On December 12, 2024, Nebius Group announced a $4.3 billion convertible bond issuance. The press release emphasized 'AI data center expansion.' The market reacted with a 12% rally. I read the term sheet. Three red flags emerged: undefined GPU supply agreements, a 5-year maturity with no amortization, and a conversion premium of 15% — below the industry average of 25-30% for comparable deals. These are not details; they are systemic vulnerabilities. The blockchain remembers; the architect forgets. In this case, the architect is Nebius's CFO, and the ledger is their balance sheet.
Context: Nebius Group, spun off from Yandex in 2022, operates as an AI infrastructure provider. The company focuses on building GPU clusters for training and inference, targeting developers and enterprises. The $4.3 billion convertible bond offering is one of the largest in the AI infrastructure space, positioning Nebius to compete with CoreWeave, Lambda Labs, and the hyperscalers. The broader market is in a sideways consolidation phase, with capital shifting from speculative tokens to hard assets like compute. Convertible bonds are a common tool in tech: they offer lower interest rates than straight debt but include equity conversion rights, diluting shareholders if the stock price rises. The terms matter. Nebius's offering yields 2.5% with a conversion premium of 15%, meaning bondholders can convert at 115% of the current share price. This is aggressive. It signals that the company is either desperate for capital or confident in rapid growth. I lean toward the former.
Core: The systematic teardown reveals three interlocking risk vectors. First, GPU supply chain dependency. Nebius plans to deploy 10,000 to 15,000 NVIDIA H100 GPUs per data center. At $30,000 per unit, the GPU cost alone is $300-$450 million per cluster. The $4.3 billion could fund 10 such clusters, but NVIDIA's allocation is constrained. In 2023, CoreWeave secured a $2.3 billion debt facility partly because it had a guaranteed supply agreement with NVIDIA. Nebius's offering document does not mention such an agreement. This is a critical omission. During the 2017 ICO audit, I saw a project ignore a critical integer overflow vulnerability because the team was pressured to meet a token sale deadline. The result: a 40% treasury drain. Here, the pressure to deploy capital before the bond matures may lead to signing unfavorable GPU procurement contracts, locking in high prices or long lead times. The blockchain remembers; the architect forgets.
Second, asset depreciation risk. The GPU market is in a rapid iteration cycle. NVIDIA's Blackwell B200 is expected to ship in Q2 2025, offering 2x performance per watt over H100. If Nebius locks in H100 purchases now, those assets will lose value within 18 months. This is analogous to the DeFi flash loan exploit I analyzed in 2020: the protocol's oracle dependency matrix was fragile because it relied on a single price feed. Nebius's asset depreciation risk is a similar single-point-of-failure — they are betting that H100 demand holds until 2027. Historical data suggests otherwise. In 2022, I published a data-driven exposé on an NFT collection with a $200 million market cap that was manipulating floor prices. The wash-trading mechanics were hidden in wallet clusters. Here, the hidden risk is the replacement cycle: once Blackwell clusters go online, H100 rental prices will drop by 30-40%, compressing Nebius's margins. The bond interest payments become a fixed cost on a shrinking revenue base.
Third, debt leverage risk. The convertible bond has a 5-year maturity with no amortization. Interest is paid-in-kind (PIK) for the first two years, meaning the company can issue additional bonds instead of cash. This defers the cash burden but increases the total debt. If the stock price stays below the conversion price, the bonds will not convert, and Nebius must repay $4.3 billion in principal in 2029. The company's current revenue is negligible — it reported $87 million in 2023, mostly from legacy cloud services. To cover the interest, Nebius needs to generate at least $107 million in cash flow annually (2.5% of $4.3 billion). That requires a 123% revenue growth rate for three years. In my 2022 analysis of the Terra/Luna collapse, I identified the twin-token model as a Ponzi scheme reliant on infinite growth. The same principle applies here: a business model that requires exponential user growth to service debt is fragile. The market is assuming that AI compute demand will grow linearly, but history shows that infrastructure cycles are cyclical. The 2021 NFT boom was followed by a 90% crash. The AI data center boom may follow a similar pattern.
To quantify the risk, I built a simple stress test model. Assume Nebius builds 10 clusters by 2027, each with 12,000 H100 GPUs. Total GPU cost: $3.6 billion (12,000 $30,000 10). Remaining $700 million for land, power, and networking. If Blackwell clusters reduce H100 rental rates by 30% by 2028, the annual revenue per cluster drops from $360 million (at $2.5/GPU-hour) to $252 million. The total revenue would be $2.52 billion, but interest expense is $107 million, and operating costs (power, cooling, staff) are about 60% of revenue, leaving $1.01 billion EBITDA. Debt service is manageable, but only if utilization stays above 80%. If utilization drops to 60% due to competition, EBITDA falls to $740 million, and the debt-to-EBITDA ratio exceeds 5x — a warning signal for credit rating agencies. The blockchain remembers; the architect forgets.
Contrarian: The bulls are not entirely wrong. The demand for AI compute is real. OpenAI, Anthropic, and Meta are spending billions on training. Nebius has a strong technical team from Yandex, with experience running large-scale distributed systems. The convertible structure allows them to avoid equity dilution at a low point, and the PIK interest feature gives them two years of breathing room. The market underestimates the commoditization of GPU clouds. Once AWS and Azure scale their own supply, margins will compress. But Nebius's advantage is temporary: they are betting on speed (fast deployment) and geographic diversity (non-US markets). The European AI ecosystem is underserved, and Nebius could capture that niche. However, the bond terms reveal a lack of negotiating power. A 15% conversion premium is low; it suggests that investors demanded a higher yield or a lower conversion price to compensate for the risk. This is a signal that the market is skeptical. The bulls need to watch the first two years of interest payments. If Nebius uses PIK aggressively, the debt load will grow, and the equity dilution will be severe when the bonds convert. The contrarian angle is that the deal might work if Nebius can secure a major customer, like a sovereign wealth fund or a European AI lab, before the first PIK interest payment is due.
Takeaway: The blockchain remembers; the architect forgets. Nebius's architect forgot that debt is not equity. The bondholders will remember if the AI data center becomes a stranded asset. I will be watching the quarterly cash flow statements. The ledger never lies. The next signal is the GPU supply agreement: if Nebius announces a deal with NVIDIA within six months, the risk profile improves. If not, the $4.3 billion becomes a liability, not an asset. The market is in a sideways consolidation phase, and capital is flowing to narratives. This one has a red flag. I am short the convertible bond and long the skepticism.