Domain-Specific Foundation Models vs. Frontier LLMs: Where's the Enterprise ROI?

Lucas (human) wrote this:

What returns more enterprise ROI over the next 2-3 years: frontier lab LLMs or purpose-built Domain Specific Foundation Models (DSFMs) like AlphaFold or BloombergGPT?

The former has much larger scope and "awareness", but is also 1-2 orders of magnitude more expensive to perform a training run (we're talking something like 50B parameters with BloombergGPT vs. hundreds of billions of parameters for the models from OpenAI, Anthropic, Google, Meta and Mistral).

Given the huge returns on successful pharma, I think (loosely held) that medical/bio DSFMs may outstrip even the code-base fine-tuned LLMs (i.e. Claude Code driving $4B revenue is different than say $26B in Ozempic sales in '24.)

Claude thought it'd expand a bit with its own 2 cents:

The Economics Don't Lie

Let's look at the math:

Frontier LLMs:

Training costs: $100M+ per run (GPT-4 class models)

Parameters: 100B-1T+

Scope: Everything

Competition: 5-6 major players burning cash

Domain-Specific Foundation Models:

Training costs: $1M-10M per run

Parameters: 1B-50B

Scope: Narrow but deep

Competition: Much less crowded

The capital efficiency story writes itself.

Why Medical DSFMs Could Win Big

The pharmaceutical industry represents a massive addressable market with economics that dwarf even the biggest tech outcomes:

Ozempic: $26B in 2024 sales

Humira (peak): $21B annually

Average blockbuster drug: $1B+ annually

Total pharma market: ~$1.5T globally

Compare this to software:

Microsoft Office: ~$50B annually

AWS: ~$90B annually

All of Anthropic's projected revenue: <$1B (for now)

A DSFM that accelerates drug discovery by even 6 months could be worth billions in NPV to a single pharmaceutical company.

The Competitive Moat Question

Frontier LLMs face a brutal reality: they're competing in a race where the finish line keeps moving and the entry cost keeps rising. Every major tech company is building essentially the same thing.

DSFMs have natural moats:

Domain expertise barriers - you need deep subject matter knowledge

Specialized datasets - proprietary data is the new oil

Regulatory approval - huge barriers in pharma/finance

Network effects - first mover advantages in specialized fields

The Attention Allocation Problem

Here's what I see in enterprise sales cycles:

CTOs are spending 80% of their AI budget discussions on "Should we use GPT-4 or Claude?" when the real value might be in "Should we build a DSFM for our specific use case?"

The mindshare imbalance is striking. Everyone talks about frontier models, but the quietly profitable companies are building specialized models for vertical applications.

Where I'm Placing My Bets

As an investor and advisor, I'm increasingly interested in companies building DSFMs for:

Life sciences - drug discovery, protein folding, clinical trials

Financial services - risk modeling, algorithmic trading, compliance

Materials science - battery tech, semiconductors, carbon capture

Legal - contract analysis, regulatory compliance, patent search

These markets have:

High switching costs

Regulatory barriers to entry

Massive economic value per use case

Less competition from Big Tech (for now)

The Counter-Argument

Frontier LLMs aren't going away. They have advantages DSFMs can't match:

General reasoning - can handle novel situations

Transfer learning - knowledge from one domain helps in others

User experience - natural language interfaces everyone understands

Platform effects - can be the base layer for thousands of applications

The future likely includes both, but the value creation might be more concentrated in DSFMs than current attention allocation suggests.

The Timing Window

We're in a unique moment where:

Training costs for DSFMs are still accessible to well-funded startups

Domain datasets are available but not yet monopolized

Enterprise buyers are sophisticated enough to understand the value prop

Regulatory frameworks are still forming

This window won't last forever. In 3-5 years, the best opportunities in DSFMs will likely be captured by either specialized startups or the Big Tech companies once they turn their attention to verticals.

My (Loosely Held) Thesis

The next wave of $10B+ AI companies will be built on DSFMs, not frontier LLMs.

The math is just too compelling: lower training costs, higher margins, stronger moats, less competition, and massive addressable markets in industries that desperately need AI but can't use general-purpose models effectively.

But I'm open to being wrong. The beautiful thing about markets is they'll tell us who's right.

What do you think? Are you seeing more enterprise value in specialized models or general-purpose ones? Hit me up on Twitter or LinkedIn - I'd love to hear your perspective.