CoreWeave's IPO was a 12-bagger in its first month. The chart didn't care about the debt. The market priced in a future where NVIDIA's silicon prints money through a thin layer of cloud middleware. Nebius, the post-Yandex resurrection, followed suit with a NASDAQ listing and a promise to build Europe's AI backbone. Everyone cheered the growth narrative. I stared at the balance sheet. The chart didn't tell you about the 7-year leases, the GPU delivery timelines, or the single customer that could gut the revenue stream overnight.
Context: The GPU-as-a-Service Ponzi
CoreWeave and Nebius are not AI companies. They are arbitrageurs of NVIDIA's supply chain. They buy GPUs at wholesale, bolt on Kubernetes, RDMA networks, and liquid cooling, then resell compute at a markup. The market values them as high-growth tech, but the underlying economics are closer to a data center REIT with a 10x burn rate. Their technical moat is not a novel algorithm or a proprietary chip. It's a contractual priority allocation from NVIDIA, a signed power purchase agreement in a county with spare transformer capacity, and a sales team that promises faster deployment than AWS. That's a moat that can be crossed by a determined competitor with a checkbook.
Core: The Order Flow Analysis
Let me walk through the capital stack. CoreWeave raised $1.2B in debt in 2024, secured by GPU hardware. The interest rate on that debt is around 8-10% in the current environment. Their gross margin on GPU compute is roughly 30-40% before depreciation, but they need to run at 80%+ utilization to cover the debt service. I backtested this against a simple model: if GPU utilization drops to 60% for two consecutive quarters, the cash flow turns negative. In a bull market for AI, utilization is high. But the macro risk is that AI training demand peaks, or that hyperscalers (AWS, Azure, GCP) drop their GPU prices to crush the upstarts. The chart didn't show the price elasticity of GPU compute. Every candle tells a story of fear, but the fear here is not about a crash—it's about a slow bleed from margin compression.
Nebius has a different wrinkle. They are betting on European AI sovereignty. The EU is pouring subsidies into domestic AI infrastructure, and Nebius is positioned to capture that government-adjacent demand. But government contracts come with red tape, data localization requirements, and slower payment cycles. Their revenue recognition is lumpy. I've seen this pattern before in the 2020 yield farming experiments: a protocol that looked like a cash cow was actually a liquidity trap. Nebius's revenue may grow, but the free cash flow will lag. Risk isn't a feeling, it's a ratio of debt to EBITDA.
Contrarian: The Retail vs. Smart Money Angle
Retail investors see CoreWeave and Nebius as pure plays on the AI boom. The narrative is simple: AI needs compute, these companies provide compute, buy the stock. Smart money sees the fragility. The real alpha is in the counter-trade: short the GPU cloud providers and long the decentralized compute networks. Let me explain.
Last year, I ran an experiment with an AI trading agent that I deployed on both AWS and the Akash Network. Akash is a decentralized marketplace for compute, where providers bid in a Dutch auction. The latency was higher—about 200ms vs 20ms on AWS for a single inference call—but the cost was 40% lower. For batch inference or training jobs that tolerate latency, decentralized compute is a viable alternative. More importantly, Akash has no debt, no single point of failure, and no NVIDIA dependency (they support AMD MI-series as well). The chart didn't show that the unit economics of decentralized compute are improving, while centralized GPU cloud margins are under pressure from rising hardware costs and competition.
There's a deeper structural flaw in the CoreWeave/Nebius model. They are essentially leveraged long NVIDIA. If NVIDIA's next-gen GPU (Blackwell, Rubin) shifts the performance curve, last-gen H100s become obsolete faster. The depreciation schedule is brutal. CoreWeave's fleet is mostly H100s. They are already seeing discounting on H100 spot instances. I bought the pixel, not the promise. The pixel here is the utilization rate, and it's flashing amber.
Takeaway: Actionable Price Levels
I don't hold a position in CoreWeave or Nebius, but I've set price alerts. If CoreWeave drops below its IPO price ($35), I'll consider a short. The risk-reward flips: the debt burden becomes a weapon of mass destruction. For Nebius, the key level is the valuation relative to its book value. If it trades above 5x book, it's a sell. The decentralized compute tokens (AKT, IO) are a different play. They are not correlated with the GPU supply chain. They are a hedge against the centralization of compute. Code is law, until it isn't. But in this case, the economic law says that leverage amplifies both gains and losses. The market is pricing in only the gains.
Every candle tells a story of fear. The next candle for GPU cloud providers might tell a story of margin calls.