When Nebius reported Q2 2025 revenue of $582.3 million—a 454% year-over-year surge—the stock promptly dropped 8%. The market wasn’t punishing success; it was pricing in the cost of that success. I’ve audited enough protocol treasuries to recognize the pattern: explosive top-line growth masks a structural fragility that only shows up in the fine print. For Nebius, that fine print is $190.4 million in net losses from continuing operations, even after claiming $236.2 million in adjusted EBITDA. The gap is the depreciation tax of the GPU economy.
Nebius is not a crypto company. It’s a neocloud—an AI infrastructure provider born from the Yandex split, selling GPU compute power to AI developers. Its business model is straightforward: buy NVIDIA GPUs, rack them in data centers, and lease them out. The revenue is real. The growth is staggering. Q1 2025 revenue was approximately $399 million (derived from $981.3 million H1 minus Q2’s $582.3 million), meaning a 45.9% quarter-over-quarter jump. That kind of acceleration in B2B cloud doesn’t happen without deep enterprise contracts and pre-committed capacity. The EBITDA margin of 40.6% ($236.2M / $582.3M) is impressive for a capital-intensive business. But the net loss tells a different story.
The core insight here is the depreciation–cash flow disconnect. Nebius’s adjusted EBITDA excludes depreciation, amortization, stock-based compensation, and restructuring costs. In the AI cloud world, depreciation is not a non-cash footnote—it’s the central cost of doing business. GPU servers depreciate over 3–5 years, but technological obsolescence compresses that to 18–24 months. When NVIDIA releases the next generation of chips, Nebius’s existing hardware loses value overnight. The $190.4 million net loss suggests that the company is expensing a massive amount of depreciation, which is a real economic cost even if it doesn’t affect cash flow immediately. Based on my experience analyzing DeFi liquidations, true sustainability is measured by the ability to cover all costs—including capital replacement—not just operating cash flow. Nebius’s adjusted EBITDA is a selective metric that hides the true capital intensity.
Now for the contrarian angle. The market’s negative reaction to a 454% revenue beat is usually dismissed as irrational. I see it as rational. The market is discounting the sustainability of Nebius’s growth. Its reliance on NVIDIA GPU supply and the cyclical nature of AI demand create a structural vulnerability. If AI model training demand slows—or if competitors like AWS, Azure, and Google Cloud lower prices to capture market share—Nebius’s pricing power evaporates. The company is essentially a GPU rental middleman with a thin software layer. Its “leading neocloud” status is a media label, not a confirmed market share. The real question is: when the GPU cycle turns, will Nebius’s EBITDA margin contract faster than its revenue? The 45.9% QoQ growth is a lagging indicator of capital deployment decisions made 12–18 months ago. Future growth depends on re-upping those investments at a time when GPU lead times are shrinking and competition is intensifying.
Speed without direction is just volatility. Nebius has speed, but its direction—commodity GPU rental—offers limited differentiation. The protocol remembers what the regulators forget: in markets, hype fades, but structural costs persist. Crisis is just code with a high gas fee, and Nebius is paying premium gas to stay ahead of depreciation. The takeaway for any crypto-native observer is this: the same dynamics that plague DeFi lending protocols—overcollateralization, liquidation risk, and fee compression—apply to AI cloud infrastructure. Nebius is a case study in how fast growth can mask a fragile unit economy. Investors should watch its capital expenditure efficiency, not just its revenue growth. When the GPU bear market comes, the only thing that will save Nebius is a software layer that commands premium pricing—something it has yet to prove.