Investors and specialty lenders are beginning to reallocate capital toward infrastructure that runs open-source AI models more cheaply than the latest large language models from frontier labs. A recent financing arrangement for General Compute, which plans an inference-focused neocloud based on SambaNova silicon, highlights how chip-backed loans and non‑Nvidia hardware are becoming central to that effort.
General Compute’s neocloud and the promise of SN50 silicon
General Compute, founded by CEO Finn Puklowski and CTO Jason Goodison, secured a $15 million seed round in May to build a neocloud tailored for inference workloads. The company’s chosen engine is SambaNova’s SN50 chip, a processor designed specifically for inference rather than training. According to the company, the SN50s are power-efficient and avoid the need for expensive water-cooling systems, enabling faster deployment across a wider range of data centers than many GPU-based solutions.
General Compute has also said the new chips can deliver substantially faster inference than traditional GPU clouds, a claim the company quantifies as a 16x improvement in inference speed versus GPU-based offerings. That performance profile is central to the startup’s pitch: most customers don’t need a supercomputer, they need low-latency, cost-efficient inference.
Chip-backed financing gains traction with lenders
Securing significant quantities of alternative chips is difficult for a new entrant. Upper90, a specialist lender co-founded by Billy Libby, has been among the firms willing to step into that gap. Libby previously financed GPU purchases for Crusoe in 2021, a deal he views as early evidence that chips could serve as collateral despite earlier reluctance from traditional banks.
At the time, lenders avoided such loans because the depreciation patterns and market dynamics for high-end accelerators were uncertain. As the market matured — with companies such as CoreWeave turning chip-backed lending into a repeatable business and then pursuing an IPO — the approach became more mainstream. Upper90 has since expanded that playbook beyond Nvidia GPUs to newer architectures and companies like General Compute that base their deployments on other vendors’ silicon.
Libby says Upper90 intentionally sought out businesses focused on inference for open models. In his view, broad demand for inference capacity makes these loans a natural fit: not every organization requires top-tier training clusters, but many need affordable inference to run models in production.
Alternatives to Nvidia and a more fragmented compute landscape
The General Compute financing is part of a broader market shift. A growing ecosystem of companies is building around open-source models and non‑Nvidia hardware. Access providers and platforms catering to open models, such as OpenRouter and Fireworks, have attracted funding, while new models like Kimi’s K3 have demonstrated competitive performance on tasks such as coding benchmarks.
At the same time, rival chipmakers — including Groq, Cerebras and SambaNova — are drawing investor and acquirer interest. Some infrastructure providers are forging partnerships with AMD or other vendors; for example, TensorWave has focused on an AMD-based proposition. These moves indicate that alternatives to Nvidia are beginning to scale and that total cost of ownership calculations may favor non‑GPU accelerators for certain inference workloads.
General Compute’s CEO Puklowski framed the Upper90 deal as more than a typical startup equipment loan. He said the transaction signals capital organizing to support a broader fragmentation of what has been a largely Nvidia-dominated market. For companies building inference platforms around alternative silicon, access to capital that understands chip economics is a crucial enabler.
Implications and near-term challenges
The immediate implication is straightforward: financing models that allow startups to acquire large quantities of non‑Nvidia accelerators lower the barrier to entry for differentiated inference services. Power-efficient chips that avoid complex cooling can be sited in more locations and may reduce deployment lead times and operating costs.
But challenges remain. New hardware ecosystems still need buyers, and manufacturers must scale supply chains. Lenders face residual risk around depreciation and secondary markets for non‑Nvidia accelerators, and startups must prove sustained demand for inference capacity built on alternative chips.
Regardless, the Upper90–General Compute arrangement points to an evolving capital market: one that increasingly recognizes value in purpose-built inference infrastructure and in the economics of open-source models. As more lenders and investors get comfortable underwriting chip-backed deals beyond GPUs, the market for AI compute could diversify in ways that reshape access, costs and the competitive landscape for inference services.