Financial services in 2036, argued from invariants — fewer, larger, more universal banks, with tokens and AI as concentrating forces
Ten-year forecasts in finance are usually technology extrapolations wearing suits: take a capability curve, bend the industry around it, publish. This essay forecasts from invariants instead — what law fixes, what balance sheets require, what politics will not permit, and what fixed costs do to populations of firms. Two companion analyses established that the blockchain’s necessity case failed on its merits (A Critical Appraisal) and that tokenised markets are nonetheless being absorbed by incumbent infrastructure (Quo Vadis). This essay extends the same logic ten years forward, to the institutions themselves.
The conclusion inverts the disruption narrative. In 2036 there are fewer banks, larger and more universal than today, because tokenisation and artificial intelligence both raise the fixed-cost share of finance and both reward licence scope. The banking bundle — deposits, payments, credit, brokerage, custody, advice — holds at the level of institutional form even where individual firms fail; challengers that win are converted into banks by the very functions they capture — because technology unbundles functions faster than it unbundles liabilities. The selection variable is not size but metabolism: the capacity to migrate a live balance sheet off a batch-cycle core. The challenger wave, seen from 2036, was a recycling mechanism — the process by which the industry’s technological backbone was forced to an efficient level. And the forecast’s own failure mode is named rather than hedged: everything here holds exactly as long as the regulatory perimeter does.
Financial futurology fails in one repeatable way: it extrapolates capability and ignores constraint. The capability curve is real — settlement can be atomic, advice can be generated, credit can be scored by machines — but industries are not shaped by what technology can do. They are shaped by what survives contact with law, funding, politics and cost structure. So the method here is deliberately austere. Four invariants, each of which has held through every technological regime since the telegraph, do all of the argumentative work:
Every claim in this essay traces to one of these four. Claims that would require guessing at technological or political choices are quarantined in Section VIII, labelled as the forecast’s failure modes rather than smuggled in as hedges. That discipline is what separates a forecast from a prophecy — a prophecy tells you what will happen; a forecast tells you what would have to break for it not to. One scope note: the unit of analysis is the bank, not the whole of financial services — insurers, asset managers and market infrastructures appear here as funding substitutes and absorption vehicles, because the contested object of the next decade is the banking bundle itself.
The object under forecast is the universal banking bundle: deposits, payments, credit, brokerage, custody and advice, held together under one licence and one balance sheet. The bundle is old, widely resented, and repeatedly declared obsolete. To ask whether it survives, ask first what holds it together — because every disruption theory of the past thirty years is, at bottom, a claim that one of three glues dissolves.
The first glue is regulatory: licence scope makes adjacency cheap. An institution already supervised for deposit-taking adds custody or brokerage at the marginal cost of a permission, while an outsider must build the compliance stack from zero. The second is funding: deposits fund loans below wholesale cost, which is why the deposit franchise and the loan book have travelled together since banking began. The third is liability: regulators prefer one consolidated, capitalised entity to supervise — one throat to choke; more precisely, regulation assigns accountability to entities that can hold capital, absorb losses and be examined — and they structure their rulebooks accordingly. Fintech unbundling was a bet against the first glue; market-based finance against the second; decentralised finance against the third. Score to date: the interface was successfully unbundled, the funding glue was partially loosened in corporate credit, and the liability glue was not touched at all.
One scope condition governs everything that follows, and stating it precisely prevents a category error. “The bundle holds” is a claim about institutional form — licensed, bundled, deposit-funded banking — not about the market shares of incumbent firms. When a challenger broker acquires a full banking licence — a licence its business model pulled it toward once custody, cash accounts and settlement were at stake — that is not the bundle dissolving; that is the bundle capturing another firm (ECB Banking Supervision, 2023). The form is conserved even where the firms are replaced. Much of what is described as disruption is, on inspection, recruitment. Two clarifications complete the frame. The bundle is an institutional architecture, not an organisational chart: a bank can outsource technology, distribution, servicing, even origination — what it cannot outsource is being the liability anchor. And beneath all three glues sits the sentence on which this essay stands: technology unbundles functions far faster than it unbundles liabilities. Who performs a function can change freely; who ultimately owes stays pinned.
The market-structure argument is made at length in Quo Vadis and imports in three sentences. The register layer of tokenised markets is going to the depositories by legal mandate, and the economics — asset servicing, collateral mobility, fund registers — to the custodian banks. On the cash leg, the official sector’s direction of travel is settled: central bank money anchoring settlement, tokenised deposits emerging as the principal wholesale private-money layer, stablecoins pushed to specialised roles (Hernández de Cos, 2026; BIS, 2026). What remains is to ask what this settlement does to the banking bundle over ten years, and the answer runs through the invariants. (The selection is a policy, not a law of nature; Section VIII prices the ways it could fail.)
First, the funding glue is strengthened by official design. Whatever a tokenised deposit’s technical format, the liability underneath it is a deposit — a claim on a regulated bank — and deposits are a licensed product: the instrument selected to carry the digital economy’s payments is one only the licensed perimeter can issue. The token era does not disrupt the deposit franchise; it re-platforms it, extends it to programmable rails and around-the-clock operation, and hands it back to the same licence. Whatever else changes by 2036, the strategic core of banking — gathering insured deposits and lending against them — emerges from tokenisation more entrenched, not less.
Second, the fixed-cost invariant starts doing its quiet work. Tokenised settlement infrastructure, digital custody, always-on payments — these are fixed costs of considerable size, and they shift bank cost curves further from variable toward fixed. The official sector has itself flagged the distributional consequence: tokenised-deposit ecosystems that develop as proprietary silos favour the institutions that can build them and squeeze those that cannot, with smaller banks at a structural disadvantage (Hernández de Cos, 2026). Concentration is not a side effect of the token era. It is the mechanism — and an unglamorous one: continuous compliance, custody, connectivity and liquidity are fixed costs; fixed costs favour balance sheets that can amortise them; everyone else shares, rents, or sells.
The disruption narrative got the direction of concentration exactly backwards: the technologies of decentralisation are centralising finance.
Artificial intelligence attacks a different layer of the bundle, and the liability invariant decides where it lands. A model can compute a mortgage decision, generate investment advice, price a derivative, and run a compliance review — but it cannot owe anyone any of these. Advice creates liability; underwriting creates exposure; execution creates obligations. Each requires a suable, capitalised, licensed counterparty at the end of the chain, and no amount of capability moves that requirement, because it was never a statement about capability — the load-bearing assumption of this entire forecast, and Section VIII prices it. The consequence is that AI does not migrate financial functions out of licensed entities; it migrates them into machines inside licensed entities. The provider of the model can perform the function without becoming the institution — what cannot move is where the obligation lands. The licence becomes the wrapper within which the models are allowed to mean what they say — the argument is made at essay length in The Refusal Layer.
The cost-curve consequence compounds Section III. AI substitutes fixed costs — models, data infrastructure, evaluation and control frameworks — for variable ones, chiefly people. By the fixed-cost invariant, minimum efficient scale rises again, and the same concentration logic runs a second time through the same population. What AI genuinely commoditises is the interface and the middle office: onboarding, servicing, first-line advice, document handling — the layers where headcount lived and where challengers competed. But the interface was never the moat. The moat was always the licence, the balance sheet and the liability, and AI touches none of them.
One forward-looking element belongs here rather than in the failure-mode section, because it follows from the invariants rather than fighting them: agentic payments. Software agents transacting on behalf of clients — continuously, at machine speed, at three in the morning — presuppose programmable money and real-time rails, which ties the AI thread back into the token thread. And agents need principals: the moment an agent moves regulated money, extends credit or gives regulated advice, the perimeter has to attach somewhere — and under current law, somewhere is a licensed entity. The 2036 bank does not compete with the agents; it is the thing the agents are agents of.
Deposits and payments. Banks — fewer and larger. Tokenised deposits and instant payment mandates turn payments into an always-on scale utility with utility margins — interchange capped by regulation, scheme fees compressed by instant-payment rails priced near cost, float thinned by remuneration; the deposit franchise deepens as a moat precisely as it thins as a profit line. The retail client of 2036 holds a bank liability whether or not she has ever entered a branch, because the official architecture of digital money was designed so that she would.
Mortgages. Still banked — with a precision the first draft of this argument lacked. The banked element is the process: origination, servicing, collateral enforcement through national property and foreclosure law, and the regulatory interface of the most politically protected credit market in every democracy. The funding behind that process is less exclusive than the deposit-advantage argument implies, and honesty requires conceding it: long-dated fixed-rate mortgages are precisely the asset that insurance and pension liabilities want, and where mortgage cash-flows are standardised enough to match them, non-bank capital already funds them at scale — the Dutch pension-funded mortgage market is the standing proof, and insurance-backed lenders are probing the same door elsewhere. So the 2036 verdict is a split within the split: origination, servicing and the client relationship stay banked; funding is increasingly shared with insurance balance sheets along the corporate-credit pattern; and the deposit advantage defends the floating-rate and non-standard segments where liability matching fails. AI takes over valuation and decisioning throughout. The interface is contestable. The process is not.
Corporate credit. The honest concession, because the bundle genuinely loses here — and lost before either technology arrived. Private credit funds and institutional investors answer the funding invariant differently, matching long-dated loans to insurance and pension liabilities rather than deposits, and that leak continues through 2036. Banks converge on originate-to-distribute: they keep the relationship, the ancillary business and the distribution, and rent out ever less of the balance sheet. This is form-level change, it is real, and it is the strongest evidence that the bundle’s glues are separable — which is exactly why the forecast must explain why the same separation does not generalise fully to the household side. As the mortgage verdict above conceded, it partially does — insurance capital funds the standardised core of household credit too. What it cannot replicate is the banked process around that funding: origination, servicing, enforcement and the political perimeter of housing have no non-bank substitute, and corporate credit never had an equivalent perimeter to lose.
Brokerage and custody. The verdict here must be stated carefully, because the sloppy version is self-contradictory. Retail brokerage interfaces were lost to the challenger cohort — neo-brokers and the exchange-broker hybrids — and with the client relationship went the retail securities account. But look at what capturing it cost the challengers: either a full banking licence — one the business model pulled them toward once custody, cash accounts and settlement were at stake (ECB Banking Supervision, 2023) — or an omnibus account at an incumbent custodian, which keeps the incumbent as invisible back end. Retail custody, in other words, remains a banked function even where it changes firms — and it was never where the money was: custody per se earns close to nothing at retail, the economics sitting in order flow, spreads, lending and float. What banks lost at retail was the client, which matters, and the custody, which does not. Wholesale custody, meanwhile, has barely moved and is far harder to displace: fund depositary law confines the role to a narrow class of regulated institutions (UCITS V; AIFMD), sub-custody networks are decade-deep scale assets, and the tokenised-collateral economics pool exactly there. By 2036 the pattern is stable: interfaces churn, custody stays banked, and the challengers that won the interface have become banks in the process.
Advice. Generated by models, owed by licences. Mass-market advice becomes an AI product inside regulated wrappers, because personalised regulated advice is a liability, and liabilities need addresses — the merely informational layer commoditises without them. Human advice survives as a luxury segment — clients paying for a person precisely because a person is scarce — and as the eventual venue of accountability when the models are wrong.
Market-making. Outside the classic retail bundle but governed by the same invariants: AI-native and more capital-hungry, therefore fewer and larger firms. If tokenised markets become genuinely continuous, inventory exposure extends around the clock — and inventory requires financing — which loops back into the intraday funding machinery described in Quo Vadis. Liquidity provision in 2036 is a scale oligopoly wearing exchange connectivity.
Compressed, the verdicts read:
Seen from 2036, the challenger wave of the 2010s and 2020s was not an assault on the banking form. It was a recycling mechanism: the process by which the industry’s technological backbone was forced to an efficient level — efficient meaning real-time in capability, continuously reconciled, and priced at software rather than headcount unit cost. The cycle runs in three strokes. Entry brings modern cores inside the perimeter: the challengers that captured real banking functions arrived with real-time stacks and no legacy. Competitive pressure forces incumbents to modernise or shrink. Exit removes the institutions that can do neither, by absorption rather than failure. At the end of each stroke, the population of licensed banks is smaller and its average technical metabolism is higher. Selection pressure runs through the form, not against it.
Which makes the survival variable worth naming precisely, because the obvious candidate is wrong. The core banking system itself is not the differentiator: modern cores can be bought or rented from a handful of vendors, and software anyone can purchase is a hygiene factor, not a moat. The differentiator is migration capacity — the organisational ability to move a live balance sheet off a forty-year-old batch mainframe without detonating, a manoeuvre whose failure mode was demonstrated publicly when a mid-sized UK bank’s 2018 core migration locked customers out for weeks and ended in regulatory censure (FCA/PRA, 2022). Processes and procedures, in other words, are the differentiating factor precisely because they are the one part of the stack that cannot be procured. The graveyard of the next decade fills with banks that could not migrate, not banks that could not buy.
The batch cycle is the deep incompatibility that makes this existential rather than optional. A core that reconciles overnight cannot natively participate in markets that settle at four-hour granularity, serve agents that transact at three in the morning, or feed AI operations that presuppose real-time positions — it can only bolt a second, real-time architecture around itself, which is how the migration bill compounds — nor, for that matter, satisfy the risk-data aggregation supervisors have expected of systemically important banks since 2013, and increasingly of everyone (BCBS, 2013). Tokens and AI do not merely raise fixed costs; both presuppose a real-time core. Core modernisation is therefore not an efficiency programme. It is the admission ticket to every revenue line in Sections III through V.
Banks will not die of size. They will die of metabolism.
Size still enters, but only as the denominator of the migration bill. This reframes the fate of the mid-tier — the quiet macro story of the decade. The euro area has lost roughly a third of its credit institutions since the financial crisis, overwhelmingly through consolidation (ECB structural indicators), and the mechanism above predicts — this is the forecast’s riskiest specific — that the middle empties first: large banks can fund the migration, small banks can rent the stack, and the mid-sized institution is caught — too small to amortise its own modernisation, too large and too proud to run on someone else’s. Two qualifications keep the account honest. The politically protected sectors — the German savings and cooperative banks above all — survive through a combination of political protection, regional mandates and mutualised structure rather than stand-alone economics — and it is no accident that they industrialised pooled cores through shared IT utilities decades ago: mutualised metabolism is a genuine third path, available to those who built it in time. And the rented-core route carries its own sting: a small bank running on a vendor’s stack survives as a distribution shell, but renting your metabolism shifts part of the operating economics to the vendor. Survival and commoditisation arrive in the same box.
Underneath the consolidation sits its monetary mechanism, worth stating plainly: everything in Sections III and IV makes the banking licence itself more valuable. Tokens entrench the deposit franchise the licence guards; AI must operate inside the liability wrapper the licence provides; the charter becomes the scarce input of the whole system. That is why the graveyard fills by acquisition rather than failure — a bank that cannot migrate may still be worth its charter, and the acquirer is buying the licence and the deposits, not the mainframe. Rising charter value is the quiet price signal behind the decade’s mergers.
What disappears by 2036: the branch as a primary distribution channel — what remains of it serves advisory, private banking and brand theatre in expensive postcodes. The human interface at scale — routine interaction replaced by service that never sleeps and never varies, human attention retreating to exceptions, complex advice and the moments when trust must be rebuilt. The genuinely independent mid-tier — the mid-sized bank and the sub-scale national champion — contracting sharply, absorbed or merged sideways into regional groups. And the technology itself as a topic: “blockchain” and “AI” vanish from strategy decks into infrastructure inventories, the way “relational database” did — the terminal stage of every successful financial technology being invisibility.
What does not disappear: the licence, the deposit, the suable entity, the bundle. Nor, it should be said plainly, the crisis — a system of always-on money, instant redemption and machine-speed decision-making is a system in which the 2036 bank run happens in minutes rather than days — a resolution-machinery speed problem rather than a will problem, since deposit insurance holds by the third invariant; what must accelerate is the machinery that makes the promise credible within minutes — and supervisory architecture spends the decade retrofitting circuit breakers accordingly. The bundle holds; it does not follow that it holds calmly.
Every claim above traces to the four invariants, and so the forecast fails exactly where an invariant breaks. The load-bearing assumption is the first one: that liability must terminate in a licensed, capitalised entity — that the regulatory perimeter holds. The genuine uncertainty of the next decade is whether agentic AI forces the creation of a new licence category: payment institutions that think, credit intermediaries that hold no balance sheet, autonomous entities granted their own legal personality and capital requirements. If that happens, the liability glue can be re-formed outside banking law, and the bundle can be reassembled around a different chassis. This essay’s bet is that legislators, facing the choice between adapting a supervised perimeter they trust and chartering one they do not, choose adaptation — that is how every previous technology was absorbed. But it is a bet on political behaviour, not an invariant, and it is stated here as such.
Two lesser falsifiers deserve their sentence. A retail central bank digital currency deployed at scale and without holding caps would attack the deposit franchise from the public side — the one actor that could disintermediate the funding glue is the state itself; current designs cap holdings precisely to avoid this, and the forecast assumes that caution persists. And stablecoin dollarisation of weak-currency economies could build a parallel dollar system outside the two-tier architecture faster than the official sector can extend it (BIS, 2026) — a regional failure of the forecast rather than a global one, but a real one. A forecast that cannot name what would falsify it is a mood. These are the falsifiers.
So: the composite, coldly. Fewer banks — on the trajectory of the past fifteen years, illustratively another third fewer in Europe, though the direction matters more than the fraction — larger, more universal, and harder to distinguish from one another, each a licence holding a balance sheet, operated by models, wearing whatever interface the client last downloaded. Deposits re-platformed onto programmable rails and more entrenched than ever. Mortgages still banked; corporate credit still leaking to funds; brokerage interfaces churning above custody that never moves. The challenger generation either converted into banks or absorbed by them, having performed its historical function: forcing the industry’s metabolism to an efficient level and retiring the institutions that could not follow. The technologies that were sold as decentralising finance will have concentrated it; the revolution’s lasting monument will be a shorter list of larger banks with better plumbing.
The future of banking, on the evidence of every previous technological wave and the invariants that governed them all, is more banking — done by fewer banks, visible to no one.
Written in a personal capacity. Analytical views only — not legal or investment advice.
Julian Gretzinger — Investor and writer on monetary history, real wealth mechanics, and financial markets. substack.com/@juliangretzinger