AI Moats Are Still Verticals - Late 2025.

By Didier Vila, PhD, Founder and MD of Alpha Matica.

The age of worshipping ever-larger general-purpose models is giving way to a far more profitable reality. By late 2025, the frontier labs still command headlines with trillion-parameter reasoning engines, but the companies actually capturing enterprise budgets—and building durable, high-margin businesses—are doing something radically different. They are going vertical and going small.

General-purpose LLMs have become commoditised infrastructure: powerful, accessible, and rapidly falling in price. The new frontier—the place where real economic moats are being dug—is in hyper-specialised models that deliver superhuman accuracy in painfully narrow domains while running at a fraction of the cost of yesterday’s giants. This is no longer a theory. It is the dominant pattern in enterprise RFPs, Series C/D term sheets, and production deployments in 2025.

Vertical Depth

In every regulated or high-stakes industry, the value curve is brutally non-linear. Jumping from 90% to 99% accuracy is not a nice-to-have; it is the difference between a model that sits in a sandbox and one that becomes mission-critical.

A frontier model can pass the Bar Exam or the USMLE in aggregate, but it will still hallucinate obscure Delaware bankruptcy precedents, misread a specific hospital’s clinical pathways, or miss subtle trading-desk jargon. Enterprises paying seven- and eight-figure ACVs demand perfection on their data and their workflows.

The new moat is built in three layers:

• Proprietary, decade-deep datasets that no one else has (sensor logs, annotated scans, trading transcripts, contract archives).

• Aggressive domain-specific continued pre-training and fine-tuning on those datasets.

• Workflow embedding so deep that the model isn’t a chatbot—it’s a new piece of core infrastructure.

  
  

The result: vertical specialists that outperform even the largest general models in their niche while being impossible to replicate without the underlying data flywheel [1, 4].

Illustrative Examples in Production (Late 2025)

The following table shows leading startups identified by investors as pioneers in the vertical trend. The metrics shown are projected or generalised achievements based on the market trajectory established in early 2025, not direct financial quotes from any single reference [1, 4].

Vertical Model / Startup	Industry / Niche	Key Features & Impact (2025)	Funding / Deployment Highlights (Projected)
Harvey AI	Legal (contracts, bankruptcy)	Fine-tuned on 10M+ proprietary contracts; <1% hallucination on case law; 70% faster review	Estimated $300M+ Series D in late 2025 based on growth;
Abridge	Healthcare (clinical notes)	Audio-to-notes, hospital-specific pathways, HIPAA on-prem (single H100), 99% transcription accuracy	Identified as a leader by BVP [1]; Projected $150M Series C in 2025; 200+ hospitals
BloombergGPT	Finance (trading & risk)	Distilled frontier reasoning on decades of market data; 40% better risk forecasts	Internal, deployed at 60%+ of top global banks
EvenUp	Legal (personal injury claims)	1M+ case archive; automates 85% of demand letters; 4× faster settlements	Identified as a leader by BVP [1]; Projected large Series B in 2025; 1K+ law firms
Articul8 DSM	Manufacturing (supply-chain compliance)	ERP-log fine-tuning for tax classification; 60% error reduction, VPC deployment	Estimated $80M Series A in 2025, based on enterprise pilot success
Outset AI	Market Research (survey synthesis)	Proprietary transcript flywheel; 95% actionable insights, 50% faster analysis	Estimated $50M Seed in early 2025, capturing major CPG/finance RFPs

These deployments routinely secure commercial partnerships because they solve the last 1% that general models cannot [1].

Efficiency: The Silent Winner

The 2023–2024 AI startup graveyard is paved with gorgeous demos that burnt cash on inference. Relying on frontier API pricing produced negative gross margins the moment any customer reached real scale.

In 2025 that equation has flipped. Models in the 3B–14B parameter range—exemplified by the Microsoft Phi-3/4 series, Llama-3.2/3.3 dense models, and Mistral Nemo—now deliver 80–95% of frontier performance on most enterprise tasks. This performance level is achieved while running at a tiny fraction (often 1/50th to 1/100th) of the inference cost [2].

The real-world goal is efficiency and security. On-premise, VPC, or edge deployment is now table stakes for any regulated customer. Latency drops from seconds to milliseconds; data never leaves the firewall; costs become predictable and tiny.

The Inescapable Convergence by Distillation

The most sophisticated teams are not abandoning frontier models—they are using them as expensive teachers rather than production servers.Classic knowledge distillation [3], turbocharged with modern synthetic data pipelines, is now standard:

• Frontier models label proprietary data, generate synthetic edge cases, and verify outputs.

• That knowledge is distilled into compact 7–14B student models that inherit the teacher’s reasoning style but cost pennies per million tokens to run.

  
  

The outcome is a specialist that thinks like a trillion-parameter model yet deploys like a spreadsheet macro. Vertical focus and radical efficiency are not two strategies—they are the same strategy. A model that only needs to master Chapter 11 plan objections in the District of Delaware does not require parameters for the history of the Roman Empire.

The 2025 Playbook

1. Choose an excruciatingly specific, high-value niche.

2. Secure proprietary data + workflow access.

3. Distill frontier reasoning into a small, domain-saturated student.

4. Deploy on a single GPU and charge like SaaS (leveraging the 65–85% gross margins observed in leading vertical AI companies [1]).

Conclusion

General-purpose models are the new electricity: foundational, ubiquitous, and best consumed indirectly. The companies that will build the next wave of enduring, multibillion-dollar AI businesses are the ones treating intelligence as precision manufacturing—distilled, specialised, and ruthlessly efficient.

References

[1] Bessemer Venture Partners. “Part I: The future of AI is vertical.” (September 2024). The core thesis for high margins, ACVs, and company selections (e.g., Abridge, EvenUp) are derived from this analysis.

[2] Microsoft Research. Phi-3 Technical Report, arXiv:2404.14219 (April 2024) and subsequent releases. This paper establishes the successful architecture for small (3B-14B parameter) models achieving frontier-level performance in a highly efficient manner.

[3] Hinton, G., Vinyals, O., & Dean, J. “Distilling the Knowledge in a Neural Network.” (2015) arXiv:1503.02531.

[4] CB Insights. “AI 100: The most promising artificial intelligence startups of 2024” (April 2024). This report provides market context and validation for the overall investment trend in specialised AI startups.