Goldman Sachs has been building internal LLM infrastructure since early 2023. What’s emerged from job postings, conference talks, and selective industry briefings is a picture of enterprise AI deployment under significant regulatory constraint — and the architectural choices that result.
The core constraint for Goldman: data cannot leave the firm’s infrastructure. No third-party API calls with client or transaction data. This immediately eliminates the standard “call OpenAI’s API” deployment pattern that most AI coverage assumes. Goldman’s AI stack is built on self-hosted models, primarily running on their own GPU infrastructure, with frontier model access through dedicated private tenancy arrangements with specific providers.
The interesting choices: Goldman runs a tiered model architecture. Commodity tasks (summarization, classification, extraction) run on smaller fine-tuned models where they control the training data and can certify the outputs don’t encode material nonpublic information in a regulatorily problematic way. Higher-complexity tasks route to larger models with more oversight in the workflow.
The workforce implications are more nuanced than press coverage suggests. AI tools in investment banking are primarily accelerating analyst work — memo drafting, data pull structuring, comparable analysis formatting — rather than replacing judgment. The work that requires human sign-off (valuations, risk assessments, client advice) remains human, with AI tools improving the throughput of the supporting work.
The lesson for regulated industry AI buyers: the architecture required for compliance in financial services adds 12-18 months to deployment timelines compared to greenfield. If you’re planning enterprise AI for a regulated industry, start the compliance work before the technical work.