Should Financial Firms Build Their Own Foundation Models?

Updated 4 November 2026 · Standard formula, shown below

Should financial firms build their own foundation models? The short answer is: usually no, but with important exceptions. Building a foundation model from scratch is a multi-year, multi-million-dollar effort that most banks, insurers, and asset managers should avoid; instead, they should fine-tune or orchestrate existing models for their specific use cases. However, a small number of very large institutions with unique data, extreme privacy requirements, or a strategic need for differentiation may justify building a custom model.

Background: What Are Foundation Models and Why Finance Cares

Foundation models are large machine learning models trained on vast, general-purpose datasets. Examples include GPT-style language models, multimodal models, and large embedding models. They are called “foundation” because they can be adapted to many downstream tasksâ€”summarization, classification, generation, coding, and moreâ€”with relatively little additional training.

Financial firms are interested because these models can read earnings calls, draft client communications, detect suspicious transactions, answer regulatory queries, and analyze unstructured data like news, contracts, and research reports. A 2024 survey by the Bank for International Settlements found that a majority of central banks and financial authorities are exploring AI, but very few are building their own models. The industry trend is toward partnerships with specialized AI vendors and cloud providers.

How Building a Foundation Model Actually Works

Building a foundation model involves several distinct stages, each with its own costs and risks:

Data collection and curation: You need hundreds of billions of tokens of high-quality text. For finance, this could include filings, transcripts, and proprietary documents, but you also need general web data to teach the model language and reasoning.
Pre-training: This is the most compute-intensive phase. Training a 7-billion-parameter model can cost around $1 million in cloud compute; a 70-billion-parameter model can cost tens of millions. Training runs can take months and require specialized clusters of thousands of GPUs.
Fine-tuning and alignment: After pre-training, the model is fine-tuned on narrower tasks and aligned with human preferences to be helpful and safe. This is where most financial applications actually live.
Inference and serving: Running the model in production requires significant infrastructureâ€”GPU servers, load balancing, and latency optimization. For a large model, each query might cost fractions of a cent, but at scale that adds up.

Most firms that “build” a model are actually fine-tuning an open-source base like Llama or Mistral. Fine-tuning adjusts weights on a smaller, domain-specific dataset, costing thousands of dollars instead of millions.

Why Most Firms Should Not Build

There are several compelling reasons to avoid building from scratch:

Cost: Pre-training a competitive model requires tens of millions of dollars in compute alone. This is beyond the budget of most financial institutions, even large ones.
Talent scarcity: You need a team of experienced ML researchers and engineers who have trained large models before. Such talent is rare and expensive, and often prefers working at AI labs.
Rapid obsolescence: The field moves quickly. A model you spend two years building may be outdated by a newer architecture or a better open-source release.
Regulatory and compliance risk: Building a model gives you full control, but also full responsibility for its behavior. You must document data provenance, test for bias, and ensure explainabilityâ€”harder with a custom model.
Opportunity cost: The same investment in fine-tuning and application development will likely yield more value faster.

When Building Makes Sense

There are a few scenarios where building a custom foundation model is defensible:

Unique proprietary data: If your firm has a massive, unique dataset that no public model has seenâ€”for example, decades of proprietary trading signals or legal documentsâ€”a custom model might capture value that competitors cannot replicate.
Extreme privacy or security: Some data cannot leave your infrastructure due to regulation or client agreements. If no existing model can be deployed on-premises with acceptable performance, you might build one.
Strategic differentiation: A few mega-banks, like JPMorgan Chase, have built their own large language models (e.g., the LLM Suite for employees). They have the scale and the data to justify the investment.
Hardware ownership: If you already own a large GPU cluster (e.g., for risk simulations), the marginal cost of training may be lower.

Alternatives to Building: Fine-Tuning, RAG, and Orchestration

Most firms should adopt a layered approach:

Use existing APIs: Start with a commercial or open-source model accessed via API. This is the fastest and cheapest way to test use cases.
Fine-tune on your data: Once you know the task, fine-tune a smaller open-source model on your proprietary documents. This often achieves 90% of the quality of a custom model at a fraction of the cost.
Add retrieval-augmented generation (RAG): Instead of retraining, you can give the model access to your internal knowledge base at inference time. This is effective for question-answering over your own documents.
Orchestrate multiple models: Use a small, fast model for simple tasks and a larger model for complex reasoning. This reduces cost and latency.

Costs and Risks: What to Budget For

Even fine-tuning has hidden costs. You need to budget for:

Data preparation: Cleaning and labeling data is often 50-80% of the effort.
Evaluation: You need a robust evaluation framework to measure performance on financial tasks, including regulatory compliance and accuracy.
Governance: You must establish model risk management, which includes documentation, validation, and ongoing monitoring. This is a significant cost in regulated industries.
Ongoing maintenance: Models drift as data distributions change. You need to retrain or fine-tune periodically.

Risks include hallucination (generating false information), bias, and regulatory penalties. In 2023, the SEC proposed rules on AI conflicts of interest, and regulators worldwide are scrutinizing AI use in finance. A custom model does not exempt you from these risks; it may increase them if you lack the expertise to validate it.

How to Decide: A Practical Framework

Before starting any build, answer these questions:

Do you have a clear, high-value use case that cannot be met by fine-tuning an existing model?
Do you have access to unique data that gives a durable advantage?
Can you afford the upfront and ongoing costs, including hiring top ML talent?
Do you have the governance and risk management infrastructure to handle a custom model?
Will the model be used for many tasks, or just one? If one, fine-tuning is almost always better.

If you answered “no” to any of these, you should not build. Instead, focus on building a strong application layer on top of existing models. The competitive advantage in finance will come from proprietary data, workflow integration, and domain expertiseâ€”not from owning the model weights.

Frequently Asked Questions

How much does it cost to train a foundation model for finance?

Pre-training a 7-billion-parameter model can cost around $1 million in compute, while a 70-billion-parameter model can cost $10-30 million. This excludes data collection, engineering, and ongoing serving costs. Fine-tuning a smaller model on financial data can cost as little as a few thousand dollars.

What is the difference between pre-training and fine-tuning?

Pre-training teaches the model language and general knowledge on massive datasets. Fine-tuning takes a pre-trained model and trains it further on a smaller, specific dataset to adapt it to a particular domain or task. Fine-tuning is much cheaper and faster.

Can we use an open-source model like Llama or Mistral for finance?

Yes. Open-source models can be fine-tuned on your data and deployed on your own infrastructure. They offer transparency and control. However, you need to ensure you comply with the license terms and have the technical expertise to deploy and maintain them.

What are the main regulatory concerns with custom foundation models?

Regulators expect firms to manage model risk, which includes documenting the model’s development, testing for bias, ensuring explainability, and monitoring performance. Custom models may face higher scrutiny because they are less understood. You must be able to explain how the model makes decisions to auditors.

Is retrieval-augmented generation (RAG) a good alternative to fine-tuning?

RAG is excellent for question-answering over your own documents because it retrieves relevant text at query time and gives the model context. It is cheaper and easier to update than fine-tuning. However, it may not perform as well for tasks that require a specific style or domain-specific language, where fine-tuning is better.

Will building our own model give us a competitive advantage?

Not necessarily. The model itself is rarely the differentiator; the data and the application are. Many firms achieve competitive advantage by fine-tuning existing models on proprietary data and integrating them into workflows. Building a model only makes sense if you have a unique data moat and the scale to sustain it.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.