Check the loan docs. $400M at what basis? Upper90 didn’t lend against cash flows. They lent against silicon. SambaNova ASICs, locked in a data center that once mined Bitcoin. That’s the asset. No revenue, no user base, no token. Just hardware that depreciates faster than a used Tesla.
I don’t trust hype, I trust the contract. So I pulled the structure. General Compute, a startup with a $15M seed round, just closed a $400M debt facility from Upper90, a fintech lender specializing in asset-backed loans. The collateral: a fleet of SambaNova RDU (Reconfigurable Dataflow Unit) chips, deployed in ex-crypto mining facilities. The pitch: build a low-cost AI inference cloud by piggybacking on existing power and cooling infrastructure, then use the chips as a currency to borrow cheap capital.
Smart contracts don’t have emotions, but they do have logic leaks. This loan is a logic leak. Let me explain why.
The Context: Mining Hardware Meets AI FOMO
General Compute is not a traditional cloud provider. They’re a hardware arbitrage shop. They buy or lease ASICs that are niche—SambaNova, not NVIDIA—and plant them in facilities that were built for SHA-256 hashing. The theory: crypto miners already solved the power, cooling, and real estate problem. AI inference has similar requirements but higher margins. By retrofitting miner sheds with SambaNova racks, General Compute claims they can offer inference at 1/3 the cost of AWS Inferentia or NVIDIA T4 instances.
The loan from Upper90 is structured as a five-year term, interest-only for the first two years. The interest rate is not disclosed, but typical Upper90 deals for hardware-backed loans hover around 12–18% APR. That’s $48M to $72M per year in interest alone. General Compute’s current revenue? Zero. They haven’t launched a public cloud yet. They’re building the platform as they burn through the loan proceeds.
This is not a growth story. It’s a leverage story. And in a sideways market where token prices are flat and liquidity is shallow, leverage kills.
The Core: Order Flow Analysis of the Collateral
Let me break down the order flow. The collateral—SambaNova chips—are not liquid. There is no secondary market for RDU units. If General Compute defaults, Upper90 cannot sell these chips quickly without taking a massive haircut. Compare this to a DeFi lending pool where you deposit ETH and borrow USDC over-collateralized. Here, the collateral is a single-vendor ASIC that only works with SambaNova’s software stack. If SambaNova goes bankrupt or pivots, the chips become e-waste.
I audited ICO contracts in 2017. I remember a token that promised “audited code” but had a backdoor that let the admin mint unlimited coins. This loan has a similar backdoor: the software dependency. The SambaNova ecosystem is controlled by one company. Their SDK is closed-source. Their support team is small. If a critical vulnerability emerges in the RDU firmware, General Compute cannot patch it—they rely on SambaNova. That’s a single point of failure.
Code is law, but human greed is the bug. Upper90 and General Compute both think they’re hedging. Upper90 gets a 12-18% yield on a loan backed by physical assets. General Compute gets cheap capital to scale. But the real yield is priced in the risk of a chip crash. I’ve seen this before—in 2021, when NFT floor sweepers bought CryptoPunks at 100 ETH, thinking they could always dump to the next sucker. They didn’t. They became exit liquidity.
General Compute is exit liquidity. They are buying SambaNova chips at presumably list price, but they’re taking on debt to do it. If SambaNova’s next-gen chip (expected Q3 2026) makes the current RDU obsolete, the collateral value drops 50% overnight. Upper90 will margin call. General Compute will sell chips at a loss to cover interest. The cycle repeats.
The Contrarian Angle: Why Retail Loves This Narrative (and Why It’s Wrong)
Retail sees “$400M loan for AI infrastructure” and thinks “bullish.” They imagine a future where cheap inference drives a new wave of DApps, AI agents, and decentralized compute networks. They buy into the narrative that ASICs will replace GPUs for inference, and that General Compute is the first mover.
Smart money watches the blockchain, not the ticker. They see a company that raised $15M in seed equity, then took on 27x that amount in debt. That’s not a venture-backed startup—that’s a hedge fund bet on hardware appreciation. The same thing happened in 2020 with DeFi yield farming. People borrowed USDC at 10% APR to deposit into SushiSwap at 200% APR, thinking the yield would last forever. It didn’t. Impermanent loss killed them.
General Compute’s impermanent loss is the difference between the chip’s value at purchase and its liquidation value. If the AI inference market shifts to a new architecture—say, Groq’s LPU or Cerebras’ wafer-scale engine—SambaNova’s market share shrinks. The chips become stranded assets. Lower90 is not stupid; they likely wrote the loan with a downside clause: if the chip’s market price drops below a threshold, General Compute must add collateral or repay principal. But General Compute has no other assets. They’re all-in on one vendor.
This is the contrarian truth: the loan is not a sign of confidence. It’s a sign of desperation. General Compute needed capital to buy chips before SambaNova increased prices or competitors like Lambda Labs and CoreWeave locked up supply. They took the cheapest money they could find—debt, not equity—to avoid dilution. But debt requires cash flow. They have none.
The Technical Experience: What I Learned From Five Battle-Tested Trades
I’ve been through five major cycles where leverage disguised as infrastructure turned into losses. Let me connect the dots.
2017 ICO Audit: I found a reentrancy bug in Project Alpha’s token contract. The team ignored my report until I made it public. Then the scam collapsed. General Compute’s loan structure also has a reentrancy bug: the dependency on SambaNova software. If SambaNova’s code contains a vulnerability that bricks the chips, the entire collateral pool becomes unseizable. Upper90 would hold worthless silicon.
2020 DeFi Farming: I deployed 50 ETH into SushiSwap. I logged every impermanent loss trade. I learned that yield farming is just latency arbitrage. General Compute is trying to arbitrage the latency between ASIC demand and AI inference demand. But the yield (revenue) is not guaranteed. The cost (interest) is. That’s a negative carry trade.
2021 NFT Sweep: I tracked whale wallets accumulating CryptoPunks. I bought 12 at 15 ETH average, sold at 60 ETH. The key was on-chain distribution. For General Compute, the on-chain distribution of SambaNova orders is opaque. We don’t know how many chips they actually bought. The $400M could be for 10,000 RDU units or 100,000. The unit economics matter more than the headline number.
2022 Terra Collapse: I shorted governance tokens after the UST depeg. I saw that withdrawal limits protect the exchange, not the user. General Compute’s loan has a similar withdrawal limit: if they fail to hit revenue targets, Upper90 can seize the chips. The company becomes a zombie, unable to generate revenue because they lost their only capital.
2025 AI Bridge Audit: I reverse-engineered an AI trading bot that promised 40% APY. Hidden slippage costs erased profits. General Compute’s hidden cost is the software license. SambaNova charges annual maintenance fees for their software stack. That’s an ongoing OpEx that the loan doesn’t cover.
The Takeaway: Watch the Secondary Market, Not the Press Release
The final piece of the puzzle is the secondary market for SambaNova chips. If you want to know if General Compute will survive, don’t look at their announcements. Look at eBay, third-party vendors, and liquidation auctions for RDU units. If the price holds steady above $50,000 per chip, the collateral is safe. If it drops 20%, the margin call is coming.
I watch the blockchain, not the ticker. But this isn’t on-chain—it’s more primitive. It’s physical. And that’s the risk.
So here’s the forward-looking judgment: General Compute will either succeed or fail based on their ability to sign a major customer in the next six months. If they secure a contract with a company like Amazon (for Alexa inference) or Microsoft (for Bing Chat), they have revenue to cover interest. If they don’t, the loan will accrue until the chips are obsolete, and the lenders will pull the plug.
Smart contracts don’t have emotions. But humans do. And right now, the emotion is hope. Hope that ASICs will replace GPUs. Hope that inference margins are fat. Hope that a $400M loan is a sign of strength. It’s not. It’s a sign that someone is willing to lend against hardware that may soon be worth less than the debt.
Code is law. But in this case, the law is physics. Chips depreciate. Debt doesn’t.
