Preface

General Compute, an AI inference cloud startup, has secured a $400 million loan from Upper90, a technology-focused investment firm. This transaction is notable because it appears to be among the first where inference-specific chips — designed to run already-trained AI models efficiently rather than to train them — are being used as collateral. The deal underscores a broader market response to the rising cost of AI systems: investors and operators are increasingly focused on infrastructure that can deliver affordable, high-performance inference for open-source and lower-cost models, rather than exclusively pursuing the most expensive training GPUs.

Lazy bag

Key takeaway: inference chips are emerging as a practical, cost-efficient collateral class for financing AI infrastructure. Upper90’s loan to General Compute highlights a market pivot toward hardware optimized for running models — which can be deployed more broadly and cheaply than water-cooled, training-focused GPUs.

Main Body

General Compute, founded by CEO Finn Puklowski and CTO Jason Goodison, is building an inference-focused neocloud using hardware from SambaNova, a chipmaker backed by Intel. Neoclouds are specialized cloud environments designed specifically for AI workloads, in contrast to generalized offerings from hyperscalers like AWS or Azure. The SN50 chips that General Compute is adopting are optimized for inference: they are power-efficient, avoid the complexity of water-cooling, and can be installed across a wider range of data centers. According to the company, these chips deliver inference throughput up to 16 times faster than traditional GPU-based clouds, which could make low-latency, cost-effective access to models vastly more accessible.

The fundamental challenge for a new entrant is acquiring enough of these chips quickly. Upper90’s involvement reflects an evolving financing playbook. Billy Libby, Upper90’s co-founder and CEO and a former quantitative trader at Goldman Sachs, previously financed GPU acquisitions for Crusoe in 2021 — a move that he regards as an early example of loans backed by advanced compute hardware. At the time, many conventional lenders avoided such deals because of uncertainties around GPU depreciation and market risk.

Over the past few years, however, the market has matured. Firms like CoreWeave turned chip-backed lending into a repeatable business model and then into a high-profile public offering, which helped normalize loans secured by compute inventory. Libby explained that being an early participant in GPU-backed finance allowed Upper90 to benefit from market inefficiencies. Now, seeing GPUs become well understood and in some cases overbought, the firm is pivoting to the next opportunity: financing inference-centric hardware for companies positioned to run open-source and cost-conscious models at scale.

This strategy reflects a broader industry shift. Rising concern about the price of AI models and access to specialized tokens has pushed attention toward infrastructure that runs open-source models more cheaply than the latest closed, frontier LLMs. Access points and platforms supporting open models, such as OpenRouter and Fireworks, have attracted significant investment. New models like Kimi’s K3 have demonstrated competitive performance on certain benchmarks relative to offerings from Anthropic and OpenAI, further validating demand for alternatives. Simultaneously, novel chipmakers — including Groq and Cerebras — have drawn strategic interest from acquirers and public markets, adding momentum to the diversity of available silicon.

General Compute’s decision to source chips beyond Nvidia’s ecosystem is consequential. Diversifying beyond a single dominant supplier can lower costs and reduce vendor lock-in. TensorWave, another infrastructure provider, is making a comparable bet through a partnership with AMD. As more non-Nvidia chips reach scale, providers that can adopt multiple architectures may gain a competitive edge in delivering cost-effective inference services.

Puklowski emphasized the potential for these newer processors to offer attractive total cost of ownership or performance advantages. Yet, he noted, demand-side adoption trails supply: many of these chips are coming to market, but buyers remain relatively few. That imbalance creates an opening for financiers and infrastructure startups to organize capital and deployment around inference hardware. By partnering with Upper90, General Compute’s financing is more than a straightforward capital infusion to purchase compute — it represents an early sign that investment capital is reorganizing to support a more fragmented, multi-vendor compute landscape and to challenge Nvidia’s dominant position.

The implications are practical. Inference-optimized chips that avoid complex cooling and electrical requirements can be hosted in a broader range of colocation facilities, accelerating time-to-deployment and expanding geographic reach. For customers, that can translate into lower latency, reduced costs, and wider availability of model-serving endpoints. For lenders and investors, chips designed for inference represent a different risk/return profile than training GPUs: they may depreciate differently and be easier to repurpose for operational workloads, making them a more attractive collateral class for certain financing structures.

Ultimately, Upper90’s $400 million loan to General Compute signals that capital markets are increasingly comfortable underwriting specialized AI infrastructure when paired with a coherent operational plan. As open-source models and alternative silicon gain traction, financing models that once focused on training GPUs are evolving to support the hardware that powers inference — the day-to-day execution layer for practical AI products. If this trend continues, expect more deals where inference hardware is the central asset securing growth capital, helping to broaden access to fast, affordable AI services outside the narrow confines of the most expensive training stacks.

Key Insights Table