Nebius Group

I added to the position after earnings strengthened my conviction that scarce GPU compute demand is translating into real financial growth. That is the specific thing the report confirmed: not that demand exists, which was already visible, but that it is arriving as revenue.

The thesis centers on very strong compute revenue growth, large contracted demand and backlog, scarce GPU capacity, strong utilization, and pricing power. Those are connected. Scarcity produces the pricing power, utilization turns installed capacity into revenue, and backlog makes the next period visible in advance.

The economics depend on keeping expensive hardware busy. A GPU fleet depreciates from the day it is installed, so utilization is the difference between an attractive return and a poor one. Pricing power holds only while capacity remains scarce relative to demand.

This is an extremely capital-intensive business. Growth requires continuous investment in hardware, power and data-center space, which has to be financed before the revenue it generates arrives. Cost of capital is therefore a first-order variable rather than a financial detail.

I keep it small because Nebius is extremely volatile and competes against hyperscalers with cheaper financing, larger balance sheets, greater infrastructure scale, and the ability to access capital at better rates. Every one of those is a durable advantage in a business where the constraint is capital and power rather than software.

There is no assumption here that independent GPU cloud providers automatically win the long-term compute market. It is entirely possible that hyperscalers ultimately dominate compute, and that independents end up serving overflow demand at thinner margins. That possibility is why I keep it small rather than why I avoid it. The return if scarcity persists is large enough to justify a controlled allocation against a real chance of being wrong.

Earnings are not the right lens yet. The comparable basis against other neoclouds is a revenue multiple, and the peers worth comparing against are the other independent AI-infrastructure businesses, with CoreWeave the closest match on business model and capital structure. The hyperscalers compete hard for the same workloads, but a diversified platform's multiple describes a different financial object and should not be averaged in.