Meditation on Banking

There is a conversation happening in finance right now that is not getting airtime amid the AI noise. The headlines fixate on which banks licensed which model, which boutique trimmed its analyst class, which workflow got automated last quarter. But this is just surface noise.. The real story is deeper: what survives, and why the core of direct lending and private credit looks archaic on purpose, and will stay that way.

The frontier AI labs are spending at a scale with no precedent in the history of capital allocation. Microsoft, Google, Meta, Amazon, and Oracle combined will spend something approaching $700B on AI infrastructure in 2026, with cumulative hyperscaler capex through 2027 projected past $1T. Inference costs have fallen roughly tenfold per year for the last three years. Context windows have gone from 8k tokens in 2022 to over a million in current frontier models. Agents are running multi-step workflows without supervision. These labs are not going to leave profitable, scalable verticals alone. Law, accounting, consulting, banking, healthcare administration, software engineering, marketing, design. Every credentialed knowledge worker sitting on top of a process that is fundamentally pattern recognition on text and numbers is in the kill zone. The personal finance apps were the appetizer. The professions are the main course.  Most of us already know this.

You can see the same drama playing out in law. Harvey AI has gone from five to $11B in valuation in a year as the legal vertical's answer to foundation model capability. The pitch was that domain expertise, workflow integration, and law firm relationships would create defensible value on top of GPT and Claude. The uneasy question is when the foundation models simply absorb it.

Every quarter the base models eat more of the wrapper's stack. The labs have orders of magnitude more capital, better researchers, and direct relationships with the same firms. The expertise wrapper was a 2023 thesis. It does not survive the next two model generations intact. The likely exit may be an acqui-hire by a foundation lab, not independent dominance. The same dynamic is coming for every vertical wrapper in every industry, and banking is no exception.

What gets eaten is the production work. Drafting, modeling, deck building, research, diligence response coordination, document review, memo writing, contract markup, summarization. This is happening now, not in 2028. The pyramids that organized white collar work for the last fifty years, the ones that took bright 22-year-old finance grads and put them through five to seven years of grunt work to produce the next generation of partners and managing directors, are being structurally dismantled. Not because anyone planned it, but because the economic logic of paying a summa cum laude Wharton graduate $200k+ a year to do work a model can do for a few dollars an hour is collapsing in real time, and any firm that does not capture the margin gets undercut by one that does. Goldman, Morgan Stanley, JPMorgan, and the major banks have all signaled reductions in analyst hiring and broader junior headcount over the last two years.

This creates a training problem that no one in the credentialed professions has a credible answer to, and it is the question I find most interesting right now. The senior partners and MDs who taught me the business were made by a decade of doing terrible work nobody else wanted to do. The first year associate who builds the cap table from scratch, the second year who sits through every diligence call, the third year who drafts the credit agreement and gets it red lined to death by the other side, that is how judgment gets built. You cannot read your way to it. You cannot prompt your way to it. You build it by being in the room a thousand times for the conversations nobody writes down, watching how senior people handle the moments that matter, getting things wrong and being corrected, accumulating the pattern recognition that lets you know which deal is real and which is noise.

If the AI does that work, you get a generation of seniors who can supervise output they cannot produce, and twenty years from now there is nobody left who can tell when the model is wrong. The bar associations and medical boards and accounting bodies have no plan for this. The likely outcome is that the current senior cohort runs out the clock on legacy training, the professions hollow out underneath them, and either deliberate apprenticeship gets rebuilt or the credential becomes ceremonial. The people who already did their reps have something the next generation, structurally, cannot acquire on the same path. That is not a small thing. It is the kind of structural advantage that compounds quietly for the rest of a career.

So that is the destruction half. The construction half, the part that does not get talked about enough, is where the moats actually are.

There is a structural feature of merchant banking, advisory, direct lending, special situations, and the entire universe of private credit and bespoke deal making that is invisible from outside the industry and obvious from inside it. Global private credit AUM has grown from roughly $400B in 2015 to over $2T+ today, with credible projections from Preqin, Morgan Stanley, and BlackRock on track toward $3T by the late 2020s. Direct lending is the dominant strategy, accounting for over half of total private credit AUM. The moat is not technological. The moat is that the participants do not want transparency, and they have the market power to prevent it.

Consider what would have to be true for AI to disrupt the placement of a $150M credit facility with an institutional lender. The borrower would need to publish their financials and structure preferences on a neutral platform. The lenders would need to bid against each other with full information. The intermediating banker would need to be replaced by a matching algorithm or direct portal. Every one of those steps requires the participants to want the outcome. None of them do.

The borrower does not want a public auction because they are revealing strategic information about their capital structure, their covenants, their growth plans, and the specific operational details that went into the diligence materials. The lender does not want a public auction because their entire value proposition to their own investors is that they see proprietary deals on proprietary terms. If a lender's LPs could see that every deal the lender did was also offered to twelve other funds on identical terms, the lender's fee structure becomes very difficult to justify. The entire industry runs on the premise that each fund has differentiated origination, differentiated underwriting, differentiated structuring. Some of that is real. Much of it is performance. The performance is required for the fee structure, and the fee structure is required for the industry to exist in its current form. And occasionally, the smartest guy in the room is smarter than the other smartest guy in the room. Or lucky.

So what you get, and what persists for as long as anyone reading this is alive, is a process that looks archaic from the outside and is preserved deliberately from the inside. Bankers email teasers to fund analysts. Fund analysts walk the teasers to their investment committees. Committees ask questions. Bankers answer them. Term sheets get drafted and redrafted. Phone calls happen on Sunday nights. Senior people fly to dinners. None of this needs to happen in the technological sense. All of it needs to happen in the institutional sense, because each participant requires the others to maintain the bilateral, opaque, relationship driven structure that is their reason for being.

I recently asked a senior guy at one of the largest private credit platforms if he thought an AI powered, more efficient market process was inevitable. His response: we would never do that.

That is the whole argument in five words.

Fees may come down. Fees probably will come down, because the production work behind the deal is getting cheaper and the lenders know it. But the process will remain archaic. Every fund has to demonstrate to its LPs that it sees better deals, structures better terms, and accesses opportunities competitors do not. The moment a fund stops being able to make that claim credibly, the fund stops raising capital. That dynamic does not change because the underlying analytical work got automated. It might actually intensify, because when the analytical work is commoditized, the only remaining differentiator is origination and judgment, and the only way to demonstrate origination and judgment is to keep running the bilateral process that demonstrates you saw something the others did not.

The principle is this: disruption happens where the seller wants transparency, and persists where the seller wants opacity. Where the asset is standardized and the seller wants maximum price discovery, the process gets platformed. Stabilized institutional real estate, single tenant net lease, performing loan portfolios, energy minerals packages, secondary fund interests at scale. These have moved to digital auctions, broker run electronic platforms, standardized data rooms. Fee compression there is real. The function survives because sellers still want a managed process, but margins are thinner and cycles faster. Speed and execution certainty become the differentiators, not relationships.

What does not get disrupted is the bilateral world. Distressed credit, growth equity, sponsor finance, project finance with structural complexity, rescue financing, anything where the deal is fundamentally one borrower talking to one or two or three potential capital providers through an intermediary who knows everyone in the room. The intermediary's value is the relationship, the judgment, the trust. None of that scales with model size. Middle market direct lending facilities still close in four to eight weeks, precisely because there is no syndication circus. The premium the borrower pays for that speed and certainty is the entire economic basis of the industry, and it is not going away.

This is the part the people forecasting the death of finance get wrong. They look at the analytical layer, see it being automated, and conclude the industry is being disintermediated. The analytical layer is not the industry. The analytical layer is the cost of goods sold. The industry is the network of relationships, the trusted intermediation between parties who do not want to share information with each other but need a conduit, the judgment about which deal is real and which is noise, the situational creativity to structure something that works for both sides when neither could have built it alone. That does not platformize. That cannot platformize, because the moment you try to platformize it, the participants who are paying for the opposite of platformization walk away.

The implication for those of us who do this work is sharper than the obvious one. The obvious one is that we need to use the new tools, accept margin compression on production work, lean into relationships and origination. That is correct but insufficient. The deeper implication is that the structure of careers in this industry is going to change in ways that have not been fully internalized. The pyramid does not just compress, it inverts. Seniors with origination relationships and judgment built over decades become more valuable per head, not less, because they can do more with fewer people supporting them. The path from college to managing director gets harder, narrower, and more dependent on getting access to the few remaining apprenticeship seats early.

For the funds, the same logic applies one level up. The originating partner, the relationship MD, the senior credit officer with judgment, those seats survive and get more valuable. Analyst and associate seats compress. The middle management layer that translated between senior partners and the junior pyramid becomes structurally redundant in a way it has never been. Operating expenses come down. LP fee pressure rises. The industry consolidates around the mega funds. Apollo, Ares, Blackstone, KKR, Brookfield, and a handful of others now control a majority of new private credit capital formation. Smaller players struggle, get acquired, or find narrow niches where their senior people still carry differentiated relationships. The surviving funds are larger, leaner, more dependent on a thin bench of rainmakers whose personal books carry the franchise.

The end state, fifteen years out, is a financial services industry that looks superficially similar to today's. Same names on similar buildings, similar deals over similar dinners. But the internal composition is radically different. The pyramids are gone, or at least radically scaled back. Production work is done by machines. The remaining humans are at the top of their fields, doing the work that cannot be automated because the participants do not want it automated. Compensation per surviving seat is higher. Number of seats is much smaller. The path in is harder.

What does this mean for the people inside the work? It means orienting toward the skills that compound rather than the ones that commoditize. Judgment under genuine uncertainty, where the data is missing or contradictory and someone has to decide. Taste, the ability to recognize quality before it can be articulated. Relationship capital built over decades that cannot be reconstructed from training data. Negotiation in high stakes adversarial settings where reading the other side matters more than the analytical substrate. Originality in the strong sense, not recombination of existing patterns but actual novel framing of problems. Moral and fiduciary accountability, someone has to be on the hook. Leadership, getting other humans to do hard things together, a function of trust and presence that cannot be delegated to a model.

For young people considering this industry, the honest advice is uncomfortable. The work that built the previous generation of partners is no longer available as a training ground. The credential is necessary but no longer sufficient. The path is through getting close to a senior person doing the unautomatable work, learning by osmosis what cannot be taught by drafting, and building the relationships and judgment that constitute the moat. There are fewer of those seats than there used to be. They are harder to get. They are also more valuable than they have ever been.

For the rest of us, here is the meditation. The technology will not save the industry from itself, and the industry will not be destroyed by the technology. What will happen instead is that the parts of the work that were always pretextual, the production layer that justified the fee but did not constitute the value, will be stripped away. What remains is the actual work, the bilateral, opaque, relationship driven activity that was always the real product. The funds will keep emailing each other. The bankers will keep flying to dinners. The deals will keep getting done one phone call at a time. Meanwhile, the pyramid will keep flattening.

The market will never be perfectly efficient. It was not built to be. It was built by people who know other people, and that is how it will continue to operate long after the models have absorbed every word that has ever been written about it. The people who can use the new tools will benefit. The people who cannot will fall behind. But in this business, when all is said and done, you still have to know who to call.

That is the moat. It is older than any technology we have invented, and it will outlast every technology we are about to invent.