We didn’t need another AI infrastructure story. But the market bought one anyway. CoreWeave and Nebius went public, and the narrative is simple: GPU demand is infinite, and these companies are the toll collectors. History doesn’t repeat, but it rhymes. And this rhyme is from the 2021 DeFi summer—when every token with a yield farm was a “protocol” and every liquidity pool was a “moat.”
Context: The Narrative Machine
CoreWeave and Nebius are not model developers. They are not chip designers. They are GPU-as-a-service resellers, packaging NVIDIA H100/H200 clusters into cloud compute. CoreWeave, fresh off its 2025 IPO, markets itself as the fastest way to get GPU time outside of AWS. Nebius, the reincarnation of Yandex’s international AI assets, positions itself as the “AI-native cloud” for Europe. Both are funded by massive debt, and both are bleeding cash. The market, however, celebrates their revenue growth as a sign of AI infrastructure’s inevitability.
I’ve been here before. In 2022, I watched LUNA’s algorithmic stablecoin narrative collapse when the real yield didn’t materialize. The GPU cloud story has the same structural flaw: it assumes exponential demand without questioning the sustainability of the underlying economics. The ETF inflow wasn’t a signal of fundamental strength; it was a liquidity game. The same applies here.
Core: The Real Barriers Are Not Technical
The alpha isn’t hidden in the technology. It’s hidden in the collective belief system that GPU clouds have a moat. The truth is, the technical barriers to entry are low. High-speed RDMA networking, Kubernetes orchestration, and liquid cooling are engineering problems, not science. Any well-funded startup can buy NVIDIA GPUs, hire a few engineers, and build a cluster. The real moat—if you can call it that—is three things: GPU supply priority, power contracts, and operational trust.
NVIDIA controls the supply. CoreWeave and Nebius get preferential allocation because they are essentially NVIDIA’s distribution arm. But that relationship is fragile. When demand softens, NVIDIA will prioritize its own cloud partners and direct enterprise sales. Power contracts are another bottleneck. Data centers in Virginia and Texas now face 3–5 year lead times for grid connections. That’s a real constraint, but it’s not a competitive advantage—it’s a shared industry pain point.
Operational trust is the only intangible. CoreWeave claims it can spin up a GPU cluster faster than the hyperscalers. That’s true in a supply-constrained market. But once supply normalizes, speed becomes a commodity. The margins will compress.
Now, the hidden variable: customer concentration. Based on public filings and industry whispers, CoreWeave’s top three customers likely account for over 50% of revenue. One is probably OpenAI. Another might be a large AI lab. If even one of them decides to build its own GPU cluster—and the hyperscalers are already offering custom silicon—the revenue shock will be severe. The market is pricing this as a zero-risk event. It’s not.
And the business model is rate-sensitive. These companies are leveraged to the hilt. CoreWeave’s debt load is reportedly in the billions. Every 50-basis-point rate hike adds millions in interest expense. The narrative of “AI infrastructure is a utility” masks the fact that it’s a high-beta financial asset.
Contrarian: The Demand Myth
The counter-intuitive view is that the AI compute demand narrative is overestimated. The market assumes that training large models will only get bigger, and inference will explode. But the data tells a different story. Inference costs are dropping faster than training costs. Edge AI and model compression are reducing the need for massive GPU clusters. And the return on investment for AI startups is still negative for most. The VC money that fueled this GPU spend is drying up.
If demand growth slows to 20% per year instead of 50%, the entire GPU cloud thesis collapses. The capex is front-loaded, and the assets depreciate fast. NVIDIA’s H100 loses 30% of its resale value every 12 months. CoreWeave and Nebius are sitting on billions of depreciating hardware. If utilization falls below 60%, the unit economics turn negative.
There’s also the regulatory risk. The EU’s AI Act and US export controls are already creating uncertainty. Nebius’s “European AI sovereignty” narrative might win it some government contracts, but it also means it cannot easily access the latest NVIDIA chips if export controls tighten. The same applies to CoreWeave if it expands into Asia.
Takeaway: The Real Trade
The market is buying a narrative of infinite demand. I’m selling the thesis that the next earnings miss will reveal utilization rates below 50%. The ETF inflow wasn’t a signal of fundamental strength; it was a liquidity game. The real alpha isn’t in buying the GPU cloud stocks. It’s in shorting the narrative when the next earnings miss reveals utilization rates.
We didn’t learn from LUNA. We didn’t learn from the 2022 crypto credit crunch. The same pattern repeats: a hot narrative, easy capital, and a belief that this time is different. It isn’t. The GPU cloud is a commodity business dressed in jewels. When the music stops, the only thing that will matter is who has the cheapest power and the longest lease terms. And that’s not CoreWeave or Nebius. That’s the hyperscalers.
I’ll be watching the November earnings. If CoreWeave doesn’t disclose GPU utilization, I’ll take that as a confirmation. The narrative is already priced in. The reality is still being built. And in the bear market of 2025, survival matters more than growth.