Nebius Borrows $775M for GPUs: The $40B Promise That Smells Like Debt Fumes

Over the past 7 days, one number has haunted my feed: $40 billion. That’s the customer backing Nebius Group claims for their latest $775 million debt raise. But here’s the thing—debt isn’t a victory lap; it’s a leash. And in a sideways market where everyone’s waiting for direction, a leash can either guide you or strangle you. Nebius isn’t a household name like AWS. It’s the AI cloud spin-off from Yandex, the Russian tech giant. Headquarters now in the Netherlands, but the roots are tangled in geopolitics. They’re building GPU capacity—think H100 clusters—to rent to AI startups. And they just borrowed $775 million to buy more GPUs. No equity dilution. Bondholders take the risk. But at what cost? Let’s break down the numbers. $775 million at, say, 10% interest means $77.5 million annual interest. That’s a heavy nut. The $40 billion customer backing is likely a mix of contracts, letters of intent, and market projections. In my hackathon days, I saw founders promise the moon to raise capital. But $40 billion is insane. Compare: CoreWeave, the GPU cloud leader, raised $2.3 billion and is valued at $19 billion. Nebius claims double that in customer interest? I’m skeptical. First, the structure. Senior secured debt means they put up assets—existing GPUs, data center equipment, maybe even their stake in Yandex derivatives. That’s fine if you have cash flow, but Nebius’s revenue isn’t public. Their predecessor, Yandex N.V., did about $4.6 billion total revenue in 2023, but the AI cloud segment? Unknown. Interest coverage ratio? Also unknown. If revenue falters, lenders seize the GPUs. That’s how debt circles back. The $40 billion figure is the real stink. It’s suspiciously round and massive. Even the largest AI cloud, AWS, doesn’t advertise a single customer backing number that big. CoreWeave’s total contracted backlog is a fraction. So where does $40 billion come from? Probably aggregate total addressable market across 5–10 years, or hyper-optimistic pipeline projections. I’ve seen this in crypto: a protocol claims $100B in TVL from “partnerships” that never materialize. Same energy. The merge wasn’t the first time we saw inflated metrics—it happened with every DeFi yield claim. Nebius is giving me that déjà vu. Now, the technical side: what GPUs will they buy? At $30k per H100, $775M minus fees and operating escape gives maybe $600M for hardware. That’s 20,000 H100s. But B200s cost $50k+. They also need networking (InfiniBand vs RoCE), cooling (liquid vs air), power (200W per GPU). Realistically, maybe 10,000 H100-equivalent units. Is that enough to serve $40B in customer demand? Only if each GPU generates $4M over its lifespan (3–5 years). Possible if rented at $4–5/GPU-hour continuously, but that requires 80%+ utilization. In a supply glut, utilization drops. The merge wasn’t a one-time event—it was a catalyst for decentralized compute. Now, Nebius is betting on centralized efficiency. Hackers don’t hack, they listen—and they’re listening to the $40B promise. But they also know history. The GPU cloud market is overheating. Every Tom, Dick, and Harry is buying GPUs. CoreWeave plans 100,000+ next year. Lambda Labs is expanding. Even Google and AWS are doubling down. Supply surplus could hit by 2026, driving rental prices down. If Nebius locked in long-term contracts at current prices, they’re fine. If not, debt service eats margins. Let me tell you what my network is whispering. At a recent meetup in Mexico City, a former Yandex engineer told me Nebius’s Russian teams are still handling hardware procurement. That’s a red flag for US export controls. Advanced NVIDIA chips can’t go to Russia-aligned entities. Even with a Dutch HQ, if procurement passes through Russia-linked hands, OFAC could block shipments. I saw a similar situation with a crypto mining firm that lost its GPU supply overnight. Nebius might have backup plans—AMD MI300X or even self-designed ASICs—but that would raise costs and delay deployment. Onto the commercial model: debt avoids dilution, sure, but it forces discipline. Bondholders want performance covenants. If Nebius misses revenue milestones, they could trigger default. In the crypto world, we call that “liquidation.” The merge wasn’t the only one in crypto’s history—AI cloud might see its own merge of hype and reality. Now, the contrarian angle I don’t see anyone else covering: the $40B customer backing is probably a masterful PR lock. They said “in customer backing,” not “signed contracts.” That could mean: “We have enthusiastic discussions with customers about potential capacity needs.” That’s like saying I have $10 billion in backing because I asked my friends what they’d pay for a pizza. It’s vapor. The real test: will they disclose the top three customers? If they name an OpenAI or Meta, I’ll eat my words. If not, it’s a red flag. Another blind spot: debt markets are tightening. Interest rates are still elevated (5%+ risk-free). Senior secured debt for a non-rated company like Nebius could carry 12–15% coupon. That’s $90–115M interest per year. Their free cash flow from AI cloud operations? Estimate maybe $50–100M if they’re running hot. That’s a razor-thin margin. Any downturn could force asset sales. But let’s be fair: if the $40B backing is even 10% real, that’s $4B in committed revenue over time, which would justify the debt. The key is the conversion rate. I reached out to a venture debt analyst friend—she says these numbers are common in PR but rarely hold up under due diligence. “We see 10x oversubscribed rounds with fairy-tale backing all the time,” she texted. “Usually 70% evaporates within 18 months.” Hard data saves us from fairy tales. What about the competitive landscape? Nebius is a mid-tier player. It has no custom silicon (unlike Google TPU), no full-stack edge (unlike AWS SageMaker), no massive ecosystem (unlike Microsoft Azure). Its selling point? Possibly European data sovereignty—but only if they can prove no Russia ties. And price? Maybe they undercut AWS by 20% on compute. But that’s a race to the bottom. In crypto, we saw this with L2s: overhyped DA layers that nobody needed. Now, GPU clouds are the new L2s—everyone wants one, but most will consolidate. I’d be remiss not to mention the human cost of this build-out. The article ignored the environmental impact. GPUs guzzle power. A 20,000-H100 cluster draws 20 MW, enough for a small city. If Nebius uses dirty energy (coal-heavy regions like parts of Eastern Europe), the carbon footprint is massive. They didn’t disclose sustainability plans. In the crypto world, proof-of-stake was celebrated for saving energy. GPU clouds are the new proof-of-work. The merge wasn’t just a consensus change—it was an environmental pivot. Nebius missed that memo. Now, the takeaway. The market is sideways, chop is king. Every position needs technical signals. For Nebius, watch three things: (1) The debt terms: filing details on interest rate, maturity, covenants. (2) GPU procurement announcements: which models, from whom, delivery timeline. (3) Customer names: even a single Fortune 500 or top AI lab would validate the $40B claim. If none materialize within 6 months, this is a debt flash in the pan. Hackers don’t hack, they listen. And right now, I’m listening to the silence from Nebius about the details. The $40 billion number echoes louder than a bull run, but silence on execution speaks volumes. In this market, the best signal is often the absence of noise. So far, Nebius is sweating words, not facts. Final thought: the merge wasn’t the only moral lesson in crypto history. In 2022, we saw Celsius and BlockFi blow up on leverage and inflated balance sheets. Nebius is offering debt to expand—same energy, different asset. The ecosystem needs more physical infrastructure, sure, but not at the cost of common sense. I’ll believe the $40 billion when I see the signed contracts. Until then, my bullshit detector is beeping at 11. Stay cheetah-fast, but human-slow on conviction. The next 90 days will tell the real story.