General Compute secures $400 million debt line for AI inference cloud

General Compute Inc. said it has secured $400 million in debt financing to expand its AI inference cloud, with investment firm Upper90 underwriting the facility. Upper90 will fund an initial $100 million, while General Compute plans to draw additional capital as demand from customers rises.

The company did not disclose the debt terms, maturity, interest rate, valuation, revenue, headcount, customer names or current deployed capacity. The structure matters because AI infrastructure companies have been using debt to buy or lease expensive compute without pricing a new equity round, a model that works only if utilization and customer commitments keep up with hardware costs.

A non-Nvidia inference stack

General Compute is building its cloud around chips from Advanced Micro Devices Inc. and SambaNova Inc., rather than Nvidia Corp. GPUs. SambaNova, another AI chip startup, recently raised $1 billion, according to SiliconANGLE.

The company splits inference work across two types of hardware. AI systems process a prompt in a prefill phase, then generate an answer token by token during decoding. General Compute uses AMD’s MI300X GPU for prefill and SambaNova accelerators for decoding.

That architecture is tied to SambaNova’s chip design. Decoding requires frequent data movement between memory and compute circuits. SambaNova’s accelerators place memory and processing circuits close together, which the company says reduces data travel time and can make decoding faster than some Nvidia GPUs.

The AMD MI300X, used for the prefill side of the workload, includes 12 dies and eight memory stacks with a combined 153 billion transistors. General Compute’s bet is that pairing AMD silicon with SambaNova chips can deliver inference capacity without joining the same Nvidia supply queue as larger cloud providers and AI labs.

Air-cooled racks are part of the cost argument

General Compute also says the AMD and SambaNova hardware can run in conventional air-cooled server racks. That is a relevant operational claim because many high-density GPU clusters require liquid cooling, which can add time and capital expense to data center buildouts.

The company says avoiding liquid cooling lets it expand infrastructure faster. It has not disclosed comparative deployment timelines, unit economics or pricing for its services, so the cost advantage remains a company claim rather than a proven operating metric.

General Compute sells access to its inference hardware through two products. One provides managed open-source language models through an API designed to resemble OpenAI Group PBC’s interface. The other gives customers dedicated infrastructure environments with more compute capacity.

According to a document on General Compute’s website, the company has agreements that give it the option to buy 15 megawatts of air-cooled rack capacity in colocation facilities. General Compute expects that capacity to cover its fourth-quarter growth plans.

The financing gives General Compute more room to buy capacity in a market where inference demand is rising, but it also raises the operating bar. Debt-funded compute providers need steady customer usage to service obligations, and General Compute has not disclosed the customer commitments behind the facility.

This story draws on original reporting from SiliconANGLE.