Comparative LLM Usage Across Sectors

Narrative

The academic literature on real-world LLM deployment has grown rapidly since mid-2025, with a cluster of empirical papers providing direct evidence on deployment shapes, cost structures, and workload patterns. A May 2025 arXiv preprint (Hou et al., arXiv:2505.02502) conducted an internet-wide scan identifying over 320,000 publicly exposed LLM services across 15 frameworks, revealing how quickly self-hosted inference has proliferated outside controlled enterprise environments, with Ollama, vLLM, and LM Studio among the most prevalent stacks. A concurrent August 2025 arXiv paper (Pan et al., arXiv:2509.18101) offered formal cost-benefit analysis for on-premise versus cloud deployment of Qwen, Llama, and Mistral variants, finding that break-even periods range from a few months for small models to roughly five years for large ones, with viability conditional on processing at least 50 million tokens per month or operating under strict data residency requirements. An SSRN working paper (Zhang, Shi, and Tang, June 2025, abstract 5296479) approached the same decision from an economic theory perspective, analysing how data privacy concerns and user heterogeneity shape equilibrium choices between cloud and on-premise deployment, and showing that open-source entrants may be structurally disadvantaged in competitive markets once incumbents offer localisation options.

On sector adoption, a January 2025 arXiv position paper (Xu et al., arXiv:2501.09906) analysed 201 foundation models and 6,198 arXiv papers, finding that closed LLMs led by GPT-4 dominate high-performance healthcare applications such as medical imaging and multimodal diagnostics, while open models including LLaMA gain traction for cost-sensitive tasks like mental health dialogue and patient communication. A broader August 2025 arXiv survey (arXiv:2508.19667) found specialised LLMs deployed across healthcare, finance, law, education, and manufacturing. The arXiv legal-agents survey (arXiv:2601.06216) and the critical-domains survey (Chen et al., arXiv:2405.01769) document active fine-tuning and RAG deployments in law and finance, but note persistent accuracy limitations that constrain autonomous deployment. The academic evidence consistently separates sectors by compliance sensitivity: healthcare and finance prefer on-premise or private-cloud closed models for regulated workloads, while research and content functions absorb open-weight alternatives.

METR's empirical evaluations supply the most controlled evidence on real-world LLM productivity and autonomous capability. Their July 2025 randomised controlled trial (Becker et al., arXiv:2507.09089) found that experienced open-source developers using early-2025 AI tooling (Cursor Pro with Claude 3.5/3.7 Sonnet) completed tasks 19 per cent more slowly than without, despite self-reporting a perceived 20 per cent speedup. METR separately evaluated GPT-5 pre-deployment (August 2025) and determined its 50 per cent time-horizon on agentic software engineering tasks to be approximately two hours and 15 minutes, placing it well above earlier frontier models but below thresholds for catastrophic autonomous risk. Their HCAST benchmark infrastructure and RE-Bench suite for ML research engineering tasks now underpin several pre-deployment assessments of Claude, GPT, and DeepSeek models, providing consistent human-calibrated comparison baselines. The SWE-Bench and SWE-Lancer lineage (arXiv:2502.12115, arXiv:2509.16941) independently demonstrates that frontier models reach around 72 to 75 per cent solve rates on verified repository-level tasks, but that saturation is driving researchers toward harder benchmarks with real economic stakes.

On inference cost, Epoch AI's March 2025 analysis (Cottier et al.) documented price declines of 9x to 900x per year depending on the performance milestone, with the steepest drops occurring in the 12 months before publication. A December 2025 arXiv preprint (arXiv:2511.23455) formalised the price-performance relationship econometrically, introducing a tiered super-Moore's law hypothesis in which economy and mid-tier segments see prices halve roughly every 1.1 to 1.55 years. The arXiv survey on LLM routing (arXiv:2603.04445) and the ICLR 2025 RouteLLM paper document a growing research programme around cascading and routing architectures that direct simpler queries to cheaper models while routing complex ones to frontier systems, with routing overhead of under 0.4 per cent of GPT-4 generation cost. An April 2025 arXiv working paper (arXiv:2604.00626) on on-policy distillation found adoption across Qwen3, DeepSeek, Gemma, and other production pipelines as a core ingredient for compressing reasoning capability into smaller, cheaper models, reducing the economic case for always-on frontier API access. Despite these advances, independent measurement of actual enterprise spend remains sparse in the academic literature: most numeric claims derive from vendor surveys, analyst estimates, or industry blogs rather than audited procurement data.