ALITEQ.

a startup just landed a $400 million loan and there's not a single Nvidia GPU behind it

General Compute just got the first-ever loan collateralized by AI inference chips instead of GPUs — a quiet signal the market is betting beyond Nvidia.

Ravi MalhotraUpdated 1h ago6 min readWeb story
Rows of AI inference servers in a data center rack

General Compute just secured a debt facility from lender Upper90 that starts at $100 million and scales up to a $400 million ceiling — and for the first time, the collateral backing a major AI infrastructure loan isn't a stack of Nvidia GPUs. It's SambaNova's SN50 chips, purpose-built silicon that does one job: running already-trained AI models fast and cheap. That's the actual story here, not the dollar figure — what lenders are now willing to bet on.

$400M

Loan ceiling

Scales up from an initial $100M facility

$100M

Starting size

Where the Upper90 facility begins

$15M

Prior seed round

General Compute's raise in May 2026

16x

Claimed speed edge

SN50 vs. typical GPU-based inference clouds

The deal, in plain terms

Who's involved

Borrower

Info
General Compute — founded by CEO Finn Puklowski and CTO Jason Goodison

Lender

Info
Upper90, led by co-founder and CEO Billy Libby

Collateral

Info
SambaNova SN50 inference chips

Loan structure

Info
Starts at $100M, scales to a $400M ceiling

Prior round

Info
$15M seed, May 2026

Why chips-as-collateral is actually the story

Lenders have historically stuck to financing Nvidia GPUs and avoided everything else, because non-Nvidia AI silicon depreciates unpredictably and has almost no resale market if a borrower defaults. Billy Libby knows that history better than most — he was among the first financiers to lend against Nvidia GPUs at all, backing Crusoe back in 2021. That he's now willing to make the same bet one rung down the stack, on inference-specific chips instead of general-purpose GPUs, is the actual signal in this deal. It suggests lenders increasingly see inference — the part of AI that runs every time someone actually uses a model — as durable enough to underwrite on its own, separate from the training compute that gets most of the headlines.

The GPU price angle this connects to

This isn't happening in isolation. Nvidia itself has leaned on its own roughly $500 billion GPU collateral arrangements to keep chip financing flowing, consumer GPU prices have jumped as much as 30% this year as DRAM and memory shortages ripple through the whole supply chain, and GPU rental prices have kept climbing even as AI stocks sold off 40-60%. General Compute's deal is a variation on the same theme, one level removed: capital is still chasing compute, it's just increasingly willing to chase compute that isn't Nvidia's. If you're pricing out your own hardware for local AI work, this is the same capital squeeze showing up from a completely different direction.

Close-up of a computer processor chip on a circuit board
Inference-specific chips are built to run trained models cheaply at scale — a narrower job than the GPUs used to train them. · Unsplash

The claimed speed gap

Typical GPU inference cloud1x baseline

Standard GPU-based inference throughput

SambaNova SN50 (claimed)16x faster

General Compute's claimed inference speed edge

The risk nobody's saying out loud

Every party in this deal has an obvious incentive to talk up how safe it is. What gets said less is that a specialized inference chip has essentially no market outside AI infrastructure. If the GPU trade sours, a Nvidia card can still find a buyer in gaming or crypto mining. An SN50 chip, if SambaNova's ecosystem doesn't keep pace with Nvidia's software moat, has nowhere else to go.

What to watch next

What is General Compute's $400 million deal?
A debt facility from investment firm Upper90 that starts at $100 million and scales to a $400 million ceiling, collateralized by SambaNova's SN50 inference chips rather than Nvidia GPUs.
What makes this the first deal of its kind?
Lenders have historically avoided financing non-Nvidia AI silicon because it depreciates unpredictably and has a thin resale market. This is reportedly the first loan to use inference-specific chips as primary collateral.
What's the difference between training chips and inference chips?
Training chips, typically Nvidia GPUs, build AI models from scratch. Inference chips like SambaNova's SN50 run already-trained models as cheaply and quickly as possible — a narrower, more specialized job.
Is SambaNova the same as Nvidia?
No. SambaNova is an Intel-backed chipmaker building silicon specifically for inference. Its SN50 chips claim up to 16x faster inference than typical GPU-based clouds, without needing water cooling.
Does this affect gaming GPU prices?
Not directly — SN50 chips serve AI data centers, not gamers. But it's part of the same broader story: capital and manufacturing capacity increasingly chasing AI compute, which is a real contributor to this year's consumer GPU price hikes.

Worth watching over the next few months: whether other lenders follow Upper90 into inference-chip-backed debt, and whether SambaNova's SN50 numbers hold up once General Compute's customers are running production workloads instead of a tech preview. If they do, this becomes the template for financing the next wave of non-Nvidia AI hardware. If they don't, it becomes a cautionary tale about how fast 'the next big collateral' can turn illiquid.

Hardware Editor

Ravi Malhotra

Ravi has been building and taking apart PCs since the single-core days — his idea of a good weekend is a repaste and a spreadsheet full of thermals. He covers GPUs, CPUs and the build decisions that actually move frame rates, and he'd rather hand you a benchmark than a press release.

Work out the hardware

The Aliteq brief

The tech worth knowing — hardware, AI, gaming, deals. No spam, unsubscribe anytime.

Keep reading