The arms race for AI infrastructure is heating up: Who can break through the capacity ceiling?

The most competitive segment of the AI chip Supply Chain has emerged — Computing Power capacity.

Take Nebius for example, the orders in the third quarter are overwhelming, and every new capacity is instantly sold out. Management acknowledges that capacity has become the biggest bottleneck for growth. What is their solution? Hard push: increase the contract electricity from 1GW to 2.5GW by 2026, and connect 800MW-1GW by the end of the year. To achieve this, capital expenditure for 2025 is directly multiplied from $2 billion to $5 billion. What is the effect? The revenue guidance for 2025 is adjusted to $500 million to $550 million, and the ARR target for 2026 is sprinting to $700 million to $900 million. Sounds aggressive, but the question is—Can electricity and the Supply Chain keep up?

The same dilemma spreads throughout the entire track:

Microsoft hasn't been idle either. Azure AI capacity is expected to increase by 80% this year, with data center footprint doubling. That's not enough, as the entire fiscal year is expected to be constrained by capacity. Q2 Azure revenue grew by 37%, but due to limited available capacity, orders are piling up.

CoreWeave is growing even faster, but is also locked by the Supply Chain — delays in delivery from data center partners have directly cut the revenue expectations for 2025 (from 5.15-5.35 billion to 5.05-5.15 billion).

The underlying logic is very clear: demand far exceeds capacity, whoever solves electricity + Supply Chain first will win. Nebius's stock price has risen 144% in six months, but its valuation PE is only 4.66 times, far below the software industry's 39.95 times. Is this a discount or a trap? The key depends on whether the capacity can go online as scheduled in 2026.