You think the AI infrastructure boom is a story of innovation. The truth is it’s a story of GPU arbitrage, debt-fueled expansion, and a ticking clock on profitability. CoreWeave and Nebius are not building the next NVIDIA H100; they are renting it out, hoping the math works before the market corrects.
Let’s start with the numbers. CoreWeave went public in 2025, riding a wave of hype that valued its GPU-as-a-service model at billions. Nebius, the reincarnation of Yandex’s international AI assets, followed suit, listing on NASDAQ with a promise to build Europe’s answer to AWS. But here’s the cold, hard fact: neither company is a technology company. They are logistics companies, optimizing for GPU delivery speed and power contracts, not chip design or model training. The market is treating them as the next great AI bet, but the underlying code is brittle.
I’ve been here before. In 2017, I spent weeks tracing Ethereum’s Geth code, finding memory leaks in the transaction pool that no one wanted to fix. The same pattern repeats: hype hides structural flaws. CoreWeave’s headline growth is real, but dig into the mechanics, and you’ll find a system built on borrowed time—literally. Their business model is a bet that NVIDIA’s GPU supply will remain tight, that interest rates will stay low, and that their clients won’t build their own clusters. That’s a fragile foundation.
Logic doesn’t lie. The core of CoreWeave and Nebius is a simple equation: revenue = GPU count utilization price. The unknowns are massive. Utilization rates are a black box. Are they running at 80% or 40%? If it’s the latter, the math collapses. The article I analyzed didn’t provide a single utilization figure. That’s not a coincidence; it’s a red flag.
Context
Let’s set the stage. The AI infrastructure market is a feeding frenzy. OpenAI, Anthropic, Google, and Meta are all racing to secure compute. NVIDIA’s H100 is the gold standard, and supply is tight. Enter CoreWeave and Nebius: they buy GPUs in bulk, lease them to AI startups and enterprises, and charge a premium for speed. Their pitch is simple: “Why wait 12 months for AWS? We can deliver in 4.”
This works in a bull market. But the structure is inherently unstable. The article breaks down four key areas: technical route, commercialization, competitive landscape, and investment logic. I’ll dissect each, adding my own forensic analysis.
Core: The Technical Teardown
First, the technical route. Both companies rely on NVIDIA GPUs, RDMA networks, and Kubernetes orchestration. That’s not innovation; it’s integration. The article rates this as “engineering-level innovation,” which is generous. The real innovation is in procurement and power contracts, not code. The technical barrier to entry is low—anyone with capital can replicate this. The moat, if one exists, is in GPU supply priority and data center power deals.
But here’s the hidden truth: the technical debt is massive. CoreWeave’s infrastructure is optimized for NVIDIA’s current generation, but what happens when AMD’s MI300 or Intel’s Gaudi 3 gain traction? The flexibility is minimal. The article doesn’t mention any multi-vendor strategy, which is a critical oversight. I’ve seen this in DeFi—protocols that lock into a single oracle (like Chainlink) face existential risk when that oracle fails. The same applies here.
Greed is the feature; the bug is just the trigger. The article’s commercialization section highlights “strong growth” but admits it’s debt-driven. CoreWeave’s debt-to-equity ratio is likely high, given their capital expenditure. The article doesn’t provide the exact numbers, but the pattern is clear: they are burning cash to build infrastructure, hoping future revenue covers the interest. This is a classic Ponzi-like structure, but for compute, not tokens.
Let’s do the math. Assume CoreWeave has 100,000 H100 GPUs, each costing $30,000. That’s $3 billion in hardware. If they finance 70% of that, their annual interest payment at 8% is $168 million. Add power, cooling, and labor, and the breakeven utilization rate is high. The article doesn’t ask the obvious question: what happens if utilization drops below 60%?
The exploit wasn’t in the code; it was in the assumptions. The competitive landscape is a three-front war: hyperscalers (AWS, Azure, GCP), other GPU clouds (Lambda, Crusoe), and in-house AI chips (Google TPU, AWS Trainium). CoreWeave’s advantage is speed, but speed is a temporary edge. The hyperscalers have deep pockets and can outspend them. The article’s comparison matrix is useful, but it misses the key point: the hyperscalers are also building custom chips, which will reduce their dependence on NVIDIA. CoreWeave and Nebius have no such leverage.
I recall my work on the Axie Infinity bridge exploit. The vulnerability was a gas optimization issue, but the root cause was a lack of diversity in the trust model. The same applies here: CoreWeave’s reliance on NVIDIA is a single point of failure. The article mentions this but doesn’t stress it enough.
Contrarian: What the Bulls Got Right
Now, the contrarian angle. The Bulls aren’t entirely wrong. The demand for AI compute is real, and it’s growing. CoreWeave and Nebius are capturing a segment that hyperscalers can’t serve quickly. Their speed is a genuine advantage in a supply-constrained market. The article also notes that these companies could benefit from “European AI sovereignty” themes, which is a real political tailwind.
But here’s the nuance: the market is pricing in a future that may not materialize. The article’s investment section warns that valuations are high, but it doesn’t go far enough. The real risk is a “gray rhino” event—a slow-moving crisis like a rate hike or a shift in AI demand. If the next generation of AI models requires less compute (e.g., through better algorithms), the entire infrastructure thesis collapses.
You didn’t check the math. The article’s investment analysis gives a confidence rating of B, but I’d lower that to C. The lack of data on gross margins, free cash flow, and customer concentration makes any valuation highly speculative. CoreWeave’s IPO prospectus, if I had access, would reveal the true picture. But the article’s analysis is based on inference, not verification.
Takeaway
The bottom line is this: CoreWeave and Nebius are not the future of AI; they are a leveraged bet on the current GPU shortage. The market’s euphoria is masking the structural risks. I don’t see a path to sustainable profitability without a dramatic improvement in utilization or a shift to higher-margin software services. The smart money is on NVIDIA, which controls the supply chain, not the resellers.
The exploit wasn’t in the code; it was in the assumptions. The AI infrastructure bubble will burst, and when it does, the casualties will be those who ignored the math. The question is whether you’ll be holding the bag.
