A lender that pioneered the original chip-backed loan in AI has walked away from Nvidia silicon for the first time — and the collateral it accepted instead may define who gets to build the next layer of AI infrastructure.

General Compute, a Boston-based inference neocloud founded earlier this year, announced on Thursday that it has secured a committed debt facility of up to $400 million from Upper90 Capital Management. The loan is backed not by Nvidia's H100s or B200s — the hardware that has anchored every major chip-backed loan in AI to date — but by SambaNova's SN50 inference ASICs: purpose-built silicon designed specifically to run already-trained models as fast and cheaply as possible. The facility begins at an initial commitment of $100 million and scales as customer demand grows, with a ceiling of $400 million.

If the deal holds as described, it is the first time inference-specific chips have served as primary collateral for a major AI loan. That distinction is not a technicality. It is a structural signal about where the debt markets now believe recurring revenue lives in the AI stack.

How Chip-Backed Loans Became Wall Street's AI Play

To understand why this deal breaks from the template, it helps to understand the template itself.

In 2021, Upper90 co-founder and CEO Billy Libby — a former Goldman Sachs quantitative trader — financed GPU purchases for Crusoe, an energy-focused data center startup, in what he describes as the first loan against the value of advanced AI chips. At the time, traditional lenders avoided the structure entirely. GPU depreciation was poorly understood, and no market consensus existed on what a rack of Nvidia chips was actually worth as a recoverable asset.

That skepticism reversed fast. In August 2023, CoreWeave borrowed $2.3 billion using Nvidia H100s as collateral in a deal led by Magnetar Capital and Blackstone. The chip-backed model became the engine of CoreWeave's entire growth strategy. Its $8.5 billion DDTL 4.0 facility, closed on March 31, 2026, received investment-grade ratings of A3 from Moody's and A (low) from DBRS — the first GPU-backed financing to achieve that status — at a fixed borrowing cost of approximately 5.9%. That rate compares with the roughly 15% Upper90 charged Crusoe five years earlier — a compression that reflects how rapidly institutional lenders reclassified AI hardware as an asset class.

CoreWeave's total debt now exceeds $21 billion. The model it built has defined how every neocloud — the category of purpose-built AI cloud operators that sit between hyperscalers and AI labs — has financed its hardware buildout. All of it has rested on Nvidia silicon. Until now.

Why SambaNova's Chips Are Architecturally Different

The reason General Compute chose SambaNova's SN50 over Nvidia GPUs is not primarily about price — it is about architecture, and specifically about what happens during inference at the chip level.

Running a large language model involves two distinct computational phases. The first is prefill: the model reads and processes the entire input prompt, building a key-value cache that summarizes what it has seen. This is a parallel computation task — it maps cleanly to the SIMT (Single Instruction, Multiple Threads) architecture that Nvidia GPUs were built around. The second phase is decode: the model generates a response one token at a time. Each output token depends on all previous tokens, making decode a sequential, memory-bandwidth-bound operation. The bottleneck is not compute — it is data movement. How quickly can the chip shuttle weights and activations between its memory and its processing units?

SambaNova's SN50 is built specifically for that second constraint. Its Reconfigurable Dataflow Unit (RDU) architecture places memory and processing circuits immediately next to one another, minimizing data travel time. Rather than using the general-purpose instruction dispatch model of a GPU, the RDU maps computation graphs directly to the most efficient hardware path for moving data — a dataflow approach that dates to MIT research in the 1970s but has found commercial relevance in the era of memory-bottlenecked inference workloads. The SN50's three-tier memory hierarchy — on-chip SRAM running at hundreds of terabytes per second, 64 GB of HBM2E at 1.8 TB/s, and up to 2 TB of DDR5 — gives the chip a deep memory pool accessible at speeds that GPU-only designs cannot match for sequential decoding.

General Compute exploits this architectural split directly. AMD MI300X cards — with 12 dies, 8 memory stacks, and 153 billion transistors — handle the compute-intensive prefill phase. SambaNova's SN50 RDUs handle decode. Independent benchmarking by Artificial Analysis in July 2026, using a heterogeneous platform combining four Nvidia H200 GPUs with 16 SambaNova SN50 RDUs, reached 763 tokens per second on MiniMax M2.7 at a 10,000-token context — several times faster than GPU-only providers in the same benchmark. SambaNova's own data, which should be read as company claims rather than independent validation until more benchmarks accumulate, reports sustained performance above 450 tokens per second at longer context lengths.

A second architectural advantage matters directly for General Compute's financing structure: the SN50 does not require water cooling. The chip operates on 20 kilowatts per rack, within the range of standard air-cooled data center infrastructure. General Compute has already secured agreements for 15 megawatts of air-cooled rack capacity in co-location facilities. This is not a minor logistical point. Liquid cooling requires specialized facilities and multi-year construction cycles. Air-cooled deployments can be installed in weeks. The faster a neocloud can add capacity in response to customer contracts, the tighter the feedback loop between revenue and capital draws — which is also how the $400 million facility is structured.

Lenders Follow Recurring Revenue

The deeper logic of the deal runs through the distinction between training compute and inference compute.

Training a large model is a one-time event, or close to it. A lab trains a new model during an intensive burst of compute that may last days or weeks, then the cluster sits at partial utilization between runs. Inference — serving that model to users, developers, and autonomous agents — is continuous. Every API call, every chatbot response, every agentic workflow step is an inference workload. As model deployment scales and AI becomes more deeply embedded in enterprise software, inference approaches something like a utility: a persistent revenue stream generated by every token produced.

"Everyone doesn't need a supercomputer, but they do need inference and AI," Libby told TechCrunch. Upper90's thesis is that lenders who followed recurring revenue into GPU loans are now following it into inference-specific infrastructure. The company that serves inference at lowest cost per token, with fastest time to first token, captures that revenue — and the hardware backing those economics makes for more bankable collateral than hardware whose utilization is tied to episodic training schedules.

The facility's draw structure reinforces the argument. Each additional tranche of the $400 million is tied to customer demand, meaning General Compute draws capital only as it secures paying contracts. The credit risk scales in step with actual revenue rather than ahead of it — a discipline that was absent from some of the more aggressive GPU financing structures of the prior cycle.

"There are a bunch of chips that are starting to scale that have amazing total cost of ownership, or that can operate much faster than Nvidia, but there's not too many buyers for them," CEO Finn Puklowski told TechCrunch. "By getting together with Upper90, this is not just, 'a cool startup got some money to buy some compute.' This is the first signal of capital organizing itself and the fragmenting of Nvidia's monopolistic dominance."

Implications Beyond One Startup's Balance Sheet

The financing template General Compute has established — ASIC inference chips as primary loan collateral — carries downstream consequences for how the broader AI infrastructure market evolves.

SambaNova closed the first tranche of a $1 billion Series F round on July 8, 2026, at an $11 billion post-money valuation, led by General Atlantic. That valuation — a roughly fivefold increase from the $2.2 billion implied by its February 2026 Series E — reflects investor confidence in inference-specific silicon as a distinct product category. Upper90 independently arriving at the same conclusion from the credit side of the ledger, within two weeks of that equity round, represents a convergence: equity markets and debt markets are now both pricing inference chip infrastructure as a legitimate asset class.

TensorWave, another neocloud operator, is making a parallel bet on AMD hardware — non-Nvidia silicon as inference infrastructure. Groq and Cerebras, two other inference-focused chipmakers, have drawn acquisition interest and public market attention. The common thread is that each represents a non-Nvidia path to serving the inference layer.

For chip startups that have historically struggled to find buyers willing to commit to large, upfront non-Nvidia silicon orders, the General Compute deal offers a new template: if inference chips can serve as debt collateral, they can anchor the same debt-funded acquisition model that has defined the GPU neocloud category. That template does not require the chipmaker to have a relationship with a lender — it requires a cloud operator willing to structure its business around the chip and a lender willing to price the chip as a recoverable asset.

What Lenders Are Pricing Risk On — and What They Cannot Yet Know

None of this is without structural risk. Understanding what lenders have accepted — and what remains unpriced — requires separating two things that the deal has conflated.

The first is performance risk. SambaNova's headline claims — 5x the compute performance of competing accelerators, 3x the total cost of ownership advantage — are the company's own benchmarks, not independently verified by MLPerf submissions or third-party laboratory evaluations. The Artificial Analysis benchmark of 763 tokens per second is real and significant, but it reflects a specific heterogeneous configuration on a specific model (MiniMax M2.7) at specific context lengths. Production deployments will encounter a wider range of models, request patterns, and concurrency levels. Practitioners should demand cost-per-token, sustained concurrency, and integration data before moving production inference traffic to any new chip platform.

The second and deeper risk is collateral marketability. The asset-backed lending framework requires that collateral be marketable — meaning it can be sold under normal conditions at current fair market value to a willing buyer. For Nvidia H100s and H200s, that secondary market exists and is increasingly liquid; Compute Exchange opened a secondary market for used Nvidia H100 and A100 GPUs on July 17, 2026. For SambaNova SN50 ASICs, no secondary market exists. Upper90 has accepted these chips as collateral before any resale history has been established — the identical structural position that made traditional lenders refuse GPU loans in 2021, except that the ASIC version carries additional risk. Unlike a used Nvidia GPU, which can serve almost any AI workload a new buyer brings to it, a used inference ASIC optimized for transformer-based decode may become worthless if the dominant model architecture shifts away from the patterns it was designed for. The SN50's advantage is architecturally specific; so is its obsolescence risk.

A high-end Nvidia GPU loses roughly half its resale value within three years, according to S&P Global Ratings analysts. ASIC depreciation curves for inference chips are not yet established, because the asset class has no secondhand market to calibrate against. Upper90's willingness to price that uncertainty as acceptable risk is the bet. Whether the inference revenue flowing through General Compute's infrastructure proves durable enough to service the loan — and whether SambaNova's chips hold enough residual value if it does not — are questions that will take years, not weeks, to resolve.

What the deal has established clearly is that inference chips are, in the judgment of at least one major AI infrastructure lender, worth treating as a separate asset class from training hardware. That reclassification will be read by competing lenders, by other inference cloud operators, and by chipmakers that have struggled to find financing pathways outside the Nvidia ecosystem. The downstream effects are only beginning to be understood.

Frequently Asked Questions

What makes an inference chip different from a GPU for the purposes of a loan?

A GPU is a general-purpose accelerator: it can run training, inference, graphics, or simulation workloads, and the secondary market for used Nvidia GPUs is now large enough that lenders can estimate its resale value with reasonable confidence. An inference-specific ASIC — like SambaNova's SN50 — is optimized for a narrower set of operations: specifically, the decode phase of large language model inference, where sequential token generation is constrained by memory bandwidth rather than raw compute. That optimization makes it faster and cheaper per token for its target workload, but it also means the chip has almost no reuse value outside that workload. Lenders accepting inference ASICs as collateral are betting that the specific workload the chip is optimized for remains economically significant for the life of the loan — a bet that GPU lenders do not have to make in quite the same way.

How does SambaNova's SN50 chip handle inference differently from Nvidia's architecture?

Nvidia's GPUs use a SIMT (Single Instruction, Multiple Threads) design that dispatches the same instruction to many processing cores in parallel — a model built for the highly parallel operations in training. SambaNova's Reconfigurable Dataflow Unit (RDU) maps computation graphs directly to hardware: processing units sit immediately next to the memory they need, so data travels shorter distances and the decode phase runs faster. General Compute uses AMD MI300X cards to handle the compute-heavy prefill phase — when the chip reads and processes the input prompt — and routes decode operations to SambaNova's SN50s, where the memory-proximity advantage matters most. Independent benchmarking by Artificial Analysis found this heterogeneous configuration reached 763 tokens per second on MiniMax M2.7, several times faster than GPU-only setups in the same test.

What happens if the SambaNova chips lose value before the loan is repaid?

This is the structural risk the deal leaves unresolved. Asset-backed loans require that collateral be sellable at fair market value under normal conditions. For SambaNova SN50s, there is no secondary market — no established pool of buyers for used inference ASICs, and no historical price data to calibrate depreciation against. If General Compute cannot service the loan and Upper90 is forced to liquidate the chips, the recovery value depends on finding buyers for specialized hardware whose commercial value is tied to a specific model architecture that may or may not remain dominant. Bitcoin mining ASICs provide a cautionary precedent: chips built for one hashing algorithm lost nearly all their value when the algorithm was superseded. Inference ASICs face analogous risk if transformer-based decode is displaced by architectures that do not benefit from the SN50's memory-proximity design. Upper90 has priced this risk as acceptable; the terms of how it has done so have not been publicly disclosed.

Why is inference considered more bankable collateral than training compute?

Training a large model is episodic: it happens during an intensive burst of compute, then the cluster idles between runs. Inference is continuous: every API call, every chatbot response, every agentic workflow step generates a token and a billable event. As AI deployment scales, inference revenue approaches a utility-like pattern — predictable, recurring, and growing. Goldman Sachs has projected that global AI token consumption will grow roughly 24-fold over the next several years. Lenders, as Upper90's Libby has noted, tend to follow recurring revenue. General Compute's deal structure links each capital draw to confirmed customer demand, which means the credit risk scales with actual revenue rather than ahead of it — a tighter coupling between cash flow and debt service than the more aggressive GPU financing structures of 2023–2025.