#08 · Foundation Models & Inference Infrastructure
Best Domain-Specific Foundation Models
What is a domain-specific foundation model?
A domain-specific foundation model is a large language or multimodal model that has been pre-trained, continued-pretrained, or extensively fine-tuned on data from a specific knowledge domain — most commonly healthcare, legal, financial services, scientific research, or coding (though coding warrants its own category given its scale). The category sits between fully general frontier models (Claude, GPT-5, Gemini) and narrow task-specific models. The argument for domain-specific foundation models is that domain-specialized training produces better fluency with the domain's terminology, better calibration on domain-specific edge cases, fewer hallucinations on jargon and rare entities, and outputs that more naturally fit the domain's accepted conventions (legal citation formats, medical SOAP notes, financial filing structures).
Why this category is contested.
Domain-specific foundation models occupy a genuinely contested middle ground. The pace of general frontier model improvement makes the case for from-scratch domain pre-training harder to defend every year — by the time a domain model ships, a new general frontier model often closes most of the gap, and a general frontier model fine-tuned on the same domain data frequently matches or beats a domain model trained from scratch, at a fraction of the cost. The category has survived in specific niches: regulated domains where *indemnification, audit posture, and explicit domain training* matter as much as raw capability (healthcare under HIPAA, legal under privilege protection, financial under model risk management); scientific domains where specialized pre-training meaningfully extends model knowledge into territory general models don't cover (protein sequences, materials science); and vertical platforms where the foundation model is bundled with domain-specific retrieval, tooling, and workflow integration (Harvey, BloombergGPT, Hippocratic AI).
What to evaluate.
Buyers should be honest about whether they actually need a domain foundation model or whether a frontier model with domain RAG/fine-tuning suffices. The case for domain models strengthens when: (1) domain terminology is heavily specialized and general models hallucinate on it; (2) audit and indemnification posture matters for regulatory reasons; (3) the vendor bundles the model with domain-specific tooling and workflows that would be expensive to build; (4) the domain is genuinely outside general pre-training distribution (e.g., proteomics). The case weakens when the domain is well-represented in general training corpora and the foundation model adds marketing positioning without meaningful capability differentiation.
Healthcare foundation model with GCP integration
Google's Med-PaLM line, now productized as MedLM within Google Cloud Healthcare, was among the first credible healthcare foundation model families — notable for being the first to clear the US Medical Licensing Examination passing threshold and for tight integration with Google Cloud's healthcare data and compliance tooling. Available via Google Cloud Vertex AI with healthcare-specific data agreements. Best for healthcare workflows already running on Google Cloud, organizations using Google Cloud Healthcare APIs, and clinical decision-support applications wanting Google's research pedigree. Strengths include medical-specific fine-tuning, USMLE-level benchmark performance, deep GCP healthcare ecosystem integration, and Google's research and compliance posture. Trade-offs are that MedLM requires Google Cloud commitment to access, the model isn't generally available outside that stack, and pricing requires direct engagement rather than self-service.
Finance-domain foundation model on proprietary corpora
Bloomberg's BloombergGPT was a landmark 2023 release demonstrating that a major financial data company could train a competitive foundation model on its proprietary financial corpora. The model is positioned not as a standalone API but as the foundation for AI features within Bloomberg Terminal — newsroom drafting, financial analysis assistance, and structured extraction from financial documents. Best for organizations using Bloomberg Terminal where AI features are integrated into the platform, financial analysts wanting AI augmentation within Bloomberg's familiar workflow, and Bloomberg-standardized financial services firms. Strengths include training on Bloomberg's proprietary financial corpora (a moat general models can't easily replicate), integration with Bloomberg Terminal, and Bloomberg's enterprise sales motion and existing relationships. Trade-offs are that the model is platform-locked (not a generally available API), use is constrained to Bloomberg subscribers, and frontier general models with retrieval over public financial data increasingly close the capability gap.
Legal AI platform with deep law-firm adoption
Harvey, founded in 2022 and now valued at approximately $11 billion after a 2026 funding round, has become the dominant legal AI platform — counting Allen & Overy (now A&O Shearman), PwC, and many of the world's largest law firms as customers. The product combines fine-tuned models on legal corpora with workflow tooling for contract analysis, due diligence, compliance, regulatory analysis, and litigation support. Harvey is not strictly a foundation model in the pure sense (the underlying models are partly from frontier labs), but the company has built a defensible position around legal-specific training, workflow integration, and enterprise sales motion. Best for large law firms and corporate legal teams, complex legal workflows (M&A due diligence, regulatory analysis, contract families), and organizations wanting enterprise-grade legal AI deployment with white-glove support. Strengths include the deepest law-firm customer base in legal AI, strong enterprise compliance posture, broad workflow coverage, and continued aggressive funding for product expansion. Trade-offs are enterprise pricing that excludes smaller firms, narrow legal focus (not general-purpose AI), and the underlying model dependence on frontier labs (rather than a fully proprietary model).
Legal AI platform integrated with Westlaw and Practical Law
CoCounsel, originally built by Casetext and acquired by Thomson Reuters in 2023, brings legal AI capability into Thomson Reuters' broader legal research ecosystem — Westlaw, Practical Law, and the company's deep catalog of legal content. The integration with Thomson Reuters' authoritative legal corpus is the key differentiator: CoCounsel produces outputs grounded in licensed primary and secondary legal sources rather than the open internet. Best for law firms and corporate legal teams using Thomson Reuters' legal research products, litigation work requiring authoritative source grounding, and organizations valuing Thomson Reuters' enterprise compliance posture. Strengths include integration with Westlaw and Practical Law, grounding in licensed legal content, mature Thomson Reuters enterprise sales motion, and broad workflow coverage (research, document review, summarization, drafting). Trade-offs are Thomson Reuters subscription requirement, narrower than Harvey on some workflow areas, and Thomson Reuters' enterprise pricing model.
Patient-safety-focused healthcare foundation model
Hippocratic AI is purpose-built for healthcare with explicit patient-safety guardrails as a foundational design constraint — positioning itself for patient-facing workflows (chronic care management, post-discharge follow-up, medication management) where general frontier models would carry unacceptable safety risk. The company has built an extensive safety evaluation framework involving thousands of medical professionals. Best for healthcare patient-facing workflows, post-discharge care management, chronic care touchpoints, and any healthcare application where patient-safety risk is the dominant design constraint. Strengths include healthcare-specific safety posture, extensive medical-professional evaluation framework, narrow patient-facing focus that allows deeper safety guarantees than general models, and strong healthcare enterprise sales motion. Trade-offs are very narrow healthcare focus (not general healthcare AI), platform-locked deployment, and limited applicability outside patient-facing use cases.
Foundation model platform for biology and drug discovery
NVIDIA's BioNeMo is a platform of foundation models for biological domains — proteins, small molecules, DNA, and increasingly cell biology — purpose-built for pharma R&D, drug discovery, and synthetic biology workflows. The category is one of the clearest cases where domain foundation models genuinely outperform general frontier models, because biological sequence data is sufficiently different from natural language that general models simply don't cover it. Best for pharmaceutical and biotechnology research and development, drug discovery workflows, protein engineering, and academic and industrial bioinformatics. Strengths include category-leading biological domain coverage, NVIDIA hardware optimization, integration with NVIDIA Clara healthcare platform, and active model release cadence. Trade-offs are very narrow biology focus (not for general healthcare AI use cases), requires NVIDIA infrastructure for optimal performance, and assumes significant in-house bioinformatics expertise.
Vertical Granite variants with IBM indemnification
IBM's vertical Granite variants extend the family's IP-indemnification and enterprise governance posture into finance, compliance, and other regulated verticals. The pitch is straightforward: regulated enterprises that need domain-tuned AI but can't accept the indemnification and governance gaps of community open-weight models get an IBM-backed alternative. Tightly integrated with IBM watsonx for governance and lifecycle management. Best for regulated financial services and compliance teams wanting IBM indemnification on domain AI, watsonx-standardized organizations, and government and public-sector domain AI workflows. Strengths include IBM IP indemnification, mature enterprise governance tooling, watsonx integration, and clear regulated-industry positioning. Trade-offs are a smaller community than vertical specialists like Harvey or BloombergGPT, narrower workflow tooling, and the broader IBM enterprise pricing model.
Frontier protein language model for biological design
EvolutionaryScale, founded by former Meta FAIR researchers behind the original ESM (Evolutionary Scale Modeling) protein language models, ships ESM-3 — a frontier-class foundation model for protein sequences with applications spanning drug discovery, enzyme design, and synthetic biology. ESM-3 is one of the clearest examples of a domain foundation model that meaningfully extends what general LLMs cover, because protein language is genuinely outside natural-language training distribution. Best for pharmaceutical and biotech research, enzyme and protein engineering, synthetic biology workflows, and academic and industrial computational biology. Strengths include category-leading protein modeling, founder pedigree from the original ESM team, active research community, and credible applications in drug discovery and design. Trade-offs are very narrow biological focus, requires substantial computational biology expertise to use effectively, and is not a general-purpose AI tool.
Enterprise knowledge platform with vertical AI tuning
Glean, while not strictly a foundation model, has emerged as a vertical platform for enterprise knowledge work — combining frontier model access with deep enterprise integrations (Google Workspace, Microsoft 365, Slack, Salesforce, Confluence, Jira, etc.) and fine-tuning on each customer's corpus to deliver organization-specific AI assistance. The platform exemplifies the "vertical-tuned stack" pattern that increasingly competes with pure foundation models in enterprise use. Best for enterprise knowledge work and search, organizations with information spread across many SaaS systems, and enterprises wanting AI grounded in their own corpus rather than the open internet. Strengths include unmatched enterprise SaaS integration breadth, organization-specific tuning, mature enterprise sales and deployment motion, and clear ROI on knowledge work productivity. Trade-offs are that it's a stack rather than a raw model, pricing is enterprise-tier, and customization requires meaningful deployment work.
Vertical AI platform for financial and professional services analysis
Hebbia, founded in 2020, has built a vertical AI platform specifically for financial services, consulting, and professional services analytical workflows — combining frontier model orchestration with structured document processing for the long-form analytical work these industries depend on (investment memos, due diligence reports, M&A analysis, consulting deliverables). Best for financial services analytical workflows (private equity, hedge funds, investment banking), consulting firms doing client analytical work, and any organization where multi-document financial analysis is the core workload. Strengths include deep financial-services workflow integration, structured document processing built for financial corpora, and clear positioning for high-value professional-services work. Trade-offs are enterprise-tier pricing, narrower than general knowledge platforms like Glean, and dependence on frontier models rather than proprietary foundation models.