AI Models as Collateral: Can You Lend Against a Large Language Model?

During my years structuring asset-backed securities at NM Rothschild & Sons, the fundamental principle was always the same: identify a cash-generating asset with predictable performance characteristics, isolate it in a special purpose vehicle, and pledge it as collateral for debt securities. The assets were initially straightforward — aircraft leases, shipping contracts, mortgages. Over time, the assets became more exotic — carbon credits, electricity futures, music royalties.

The question I have been exploring for the past 12 months is whether the same principles can apply to the most valuable assets in the modern economy: proprietary AI models, training datasets, and the customer relationships they support.

The answer is yes. But it requires rethinking what makes a credible collateral asset in an intangible-asset economy.

What Makes Credible Collateral

Before addressing AI models specifically, it is worth recalling what lenders fundamentally require from collateral. From structured finance perspective, credible collateral has four attributes:

1. Identifiable: The asset can be precisely defined and located. In an enforcement scenario, the lender knows exactly what it owns.

2. Separable: The asset can be transferred to a new owner without losing its fundamental value. Real estate is separable (you can transfer title). A relationship with a specific customer may not be (the customer relationship dies when the original owner changes).

3. Measurable: Performance can be quantified over time. Cash flows are measurable. Intangible assets are measurable if their value correlates to specific business metrics.

4. Enforceable: The lender has a clear path to realise the value of the asset if the borrower defaults. Enforceability often means there is a secondary market for the asset, or specific mechanisms for the lender to extract cash flows if default occurs.

Under these criteria, physical collateral (real estate, equipment, vehicles) has very high marks across all four dimensions. IP collateral (patents, trademarks, software IP) scores well on identifiable and separable, moderately on measurable, and lower on enforceable (IP is valuable to specialists but hard to sell quickly).

The question: Do AI models meet these criteria sufficiently well to serve as loan collateral.

AI Models as Collateral: The Case

On the positive side, AI models and the assets that support them have several characteristics that make them credible collateral.

Identifiable

A proprietary AI model can be precisely defined: the model architecture (transformer, LSTM, custom), the training dataset, the hyperparameters, the performance baselines. At close, all of this can be documented, versioned, and held by a security trustee. This is not materially different from how IP is held in IP-backed lending structures.

Proprietary datasets can be catalogued, versioned, and held in a secure data environment. Training data for a large language model, or a proprietary customer dataset used for recommendations, can be precisely defined and secured.

Separable

This is where AI models introduce complexity. The model itself is separable — you can transfer ownership of the code and the weights. But the model's value often depends on context: the specific customer relationships it serves, the operational workflows it supports, the team's understanding of how to deploy it.

However, the key insight is that successful IP-backed lending has always dealt with this separation problem. A patent is valuable to its owner because of the company's capability to commercialise it. If you transfer the patent to a different owner, the value may change. But the patent is still collateral because it has an identifiable value independent of its current owner.

The same applies to AI models. A proprietary recommendation model trained on a retail company's customer data has specific value to that retail company. But it also has an identifiable value to other retail companies (who could retrain it on their data) and to potential acquirers (who could integrate it into their platform).

Measurable

This is the critical advantage of AI assets relative to many traditional IP assets. AI models produce measurable, quantifiable performance: accuracy, latency, throughput. Customer datasets are measurable by volume, freshness, and structure. Revenue generated by AI-augmented products is measurable.

A lender evaluating a property takes a survey and values the property based on comparable transactions. A lender evaluating an AI model should require:

• Model performance baselines (accuracy on test datasets)
• Performance in production (live performance data)
• Sensitivity to data degradation (what happens if training data quality declines)
• Alternative valuation (what is the cost to retrain the model on a different customer base)

These are all measurable, and they are the inputs to model valuation.

Enforceable

This is the most challenging dimension for AI collateral. If an AI company defaults on a loan and the lender needs to realise the collateral, what exactly happens.

Option 1: The lender enforces against the intellectual property and sells it to a third party. This is analogous to IP licensing in traditional structured finance. The lender (through a security trustee) can license the model to a third-party operator, who uses it to generate cash flows that service the loan. This requires the IP to be transferred to an SPV (special purpose vehicle) and licensed back to the original company.

Option 2: The lender enforces against the customer revenue that the model generates. If the model is embedded in a customer-facing product, the revenue flows from that product can be assigned to the lender. This is analogous to revenue securitisation — cash flows from the product secure the loan.

Option 3: The lender enforces against the data assets that support the model. In a default scenario, the lender can assign the data to a new operator, who retrains the model and generates revenue.

The enforceability path depends on the specific model, the customer relationships, and the operational integration. But for mature, revenue-generating AI models, enforcement is achievable.

Three Collateral Structures for AI Models

Given these characteristics, I would propose three distinct financing structures for different classes of AI collateral.

Structure 1: IP Holdco Model (Most Established)

The company's core proprietary models and IP are transferred to a special purpose holding company. The holdco licenses the IP back to the operating company and pledges the IP as security for a lender.

How it works:

• The proprietary AI models, training datasets, and associated IP are transferred to an IP holdco
• The holdco grants an exclusive, perpetual licence back to the operating company for a licence fee
• The lender takes first security over the IP in the holdco
• In a default scenario, the lender can enforce against the holdco, hire an operator to continue running the licensed models, and extract cash flows from the licence fee

Advantages:

• Well-established in IP-backed lending
• Clean separation between operating risk (company's performance) and IP risk (model's value)
• Lender has clear enforcement path
• Operates independently of specific customers or use cases

Challenges:

• Requires transfer of IP ownership to holdco, which can be legally complex
• Licence-back arrangement can create tax complications
• Dependent on the IP being truly separable from the operating company

When to use: For companies with defensible proprietary models that would have value to other operators (e.g., recommendation models, forecasting models, language models fine-tuned on proprietary data).

Structure 2: Revenue Securitisation (Developing)

Customer revenue generated by AI products is assigned to a special purpose vehicle (SPV) that issues debt securities backed by those cash flows.

How it works:

• The operating company identifies revenue streams that are directly generated by AI models (e.g., AI-powered recommendations, AI-driven customer service, AI-enabled decision support)
• Those revenue streams are assigned to an SPV
• The SPV issues debt securities backed by the assigned revenue
• As customers pay for AI-enabled services, the revenue flows to the SPV and services the debt

Advantages:

• Revenue is easier to value and enforce against than IP
• Analogous to established asset-backed securities structures (lease receivables, mortgages)
• Does not require IP transfer, just revenue assignment
• Cash flow-based pricing is more straightforward than asset-based pricing

Challenges:

• Requires revenue to be truly attributable to AI (hard to separate from brand, customer relationships, and sales effort)
• Model performance risk carries through to revenue — if the model degrades, revenue declines
• Customer concentration risk (if revenue is from a small number of accounts, the securitisation is at risk)

When to use: For SaaS or recurring-revenue businesses where AI generates a specific, measurable portion of revenue (e.g., "our AI features contribute £2M of the £5M annual revenue"). Requires customer concentration to be manageable (<20% from largest customer).

Structure 3: Hybrid Structure (Frontier)

The most sophisticated approach combines IP collateral with revenue securitisation in a single facility. The lender takes security over both the IP and the revenue streams it generates.

How it works:

• Core proprietary models are transferred to an IP holdco (as in Structure 1)
• Revenue streams generated by those models are assigned to an SPV (as in Structure 2)
• The lender has first security over both the IP and the assigned revenue
• In a default scenario, the lender can either a) enforce against the IP and license it to a third-party operator, or b) enforce against the revenue streams directly
• The dual security provides optionality and improves the lender's recovery position

Advantages:

• Provides lender with two enforcement paths, reducing risk
• Improves advance rate (70-85% LTV vs 60-70% for single-security structures)
• Reduces pricing (5-7% above prime vs 6-9% for single-security)
• Optimal for mature AI companies with both defensible IP and measurable revenue contribution

Challenges:

• Complex documentation and legal structure
• Requires careful tax structuring to avoid double taxation
• Dependent on both IP transferability and revenue attributability being clear

When to use: For mature AI companies that have built defensible proprietary models AND have clear revenue streams attributable to those models. Requires 18+ months of production performance history.

Valuation Approaches for AI Collateral

Pricing an AI collateral facility requires valuation methodologies that capture both the asset's current value and its depreciation risk.

For proprietary models:

• Cost approach: What would it cost to build an equivalent model from scratch. Data scientists, compute resources, time. For a mature proprietary model, this could be £500k-£2M. This sets a floor.
• Income approach: How much revenue does the model generate, and what percentage of that is attributable to the model versus other factors. A model that contributes £500k annually might be valued at 3-5x that (£1.5M-£2.5M) depending on model life and depreciation risk.
• Relief-from-royalty: If the company did not own the model, what would it have to pay to license an equivalent. Industry standard licensing rates are 5-15% of revenue for software IP. A model generating £500k in revenue might license for £25k-£75k annually, implying a capitalised value of £250k-£750k.

For proprietary datasets:

• Valuation is typically tied to revenue contribution and durability of competitive advantage
• A dataset that improves with use (network effects) is worth more than a static dataset
• Valuation approach: What is the customer lifetime value improvement that the data enables. If proprietary data improves customer LTV by 20% and annual customer value is £5k, and the dataset covers 10,000 customers, the data is worth approximately £10M
• Depreciation: Data assets depreciate if new competitors can accumulate similar data quickly. A dataset with high accumulation barriers is more stable collateral

For recurring revenue:

• Standard SaaS valuation methodology: annual recurring revenue times a multiple (typically 5-10x for SaaS companies, lower multiples for AI-specific revenue)
• Discount for concentration risk: if a single customer represents >20% of AI-driven revenue, apply a concentration discount
• Discount for model risk: if model performance has degraded in the past 12 months, apply a risk premium

The Market Opportunity

The opportunity for AI-based lending is substantial. Consider a typical scenario:

A software company with £10M ARR, 90% recurring revenue, has built a proprietary AI-powered recommendation engine that generates £2.5M of the annual revenue. The company needs £2M in growth capital.

Under traditional lending, this company has limited options: unsecured debt (expensive, limited availability), equity dilution (expensive to the founders), or venture debt (short-term, expensive). Most likely, the company raises equity and dilutes itself.

Under AI-collateral-backed lending:

• The proprietary recommendation model can be valued at £3M-£5M (using relief-from-royalty or income approaches)
• The £2.5M annual recurring AI revenue can support an additional securitisation of £2M-£3M
• Combined, the company could borrow £3M-£5M backed by AI collateral
• Cost of capital: 6-8% (below venture debt, above unsecured institutional lending, but with less dilution than equity)

The result: the company preserves equity value for founders while accessing growth capital at reasonable cost.

What Lenders Need to Get Started

Banks and alternative lenders interested in AI collateral-backed facilities need to develop three capabilities:

1. Valuation expertise: Lenders need to hire data scientists and AI specialists who can assess model quality, data asset defensibility, and revenue attribution. This is not traditional loan underwriting — it requires technical depth.

2. Structural expertise: Lenders need to develop IP security structures, revenue securitisation mechanics, and data asset assignment frameworks. This requires involvement of sophisticated legal counsel and structured finance expertise.

3. Portfolio management: AI collateral will be riskier and less liquid than traditional collateral. Lenders need to build portfolio management capabilities to monitor model performance degradation, data quality, and revenue sustainability over time.

These are material investments. But for lenders who make them, the opportunity is significant: a multi-billion-pound market of AI companies that need capital but do not want to dilute equity.

Why This Matters for Opagio

Opagio's intangible asset valuation framework is designed precisely to support this kind of structured lending. Lenders need a systematic way to value AI models, data assets, and customer relationships. Our framework provides exactly that — structured, defensible valuations based on standard financial data.

For AI companies, the opportunity is equally significant. A company with defensible proprietary AI capability should be able to raise capital against that capability at reasonable cost. Today, that is not possible because the valuation and structural frameworks do not exist. In three years, it will be as normal as IP-backed lending and revenue securitisation are today.

The transition from equity dilution to asset-backed lending for AI companies is not inevitable — it requires lenders to build new capabilities and take on new risks. But the economics are powerful enough that the transition will happen. The question is which lenders will lead it.

Tony Hillier is co-founder of Opagio. He holds an MA from Balliol College, Oxford and an MBA with distinction. Tony held executive board positions at NM Rothschild & Sons and GEC Finance, and a non-executive directorship at Financial Security Assurance in New York, where he specialised in structured finance, asset-backed securities, and cross-border tax-leveraged transactions.