CoreWeave's IPO prospectus showed $1.2B in revenue. Net loss? $500M. That's not a bug โ it's the feature of the AI GPU-as-a-service playbook. Nebius, the former Yandex AI spin-off, just announced a $1B infrastructure investment. Both are public. Both are bleeding cash. The market cheers. I see a pattern from 2022: Terra's code was poetry; Luna's exit was prose. Here, the revenue is poetry, but the balance sheet is prose. Let's read the fine print.
Context: What They Actually Do
CoreWeave and Nebius are not AI model builders. They are not chip designers. They are GPU middlemen. They buy NVIDIA H100s and H200s at scale, plug them into data centers, and rent them out. The technology is commodity โ anyone can stack GPUs. The real moat? Access to NVIDIA's allocation queue, long-term power contracts, and the ability to borrow billions at low rates. They are leveraged plays on NVIDIA's supply chain. The market treats them as AI infrastructure pioneers. I treat them as high-beta debt instruments with a GPU wrapper.
Both companies have secured multi-year contracts with AI labs. CoreWeave counts Microsoft and OpenAI among its clients. Nebius targets European AI startups riding the "AI sovereignty" wave. Revenue growth is real โ 200% year-over-year for CoreWeave. But growth is not profit. And in this sector, profit is a second-order effect after capital expenditure.
Core: The Mechanics of the Debt Trap
The core insight is simple: these companies are not selling technology. They are selling financing. A customer signs a three-year contract for GPU capacity. CoreWeave borrows money to buy the GPUs, installs them, and bills the customer monthly. The margin is the spread between the rental price and the cost of debt plus depreciation. That spread is razor-thin if rates rise or GPU utilization drops.
Based on my audit experience during the 2020 DeFi yield harvest, I saw the same structure in liquidity pools. High advertised yields masked the underlying risk of impermanent loss. Here, high revenue growth masks the risk of interest rate sensitivity and customer concentration. CoreWeave's top three customers represent over 60% of revenue. If one bolts โ or builds its own GPU cluster โ the whole house of cards wobbles.
Nebius faces a different risk: geographic concentration. Its European focus is a differentiator, but also a liability. European power grids are slower to expand. ESG scrutiny is higher. The EU AI Act may impose due diligence on infrastructure providers. Nebius's compliance costs could rise faster than its revenue.
Let's talk about the numbers. The market values CoreWeave at $19B after its IPO. That's 15x trailing revenue. For a company with negative net income and negative free cash flow. The valuation assumes that GPU utilization stays above 80% and that financing costs remain low. But the Fed is not cutting rates aggressively. Bond yields are sticky. The debt on CoreWeave's books is floating-rate. A 100 basis point hike adds $50M in annual interest expense โ that's nearly 10% of revenue. Options don't lie; the implied volatility on CoreWeave's stock is pricing in a 40% chance of a 50% drawdown within 12 months. The market is hedging its bets.
I ran a simple scenario: if GPU prices drop 20% (NVIDIA's next generation Blackwell), the residual value of CoreWeave's H100 fleet collapses. Depreciation charges spike. The rental price must drop to compete. Margins get squeezed. This is not a technology company โ it's a commodity reseller with a balance sheet.
Contrarian: The Blind Spots
Conventional wisdom says: AI demand is infinite, so GPU clouds are a sure bet. That's the gap between belief and reality. The reality is that hyperscalers โ AWS, Azure, GCP โ are the 800-pound gorillas. They have their own GPUs, their own chips (Trainium, TPU), and their own power contracts. CoreWeave's advantage is speed: it can get you H100 clusters faster than AWS. But that advantage is temporary. As soon as NVIDIA supply normalizes, the hyperscalers will match or beat the delivery time. And they will undercut on price because they can subsidize GPU cloud with other services.
Arbitrage doesn't last forever. The current arbitrage is between the narrative of AI scarcity and the reality of impending oversupply. Look at the data: global GPU shipments are up 300% year-over-year. Data center construction is at an all-time high. The lead time for a new data center is 18-24 months. The GPUs being ordered today will come online in 2026. By then, demand may have cooled. The market is building for a demand curve that assumes linear growth. But AI adoption may follow an S-curve โ rapid initial growth, then plateau. The infrastructure bets are all on the steep part of the curve. If the plateau comes earlier, these companies will be left with idle GPUs and massive debt payments.
Nebius's European focus adds a geopolitical twist. The US export controls on NVIDIA chips to China are not directly affecting Europe, but the next round of controls could target "high-performance computing" broadly. If Nebius's software stack is deemed too advanced, it could face restrictions. Or if the EU mandates data localization, it could limit Nebius's ability to sell to US clients. The regulatory environment is a wildcard that most analysts ignore.
Takeaway: Actionable Levels
For traders, the signal is clear: watch the GPU utilization metrics. If CoreWeave reports occupancy below 75% in its next earnings, sell the stock. The debt covenants will tighten. The margin of safety is thin. For investors, the risk/reward is asymmetric to the downside. The upside scenario requires perfect execution: low rates, high utilization, and no new competition. The downside scenario is a debt spiral.
I'm not calling for a crash. But I am saying: the market is pricing these companies as if they are the next AWS. They are not. They are leveraged commodity plays. And when the commodity cycle turns, the exit will be messy. Risk is the gap between belief and reality. Right now, the market believes in infinite AI demand. The reality is that infrastructure is a race to the bottom on price. The only winners are NVIDIA and the bondholders. The rest are just exit liquidity.
Remember: Terra's code was poetry; Luna's exit was prose. CoreWeave's revenue is poetry; its balance sheet is prose. Read the prose before you buy the poetry.