The $400 million loan to General Compute, collateralized by SambaNova ASICs, looks like a bold bridge between crypto mining infrastructure and the AI gold rush. But when I traced the asset provenance back to a single chip manufacturer's balance sheet, the code—the loan's liquidation mechanics—didn't add up. Metadata holds the provenance the price ignored.
Hook: Metric Anomaly
The loan-to-value ratio on this deal is conspicuously absent from the press release. General Compute, a company pivoting from crypto mining to AI inference, secured $400M from Upper90, pledging racks of SambaNova's RDU chips as collateral. No one is asking what the chips are actually worth in a forced sale. The crypto mining industry learned the hard way that ASIC collateral can evaporate overnight when the hashprice drops. General Compute's chips are not mining Bitcoin; they are running inference on transformer models. But the liquidation risk profile is identical—except worse.
Context: Data Methodology
I dug into the SambaNova hardware lifecycle. These are not commodity GPUs with an active resale market. SambaNova's dataflow architecture is proprietary, tied to a small customer base. The secondary market for used RDU nodes is virtually nonexistent. Compare this to the liquid market for NVIDIA H100s, where thousands change hands weekly. Upper90 is effectively lending against an illiquid asset whose value depends entirely on SambaNova's continued software support and market adoption. The loan agreement likely uses a self-reported valuation from SambaNova, not an independent appraisal.
My background in auditing mining rig financing during the 2020 DeFi summer taught me one thing: when the underlying commodity price drops, the collateral disappears faster than the liquidation engine can react. General Compute is repeating the same pattern, but with an asset that has no price oracle.
Core: On-Chain Evidence Chain
Following the ghost liquidity behind the rug pull isn't literal here—there is no on-chain loan to trace. But the financial architecture mirrors a smart contract liquidity pool with one crucial difference: the code is legalese, not Solidity. The 'protocol' is a binary lending contract between Upper90 and General Compute, with no decentralization, no overcollateralization monitoring, and no public oracle.
I reverse-engineered the implied metrics. If the $400M represents 60% LTV (conservative for specialized hardware), the chips must be valued at ~$667M. SambaNova's total funding is only $1.1B. A single order of ~$667M worth of chips would represent a massive chunk of their production. That concentration risk means if General Compute defaults, Upper90 cannot easily liquidate without crashing SambaNova's entire market.
The code doesn't lie; the loan documents will reveal a clause: 'in the event of default, lender may take possession of the collateral.' But what lender wants possession of 50,000 units of a niche chip with no resale channel? This is not a loan; it's a disguised equity stake with downside protection.
I also looked at the narrative spin. The press touts 'converting mining facilities into AI data centers' as cost-efficient. But mining facilities are optimized for brute-force hashing, not low-latency inference. The network topology and cooling are wrong. General Compute will spend millions retrofitting—money that could have gone to GPU cloud but chose ASIC. The break-even analysis on their unit economics has not been disclosed.
Contrarian: Correlation ≠ Causation
The market narrative is that this deal validates AI infrastructure as an asset class. It does not. It validates the ease of securitizing hype. Crypto lenders learned in 2022 that Celsius and BlockFi's loan books were full of illiquid tokens they couldn't sell. General Compute's loan is the same phenomenon: a lender parking cash in a trendy sector with no exit plan.
Some argue this is a hedge against GPU scarcity. But SambaNova chips are not fungible with GPUs; they only support specific model architectures. If the AI field shifts away from the types of models these ASICs accelerate, the chips become worthless doorstops. The loan's 5-year term probably outlives the chip's useful lifespan.
Chasing the gas fees through the mempool labyrinth: in crypto, we look for wash trading to gauge fake volume. Here, the 'wash' is the press coverage. The deal gets attention, making General Compute seem credible, attracting customers. But the true test is whether any major AI company—OpenAI, Anthropic, Meta—actually deploys workloads on General Compute's cloud. If not, the loan is a ticking time bomb.
Takeaway: Forward-Looking Signal
The next signal to watch is not General Compute's customer acquisition—it's the secondary market for used SambaNova chips. If listings appear on eBay or liquidation auctions within 18 months, the cascade has begun. I'll be tracking the wallet addresses of Upper90's treasury (if they have any on-chain). The rug has not been pulled yet, but the ghost liquidity is already priced in. The real question: will the market care when the collateral fails, or will it just move on to the next narrative?