Inference Chips Are Becoming Loan Collateral in a $400M Bet

A lender just agreed to hand an AI cloud startup up to $400 million and hold inference chips as the security. Not Nvidia GPUs, which lenders have quietly financed for years, but SambaNova’s purpose-built inference silicon. According to TechCrunch, it might be the first deal to put up inference-specific chips as collateral, and the firm writing the check is the same one that financed the very first loans against advanced GPUs.

The borrower is General Compute, an inference cloud operator. The lender is Upper90 Capital Management, the investment firm that backed Crusoe’s GPU buys back in 2021. The structure looks small next to the tens of billions in GPU-backed debt already floating through the AI economy, but the collateral is the interesting part. When a lender accepts a chip as security, it is making a claim about what that chip will be worth if the borrower stops paying. Inference silicon has almost no track record on that question, and the deal is essentially a wager that it will hold.

What Upper90 actually agreed to fund

The company’s own announcement describes a committed debt facility of up to $400 million. It does not arrive all at once. The financing begins at $100 million and scales with customer demand, a detail SiliconANGLE confirmed, meaning General Compute draws more only as it signs paying tenants for the capacity. That is a very different animal from a lump-sum term loan. The lender is underwriting a business that grows into the money, not a fixed pile of hardware bought on day one.

The collateral is SambaNova’s SN40 and SN50 inference accelerators, chips designed to run already-trained models rather than train new ones. General Compute says it has more than $300 million of secured, price-protected supply of that silicon lined up. Its serving stack pairs the SambaNova parts, which handle the decode phase where tokens are generated one after another, with AMD’s MI300X for the prefill phase where a prompt is first ingested. The company has also lined up the option to buy 15 megawatts of air-cooled rack capacity in colocation facilities, which matters because these inference parts run without the water cooling that frontier GPU clusters increasingly demand.

The performance pitch is aggressive. General Compute claims its setup delivers results 16 times faster than current GPUs, with 7x faster time-to-first-token and 8.5x higher output throughput than standard GPU clouds. Those are vendor numbers, not independently benchmarked, and they describe the product rather than the loan. The loan rests on something quieter: the belief that if this all falls apart, the chips can be sold, redeployed, or repossessed at a value close to what was borrowed against them.

Why a lender is willing to call inference silicon bankable

Billy Libby, Upper90’s chief executive, has a specific reason for treating these parts as an asset rather than a depreciating expense. He financed Nvidia GPUs when nobody else would, and he frames inference chips as the next inefficiency to exploit. In the announcement he points to power: SambaNova’s silicon, he says, is up to six times more power efficient than traditional GPUs. Efficiency is a proxy for staying power. A chip that produces more tokens per watt keeps earning its rack space even as newer parts arrive, because the binding constraint in most data centers is electricity, not the age of the accelerator.

There is a cleaner version of this argument buried in how training and inference hardware age differently. A GPU cluster bought to train one frontier model can be functionally obsolete for that job inside eighteen months, because the next model wants more memory bandwidth and a bigger interconnect. Inference is more forgiving. Once a model is trained, serving it is a steadier workload, and a chip that serves it economically this year will usually still serve it economically next year. That relative stability is what lets a lender squint at an inference accelerator and see something closer to an income-producing asset than a melting ice cube.

This is not a brand-new idea so much as a migration of one. CoreWeave turned GPU-backed borrowing into an entire business model, financing Nvidia hardware against customer contracts and riding that structure to a public listing. TECHi has tracked how much weight that debt places on the backlog holding up, an anxiety laid out in the analysis of CoreWeave’s backlog and debt test. The General Compute deal takes the same collateral logic and points it at a different chip. What changed is not the financial engineering. It is the willingness to accept silicon that does not carry Nvidia’s name or Nvidia’s resale market.

The collateral question nobody is pricing yet

Here is where the marketing and the mechanics separate. A Nvidia H100 or MI300X is close to a commodity. If a neocloud defaults, a lender can pull the cards, and there is a deep, liquid, global market of buyers who want that exact part at a knowable price. That secondary market is the real collateral. The chip is bankable because thousands of other operators will pay cash for it tomorrow.

A SambaNova SN50 does not have that market, at least not yet. It is a purpose-built accelerator sold and supported by one company, tuned for one serving pattern, and deployed by a small number of operators. If General Compute stops paying, what Upper90 repossesses is a warehouse of specialized parts whose resale value depends on finding another buyer who runs the same stack, wants the same architecture, and trusts the same vendor to keep shipping firmware and replacements. The number of such buyers is not large. The collateral is only as liquid as the ecosystem around it, and that ecosystem is early.

So the loan is not really secured by resale value in the way the framing implies. It is secured by utilization. As long as the chips are serving paying customers, the facility performs and the collateral question stays theoretical. The structure even reinforces this, since the money scales with demand rather than landing upfront. But that also means the security and the revenue are the same bet. In a stress scenario where customer demand for General Compute’s premium tokens softens, the chips lose their tenants and their resale market at the same moment, because both depend on the same thin pool of operators wanting this specific silicon. Nvidia collateral decouples those risks. Single-vendor inference collateral fuses them.

Vendor concentration cuts the other way too. The debt is implicitly a bet on SambaNova as a going concern. If the chipmaker stumbles, gets acquired, or shifts its roadmap, the value of a warehouse full of its accelerators moves with it. That is a risk a basket of Nvidia GPUs simply does not carry, because Nvidia’s install base is too large to orphan.

None of this makes the deal reckless. Lenders price these gaps for a living, and a facility that meters out cash against signed demand is already doing a lot of the work. But it does mean the headline framing — inference chips are more bankable than GPUs because they depreciate slower — is only half the sentence. Slower depreciation helps if there is a market to sell into. Without one, a slowly depreciating asset that nobody wants to buy is still worth very little in a hurry. The value that protects Upper90 is the stream of tokens the chips produce, and the resale story is a backstop that has never actually been tested at this scale.

What it signals for the neocloud economy

Strip away the collateral debate and the deal still says something real about where capital is flowing. Finn Puklowski, General Compute’s chief executive, called the financing the first signal of capital organizing itself around the fragmenting of Nvidia’s dominance. That is self-serving, but it is not wrong in direction. Money follows conviction, and a lender accepting non-Nvidia silicon as security is a small, concrete vote that the inference layer is diversifying away from a single supplier.

The broader neocloud trade has been built almost entirely on Nvidia so far, as TECHi’s look at the neocloud stocks powering the AI cloud market laid out. Alternatives are starting to get funded on their own terms. General Compute’s own decision to split serving between SambaNova and AMD echoes the disaggregation push that surfaced when AMD and Cerebras split inference into separate phases, each chip doing the part it is best at rather than one GPU doing everything. Inference is being unbundled at the hardware level, and now the financing is following that unbundling.

Pricing opacity remains the awkward part of this whole layer. It is hard to underwrite an asset whose list price nobody publishes, a problem visible even at the largest scale, where Google’s Ironwood TPU still has no public price months after launch. Custom inference silicon is powerful and, increasingly, financeable. It is also harder to value on the open market precisely because so little of it trades on the open market.

The risk map, stated plainly

This can work. If General Compute signs enough paying customers, the facility never draws past what its revenue supports, and inference demand keeps climbing, then Upper90 has found a genuinely underpriced asset class and everyone else will pile in behind it. The scale-with-demand structure is a real safeguard, because it stops the borrower from loading up on hardware it cannot fill. That is the optimistic read, and it is plausible.

The thesis breaks if two things move together: inference demand for premium, large-model serving cools, and a new chip generation leaps far enough ahead to make the SN50 look slow rather than efficient. In that case the collateral loses its tenants and its resale market at once, and a lender discovers that single-vendor inference silicon is worth far less in a fire sale than a spreadsheet assumed. The number that matters is not the $400 million headline but whether General Compute actually draws the facility up toward the cap, because a draw only happens when real customers show up. A committed line that stays mostly undrawn would say the demand is not yet there, no matter how fast the chips run.

For now, the deal is best read as an experiment in what counts as collateral in the AI economy, priced by a lender who has been early before. It is not proof that inference chips are the new gold. It is proof that at least one firm is willing to bet they might be.