The market loves Nebius’s 454% revenue growth. It ignores the $190M net loss buried in the same report.
That loss is not a bug. It’s the structural fingerprint of a capital-intensive business masquerading as a high-margin tech play. I’ve seen this pattern before — in 2021’s mining farms, in 2023’s GPU rental startups, and now in the neocloud hype cycle. The bull market euphoria is blinding investors to the arithmetic of hardware depreciation.
Let’s cut through the narrative.
Context: The Neocloud Mirage
Nebius (NBIS) is the neocloud poster child. Spun out of Yandex, it rents GPU compute to AI developers. Q2 revenue hit $582.3M — up 454% year-over-year. Adjusted EBITDA came in at $236.2M, a 40.6% margin. On the surface, this is a machine printing cash. But the machine is running on borrowed time.
The company’s own filings tell a different story. Adjusted net loss was $33.2M. Continuing operations net loss was $190.4M. The gap between EBITDA and net income — roughly $157M — is the elephant in the server room. That gap is almost entirely depreciation and amortization of GPU servers. In plain English: Nebius is spending billions on hardware, and the wear-and-tear is eating its profits alive.
Core: The Capital Expenditure Time Bomb
Here’s the structural deconstruction. Nebius’s revenue growth is a lagging indicator of past capital expenditure. The $582M in Q2 revenue reflects GPU clusters ordered and deployed in 2024 — when NVIDIA H100s were hard to get and prices were sky-high. The revenue is real. But the cost of that hardware is only partially captured in the current P&L. The depreciation charge will persist for 3-5 years, while the revenue from that hardware may peak and decline as competitors flood the market with cheaper compute.
Based on my experience auditing tokenomics for AI infrastructure projects, I’ve learned one hard rule: GPU rental is a commodity business with a fixed cost structure and a variable revenue stream. The moment supply catches up, unit economics collapse. Nebius’s 40.6% EBITDA margin looks healthy, but it’s an adjusted number. The real cash flow margin — after mandatory capex — is likely negative. Let me explain.
H1 revenue was $981.3M. Q1 implied $399M, Q2 $582.3M — a 46% sequential jump. That’s explosive. But it also means the company is racing to deploy more GPUs just to maintain that growth rate. The capex-to-revenue ratio is probably above 100% right now. In a bull market, that’s fine. Investors fund the growth. But the moment the narrative shifts — when the market starts asking about free cash flow — the stock will reprice violently.
The market doesn’t care about your adjusted EBITDA. It cares about what’s left after you replace your depreciating assets. Nebius is a toll booth on the AI highway, but the toll booth is made of silicon that rusts every 18 months.
Let’s zoom into the unit economics. A single H100 GPU costs roughly $30,000. Rental yield is about $1-2 per hour, depending on utilization. At 70% utilization, that’s $12,264 annual revenue per GPU. Subtract power, cooling, and labor — about 30% of revenue — and you get $8,585. Depreciation over 4 years is $7,500 per year. That leaves a gross margin of ~$1,085 per GPU, or 8.8% net margin. And that’s before corporate overhead.
Nebius’s 40% EBITDA margin looks impressive only because they’re not counting the full replacement cost. The adjusted EBITDA excludes stock-based compensation, which for a tech company can be 15-20% of revenue. It also excludes one-time restructuring costs from the Yandex spin-off. The true economic profit is razor-thin.
Contrarian: The Blind Spot Everyone Misses
Here’s the contrarian angle. The market is pricing Nebius as a growth story, but it’s actually a capital allocation story. The real question is not “how fast can they grow revenue?” but “how efficiently can they deploy capital?” The 454% growth is a backward-looking metric. The forward-looking metric is the marginal return on incremental invested capital.
Based on the Q2 run rate, Nebius would need to invest roughly $2-3B in capex every year to sustain 50% sequential growth. Where is that money coming from? Debt? Dilution? The balance sheet is not disclosed in the article, but the net loss of $190M suggests they are burning cash. In a rising interest rate environment, that leverage is a ticking time bomb.
We didn’t see the capex cycle coming. In 2023, every neocloud was a rocket ship. In 2025, the rockets are running out of fuel. The bottleneck is not demand — it’s the ability to raise capital at attractive terms. Nebius’s Yandex heritage adds a geopolitical discount: European investors are wary of Russian-linked entities, even after the spin-off. The regulatory bifurcation is real. Tornado Cash sanctions set a precedent that code is crime; Yandex’s history with Russian government means Nebius faces extra scrutiny.
The market’s blind spot is the lag between capex and revenue. By the time you see the revenue decline, the capex has already been spent. Investors will look at trailing twelve months and think the company is still growing, while the order book is shrinking. That’s the classic value trap in hardware-heavy businesses.
Takeaway: The Next Narrative Shift
The next narrative in AI infrastructure will not be about revenue growth. It will be about capital efficiency. The market will start rewarding companies that can generate positive free cash flow — not just EBITDA — from their GPU fleets. Nebius will need to prove that its unit economics are sustainable. If they can’t, the stock will re-rate to a multiple that reflects commodity risk, not tech growth.
I’m not shorting Nebius. But I’m watching the capex-to-revenue ratio like a hawk. The moment that ratio ticks above 1.5x, the structural flaw becomes a structural crisis. Follow the liquidity, ignore the noise. The liquidity is draining out of capital-intensive plays into software-defined AI platforms. That’s where the real alpha is.
