Fine-Tuning

Fine-tuning begins with a foundation model that has been trained at large scale on broad, general data. The foundation model has learned general patterns of language, vision, or reasoning, but it does not perform well on narrow, specialist tasks without further training. Fine-tuning takes that foundation model and trains it further on a curated dataset specific to the target domain: clinical notes for a medical AI, contract language for a legal AI, transaction records for a fintech AI, or video frames from a specific style or aesthetic for a video generation model.

The result is a model that retains the general capability of the foundation model while gaining measurably better performance on the target task. For enterprise AI companies, fine-tuning is the primary mechanism for creating capability that competitors cannot easily replicate, because the fine-tuning dataset itself is typically proprietary and takes years to accumulate.

---

How Fine-Tuning Works

Modern large-scale fine-tuning uses one of three main techniques.

Full fine-tuning updates all parameters of the foundation model on the new dataset. This produces the highest-quality results but requires substantial compute and is typically only feasible for companies with significant GPU infrastructure. The resulting model is a completely new artifact that no longer depends on the original foundation model for serving.

Parameter-efficient fine-tuning (PEFT) updates only a small subset of model parameters, most commonly using LoRA (Low-Rank Adaptation) or QLoRA (quantised LoRA). PEFT is significantly cheaper to compute and can produce near-full-fine-tuning quality on many tasks. The resulting adapter layers can be applied on top of a frozen foundation model at inference time, which means the company can update its proprietary capability without retraining the entire model base.

Instruction fine-tuning and RLHF (Reinforcement Learning from Human Feedback) are used to shape model behavior rather than expand domain knowledge. These techniques train the model to follow instructions in a particular style, maintain a persona, or prioritise certain types of output. They are widely used in enterprise AI products where the interface between the model and the end user needs to be carefully controlled.

---

Fine-Tuning in AI Company Valuation

Fine-tuning is the mechanism that separates AI companies that own defensible, compounding capability from those that are thin wrappers around third-party foundation model APIs.

The fine-tuning dataset is the moat. A foundation model is increasingly a commodity: OpenAI, Anthropic, Google, Meta, and Alibaba all publish competitive foundation models, and the gap between them has narrowed substantially since 2022. What differentiates AI companies at similar stages is not which foundation model they use but what proprietary data they have fine-tuned on. A fintech company that has fine-tuned on five years of proprietary transaction data has an asset that a competitor cannot replicate in two years, regardless of which base model they start with.

Fine-tuning quality is measurable. In M&A diligence, acquirers will test the fine-tuned model against the base foundation model on domain-specific benchmarks. If the fine-tuned model shows 20-30% performance improvement on the target task, the fine-tuning dataset is likely to be a genuine moat. If the improvement is 2-5%, the fine-tuning data may not be defensible against a competitor who fine-tunes a better base model on comparable data.

APAC-language fine-tuning carries a valuation premium. Foundation models are predominantly trained on English-language data. Fine-tuning on Korean, Japanese, Mandarin, Bahasa Indonesian, or Thai language data produces performance improvements that are much larger than English-domain fine-tuning, because the base model underperforms on APAC languages relative to English. An AI company that has fine-tuned on proprietary APAC-language data commands a meaningful premium in transactions where the acquirer is an APAC enterprise buyer.

---

Fine-Tuning Due Diligence Questions for Acquirers

When evaluating an AI company’s fine-tuning capability, acquirers should ask:

1. What is the fine-tuning dataset, and who owns it? The dataset must be either proprietary data generated by the company’s operations, data licenced from third parties with clear commercial use rights, or synthetic data generated through a documented process with known provenance. If the dataset includes scraped web data, user-generated content, or third-party data with unclear licencing, the acquirer assumes copyright and regulatory risk.

2. Can the fine-tuning be reproduced? A fine-tuning process that depends on a specific research team’s undocumented knowledge is a key-person risk, not a defensible technical asset. The company should have documented training pipelines, reproducible fine-tuning runs, and version-controlled datasets and hyperparameters.

3. What happens if the base foundation model changes? If the company fine-tunes on GPT-4, Llama 3, or another specific foundation model, the proprietary fine-tuning may need to be redone when the foundation model is updated or retired. Acquirers assess whether the fine-tuning process is foundation-model-agnostic (that is, whether the same fine-tuning data could be applied to a different base model with similar results) or whether it is tightly coupled to a specific model architecture.

4. How does fine-tuning performance degrade over time? AI models experience performance degradation when the world changes and the training data no longer reflects current patterns (sometimes called model drift). An AI company should have a monitoring and retraining cadence that maintains fine-tuning quality over time. Acquirers look for evidence of systematic retraining with updated data, not a one-time fine-tuning event.

5. Is the fine-tuning compute cost sustainable? Full fine-tuning of a large model can cost USD 50,000-500,000 per training run on current hardware. PEFT methods reduce this significantly, but the compute cost of maintaining fine-tuning quality as the dataset grows is a real operating expense. The company should model compute costs at 2x and 5x current data scale.

---

Fine-Tuning and Foundation Model Dependency

One of the most important structuring questions in an AI company acquisition is the relationship between the acquired company’s proprietary fine-tuning and the underlying foundation model.

If the company has fine-tuned on a closed-source foundation model (GPT-4, Claude 3, Gemini) via API access, the acquirer is inheriting a product whose core capability can be disrupted by a pricing change, API deprecation, or terms-of-service update from a third-party provider. This is not necessarily a dealbreaker, but it requires representations and warranties about the API relationship and typically warrants escrow provisions covering the risk of API disruption within a defined window.

If the company has fine-tuned on an open-source foundation model (Llama 3, Mistral, Wan2.1, Qwen), the acquirer inherits the fine-tuning artifact and is not dependent on a third-party API for serving. The model can be deployed on the acquirer’s own infrastructure, modified, or migrated to a future open-source model. This is the more defensible architecture from an acquisition perspective.

Companies that have invested in foundation-model-agnostic fine-tuning pipelines, capable of adapting to new base models as they become available, command the highest diligence ratings. The pipeline itself is the strategic asset, not any single fine-tuned model instance.

---

Fine-Tuning in APAC AI M&A

APAC AI companies that have accumulated proprietary fine-tuning datasets across the region’s language diversity (CJK scripts, Southeast Asian languages, Indic languages) have a structural advantage over Western AI companies attempting to serve APAC markets. The investment in APAC-language fine-tuning is not trivially replicable because it requires native-language data annotation expertise, cultural context review, and evaluation frameworks that Western companies have not built.

For corporate development teams at APAC acquirers considering AI company acquisitions, the APAC-language fine-tuning dataset should be a specific diligence workstream. Acquiring a company with three years of proprietary Korean-language financial data fine-tuning, or a company that has fine-tuned on Japanese regulatory text at clinical-grade quality, creates a capability advantage that could take years to replicate internally.

Amafi Advisory works with AI company founders preparing for sale processes and with corporate development teams evaluating AI acquisitions across Asia Pacific. If you are considering a transaction where fine-tuning data ownership or foundation model dependency is a key question, contact our team.

---

Related Terms

• Inference Cost: The compute cost of running a fine-tuned model in production; a key operating expense that affects gross margin analysis in AI company valuation.
• Synthetic Data: Artificially generated data used to augment fine-tuning datasets where real-world labelled data is scarce or expensive to collect.
• Red-Teaming: Testing methodology used to probe fine-tuned model behavior for safety failures, jailbreaking vulnerabilities, and off-policy outputs before deployment.
• Context Window: The token limit that determines how much input a fine-tuned model can process in a single inference call; affects enterprise use case feasibility.
• Due Diligence: The investigation process through which an acquirer validates claims about a target company’s technology, data assets, and financial performance.
• Foundation Model: The base model that fine-tuning adapts to a specific domain; understanding open vs closed foundation model choices is central to AI acquisition diligence.
• ARR: Annual recurring revenue; the financial metric used to benchmark AI company valuations and fine-tuning efficiency investments against revenue output.