AI infrastructure buildout with about $25 billion of debt. AI data center costs are estimated at $15 million to $25 million per megawatt, with a fully loaded 1-GW facility now cited near $100 billion, raising concerns that profitability could be pushed further out. The article is constructive on demand but cautious on the economics and balance-sheet burden.

Analysis

The market is underpricing the difference between backlog visibility and real equity value when the asset base is still being built. In neoclouds, the first derivative is not revenue growth; it is funded capacity delivered on time and at an acceptable return on incremental capital. That means the real equity bottleneck is not demand, but the spread between contracted pricing and a rapidly inflating all-in cost of power, land, interconnect, labor, and GPUs — a spread that can compress even if utilization stays high.

Second-order beneficiaries are the upstream bottlenecks, not just the obvious GPU supplier. Power equipment, grid interconnect, transformers, cooling, and electrical engineering capacity should remain tight for 12-24 months, which favors suppliers with pricing power and punishes any AI infrastructure name that depends on third-party execution. Meta is an indirect winner if this dynamic suppresses model-hosting costs at scale, while NVDA benefits in the near term from continued capex pull-forward, but eventually higher infrastructure friction can slow the cadence of incremental GPU orders if returns on deployed capital fall below hurdle rates.

The key risk is not a near-term demand collapse; it is financing reflexivity. If credit markets decide the industry is more levered than the backlog implies, funding windows can widen quickly, forcing slower build schedules and lower equity optionality. A 6-12 month catalyst path exists around build delays, margin commentary, and any evidence that inference pricing is not holding up — that would re-rate the entire neocloud complex before it shows up in reported backlog attrition.

Consensus is treating this like a secular winner-take-most infrastructure story, but the more likely intermediate outcome is oligopoly economics with brutal capital intensity. The market may be overpaying for gross bookings visibility and underpaying for depreciation risk, especially if older GPUs retain value longer than expected only in a narrow inference regime. The hidden question is whether customer concentration is actually a credit enhancement or a negotiating trap: when the largest customers are also the only ones with enough scale to pressure pricing, the backlog can be durable while shareholder returns remain mediocre.

Key Decisions for Investors

• Short CRWV on strength over a 1-3 month horizon; use any rally tied to backlog headlines to fade exposure, with a thesis that rising build costs and leverage drive multiple compression before revenue converts to durable FCF.
• Pair trade: long NVDA / short CRWV for 3-6 months; NVDA captures the semiconductor demand leg without taking direct balance-sheet and execution risk, while CRWV remains exposed to cost inflation and financing risk.
• Long META vs short a basket of AI infrastructure names for 6-12 months; META benefits if cloud economics remain tight, while hyperscaler internalization reduces pricing power for third-party AI cloud providers.
• Buy a downside put spread in CRWV dated 4-6 months out; structure for a thesis that any guidance miss or funding-market wobble triggers a fast de-rating, with limited premium at risk.
• Overweight infrastructure-enablers rather than operators: use pullbacks to add exposure to NVDA and select power/grid suppliers, as they capture the capex cycle without bearing the same depreciation and leverage burden.