The numbers hit my screen at 6:13 AM Tallinn time. General Compute — a name I'd only seen in a seed round filing eight months ago — just closed a $400 million loan. Not equity. Not convertible notes. A loan secured against SambaNova ASICs. Chips that most of the market hasn't even heard of. Chips designed not for training, but for inference. The specific event: Upper90, a lender known for esoteric collateral, wrote the check against physical hardware stacked in repurposed crypto mining facilities. No one has done this at scale. Not CoreWeave. Not Lambda Labs. The hook is simple: the GPU monopoly just got challenged by a balance sheet.
Speed was the only asset that didn't depreciate in this market — until now, General Compute is betting that their chips will hold value better than the narrative. And that's where the story gets ugly.
Context: Why Now?
The AI inference market is a paradox. Demand is exploding — every chatbot, every code assistant, every AI agent needs inference compute. But supply is trapped. NVIDIA’s H100s are allocated years out. Cloud providers charge premium margins because they can. The GPU shortage isn't a technical problem; it's a production bottleneck that benefits incumbents. Meanwhile, the 2022 crypto winter left hundreds of thousands of square feet of mining data centers empty. Stranded power. Optimized cooling. Racks waiting for purpose. General Compute saw the arbitrage: take the real estate, fill it with non-GPU inference silicon, and offer pricing that undercuts AWS by 80%. The seed round of $15 million was a toe dip — enough to test the thesis. The $400 million loan is a declaration.
This isn't a company raising capital. This is a market structure bet. The terms: 4-year loan, collateralized by SambaNova RDU processors, interest rate estimated in the 8-12% range (unconfirmed but standard for asset-backed tech loans). The lender, Upper90, specializes in revenue-based and asset-backed financing — they're not traditional VCs. They're betting that the chips themselves will retain value, either as compute hardware or as liquid assets in a secondary market. If General Compute defaults, Upper90 seizes the ASICs. That's the risk calculus.
Core: The Machinery of the Bet
Let me go deep on the technical architecture, because that's where the real insight lies. SambaNova’s chip is not a GPU. It’s a reconfigurable dataflow unit (RDU). Instead of fetching instructions from memory every cycle, it maps the entire model graph onto the chip’s fabric. Data flows through pre-arranged pathways. This eliminates the von Neumann bottleneck — the constant shuttling between memory and compute that plagues GPUs. For inference workloads, where the model is fixed and you're just feeding inputs, this is massive. Lower latency. Higher throughput per watt. SambaNova claims up to 10x efficiency versus equivalent GPU solutions. I've audited enough hardware specs to know that vendor benchmarks are taken against ideal scenarios — but even at 3x, the economics shift.
General Compute’s deployment strategy is equally aggressive. They're retrofitting existing crypto mining facilities. Why? Cost. A greenfield data center costs $10-15 million per megawatt to build. A mining facility? You’re paying for the shell, power infrastructure, and cooling. The owner is desperate to offload. General Compute can secure long-term leases at distressed prices. The operational expense drops. But there's a catch: mining farms were designed for hashing, not for high-bandwidth inter-chip communication. Bitcoin ASICs talk to the network, not to each other. SambaNova RDUs need to be clustered — model parallelism requires low-latency connections between chips. If the internal network is a flat Ethernet topology, latency spikes will kill performance for large models like Llama 3 70B. General Compute claims they've deployed custom fabric — but that's engineering money they didn't budget for in the seed round.
Volume tells the truth when price tries to lie. The loan amount is $400 million. Assuming 50% of that goes to hardware procurement (the rest for retrofitting, networking, working capital), they can buy roughly 4,000 SambaNova RDU nodes at $50,000 each. Each node is roughly equivalent to an H100 in inference throughput for certain models. That's 4,000 H100-equivalents — a meaningful cluster but still a fraction of CoreWeave's 50,000+ H100 fleet. The difference? General Compute’s cost per inference could be 60-70% lower due to chip efficiency and lower overhead. That's their wedge.
But here's the unglamorous reality: the SambaNova ecosystem is a desert. NVIDIA has CUDA, cuDNN, TensorRT, Triton Inference Server, and a decade of operator optimizations. SambaNova has a proprietary SDK, a handful of supported models, and a community of maybe a few hundred developers. General Compute must port every model their customers want. That means engineering time. And engineering time is expensive. During my PhD, I spent months optimizing a single cryptographic hash on an FPGA — the cost of non-standard hardware is not just purchase price, it's the labor to make it useful. General Compute is betting that the volume of inference requests will justify that labor. But volume takes time to build. And time accrues interest.
Contrarian: What Everyone Is Missing
The mainstream take is that this is a clever financial innovation — turning hardware into a bankable asset. I see a different story. This is a leveraged bet on a single chip vendor executing a platform transition. If SambaNova’s next generation falls behind, or if NVIDIA releases a dedicated inference chip (which they will — the H200 NVL is already hinting at it), the collateral value of those RDUs plummets. Upper90’s loan-to-value ratio was likely set at 60-70%. If the chips depreciate faster than expected, General Compute faces a margin call. That would trigger a fire sale of their only asset. The company would collapse.
Moreover, the interest burden is real. At 10% on $400 million, that's $40 million per year. Add operating costs, personnel, retrofitting amortization — the burn rate likely exceeds $15 million per quarter. Can they generate that revenue from inference customers? A typical mid-tier AI startup spends $50,000/month on inference. To hit $15M quarterly revenue, they'd need 300 such customers. That's possible, but it requires closing enterprise contracts with 6-month sales cycles. Meanwhile, AWS is already slashing prices on Inferentia. Google has TPU v5e for inference. The incumbents can afford to match any price drop because they have diversified revenue streams. General Compute cannot.
Arbitrage isn’t free; it’s the market correcting its own soul. The crypto mining connection is also a double-edged sword. Yes, cheap power. But many mining farms are in jurisdictions with unstable grids or regulatory ambiguity. Environmental scrutiny is rising. If a local government decides that AI compute licenses are required, or that energy consumption for inference is too high, General Compute could face shutdowns. The legal overhead is nontrivial.
Another blind spot: the secondary market for SambaNova chips. If General Compute defaults, Upper90 will try to liquidate the hardware. Who buys used ASICs that only run a proprietary software stack? No one. The resale value of non-NVIDIA AI accelerators is near zero. Compare that to GPUs — used H100s still trade at 80% of retail. The liquidity of the collateral is an illusion. This loan is effectively unsecured in a fire sale scenario.
We didn’t come this far to only come this far. But General Compute’s survival hinges on customer acquisition velocity. They need to win at least one hyperscaler or Fortune 500 inference contract within 12 months. A single large customer like Salesforce or Uber, committing to $5M/month, would cover half the interest. That’s the math.
Takeaway: Watch These Signals
The next 90 days will tell the story. First, benchmark results. If General Compute publishes a head-to-head comparison showing 3x cost improvement over H100 on Llama 3 70B inference, the narrative flips. Second, customer logos. A single name-brand customer validates the platform. Third, SambaNova’s roadmap. If they announce a next-gen chip with 2x performance, the loan’s collateral strengthens. If not, the clock starts ticking.
Efficiency is the price we pay for speed. General Compute is moving fast. They’re breaking the GPU hegemony, one ASIC at a time. But speed without a moat is just a fast exit. The $400M bet is either the birth of a new asset class — chip-backed compute lending — or a cautionary tale about leverage in a market that still bows to NVIDIA. I know where my money would sit: not on the loan, but on the outcome. And the outcome is binary. Either they become the AWS of inference, or they become a footnote in the next bear market.
Survival is a strategy, but leverage is a mindset. General Compute has both. Let’s see if they have the arithmetic to back it up.