We didn’t see it coming. While the crypto world was busy debating rollup decentralization and MEV, a different kind of consolidation was quietly unfolding in the AI compute layer. CoreWeave and Nebius—two GPU cloud providers—are now the fastest-growing infrastructure on the planet, and they’re doing it with a playbook that looks eerily familiar: massive capital, centralized control, and a single chip supplier. If you’re a blockchain believer, this should keep you up at night.
Context: The GPU-as-a-Service Boom
Let’s strip the hype. CoreWeave and Nebius don’t build models or design chips. They are NVIDIA GPU resellers with engineering muscle. They deploy clusters of H100/H200 accelerators, connect them via RDMA networks, and wrap them in Kubernetes-based schedulers. Then they rent that compute to AI startups and enterprises. The model is simple: buy GPUs in bulk, build data centers, sign multi-year contracts, and hope utilization stays high.
Both companies went public in 2025—CoreWeave on NASDAQ, Nebius as a reborn Yandex spin-off. Their revenue growth is explosive. But the financials tell a story of “revenue at all costs.” Net losses, negative free cash flow, and debt-to-equity ratios that would make a traditional CFO faint. Yet the market rewards them because AI demand is insatiable.
Based on my 2017 ICO audit experience, I’ve seen this pattern before. When a sector’s supply chain is controlled by one dominant player, the middlemen grow fast but remain fragile. NVIDIA holds the keys to the GPU kingdom. CoreWeave and Nebius are just the fastest horses in a race where the track is owned by a single stable.
Core: The Hidden Signals Under the Growth
Let’s dig into the data that the mainstream articles miss. First, customer concentration. CoreWeave’s S-1 filings (publicly available) reveal that its top 3 customers account for over 70% of revenue. That’s a single point of failure. If OpenAI or Microsoft shifts its compute strategy, CoreWeave’s revenue evaporates. Nebius is better diversified geographically, but its European focus means it’s exposed to regulatory whiplash from the EU AI Act.
Second, the debt trap. GPU cloud expansion is 100% debt-financed. CoreWeave’s balance sheet shows over $8 billion in long-term debt, with interest coverage ratios below 1.5x. In a rising-rate environment, that’s a ticking bomb. The company’s entire business model is a bet that GPU utilization stays above 70% and that NVIDIA keeps supplying chips at favorable terms. Both are uncertain.
Third, the “hidden” cost of depreciation. A GPU cluster depreciates over 3-5 years, but the revenue from those GPUs is recognized over the contract term. This creates an accounting mismatch: the income statement shows massive depreciation expenses, while cash flow from operations is positive only if contracts are signed before chips are deployed. The moment new orders slow, the cash burn becomes visible.
From my 2020 DeFi workshops, I remember explaining that liquidity mining APY was just a subsidy for TVL. The same applies here: the “growth” in GPU cloud revenue is subsidized by cheap debt and NVIDIA’s willingness to allocate chips. When the subsidy stops, the real users vanish.
Fourth, utilization rates are the single most important metric, yet they are never disclosed. Any GPU cloud provider will tell you that idle capacity is a killer. My industry contacts suggest that average utilization for third-party GPU clouds is around 60-70%, compared to 85%+ for hyperscalers like AWS. That 15-point gap is the margin that gets eaten by interest payments.
Contrarian: The Centralization Threat Nobody Talks About
Here’s what the growth narrative glosses over: the AI compute layer is becoming more centralized than Bitcoin mining ever was. CoreWeave and Nebius rely on NVIDIA for 100% of their chips. NVIDIA controls the supply, the pricing, and the roadmap. If NVIDIA decides to prioritize its own DGX Cloud or tighten allocation to third parties, both companies are dead. This is not a healthy ecosystem; it’s a feudal system where the only lord is NVIDIA.
From a blockchain perspective, this is a nightmare. We champion decentralization, but the very infrastructure that powers the next generation of AI—and possibly the next generation of crypto—is controlled by a single company. The irony is painful. We built trustless networks to escape centralized control, yet we’re building the tools of the future on top of the most centralized compute stack since mainframes.
Moreover, the energy and ethical risks are real. Data centers for CoreWeave and Nebius consume gigawatts of power. In Virginia’s data center alley, transformer lead times are already 3-5 years. That’s a bottleneck that will slow AI development, but it won’t stop it. The real question is: who pays for the externalities? Local communities, the grid, and the planet. The ESG backlash is coming, and it will hit GPU cloud providers first.
Takeaway: A Call for Compute Sovereignty
We need to rethink the assumption that AI compute must be centralized. Just as we championed decentralized finance, we must champion decentralized compute. I’m not talking about “GPU mining” or “compute tokens” as a speculation vehicle. I’m talking about a real alternative: open-source scheduling, community-owned clusters, and protocols that let anyone contribute GPU idle cycles to a global pool. The technology exists—Kubernetes, Ray, and even some blockchain-based compute networks. But the capital is all flowing to centralization.
If we don’t act, the AI divide will be deeper than the digital divide. The few who control the GPUs will control the future. And that’s a future we didn’t choose.