Something shifted in AI over the past year, and most people outside the industry haven’t noticed yet. We’re moving away from the “one model to rule them all” era into something that looks more like the early internet—lots of specialised tools for specialised jobs.

The big foundation models—GPT-4, Claude, Gemini—are still improving. But they’re hitting a kind of ceiling on certain tasks. Not because they can’t get smarter in general, but because being good at everything means you’re not optimised for anything in particular.

Enter domain-specific foundation models. These aren’t fine-tuned versions of existing models. They’re built from scratch with specialised training data, architectures, and objectives. And in their domains, they’re starting to absolutely smoke the general-purpose competition.

Why General Models Plateau

Think about it this way: if you’re training a model on the entire internet, you’re teaching it to be conversational, to write code, to explain science, to generate creative content, to reason about philosophy—the list goes on.

That breadth is genuinely useful. But it comes at a cost. The model’s capacity gets spread across millions of different skills and knowledge domains. For any specific task, a huge portion of what the model “knows” is just dead weight.

Medical diagnosis doesn’t need knowledge of JavaScript frameworks. Legal contract analysis doesn’t benefit from understanding meme culture. Code generation doesn’t need an opinion on 18th-century literature.

General models carry all this baggage because they don’t know in advance what you’ll ask. Domain-specific models can dump 90% of that and go deep on what actually matters for their use case.

The Performance Gap Is Real

The data on this is starting to get hard to ignore. Bloomberg’s BloombergGPT, trained specifically on financial data, outperforms GPT-4 on financial analysis tasks by meaningful margins. Not 5% better—we’re talking 30-40% improvement on specialised benchmarks.

Hippocratic AI’s healthcare model is showing similar gaps in clinical reasoning. When tested on medical licensing exam questions and diagnostic scenarios, it’s not just beating general models, it’s approaching human specialist performance in ways that general-purpose AI hasn’t.

We’re seeing it in scientific research too. Models trained specifically on protein sequences and biological data are predicting protein structures and drug interactions with an accuracy that generic models can’t match, even when prompted with extensive domain knowledge.

This isn’t surprising if you think about it. We don’t train general surgeons and expect them to be as good as neurosurgeons at brain surgery. Specialisation matters in human expertise. Turns out it matters in AI too.

The Economics Are Shifting

For a long time, building a foundation model from scratch was prohibitively expensive. You needed hundreds of millions of dollars, access to massive compute clusters, and teams of ML researchers. Only tech giants could play.

That’s changing for a few reasons. First, training techniques have improved. Methods like mixture-of-experts architectures and more efficient optimisation mean you can train powerful models with less compute than you could three years ago.

Second, if you’re going domain-specific, you need less data. A medical model doesn’t need to learn the entire internet—it needs high-quality medical literature, clinical notes, and diagnostic data. That’s a smaller, more curated dataset, which means cheaper and faster training.

Third, the value proposition is clearer. If you’re a pharmaceutical company, a model that’s 40% better at drug discovery isn’t just nice to have—it could be worth billions in faster development cycles and higher success rates. The ROI on building a specialised model starts to make sense in ways it never did for general-purpose AI.

Industry Vertical Examples

Let’s get concrete about where this is happening:

Healthcare: Beyond Hippocratic AI, we’re seeing specialised models for radiology, pathology, and clinical decision support. These aren’t just better at medical knowledge—they’re trained with clinical workflows in mind, understanding the difference between academic knowledge and bedside decision-making.

Legal: Models trained on case law, contracts, and legal procedures are getting genuinely useful for contract analysis and legal research. They understand legal reasoning patterns that general models miss, like how precedent works or how to interpret ambiguous statutory language.

Finance: Trading firms and banks are building models that understand market microstructure, financial regulations, and risk management in ways that ChatGPT never will. These models can reason about market dynamics, not just regurgitate financial definitions.

Code: We already saw this with Copilot and CodeLlama, but it’s going further. Models trained specifically on high-quality codebases with verified correctness are outperforming general models on complex programming tasks.

Scientific Research: Materials science, chemistry, climate modelling—anywhere you have deep domain expertise and specialised data, vertical models are emerging.

The Open Questions

This shift raises some interesting questions we don’t have answers to yet.

How narrow should these models be? Is there a “heart surgery AI” or just a “medical AI”? The economics push toward broader domains, but performance improves with narrower focus. Finding the right granularity is still being figured out.

Can you composite them? If you have a legal model, a financial model, and a general reasoning model, can you route questions intelligently between them? Early experiments suggest yes, but orchestration is hard.

What about regulation? If domain-specific models become the standard in healthcare or finance, do they face different regulatory scrutiny than general-purpose ones? Probably, and that framework doesn’t really exist yet.

What This Means Practically

If you’re in an enterprise trying to deploy AI, this trend matters. The “just use GPT-4 for everything” strategy might not be optimal much longer. For high-value, specialised tasks, vertical models are increasingly going to be the better choice.

But there’s a timing question. Most industries don’t have mature vertical models available yet. Building one yourself is still expensive and hard. So you’re in this awkward transition period where the future is visible but not quite here yet.

My read is that 2026 is the year this tips. We’ll see domain-specific models become available as commercial offerings in enough verticals that enterprises will start treating them as the default for specialised work, with general models as the fallback for everything else.

The age of AI-for-everything is ending. The age of AI-for-something-specific is beginning. And if the early results hold, that’s going to make AI a lot more useful for a lot more real work.