USD.AI
Financing the Compute Middle Market: Conor Moore on theCUBE + NYSE Wired
USD.AI Team

Financing the Compute Middle Market: Conor Moore on theCUBE + NYSE Wired

USD.AI co-founder and COO Conor Moore joined John Furrier on theCUBE's AI Factory series at the New York Stock Exchange to discuss GPU financing, the rise of distributed inference, and the financial infrastructure needed to support the next phase of AI growth.

The AI infrastructure buildout is often measured in gigawatts and multibillion-dollar deals. Beyond the hyperscalers, a much broader market is emerging: AI startups, inference providers, and enterprises investing in their own compute. These companies need access to capital, yet many fall outside the lending structures designed for the largest AI infrastructure projects.

During the interview, Conor Moore explained how USD.AI is addressing this gap by standardizing GPU-backed financing and connecting smaller infrastructure deployments with institutional capital.

The Compute Middle Market Is Underserved

Much of the capital flowing into AI infrastructure today is concentrated in investment-grade transactions backed by major technology companies. For lenders, these deals are relatively straightforward, because the creditworthiness of a large counterparty such as Microsoft or Meta provides the foundation for underwriting.

Conor pointed out that only about seven companies operate at hyperscaler scale, while hundreds, if not thousands, of AI startups, inference providers, and enterprises are purchasing compute themselves. Their clusters are smaller: in Conor's words, "when I say small, what I'm talking about is $50 to $150 million." These are substantial investments, but relatively small compared to the largest AI infrastructure transactions, and the Compute Middle Market tracker follows GPU-backed financings of $500 million or less across this segment.

Traditional lenders often lack the underwriting frameworks needed to evaluate these businesses. Established credit markets typically price investment-grade and non-investment-grade borrowers differently, but in compute financing, as Conor described it, "capital is available, capital is not available." USD.AI is working to change that by making GPU-backed financing more standardized and easier for institutional lenders to evaluate.

Standardization Makes Smaller Deals Financeable

A $50 million GPU deployment can be too small to attract the attention of major institutional asset managers on its own, but a portfolio of similarly structured transactions presents a different opportunity. Conor explained that institutional lenders managing tens of billions of dollars across multiple industries cannot underwrite every small deployment from scratch, so they need consistent structures that can be evaluated using a repeatable framework.

"If I know that every deal looks exactly the same, I underwrite the heuristic of how they're structured and then I can deploy capital."

USD.AI is developing that standardized approach. In November 2026, Conor shared that USD.AI had evaluated approximately 700 deals representing $25 billion in assets, which gives it market data on GPU pricing, compute rental rates, and transaction structures.

Compute IQ, USD.AI's agentic underwriting system, scores every deal against published underwriting standards, so operators financing through USD.AI can follow established criteria from the start. Conor also shared that the process can move from an initial conversation to a closed financing in under 60 days, with the goal of shortening that timeline further.

Looking ahead, he sees asset-backed securitization (ABS) as an important next step. Pooling standardized GPU financings into larger products would let the market attract a wider range of institutional investors and lower the cost of capital. He also expects the market to start preferring smaller deals over the next year or two, because they are less complicated at the project level.

Distributed Inference Changes the Economics of AI Infrastructure

Unlike large-scale training clusters, inference workloads do not need a single massive data center. Smaller deployments can be distributed across multiple locations, bringing compute closer to the products and customers using it, which matters for latency-sensitive uses such as robotics.

Conor illustrated the economics with a simple example: a one-megawatt GPU deployment represents approximately $50 million in hardware investment. That is a relatively small infrastructure project that still requires substantial financing, and the space for it is often available in a corner of an existing colocation facility.

Smaller deployments also make supply and demand easier to manage. A gigawatt-scale data center can take around 36 months to come online, making it difficult to predict market conditions when the capacity arrives. By comparison, Conor described how a one-megawatt deployment can move from a signed term sheet to operating capacity in about two months.

"If that 1 megawatt is the straw that breaks the camel's back in terms of supply demand, that's unfortunate, but that's a $50 million mistake. That's not a $50 billion mistake."

Smaller projects let operators add capacity incrementally, respond to demand more quickly, and limit the exposure tied to any single deployment. For lenders, a diversified portfolio of smaller projects also improves pricing. As Conor put it, "diversity buys me discounted interest rates."

From Neoclouds to Enterprise-owned Compute

Today, neoclouds are among the primary customers for GPU financing, but Conor expects the market to expand as more enterprises purchase and operate their own compute. The growth of open-source AI models makes this increasingly practical, since companies can run existing ones on their own hardware while keeping control over their data and intellectual property.

That control matters most for businesses whose value depends on proprietary information. Conor pointed to law firms as an example: AI is particularly good at working with legal contracts, and a firm with a long operating history has strong reasons to keep that information away from the AI labs. He expects enterprise compute ownership to become increasingly common, with deployments sized to each company's needs. In his words, enterprises "aren't going to spend $10 billion," and their budgets look closer to $10 million.

The financing market will need to accommodate that shift, and Conor compared where it is heading to the largest consumer credit markets.

"I think it looks something like Fannie Mae or auto loan securitization at scale. Everyone will own their own GPUs."

Building the Financial Infrastructure for Compute

Throughout the conversation, Conor emphasized that standardization is a prerequisite for a more mature compute market. He pointed to Standard Oil as a historical example: once the oil market was standardized, products such as futures and insurance could be built on top of it.

Today's GPU financing structures largely rely on offtake contracts that match the term of the debt, with fixed payments on both sides, which makes transactions understandable to existing private credit investors. Over time, Conor expects the market to develop more flexible products that reflect the stronger unit economics of short-duration compute, most likely something closer to a profit share than traditional debt.

As lenders gain experience with GPU-backed assets and transaction performance becomes easier to measure, the market should be able to price, finance, and manage compute-related risk with more precision. GLX, USD.AI's institutional marketplace, is part of that foundation: once a financing is seasoned and performing, it may be sold to an institutional buyer, and the capital returns to fund new deployments. USD.AI's goal is to help build this infrastructure, starting with the Compute Middle Market.

Frequently Asked Questions

What is the Compute Middle Market?

The Compute Middle Market is the segment of GPU cloud operators between startups and hyperscalers, including neoclouds, inference providers, and enterprises with real customer demand that fall below hyperscaler scale. USD.AI's tracker counts GPU-backed financings of $500 million or less in this segment.

Why is it hard for smaller GPU operators to get financing?

Much of the capital in AI infrastructure is underwritten against investment-grade counterparties such as Microsoft or Meta, so many lenders lack a framework for evaluating smaller, non-investment-grade operators. As Conor Moore put it on theCUBE + NYSE Wired, for these businesses "capital is available, capital is not available."

How does standardization help GPU financing?

Standardization lets institutional lenders underwrite one repeatable structure and apply it across many similar deals, which makes $50 million deployments practical to finance at scale. Over time, a diversified pool of standardized financings can also support asset-backed securitization and a lower cost of capital.

Why does AI inference favor smaller, distributed GPU deployments?

Inference workloads can run across many locations, so they do not depend on one massive data center the way large training runs do. A one-megawatt deployment, roughly $50 million of GPUs, often fits in an existing colocation facility, can go live in about two months, and sits closer to the products and users it serves, which matters for latency-sensitive uses such as robotics.

Why are smaller GPU deployments less risky than gigawatt-scale data centers?

Smaller deployments come online faster and cost less, so a misjudgment about supply and demand has a far smaller impact. A gigawatt-scale data center can take around 36 months to build, while a one-megawatt deployment can be serving customers in about two months, or as Conor Moore put it, "that's a $50 million mistake. That's not a $50 billion mistake."

Watch the full interview with Conor Moore on theCUBE + NYSE Wired's AI Factory series.

Financing a GPU deployment, or looking to put capital to work in compute? Talk to USD.AI.

Earn from AI infrastructure Finance your next GPU deployment