Models & Foundation Labs

Labs developing state-of-the-art proprietary models without releasing weights. API access gates the most capable reasoning, multimodal and agentic functions. OpenAI, Anthropic, Google DeepMind and xAI set market terms; their pricing and rate limits constrain what applications can afford.

Organisations that train and publicly release model weights, enabling self-hosting, fine-tuning and redistribution. They erode pricing power of closed-model providers and complicate US export controls on AI. Value capture is indirect—structural disruption rather than direct margins.

Providers that host third-party models, offer fine-tuning, and serve via API without developing the models themselves. Inference prices fell roughly 90% from 2022 to 2025, forcing competition on latency, hardware efficiency, and enterprise SLAs. Survivors win on operational execution, not model ownership.

Curated corpora, human annotation pipelines, and synthetic data for training and fine-tuning models. Public web data is saturating; proprietary and synthetic sources increasingly determine model quality. Data owners and synthetic-generation platforms are capturing emerging value.

Labs building models specialised by modality or scientific domain, not generalist frontier labs. Defensible where proprietary domain data, simulation environments, or expert feedback loops create moats. Biology and robotics are prime examples.

Services that test models for capability, failure modes, and adherence to human intent before deployment. Regulatory mandates in the EU and elsewhere are making this compulsory. Value flows to accredited testing firms and consultancies with government or enterprise contracts.