Bankers Put $400 Million on Inference Chips' Payoff

$400 million is now being lent against inference chips, not just Nvidia training GPUs, and that shift says more about AI economics than another startup financing headline. General Compute, an AI inference cloud startup, has secured the loan from Upper90, according to TechCrunch, in what may be the first deal using inference-specific chips as collateral.

The signal is simple: capital is moving from the training boom toward the machinery that runs already trained models. That means faster, cheaper model serving, especially for open source models, is starting to look bankable.

AI lenders are treating inference chips as the next bankable asset class

Upper90 is not just funding another AI cloud. It is backing a thesis that inference chips can become a durable collateral class, the way scarce GPUs did during the first AI infrastructure surge.

General Compute, founded by CEO Finn Puklowski, raised a $15 million seed round in May to build an inference neocloud around SambaNova silicon. A neocloud is purpose-built for AI workloads, unlike the broader infrastructure sold by hyperscalers such as AWS or Azure.

The company’s SN50 chips are designed for inference. General Compute says they are power-efficient, do not require expensive water-cooling systems, and can be deployed faster than GPUs across a wider range of data centers. It also says the chips will provide 16 times faster inference than GPU-based clouds.

That claim is central to the financing logic. If the chips run useful workloads cheaper and faster, they can support repeated usage. XOOMAR analysis: that is why this deal is about cash-flow visibility as much as hardware value.

Inside the $400 million loan and the collateral logic behind AI chips

Chip-backed lending turns expensive processors into financeable assets. The borrower uses the loan to acquire compute hardware, while the lender underwrites the chips, deployment plan, and expected demand for the compute they will provide.

With GPUs, lenders learned to think about depreciation, scarcity, resale value, and customer contracts. With inference chips, the underwriting question gets narrower: will these specialized accelerators stay useful as models, software stacks, and buyer preferences change?

Upper90 co-founder and CEO Billy Libby, a former Goldman Sachs quantitative trader, has history here. In 2021, Upper90 financed GPU purchases by Crusoe, the energy-focused data center startup. Libby believes that was the first loan against the value of advanced chips.

“When we financed Nvidia GPUs as the first group to do that, the market was inefficient,” Libby told TechCrunch. “We could really put together something as an early participant, and kind of get compensated for the risk.”

Now GPUs are “comparatively well understood,” per the source material. Upper90 is hunting for the next inefficient pocket: specialized inference infrastructure.

The numbers that make inference-chip financing attractive to private credit

The headline number is $400 million, but the contrast matters more. General Compute’s loan dwarfs its $15 million seed round, showing how debt can scale an infrastructure company faster than equity alone, if lenders believe the assets can produce revenue.

Training and inference carry different financial profiles:

XOOMAR analysis: lenders in deals like this will care less about generic chip scarcity and more about workload economics. The important metrics are likely to include chip utilization, revenue per accelerator hour, energy cost, depreciation assumptions, customer concentration, and loan-to-value discipline.

The risk is just as clear. If newer accelerators outperform the SN50 chips, or if buyers stay loyal to more familiar GPU-based infrastructure, the collateral could become harder to value.

From GPU gold rush to inference-chip credit

Upper90’s move follows a path already proven by GPUs. Traditional lenders initially avoided loans against advanced chips because GPU depreciation was hard to model. Then CoreWeave turned chip-backed loans into a business model and later into the basis of a major IPO, according to the source material.

That history matters because capital markets often start with the asset everyone wants, then move into the next layer once the first trade gets crowded. In AI, the first layer was Nvidia GPU scarcity. The next layer may be lower-cost inference capacity.

Libby framed the shift bluntly:

“We think open source models are going to be important, and we went and looked for a player last year that was in inference,” Libby said. “Everyone doesn’t need a supercomputer, but they do need inference and AI.”

That sentence captures the deal. The market does not need every company to train frontier models. It needs places to run models, repeatedly, cheaply, and at scale.

For related XOOMAR coverage on AI moving beyond software demos into operational and hardware decisions, see OpenAI First Hardware Snubs AI Companion Hype for Coders and 400,000 Daily Prompts Put Bank of America AI on Trial.

Borrowers, lenders, chipmakers, and cloud customers face different stakes

For General Compute, chip-backed debt offers a way to buy hardware without relying only on equity fundraising. The catch is pressure. Hardware sitting idle does not repay a loan.

For Upper90, the deal opens a new private-credit lane. But the firm is not just betting on General Compute. It is betting that inference-specific silicon can hold value long enough to support lending.

For SambaNova, financing can pull more SN50 chips into the market. That helps prove demand outside the Nvidia orbit.

For customers, the potential benefit is cost-efficient inference. The source material ties the deal to concern over the price of AI tools and tokens, and to the rise of infrastructure that runs open source models more cheaply than the newest LLMs from frontier labs.

Nvidia alternatives are moving from pitch decks into loan documents

General Compute’s use of SambaNova chips matters because the deal is outside Nvidia’s orbit. The source also points to TensorWave, which is making a similar bet through a partnership with AMD.

Other signals support the broader thesis. Companies offering access to open models, including OpenRouter and Fireworks, have raised new rounds at large valuations, according to the source material. New models such as Kimi’s K3 have competed with recent releases from Anthropic and OpenAI on coding benchmarks. Chipmakers such as Groq and Cerebras have drawn interest from acquirers and public markets.

Puklowski cast the deal as a capital-markets break from Nvidia dependence:

“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,” Puklowski said. “By getting together with Upper90, this is not just, ‘a cool startup got some money to buy some compute.’ Like, this is the first signal of capital organizing itself and the fragmenting of Nvidia’s monopolistic dominance.”

That is the sharpest claim in the story. The loan is not proof that Nvidia’s position has cracked. It is proof that lenders are willing to finance a possible crack.

Inference chips will test whether AI credit can survive beyond scarcity pricing

The next phase of AI infrastructure finance will be judged by evidence, not slogans. More loans may be structured around contracted inference revenue rather than simple chip scarcity, but lenders will need proof that specialized chips can stay busy and retain value.

The confirming signs are clear: high utilization, repeat customers, strong performance on real workloads, and evidence that open source model demand keeps flowing through inference clouds. The weakening signs are just as clear: idle inventory, weak resale markets, software friction, or a fast upgrade cycle that makes today’s inference chips look stale.

The core tension remains unresolved. Inference chips may be the more durable AI workload bet, but durability in compute finance only counts when the hardware earns through the cycle.

The Bottom Line

• The deal suggests AI infrastructure financing is shifting from training hardware toward inference capacity.
• If inference chips deliver cheaper and faster model serving, they could become a new bankable asset class.
• The $400 million loan signals growing confidence that AI usage revenue can support hardware-backed lending.