10,000 Neoclouds: Building the Financial Network for Compute USD.AI has reviewed more GPU financing opportunities than almost any other organization in the world. This year to date (October 13th), USD.AI has reviewed nearly 700 deals totalling $25bn worth of loan demand, and is currently scaling its originations volumes with the help of GLX to meet the capital needs of borrowers. Just for reference here’s the same pipeline chart, released only 3 months ago in the Lighthouse YTD Report. Across the board, USD.AI has seen a massive increase in inbound demand. Since USD.AI launched, the borrower mix has shifted twice. In late 2025, most of the inbound was Bitcoin miners turned AI providers. This year most of the demand has come from general-purpose inference sellers, aka neoclouds, looking to scale or start up. Now a third group is emerging: law firms, biotech labs, research groups, or any companies buying GPUs for their own use to stay strategic with intelligence across data controls to spend management. These companies have hit the scale where owning beats renting. That’s what the USD.AI pipeline shows, and it appears to reflect a wider trend across the industry, as compute becomes a core input to how these companies work, rather than a product they sell. These are likely the borrowers of the future, and the clearest picture of the Compute Middle Market. These end users are quasi-neoclouds. They run chips to serve inference, and they are also the core offtaker. USD.AI’s bet is that there will be 10,000 of these. Not 10,000 copies of today’s neoclouds, but 10,000 companies that own the compute their business runs on. Each is too big to keep renting and too small for anyone on Wall Street to notice. GPU capital markets aren’t built for them. They aren’t hyperscalers; they aren’t building gigawatt-scale deployments with 100,000 GPUs. They need inference, and their biggest problem is the friction around financing a deployment that the market considers small. Small is relative: these deals are $25 million to $250 million, which is a lot of money everywhere except in a GPU capital market sized for hyperscalers. So USD.AI was rebuilt around one mission: making compute loans cheaper and faster to make, move and settle. The rest of this post walks through what that mission changes, from who the borrowers are now and why the bottleneck is credit friction rather than capital, to how the business moves from holding loans to moving and settling them, and finally what it all means for the protocol and sUSDai holders. I. The borrower patterns are changing: 10,000 Neoclouds The Compute Middle Market: from 300 neoclouds to 10,000 borrowers The first wave of neoclouds came out of bitcoin mining, because miners had the one thing AI needed most and couldn’t get quickly: power. The second wave was the general-purpose inference sellers that make up the largest part of the USD.AI pipeline today, companies built from the start to sell GPU capacity, and helped along by modular data centers that make power easier to reach and faster to bring online. Between them, they lowered the bar for running a cluster far enough that a third wave is now showing up in USD.AI’s inbound: end users. So far they are mostly AI-native software and product companies, but the conversations underway reach well beyond that, to law firms, biotech companies, materials science labs and AI safety teams. Companies whose product depends on compute, deciding to own it. For an AI-native company, compute is often the biggest line on the P&L, and when rent is that expensive, owning with a loan starts to look like buying a house instead of renting one. A company running open-weight models on its own hardware also keeps its data in-house, which for a law firm or a lab is not a small consideration. That is the Compute Middle Market: a few hundred neoclouds today, typically borrowing $10-$100 million at a time. Further down the line, there will be 10,000 end-user borrowers at $5-$50 million each. After that come the microclouds, the 100,000 companies for whom a single rack is a cluster. II. Power isn't the bottleneck, it’s the cost-per-loan. The fixed-cost problem in compute credit People assume compute financing is limited by power, or by what a GPU will be worth in three years. Both of those matter, but the actual binding constraint is the cost of making a loan. There is roughly $2 trillion in private credit looking for yield, across banks, credit funds, insurers and asset managers. The problem is getting it to a $20 million borrower at a cost that makes sense. Today a bespoke GPU loan costs $5 million to $7 million to put together, across origination, underwriting, legal, structuring, settlement and distribution. That cost stays about the same whether the loan is $500 million or $20 million, because each deal is diligenced from scratch and negotiated line by line. On a $500 million loan, it's a rounding error. On a $20 million loan, it’s 15 to 25 cents of every dollar, which is not feasible at all for the 10,000 borrowers. GPU ownership is growing faster than credit can be issued against it. SemiAnalysis recently made the same argument from the other end of the market. In their view, the constraint on AI infrastructure has moved from data centers to chips and now to financing. Lenders today will underwrite almost nothing that isn’t investment-grade or backed by an Nvidia or hyperscaler guarantee, and by SemiAnalysis’s estimate, Nvidia’s balance sheet can support around 46 GW of capacity against roughly 240 GW coming online. The top of the market is covered. The compute middle market isn’t, and no amount of backstop capacity will reach a $20 million borrower. Private credit as it exists today is a tailored suit: beautifully made, cut one at a time, and right for the largest borrowers, but impossible to make 10,000 times. Housing solved the same problem with the mortgage backed security. Every house is different, but the mortgage is a standardized debt instrument that can be pooled, and pooling turns a thousand unremarkable borrowers into a security that a bank, an insurer and a pension fund can all buy. And more importantly, pooling loans together is how non-IG credit becomes IG through diversification. A billion dollars lent to one borrower needs Nvidia behind it. A billion dollars lent to a hundred borrowers doesn’t. The end state for what USD.AI is building is the creation of a new ABS market for GPUs, where 50 performing GPU loans can be packaged up and sold to institutional investors through GLX. It’s the natural state of where the market is heading. Power constraints are pushing deployments toward 2 to 5 MW instead of 250 MW, and financing constraints are pushing toward many small loans instead of a few enormous ones. Smaller deployments with lead times of a couple of months can also track actual demand more closely than gigawatt campuses planned years ahead, which makes the whole industry less likely to overbuild. Standardize the loan, and the capital follows. That’s the GPU mortgage. The AI Loan Factory If the binding constraint is the cost of making a loan, then the company has to be built around cutting it. USD.AI broke the life of a loan into five stages and gave each stage one team with one job: make that stage cheaper and faster, and prove it with a number. Origination: gpuloans.com. Inbound demand, qualified before anyone gets on a call, so the team only spends time on borrowers who can close. Underwriting: ComputeIQ. A 100-point rubric calibrated on a pipeline of more than 600 GPU financing deals, which scores operator, offtaker and contract in days rather than weeks. Servicing: the borrower portal on usd.ai. The documents, covenants and payments that run a loan for three years, tracked against live collections and telemetry instead of quarterly spreadsheets. Funding: sUSDai. The capital that funds a loan the day it closes, and the integrations that keep that pool growing. Distribution: GLX. The USD.AI loan marketplace, where seasoned loans are sold on to +8 credit funds with more than $105 billion in combined AUM, so the capital that funded them comes back to fund the next one. Together, that’s the AI Loan Factory, and its North Star is cost & speed: taking what it costs to put a GPU loan together from $5-$7 million down to under $500k. Data centers compete on time to power, and neoclouds compete on time to token. USD.AI competes on cost per loan. What is the final output when the velocity of credit reaches its theoretical limit? The GPU Mortgage: a standardized loan that turns GPUs into an asset class, enabling the credit to reach every pool of capital on Earth. Every product USD.AI builds now has to answer one question: which of those costs does it lower, and by how much? Each loan makes the next one cheaper. Each one adds a borrower, a contract, a set of covenants and a repayment history to what USD.AI knows and uses in ComputeIQ. Over time, the legal documents become templates and the underwriting becomes repeatable, and the work done by hand today becomes the data used to automate it tomorrow. III. Hold, move, settle From balance sheet to capital velocity to network: the three models As the factory gets cheaper to run, the business itself changes shape. Model 1 is the business USD.AI built. For the last year, USD.AI has made loans into sUSDai and held them to maturity. It works, and it’s profitable, but it’s a balance-sheet business: every dollar lent is locked up until the borrower repays it, so growth depends on raising more capital. Model 2 starts now. USD.AI makes the loan, holds it for a few months until it has a payment history, then sells it to an institution through GLX, takes the capital back and makes the next loan. The same $1 billion pool can turn over four times a year, generating $4 billion of loans. At a 2% origination fee, that’s $80 million of revenue a year instead of $20 million once, from the same capital base. The protocol gives up most of the interest on a loan it sells, but earns on it twice in return: once when the loan is made, and again when it is sold. The second payment exists because most of a loan’s risk sits at the start. After a few months of clean payments, the same loan looks safer, and cheaper capital will buy it at a lower yield. That gap is the gain on sale. Selling performing loans also changes the conversation with buyers, because a loan that’s already paying gets priced on its track record rather than on a promise. The first loan sale through GLX has closed: the $98.1 million facility made to Duos Edge AI in June, financing 2,304 NVIDIA B300 GPUs, which was seasoned in sUSDai and sold on to a new institutional buyer at par. This is what makes the business less dependent on the size of the pool. In Model 1, revenue tracks how much capital the protocol holds. In Model 2, it tracks how many loans the factory can make and sell, which means every dollar of cost taken out of the factory comes back as volume. Model 3 is where this goes. When a loan is sold, it doesn’t disappear. The loan leaves the balance sheet, the relationship doesn’t for the next 36 months. IV. 36 payments, 3 years, a permanent relationship Credit as an installed financial relationship. When a borrower takes a loan from USD.AI they become locked into a 36 month, 3 year relationship with USD.AI. Selling the loan doesn’t end the relationship, it just moves the credit risk somewhere else. All payments for the loan have to come through the protocol denominated in USDai. And so the settlement of compute commerce points to a future where payments, treasury and refinancing around compute credit are all settled in USDai, no matter who owns the loan. The relationship has four layers. A network. Every loan installs a three-year connection between a borrower and the capital behind it. A stablecoin. More credit outstanding means more USDai in use, means more treasury income flowing to sUSDai. Data. Every payment, covenant test and refinancing is a datapoint sharpening ComputeIQ and makes the next loan cheaper, and increases the velocity of the AI Loan Factory A platform: place to provide value-add services from 3rd party providers: hedging, telemetry, payments, model routing, and other services. Think about the iPhone. Apple sells you the phone once, at a lower margin than everything that comes after it. Once you own it, it becomes the way you buy apps, store photos, pay for things and subscribe to services, and Apple earns roughly twice the gross margin on those services that it earns on the hardware. The phone is the way in. The relationship is the business. The GPU mortgage works the same way. It sits on a company’s balance sheet for three years and becomes the natural entry point for everything around it: payments, treasury, cash management and refinancing. USD.AI doesn’t need to hold the loan to sit at the center of that. It needs to be where the loan starts and where its money settles. This is also why TVL stops being the right headline number to judge USD.AI’s success. TVL measured Model 1 well, because the size of the pool was the capacity of the business. Now in Model 2 & 3, success will be judged by a new set of numbers: Capital turns. How many times a dollar in sUSDai finances a loan each year. Originations per dollar of TVL. How much credit the same capital base creates. USDai-denominated credit outstanding. How much credit stays on the network after the loans are sold. Settlement volume. How much money moves through USDai around that credit. The goal is a very large network built on a relatively small balance sheet, growing from $1 billion of credit to $10 billion, then $50 billion, then $100 billion. What this means for sUSDai Moving from holding loans to selling them changes three things for sUSDai holders. The first is duration. In Model 1, every loan was held to maturity, which tied up sUSDai capital for three years. Amortization returned part of the principal each month, but the rest stayed exposed to a single borrower for the full term. GLX gives the protocol a way to sell loans held in sUSDai, so a seasoned loan returns all of its remaining capital to sUSDai when it sells. That answers the request capital allocators have made of USD.AI most consistently: a deeper secondary market for these loans. It also opens the door to forward-flow agreements, delayed-draw facilities and other arrangements where larger counterparties buy loans on a programmatic basis. The second is how loans leave the book. Every loan enters sUSDai through ComputeIQ and the same buy box, and each one now also needs a reliable way out, with a defined process for when and how it is sold. ComputeIQ and that sale process both get better with data, and that data only accumulates with volume. Mortgages and home equity lines have a century of performance history behind them. GPU loans are building theirs now, one sale at a time. The third is yield stability. With GLX, there will be gaps between selling one loan and funding the next, and sUSDai yield can dip during those gaps while capital waits to be lent again. Season 3 adds the CHIP Hurdle Rate, a 7% minimum sUSDai yield, for exactly these moments: if underlying sUSDai yield falls below 7%, the protocol’s Yield Reserve tops it up. That capital is what lets the protocol fund the next loan when it’s ready. Season 3 focuses on subsidizing various gaps as this product is built, just like how Season 2 focused on sUSDai getting broadly adopted in DeFi when that was the core growth constraint.IQ When a loan sells, its principal and the interest it earned while in sUSDai go back to sUSDai. The gain on sale goes to the protocol. More on where it goes from there, and on CHIP economics, coming soon. Where this goes USD.AI started by holding loans. Now it moves them. Next, it settles them. The scale of the opportunity is simple to state: a few hundred neoclouds today, 10,000 companies in the compute middle market next, and after that, 100,000 microclouds, with a network of credit growing from $1 billion to $100 billion to serve them. The impact is easier to see one borrower at a time. Picture a 40-person biotech lab that gets its first GPU cluster financed, on terms set by the same standard as a neocloud ten times its size, and an insurer on the other side of the world that ends up owning part of that loan, without anyone having to pick up the phone. Now multiply that. Eventually, every company that uses AI will have compute on its balance sheet, and many will have a loan against it. Every one of those loans has to be made, moved and settled somewhere, and USD.AI wants that to happen on one network. Every era’s money is a claim on its binding constraint. In the last century, oil made the petrodollar. In this one, the constraint is compute, and USDai can become the dollar the compute economy settles in. USD.AI finances the machines. The machines adopt the money. The money settles everything. Intelligence runs on compute. Compute runs on capital. The goal is for that capital to run on USDai.(Historical earnings: For 2026Q2 (period ended 2026-03-28), Apple reported basic EPS of 4.87, diluted EPS of 4.85, and net income of USD 71.675 billion. In the prior-year 2026Q2 comparison period (ended 2025-03-29), basic EPS was 4.06, diluted EPS was 4.05, and net income was USD 61.110 billion. For 2026Q1 (period ended 2025-12-27), basic EPS was 2.85, diluted EPS was 2.84, and net income was USD 42.097 billion, compared with the prior-year 2026Q1 period (ended 2024-12-28) basic EPS of 2.41, diluted EPS of 2.4, and net income of USD 36.330 billion. Consensus expectations: For 2026Q4 (event date 2026-11-02), consensus EPS estimate is 2.0214 and revenue estimate is USD 115.327 billion. For 2027Q1 (event date 2027-01-27), consensus EPS estimate is 2.9512 and revenue estimate is USD 156.464 billion.)