Hook
A $400M loan backed by AI chips. That's not a trade—that's a synthetic CDO on compute. General Compute just closed a $400M debt facility from Upper90, collateralized entirely by SambaNova ASICs. The pitch: turn crypto mining sheds into inference factories. The reality: they're betting the farm on the resale value of a niche chip that has zero secondary market liquidity. I've seen this movie before. In 2022, I watched Terra's algorithmic stablecoin collapse because the collateral was a narrative, not cash. Here, the collateral is a chip that has never been stress-tested under fire. Let's run the numbers on what this loan actually costs.
Context
General Compute is a startup that raised only $15M in seed equity. Then they levered up 26x to buy a bunch of SambaNova RDU (reconfigurable dataflow units) and retrofit ex-crypto mining data centers. Their target: pure AI inference workloads—not training. They argue that for inference, ASICs beat GPUs on price-per-token. The lender, Upper90, is taking the chips as collateral, essentially betting that these ASICs will hold value or appreciate. The market context: we're in a bull market for AI compute, but GPU supply is still tight. SambaNova is a David against Nvidia's Goliath. This loan is a bet that David wins—and that his slingshot (the chips) can be resold if the borrower defaults.
But here's the kicker: SambaNova chips are not liquid. They're not traded on any exchange. There's no open interest, no bid-ask spread. If General Compute defaults, Upper90 doesn't get cash—they get a pile of custom silicon that only works in specific clusters. That's a liquidity event waiting to happen.
Core
Let's break down the loan dynamics. Assume the $400M loan has a 5-year term at 12% interest (reasonable for a asset-backed loan in this rate environment). That's $48M in annual interest payments. General Compute needs to generate $48M in free cash flow just to service the debt, before they earn a penny for equity. On top of that, they have operating costs: power, cooling, networking, staff. Let's say they have 10,000 SambaNova RDUs deployed (a wild guess based on $400M spend on chips at ~$40k each). Each RDU must generate $4,800 per year in net profit after OpEx just to break even on interest. That's a lot of tokens.
But the real risk is collateral depreciation. GPUs lose value fast—Nvidia's H100 went from $30k to below $20k in 18 months as supply caught up. SambaNova chips, being niche, could lose value faster if the market doesn't adopt their architecture. A 50% haircut in 2 years would leave Upper90 holding $200M in illiquid hardware. And here's where the real trap lies: the loan is structured as a covenant-lite deal typical of crypto-style lending. No maintenance margin calls. The lender has only two options: hold the chips to maturity or sell them at a loss. This is exactly the same dynamics that killed many DeFi lending pools in 2022—overcollateralization with assets that become undercollateralized when price drops.
We don't trade narratives; we trade liquidity. The only way this loan works is if there's a vibrant secondary market for SambaNova chips. Without that, General Compute is holding a leveraged long on an illiquid asset. Smart money doesn't chase yield without understanding the collateral. Here, the collateral is a short on Nvidia's dominance—if Nvidia releases a better inference chip (like their rumored L40S successor), the value of SambaNova chips plummets. The lender is effectively short Nvidia innovation. That's a bet I wouldn't take.
Contrarian
Everyone is bullish on this pivot: "Crypto miners find AI use case—bullish!" "Real-world assets backed compute—bullish!" That's the retail narrative. But retail doesn't read loan docs. The contrarian angle: this is a sign of desperation in the crypto mining industry. Many mining operators are underwater or looking for a story to raise capital. General Compute is a re-branding of existing mining infrastructure, not a greenfield AI-first data center. The engineering challenge of running AI workloads in mining sheds is non-trivial: cooling requirements differ, networking demands are higher, and the software stack for SambaNova is immature compared to CUDA. Yield is the rent you pay for holding someone else's promises. The promise here is that General Compute can achieve better price-performance than AWS Inferentia or Google TPUs. But those hyperscalers have decades of optimization. General Compute has a loan and a hope.
Moreover, the loan terms likely include warrants or equity kickers for Upper90—meaning the lenders get upside if the equity succeeds. That dilutes the equity holders further. The seed investors are now subordinated to $400M in debt with potential conversion rights. This is a lose-lose for retail bag-holders if the company ever IPOs. The real trade is not the AI compute narrative—it's the debt restructuring trade. Watch for secondary market sales of SambaNova chips by Upper90 if General Compute misses a payment. That's the canary.
Takeaway
Actionable price levels: if you see any news about SambaNova chips being listed on eBay or a forced auction from Upper90, short the narrative. Short the sector. The floor for this trade is the liquidation value of the chips minus legal costs. My guess: 30 cents on the dollar within 18 months. That's a 70% drawdown from the implied collateral value. Are you willing to be the exit liquidity for someone else's synthetic compute CDO? I'm not.