I trace the balance sheet, not the press release. When Upper90 announced a $400 million loan to General Compute, backed by SambaNova ASIC chips, the crypto and AI press erupted in applause. Another innovation in hardware financing, they wrote. Another step towards cheap inference. But when I follow the numbers, I see a different story: a leveraged bet on unproven technology, held together by a narrative that ignores basic financial and engineering reality.
The deal is simple in structure. General Compute, a startup that raised only $15 million in seed funding, secured a massive debt facility to purchase SambaNova’s specialized inference chips. These chips will be deployed in repurposed cryptocurrency mining data centers, and the machines themselves serve as collateral. The pitch: cheap inference for AI workloads, bypassing NVIDIA’s expensive GPU ecosystem. The reality: a fragile stack of assumptions that could collapse under the weight of its own leverage.
Context General Compute positions itself as a disruptor. Instead of using NVIDIA GPUs for AI inference, it bets on SambaNova’s dataflow architecture, an ASIC designed to outperform general-purpose silicon on specific tasks. The company also claims to reuse existing infrastructure—former crypto mines—to cut capital costs. Upper90, a lender specializing in asset-backed loans, structured the $400 million facility with the chips as collateral. It’s a clever financial engineering move, but it’s also a textbook example of what happens when hype meets zero technical validation.
Core: Systematic Teardown First, the technical risks. SambaNova’s ecosystem is a desert compared to NVIDIA’s CUDA oasis. The chips may excel on paper, but in the real world, model compatibility is the bottleneck. To run a mainstream model like Llama 3 or Qwen2.5 on SambaNova hardware, General Compute must invest significant engineering resources to port and optimize software. There is no guarantee of seamless performance. Even if the chips work, the internal network of a repurposed mining facility is a joke compared to the high-bandwidth interconnects of modern AI clouds. Model parallelism across multiple chips will suffer latency penalties. The promise of “cheap inference” evaporates when you factor in the cost of adaptation and the reality of poor scaling.
Second, the financial structure is alarmingly fragile. A 26x debt-to-equity ratio is not innovation; it’s a margin call waiting to happen. The loan interest will burn cash faster than the company can onboard customers. And the customer base is unproven. Who will trust their inference workloads to an unproven platform with a niche chip? The market for cost-sensitive inference exists, but it’s dominated by established players like CoreWeave and Lambda Labs, who already have NVIDIA hardware and proven reliability. General Compute’s only differentiator—lower price—will be erased if the first major customer discovers unpredictable performance or downtime.
Third, the collateral is a ticking time bomb. SambaNova’s corporate health is uncertain. If the company falters, the chips lose all software support, turning them into expensive paperweights. Upper90 knows this, but the loan terms likely include provisions to dump the hardware if defaults occur. In a worst-case scenario, the AI chip secondary market for a niche ASIC is essentially zero. The lender’s risk is the startup’s lifeblood.
Contrarian Let me give credit where it’s due. The bulls have a point: this structure could pioneer a new asset class for AI hardware financing. If General Compute actually delivers on performance and cost, it will prove that alternatives to NVIDIA exist, and that financial innovation can accelerate deployment. The reuse of mining data centers is environmentally efficient in theory. And the first-mover advantage in the “cheap inference” niche could lock in price-sensitive customers before competitors adjust.
But these positives are conditional on execution, and execution is where the story breaks down. The onus is on General Compute to release independent benchmarks, disclose loan terms, and show a working product with real latency and throughput numbers. Until then, the bull case is a bet on the best possible outcome, not a sober analysis of probability.
Takeaway Hype is the only asset in a vacuum mint. General Compute has minted a $400 million narrative without a single verified performance metric. The industry should demand transparency, not celebrate the loan size. I trace the balance sheet, and I see a spreadsheet of risks dressed up as innovation. Prove me wrong—show the tests, show the customers, show the cash flows. Until then, this is a leveraged gamble on an ASIC that hasn't proven itself in the wild.