A $4.3 billion convertible bond. Not for a DeFi protocol, not for a Bitcoin miner, but for an AI data center builder. The market cheered. I saw a familiar pattern: capital chasing infrastructure before demand, debt disguised as growth. This is not innovation. This is a leveraged liquidity play.
Context
Nebius Group, the AI infrastructure arm spun out of Yandex, just raised $4.3B through convertible bonds. The stated goal: build massive GPU clusters to compete with AWS, Azure, and Google Cloud. The narrative is appealing—AI needs compute, and compute needs capital. But the structure matters. Convertible bonds are debt that can convert to equity at a future price. They give the issuer low interest today, but they stack a ticking clock on the balance sheet. I’ve seen this before. In 2021, miners used similar debt to buy ASICs. When Bitcoin dropped, they liquidated. The same logic applies here, except the collateral is GPUs, not hashrate.
Core
Let’s quantify the risk. $4.3B at a typical 2% coupon means $86M in annual interest payments before any principal. To service that, Nebius needs consistent revenue from GPU rentals. The current market rate for an H100 is around $2.50 per hour. To break even on interest alone, they’d need to rent out 11,000 GPUs 24/7 for a year. That’s not accounting for operational costs, depreciation, or the fact that H100 will be obsolete in 12 months when Blackwell volumes ramp.
Based on my experience managing leverage during DeFi Summer, I know that debt without matched cash flows is a liability. I borrowed ETH against ETH to chase yield, but I hedged with liquidation thresholds. Nebius has no such hedge. They are betting that AI compute demand will grow exponentially, and that their capacity will be rented at premium prices. But the data tells a different story. The GPU oversupply is already visible in spot markets. On-chain rental platforms like Vast.ai show a 20% drop in H100 rental prices since Q1 2024. The demand is there, but supply is growing faster.
Contrarian
Smart money is not buying this narrative. Look at the bond structure: it’s convertible, meaning investors want upside protection. They are not betting on Nebius’ equity; they are betting on a floor. The conversion price is likely set at a premium to current valuation, but if the stock drops, the bond becomes a debt overhang. This is the same dynamic that killed Celsius. They borrowed against assets that lost value, and the debt spiral crushed them.
Retail sees a $4.3B vote of confidence. I see a $4.3B short position on AI compute demand. If the next AI breakthrough reduces training costs—say, a more efficient architecture—the need for massive clusters drops. Then Nebius is left with billions in depreciating hardware and no way to service debt. The only exit is to sell the GPUs on the secondary market, flooding supply and crashing prices. It’s a classic liquidity trap.
Takeaway
Watch the bond conversion terms. If they are tight (e.g., 30% premium, 3-year maturity), Nebius is betting on flawless execution. I’m not taking that bet. Liquidity dries up when fear sets in, and this debt is the kindling. Gas is the toll for chaos.