A $400 million loan secured against chips that are not even the market standard.
The code does not lie; only the auditors do. And in this case, the auditor is the market itself.
General Compute just closed a debt facility that turns crypto mining real estate and SambaNova ASICs into a financial instrument. A $4 million seed round followed by a $400 million loan. The leverage ratio is 100:1. That is not a business plan. That is a leveraged bet on a single asset class: non-GPU AI inference chips.
Let me reconstruct the ledger.
Context: The Mining Farm Rebrand
General Compute positions itself as an AI inference cloud. It buys SambaNova’s dataflow ASICs (not NVIDIA GPUs), installs them in repurposed cryptocurrency mining facilities, and sells compute time. The novelty is the collateral: the chips themselves back the loan. Upper90, the lender, takes a first lien on the hardware.
The narrative is seductive. Avoid the GPU tax. Use stranded mining infrastructure. Finance through asset-backed debt rather than dilutive equity. It sounds like a financial engineering masterpiece. But I trace the flow, you trace the lies.
Core: The Systematic Teardown
The thesis rests on three unverified assumptions:
1. SambaNova’s ASIC will outperform NVIDIA on inference at a lower cost.
SambaNova’s RDU (Reconfigurable Dataflow Unit) eschews the Von Neumann bottleneck. In theory, it delivers higher throughput per watt for transformer-based inference. In practice, the ecosystem is barren. CUDA has 20 years of libraries, debuggers, and community tools. SambaNova has a proprietary SDK and a small set of supported models. Every model port is a custom engineering job. General Compute must spend tens of millions in software integration just to make the hardware usable.
2. The repurposed mining sites are fit for AI workloads.
Mining farms are optimized for cheap power and high ambient temperature tolerance. AI inference requires low-latency networking, redundant cooling, and proximity to demand centers. A former Bitcoin mine in rural Montana cannot serve a real-time chatbot in New York. Network latency alone will kill any low-latency use case. The infrastructure play is a cost-saving story that ignores the physics of data transmission.
3. The debt can be serviced before the collateral depreciates.
$400 million at, say, 10% interest is $40 million a year. Typical enterprise inference clouds earn 2-3x hardware cost annually in revenue. That means General Compute needs to generate roughly $13-14 million per month in gross profit just to cover interest. Even with stellar utilization, that is a tall order. If the ASIC market prices drop (NVIDIA releases a cheap inference chip, or SambaNova fails to deliver), the collateral value collapses and the loan accelerates.
Based on my audit experience, this is a smart contract with a single point of failure dressed in financial engineering. The real risk is not technical. It is market timing.
Contrarian: What the Bulls Got Right
I am not here to dismiss the entire thesis. The contrarian angle has merit.
First, the financial innovation is real. Creating a debt market for non-GPU AI chips could unlock capital for alternative architectures. If General Compute proves that SambaNova ASICs hold value as collateral, it opens the door for Cerebras, Groq, Graphcore, and others to access cheap capital. That weakens NVIDIA’s stranglehold on AI hardware financing.
Second, inference is a volume game. If General Compute can offer inference at 60-70% of the cost of a comparable GPU instance, it will capture the long tail of low-margin, high-volume applications: text generation, content moderation, chatbots. These customers are price-sensitive and less loyal to CUDA.
Third, the mining assets are sunk costs. The land and power infrastructure are already paid for. General Compute is essentially arbitraging stranded assets. In a capital-constrained world, that is a legitimate competitive advantage.
The bulls see a capital-efficient path to building a cloud. I see a blade running at 90% utilization with no safety margin.
Takeaway: The Contract Has a kill Switch
The loan documents will determine the outcome. Does the lender have the right to repossess and sell the ASICs in a fire sale? Can they force a liquidation if the hardware loses 30% of its value? These terms are not public, but they are the only on-chain evidence that matters.
General Compute is a test case for an industry that wants to decouple from NVIDIA. If it fails, the lesson is that alternative hardware needs more than a financial wrapper. If it succeeds, we will see copycats with every abandoned mining farm and every niche ASIC.
Promises are encrypted; data is decrypted. I will be watching the collateralization ratio.
Silence is the loudest admission of guilt.