General Compute Secures $400 Million to Scale a GPU-Free Inference Cloud

General Compute, a startup that runs AI inference workloads entirely on non-Nvidia chips, has secured a committed debt facility of up to $400 million from Upper90 Capital Management. The facility starts at $100 million and expands as customer demand grows, according to a joint announcement from the two companies. It is one of the first large loans structured around inference-specific chips as collateral, rather than the general-purpose GPUs that have backed most AI infrastructure financing to date.

A Loan Built Around a Different Kind of Chip

General Compute describes itself as a “neocloud,” industry shorthand for a cloud provider built specifically for AI workloads rather than the general-purpose computing that AWS, Azure, and Google Cloud were originally designed around. What sets General Compute apart from earlier neoclouds like CoreWeave is its hardware choice: instead of buying Nvidia GPUs, it runs production workloads on inference chips from SambaNova, specifically the SN40 and SN50 lines.

Upper90 is both the new loan’s lender and an existing equity investor in General Compute, a structure the firm says aligns its debt and equity behind the company’s growth. General Compute says it already has more than $300 million of secured, price-protected chip supply lined up, and expects to be the first neocloud to deploy this class of ASIC (application-specific) hardware at scale.

Training Chips Versus Inference Chips

The deal turns on a distinction between two jobs a data center chip can do. Training chips, the expensive Nvidia GPUs that dominate headlines, are built to teach a model by crunching through enormous datasets. Inference chips do a different job: taking a model that has already been trained and using it to answer a live query, the step that happens every time someone sends a prompt to a chatbot or an AI agent.

General Compute and Upper90 say that inference step is where GPU-based clouds struggle. The newest GPU racks can draw more than 120 kilowatts of power and need specialized liquid-cooling retrofits, tying new capacity to multi-year data center construction timelines. General Compute’s SambaNova-based racks run on air cooling at about 20 kilowatts per rack and can be installed in existing colocation facilities in weeks instead of years.

The companies say the resulting performance gap is substantial: General Compute’s inference cloud delivers 16 times faster inference performance and 7 times faster time-to-first-token than standard GPU clouds, with 8.5 times higher output throughput, reaching 1,000 tokens per second, while using about 6 times less energy than current GPUs.

What SambaNova’s SN50 Chip Adds

SambaNova’s SN50, the newer of the two chip generations General Compute runs, is the company’s fifth-generation inference processor. According to SambaNova’s own product pages, the SN50 delivers 5 times more compute and 4 times more network bandwidth than the prior SN40 generation, using a three-tier memory system, SRAM, HBM, and DDR5, built to keep several models loaded at once and switch between them quickly.

A single cluster can combine up to 256 of SambaNova’s chips (the company calls each one an RDU, for reconfigurable dataflow unit), and a full scaleout deployment can link as many as 32,768 of them, a configuration SambaNova says is built to run models as large as 10 trillion parameters with context windows of up to 10 million tokens. SambaNova has its own financial momentum behind that hardware: its website states the company has completed the first close of a $1 billion financing round at an $11 billion valuation.

The Same Playbook That Financed Crusoe

Upper90 co-founder and CEO Billy Libby, a former Goldman Sachs quantitative trader, has run a version of this deal before. In 2021, Upper90 financed GPU purchases for Crusoe, the energy-focused data center startup, in what Libby believes was the first loan made against the value of advanced chips. Traditional lenders avoided that kind of deal at the time because of uncertainty over how fast GPUs would depreciate.

“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.”

That risk premium shrank once CoreWeave built chip-backed loans into a business model that carried it to a landmark public offering, and the practice became common across the industry. Upper90 is now applying the same underwriting approach to a class of chips most lenders have not financed before. “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.”

Why Lenders Are Looking Past Nvidia

General Compute co-founder and CEO Finn Puklowski frames the deal as a signal about where AI spending is headed, not just a funding round for his own company. “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 told TechCrunch. “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.”

In the companies’ announcement, Puklowski added: “We are the only neocloud that can serve premium tokens: frontier-level intelligence on the largest models, served fast. That is where the workloads are moving. Because we are new, we are not handcuffed to a single chip supplier. This facility lets us deploy the fastest inference silicon into existing data centers and meet that demand now.”

The broader trend, according to TechCrunch’s reporting, includes rising valuations for open-model infrastructure providers such as OpenRouter and Fireworks, new open models like Kimi’s K3 that reportedly compete with the latest releases from Anthropic and OpenAI on coding benchmarks, and growing investor interest in alternative chipmakers such as Groq and Cerebras. TensorWave, another AI infrastructure company, is pursuing a similar strategy built around a partnership with AMD. General Compute’s own announcement cites a demand forecast from Goldman Sachs projecting that AI token consumption will grow 24-fold over the next three and a half years.

What to Watch Next

General Compute closed a $15 million seed round in May, backed by Upper90 among other investors, before this much larger debt facility. Three things will show whether this financing model spreads beyond one deal: whether General Compute’s supply commitments with SambaNova hold up as customer demand grows, whether other lenders extend similar chip-backed credit to non-Nvidia hardware, and whether rivals betting on other chipmakers, like TensorWave’s AMD-based approach, can match General Compute’s deployment speed. For now, the deal is a concrete data point in a broader shift: as inference, rather than training, becomes the larger and more persistent AI compute cost, financiers are starting to underwrite the chips built specifically for that job.