Citigroup assembled a new AI Infrastructure Banking team last month, staffed with investment bankers whose job is not to use AI internally but to win advisory mandates from AI companies raising capital, doing M&A, and building data centre infrastructure. The team is positioned to capture fees from what is effectively an entirely new industry's capital markets activity — the OpenAI $110 billion rounds, the data centre construction bonds, the infrastructure SPACs. It is a classic Wall Street move: when a new industry generates enough transaction volume, build a specialist team to advise it. That Citigroup is doing this now, at this scale, is a useful signal of how large the AI capital markets opportunity has become.
The mirror image of that move is happening on the other side. xAI recently went to Bloomberg with news that it is hiring Wall Street credit analysts, portfolio managers, distressed investors, and leveraged loan specialists — not to build AI products for finance, but to teach Grok how finance actually works. The goal is genuine domain expertise in the structures and conventions of credit markets, which are notoriously opaque, heavily reliant on unstructured data, and resistant to the kind of automation that has transformed equities trading. Teaching an AI model to understand a leveraged loan covenant package or a distressed bond situation requires the people who know those things at a granular level. Hence the hiring.
JPMorgan's head of fixed income, Pranav Jhamna, made the case in March for why credit specifically is the right frontier: credit markets are "the last frontier when it comes to market automation," characterised by bespoke documents, private negotiations, and data that exists in PDFs and emails rather than in structured feeds. Equity markets have been largely automated for decades. Credit has resisted because the relevant information is unstructured and context-dependent in ways that previous generations of automation could not handle. Generative AI, which can parse and reason over unstructured text, changes that calculus — potentially dramatically. JPMorgan is positioning itself to be the institution that captures that shift internally, not just through advisory mandates.
Goldman Sachs is already further along. The bank has been working with embedded Anthropic engineers to build autonomous agents for trade accounting, transaction reconciliation, and client onboarding — areas where the work is high-volume, rule-governed, and currently requires significant human labour that could plausibly be automated. The framing in Goldman's communications has been careful — "early stages," "co-development," gradual deployment — but the direction is clear. Goldman is not waiting to see how AI changes finance. It is trying to be the institution that uses AI to reduce the cost structure of its own operations before its competitors do.
What is interesting about these four moves in parallel — Citigroup advising AI companies, xAI learning from Wall Street, JPMorgan targeting credit automation, Goldman deploying Claude in back-office operations — is that they describe the same phenomenon from four different vantage points. The boundary between financial services and AI is dissolving. AI companies are becoming large enough to be significant financial market participants (they raise more capital than most industries). Financial institutions are becoming significant AI deployers (they have the data, the capital, and the incentive). And the expertise required to operate in this boundary zone — understanding both how credit markets work and how large language models process unstructured text — is scarce and being competed for aggressively.
The practical question is which institutions end up capturing the most value from this transition. Citigroup's advisory play is the most defensive: fees from advising the winners, regardless of who wins. Goldman's internal deployment play is the most aggressive: using AI to reduce operating costs before the cost savings become table stakes. JPMorgan's credit automation ambition is the longest-horizon bet, but credit markets are large enough that even partial automation would be enormously valuable. xAI's move is the most uncertain: hiring Wall Street expertise to teach a model finance is a plausible strategy, but "teaching" a model domain expertise is not the same as actually having it, and the models trained this way will face competitors who have been operating in credit markets for decades. The race is real; the outcome is genuinely uncertain.