Everyone is selling you a solution. No one is showing you the failure mode.
Last week, Nebius (NBIS) reported Q2 earnings: revenue of $582.3 million, up 454% year-over-year. The stock initially popped, then sold off. Analysts called it a “miss on expectations.” I call it a classic case of the market pricing the pitch, not the protocol.
Let me be clear: I’m not here to trash Nebius. I’ve spent years auditing the code that powers decentralized infrastructure—first Ethereum Classic’s immutability, then DeFi lending pools, then the GPU schedulers that underpin neoclouds. I’ve seen what happens when the narrative outruns the architecture. This is a story about the quiet audit that the earnings report didn’t provide.
Context: The Neocloud Promise
Nebius emerged from the ashes of Yandex’s Russian web empire, spun off its AI and cloud division into a standalone entity listed on Nasdaq. The narrative is seductive: a pure-play neocloud for AI training, tapping into the insatiable demand for GPU compute. They’re not just a “GPU landlord”—they claim to offer orchestration, developer tools, and a platform that differentiates them from the likes of CoreWeave or Lambda.
But the financial report tells a story heavier than the press release. Let’s walk through the numbers with the same rigor I’d apply to a smart contract audit.
Core Analysis: The EBITDA Mirage
Revenue: $582.3M in Q2, up from an estimated $399M in Q1 (derived from the half-year figure of $981.3M minus Q2). That’s 46% quarter-over-quarter growth. Impressive, right? Then why did the stock drop?
Because the market is finally reading the fine print.
Adjusted EBITDA came in at $236.2M, a 40.6% margin. That’s healthy for a cloud provider at scale. But the adjusted net loss was $33.2M, and the continuing operations net loss was $190.4M. The gap between EBITDA and net income is a jaw-dropping $157.2M.
What’s in that gap? Depreciation, amortization, and stock-based compensation (SBC). In the cloud world, GPUs are depreciated over 3-5 years, but AI chips like H100s are being obsoleted faster than that. Nebius is likely accelerating depreciation—a prudent move, but one that masks the true cash cost of maintaining a competitive hardware fleet. The SBC component is also a red flag. If the company is paying employees in stock that isn’t fully reflected in adjusted EBITDA, the real economic cost is higher.
Silence is the loudest audit. Nebius didn’t break out SBC separately in the headline. I’d bet it’s a significant figure—many neoclouds use equity to attract talent, and that dilution is a cost to shareholders, not a non-cash fairy tale.
The Capital Expenditure Cycle
Quarterly revenue of $582M implies a massive installed base of GPUs. To sustain this growth, Nebius has to keep spending on hardware. The report didn’t provide capex figures, but from my experience with DeFi protocols that rely on locked liquidity, I know that high growth often masks a dependency on continuous capital injection.
Let’s do a back-of-the-envelope: If Nebius is running, say, 50,000 H100s at $30/hour, that’s roughly $1.1B per month in revenue potential. But utilization rates matter. A 40% utilization would yield $440M per month, which aligns with their quarterly run rate. But the capital cost to acquire those GPUs is around $3B upfront. That’s a capital intensity that requires constant fundraising or debt.
Nebius’s growth is a reflection of past capital decisions, not future efficiency. The Q2 2025 revenue explosion is actually the result of GPU orders placed in 2024. The question is: what happens when the next generation of chips (Blackwell, etc.) requires a new round of massive capex, and the competition is doing the same?
Code doesn’t lie. Balance sheets don’t lie. But narratives do.
Contrarian View: The Hidden Strength
I’m not a permabear. There’s a genuine thesis here: Nebius has a team that built Yandex’s infrastructure, which was one of the most sophisticated in Europe. They understand distributed systems at scale. Their developer experience might be genuinely better than the American hyperscalers. And the 454% growth is real—it’s not made up.
But the market is pricing Nebius as a growth stock with a 40% EBITDA margin, ignoring that the net loss is burning cash and the capex cycle is a treadmill. The contrarian view is that if they can reach a scale where depreciation stabilizes and SBC becomes a smaller percentage of revenue, the EBITDA margin could translate into real free cash flow. That’s the bull case.
Trust the protocol, not the pitch. The protocol here is the business model: renting GPUs is a commodity business unless you build a software layer that makes customers sticky. Nebius’s platform play is the key to escaping the “GPU landlord” trap. But the earnings report didn’t provide metrics on platform revenue vs. pure compute. Without that, I’m skeptical.
Takeaway: The Quiet Audit
When I audited that DeFi protocol in 2020 and found the reentrancy bug, no one wanted to hear it. The yields were too high. The same is true now: the growth is too sexy. But the failure mode is predictable: a capital crunch, a slowdown in chip procurement, or a write-down of excess inventory.
Nebius’s stock price volatility is the market’s way of saying, “I’m not sure I trust the code.” The code is the business model. The pitch is the 454% growth. The protocol is the capital intensity, the depreciation, and the dilution.
As builders, we need to look past the noise and ask: what happens when the capex cycle turns? Can the platform layer generate enough margin to offset the hardware treadmill? I don’t have the answer, but I know the question is more important than the stock price.
Silence is the loudest audit. Nebius is a fascinating experiment in the neocloud thesis. But I’ll wait for the next report—one that breaks out SBC, capex, and platform revenue—before I trust the pitch.
