OpenAI’s latest job listing is hardly an ordinary one. The company is seeking a subject matter expert – an investment banker with at least two years of experience – to join its Applied AI team in San Francisco. The base salary ranges from $185,000 to $205,000, plus equity that, given the company’s current growth trajectory, carries significant weight, as Business Insider reports.

The move signals a precise shift: large language models (LLMs) are entering a phase where deep specialization becomes the real competitive differentiator. A generic LLM answering broad questions is no longer enough; to create value in high-stakes fields like investment banking – with its niche jargon, multi-million-dollar transactions and hypersensitive data – top-tier human expertise is essential for fine-tuning and model alignment.

Hiring a domain professional isn’t an isolated case. In recent months, several AI developers have been recruiting doctors, lawyers, and financial engineers to train their systems. The logic is clear: to automate tasks currently performed by highly paid specialists, you must first encode their tacit knowledge through high-quality feedback. In technical terms, this is Reinforcement Learning from Human Feedback (RLHF) applied to vertical domains.

But this evolution has a second, less visible consequence that directly impacts infrastructure deployment decisions. An investment bank wanting to use an AI assistant to analyze confidential documents, craft pitch books, or assess market risks would hardly trust a third-party cloud API. Financial data is subject to strict regulations (GDPR, internal policies, compliance mandates), and sending it to an external server is an unacceptable risk for many institutions.

That is why increasingly powerful financial-domain models will drive demand for self-hosted, on-premise, or private cloud solutions. Enterprises and financial institutions will need to run inference and fine-tuning on hardware under their direct control, preserving data sovereignty. This scenario isn’t far-fetched: today, frameworks and toolchains already exist to quantize open-source models (like Llama 3 or Mistral) and deploy them on local servers with consumer or enterprise GPUs. The main hurdle is less about technology and more about in-house expertise and Total Cost of Ownership (TCO) compared to API consumption.

Hardware vendors – NVIDIA with its datacenter GPUs, but also AMD and emerging dedicated accelerators – stand to benefit. Meanwhile, interest will grow in orchestration solutions that simplify on-premise LLM deployment, from Docker containers to Kubernetes clusters. For those evaluating on-premise deployment, significant trade-offs exist between capital costs, maintenance, and control benefits; our analysis on /llm-onpremise examines these aspects in detail.

OpenAI’s move, then, should be read as a wake-up call for the entire financial ecosystem: when a leading AI company hires a banker to teach the trade, the industry must brace for a radical change in how work gets done. It’s no longer a question of whether AI will enter the trading floor, but who will control the data and the models that power it.