The question of whether a large language model is a business asset — and if so, how to value it — has moved from theoretical curiosity to practical urgency. OpenAI's valuation reaching $500 billion, Anthropic's at $60 billion, and a wave of enterprise-grade LLM deployments have forced investors, auditors, and CFOs to grapple with the economics of language models as intangible assets.
The challenge is that LLMs have economic characteristics unlike any previous asset class. They do not wear out like physical assets. They do not have clear useful lives like patents. Their value depends not on what they are but on how they are deployed. And their replacement cost changes radically every 12-18 months as the technology advances.
A trained LLM can be copied at negligible cost. The model weights — the actual "knowledge" — are a file that can be duplicated in seconds. This means the value of an LLM is not in the artifact itself but in the training data, training process, and deployment infrastructure that produced it.
LLM capability depreciates faster than almost any other asset. A state-of-the-art model from 18 months ago is now a mid-tier model. This rapid depreciation has profound implications for valuation — the useful economic life of an LLM is extremely short by traditional standards.
An LLM serving one customer does not reduce its ability to serve another. Unlike physical assets (and even most software licences), LLMs exhibit non-rival consumption: the same model can serve millions of users simultaneously without degradation.
LLMs exhibit capabilities that were not explicitly trained, making their future value difficult to predict. A model trained for text generation may also prove capable of code review, reasoning, or scientific analysis — capabilities discovered after deployment rather than designed in advance.
LLMs challenge traditional asset valuation frameworks because they combine near-zero marginal cost, rapid depreciation, non-rival consumption, and emergent capabilities. No existing asset class shares all four characteristics, which means valuation approaches must be adapted rather than applied directly.
The distinction between foundation models and fine-tuned models is critical for valuation because they have fundamentally different economic characteristics.
Foundation models (GPT-4, Claude, Gemini, Llama) are general-purpose LLMs trained on broad datasets at enormous cost ($100M+ for frontier models). They are built by a small number of well-capitalised organisations and are typically accessed via API or open-source licence.
For most businesses, foundation models are infrastructure — like electricity or cloud computing. They are necessary inputs to AI applications but do not constitute a competitive advantage because every competitor has access to the same models.
Fine-tuned models are created by adapting a foundation model to a specific domain, task, or dataset. The fine-tuning process adds domain-specific knowledge, proprietary data, and task-specific behaviour that creates differentiated capability.
| Characteristic | Foundation model | Fine-tuned model |
|---|---|---|
| Development cost | $100M+ | $10K-$1M |
| Competitive advantage | None (commodity) | Potentially significant |
| Defensibility | Zero (accessible to all) | Data-dependent |
| Useful life | 12-18 months | Shorter (depends on base model) |
| Valuation approach | Cost-based | Income or market-based |
| Balance sheet treatment | Typically expensed | Potentially capitalisable |
A legal technology company fine-tunes GPT-4 on 2 million proprietary contract documents to create a contract analysis system that outperforms general-purpose models on legal reasoning. The foundation model (GPT-4) has no scarcity value — any competitor can access it. The fine-tuned model has significant value because it embodies proprietary legal data that competitors cannot replicate. The fine-tuning cost was £200,000, but the revenue-generating capability is worth substantially more.
The cost approach values an LLM at its replacement cost — what it would cost to train an equivalent model from scratch. For foundation models, this can exceed $100 million. For fine-tuned models, it is typically $10,000-$1,000,000.
The cost approach is appropriate when the LLM has been recently developed and market/income data is unavailable. Its weakness is that it does not capture the value created by the model's deployment — only the cost to produce it.
The income approach values an LLM by the future cash flows it generates (or the costs it saves), discounted to present value. This is often the most appropriate approach because it captures the business value of the model rather than just its production cost.
Apply the income approach by:
The market approach values an LLM by comparison to comparable transactions — what similar models have sold for or what similar AI companies are valued at. This approach is constrained by the limited number of comparable transactions, but AI M&A data is growing.
LLM depreciation is unlike traditional asset depreciation because it is driven by technological obsolescence rather than physical wear. A model does not degrade through use — it degrades because newer, better models become available.
For accounting and valuation purposes, estimate useful life based on:
Under IAS 38, if an LLM qualifies as a recognised intangible asset, it must be amortised over its useful life. Given the rapid depreciation of AI models, a 2-4 year amortisation period is typically appropriate — far shorter than the 10-20 year periods common for traditional intangible assets like patents or customer relationships. See our detailed guide on capitalising AI development costs.
For investors evaluating companies that use or develop LLMs, three questions are critical:
Is the value in the model or the data? Companies whose value derives from a fine-tuned model are vulnerable to the next foundation model release making their fine-tuning obsolete. Companies whose value derives from proprietary training data have a more durable asset because data retains its value across model generations.
What is the retraining cost trajectory? As foundation models improve, fine-tuning becomes cheaper. A company that spent £500,000 to achieve current capability may face competitors who achieve equivalent capability for £50,000 in two years. The valuation must account for this deflation.
Is there a data moat? The most valuable LLM businesses are those where the model improves as customers use it — creating a data flywheel that widens the competitive gap over time.
The Opagio Growth Platform provides tools for assessing AI asset value, including LLM capability maturity and data asset defensibility.
LLMs are genuine business assets, but valuing them requires frameworks adapted to their unique economic characteristics: rapid depreciation, near-zero marginal cost, and value that depends more on data and deployment than on the model itself. For investors, the key insight is that the most valuable LLM assets are not the models — they are the proprietary data and integration depth that make the models uniquely useful. Value the data, not just the model.
David Stroll is Co-Founder and Chief Scientist at Opagio. His research focuses on the economics of AI assets, productivity measurement, and the valuation of technology-driven intangible capital. Learn more about the Opagio team.
The choice between open-source and proprietary AI has profound implications for intangible asset value, competitive defensibility, and company valuation. This article examines how each approach creates — or fails to create — durable intangible assets.
Read more →
Every significant AI investment creates or enhances the same six categories of intangible assets I built during 15 years at IG Group. The problem is not that these assets do not exist — it is that no one is measuring them. Here is the framework that connects AI spending to enterprise value creation.
Read more →Most organisations cannot quantify the return on their AI investments. This practical framework breaks AI ROI into four measurable dimensions and provides concrete methods for calculating returns across cost reduction, revenue growth, risk mitigation, and strategic optionality.
Read more →Get the latest insights on intangible asset growth and productivity delivered to your inbox.
Book a free consultation to see how the Opagio Growth Platform can help your business.