The Core That Cannot Hold
Core banking is the largest, least loved, and least replaced category in enterprise software. The conditions that protected it for forty years look like they will finally start breaking.
The Big Three core banking providers, Fiserv, FIS, and Jack Henry, sit underneath roughly 70% of US banks. Fiserv alone runs the core for more than 40% of US banks and 30% of credit unions. Between them, these three companies process the deposit ledger, the loan ledger, the GL postings, and the overnight batch that decides whether your debit card works in the morning. Two years ago they were worth $140 billion combined. Today they are worth $68 billion. The market may not be pricing a clean “core banking disruption” story yet, but it is clearly no longer giving the incumbents the benefit of the doubt. I think that matters because the next renewal cycle is arriving just as the definition of money is changing.
I suspect what the market has begun to notice is that the Big Three are sitting on platforms built between 1977 and 1989, before the world wide web was invented, and now money is becoming programmable. JPMorgan has cleared $2 trillion in notional value through Kinexys. Citi runs Token Services across its own branches. Partior, the bank owned consortium platform, has been doing atomic cross bank settlement of tokenized deposits in multiple currencies since 2021. The GENIUS Act, signed in July 2025 and taking effect in January 2027, gave payment stablecoins their own federal regulatory category. None of the dominant US cores were built for any of this. Most legacy cores were not designed to represent tokenized instruments as first class objects, with token specific lifecycle states, wallet ownership, rail aware settlement, programmability, and real-time control logic.
What stands out to me more than anything is this is the last large enterprise software category without a serious AI native challenger at scale, and that is surely going to change. But I doubt AI alone is the forcing function. Programmable money creates the architectural gap. AI may simply change the economics of closing it.
The cloud shift that never quite happened
The Big Three platforms in production today were built between the late 1970s and the late 1980s. Fiserv Premier launched in 1978. FIS Systematics in 1977, FIS IBS in 1980. Jack Henry SilverLake came in 1988. They were the wave that brought banks off the central mainframe and onto IBM mid range computing, and they have been the ledger of record for most US community and regional banks ever since.
Those platforms have run for forty plus years and absorbed every wave of new technology since. They absorbed Y2K, online banking, the 2008 crisis, and the early cloud era as integration layers rather than as triggers for replacement. The Big Three moved hosting models into private cloud, then started talking about AWS and Azure, then acquired or built next generation platforms: Finxact at Fiserv, Modern Banking Platform at FIS, the cloud native deposit core Jack Henry has been building since 2022. Challengers have appeared. Thought Machine won real reference customers including some GSIBs, but none of them broke through to scaled deployment across US community and regional banks. Temenos has won a handful of US deals after probably fifteen years of trying.
Tokenization is a different order of problem. Every previous wave changed how customers interacted with the ledger. Online banking added a web channel. Mobile added a phone channel. FedNow accelerated when the ledger had to update. The ledger itself stayed the same: a row in a database, denominated in dollars, posted overnight in batch. Tokenized money changes what the ledger represents. A tokenized deposit is not just a faster account entry. It is a tokenized representation of commercial bank money, recorded on ledger infrastructure that can support transfer, settlement, programmability, and audit in ways legacy account systems were not designed to handle. It clears in seconds against external rails rather than overnight in batch, carries conditional execution logic the legacy core cannot interpret, and moves in unit values legacy COBOL packed decimal fields cannot natively express. Adding a web channel to a 1980s core is an integration project. Making a 1980s core understand tokenized deposits is basically a rebuild.
The reason core banking has absorbed every previous wave without losing its franchise is that the moat was never really the product. The moat was everything around it. A typical US community bank core sits behind 30 to 60 vendor integrations, none of which migrate cleanly. Big Three contracts run seven years with termination fees that can run into eight figures. Working with Premier or Horizon requires platform specific knowledge that has no market outside it. The bank technology consultancies have hundreds of millions of dollars of practice revenue tied to Big Three migrations, and the analysts who cover the category are funded in part by the same vendors.
The moat works for the customer too. Many bankers have scare tissue since roughly only 30% of full core migrations have succeeded over the past decade. A regional bank CIO who renewed Premier or IBS in 2023 made a defensible call. Stable operations was the right answer for the conditions of the time. However, the conditions of the time are now changing.
What the Big Three just admitted
The clearest evidence that the conditions are changing is that the incumbents are themselves announcing it. Each of the Big Three has now shipped, or is shipping, a program that signals its existing platforms cannot do what banks need them to do over the next decade.
Start with Fiserv. In January 2026, Fiserv announced a strategic collaboration with Microsoft. Microsoft 365 Copilot rolls out across the Fiserv workforce. Microsoft Foundry on Azure becomes the platform for embedded AI, with stated targets across fraud detection, customer service, and operational efficiency. Fiserv has run more than 100 billion tokens through Foundry already, and 8,000 of its engineers are using GitHub Copilot. None of that is fake. Also none of it is a description of agents running production workloads inside customer cores. It is a platform partnership and an internal productivity push, with the client facing AI workloads on the future deliverables side of the slide.
FIS is a little more interesting. FIS has run two tokenization related plays in 2026, and they really tell you everything about the architectural problem. The first, in January, was an agentic commerce offering with Visa and Mastercard, tied to FIS’s $13.5 billion acquisition of Global Payments’ Issuer Solutions business. The pitch was that issuing banks can stay top of wallet when AI agents start initiating transactions on customers’ behalf. Real product, but card network side, and orthogonal to the core ledger.
The second came on April 29, when FIS announced Lyriq, a platform purpose built for regulated banks to issue, manage, and settle tokenized deposits, stablecoins, and CBDCs while keeping the deposits on the bank’s own balance sheet. The next day, FIS announced Project Keystone, the first major Lyriq deployment: a bank administered network with five super regional founding members (Citizens, Fifth Third, Huntington, KeyBank, M&T) plus an unnamed sixth. Lyriq feels like the most consequential thing FIS has done in years, and it is also a sidecar. The single most important sentence in the Lyriq announcement is the one that says
“works with existing core banking systems, regardless of technology provider.”
To me the architectural signal is clear that the fastest path to tokenized deposits is not a full replacement of the existing core. It is a parallel digital money layer that works across cores. That may be commercially smart, but it also tells you where the legacy architecture ends. You have to smile at how the architecture admission is the platform itself.
Jack Henry have perhaps been the most disciplined of the three. Shanon McLachlan, the COO, described their recent work as “below the waterline,” foundational engineering rather than visible AI features. He also flagged a fact that’ll make every Big Three customer take a pause. Most legacy core systems use COBOL packed decimal fields fixed at two decimal places. The Jack Henry deposit core was rebuilt with nine decimal precision, which lets it natively express USDC and tokens that move in finer denominations. The bigger issue is what sits underneath the decimal point. Legacy cores treat balances as dollar denominated rows. They have no concept of a token as a distinct data type. JPMorgan, when it built Kinexys, did not extend its existing core. It built a new ledger, I suspect because the old one could not represent the instruments it needed to issue.
There’s a consistent pattern here across all three. The procurement framework gets built first, because both sides need it. So the vendor needs AI revenue for investors, the customer needs AI spend for the board. The actual deployment of agents into core workflows comes later, and “later” is not really a defined date. The Big Three’s AI work is not fake. Jack Henry’s foundational engineering in particular feels real. But bolting AI on top of an engine designed for nightly batch posting does not make it AI native. And building a sidecar around a core that cannot handle tokenization does not make the core able to handle tokenization. It just kicks the can down the road.
Why the gap is now urgent
I don’t think the urgency is that every community bank will suddenly lose retail deposits to stablecoins. To me, the first order threat is commercial operating balances, treasury flows, escrow balances, settlement accounts, liquidity sensitive corporate relationships, and tokenized asset settlement activity moving toward institutions that can operate in programmable settlement environments.
For the last two decades of my career, the answer to “when will core banking get rebuilt?” was always “soon.” Soon has kept slipping, but I think this time is structurally different, for three reasons.
First, programmable money is no longer hypothetical, and it is no longer one thing. Tokenized deposits, onchain claims against an FDIC insured bank balance, are running at JPMorgan, Citi, and Partior. Payment stablecoins, bearer instruments backed 1:1 by reserves, are running at Circle (USDC, $78 billion), Tether, Paxos, and Ripple. Different instruments with different legal characters, but both rely on a regulated bank somewhere underneath, either as the issuer of the tokenized deposit or as the holder of the stablecoin reserves. The GENIUS Act requires every payment stablecoin issuer to operate under either a bank charter, a federal trust charter, or an approved state regime. In December 2025, the OCC conditionally approved national trust bank charters for Circle, Ripple, Paxos, BitGo, and Fidelity Digital Assets. Coinbase and Stripe’s Bridge are pending. The stablecoin issuers are not bypassing the banking perimeter. They are entering it.
Yes today’s volume is still small. Fewer than 5% of corporate treasurers reported investing in digital assets in 2025. But Tradeweb’s 2026 corporate treasurer survey found 25% moderately to very interested in tokenized money market funds, 19% in stablecoins. McKinsey’s base case puts tokenized market cap at $2 trillion by 2030. A regional bank with $50 billion in deposits that loses 2% of its corporate balance over a five year window is staring at a billion dollar deposit migration. This is my back of envelope sizing of the forcing function, and I think the trajectory is only going one way.
Second, the Big Three have telegraphed that the next migration cycle is open. Once Fiserv, FIS, and Jack Henry have all said publicly, in different ways, that their existing cores need to be augmented, replaced, or sidecar’ed, the customer’s question changes. It is no longer “do I migrate?” It is “to what?” That question has been closed since 1989. It is open now in a way it has not been since the late 1970s.
Third, the tools to actually rebuild this now exist. Tessera has compressed SAP S/4HANA migrations that historically ran $250 million and three years into months. DataMigration.ai took an 18 month bank data migration and shipped it in four. Neither is a full stack core banking replacement, and no question core banking is harder than ERP, harder than data migration, harder than enterprise software broadly. But AI has demonstrably collapsed the cost and timeline of analogous enterprise modernization work in the last two years, and my bet is that the same pattern will transfer to banking. The proof point of an AI native Tier 2 bank core in production is still ahead, not behind. A serious challenger has to deliver one to make the rest of this real.
What the new core has to be
So in my mind, the AI native US core banking platform of 2030 is going to need five things. I don’t believe any are particularly speculative in isolation. JPMorgan has assembled the closest working version inside its own walls. Building it as a product the rest of the industry can buy is a big venture opportunity.
The first is a programmable balance sheet. By that I mean a ledger that represents tokenized instruments as a first class data type, can execute conditional settlement logic at the core layer, and clears coherently across two distinct kinds of external rail: real-time payment networks like FedNow and RTP on one side, programmable chains like Base, Ethereum, Solana, and Canton on the other. JPMorgan has built this for internal flows on Kinexys. Partior has built it for cross bank settlement among consortium members. What is not yet solved at production scale is atomic settlement between an arbitrary bank’s deposit ledger and a public chain without a trusted intermediary. The current options, HTLCs, trusted bridges, notary schemes, cross-chain protocols like Chainlink CCIP, all have tradeoffs. The institutional alternative is a unified ledger model, which is what the BIS is exploring with Project Agorá. Whichever architecture wins, the bank with a programmable balance sheet becomes a settlement endpoint for the tokenized economy. The bank without one accepts being routed around as the price of the existing architecture.
The second is a real-time evented ledger. The deposit ledger has to post in real time, not in nightly batch. Every state change is an event with a stable schema, durable history, and reversibility. This is what makes everything else possible. Programmable money does not work in batch, intraday liquidity is invisible without continuous balance visibility, and agents cannot react to a ledger that updates once a night. Today, most US bank cores simulate real-time on top of batch through caching layers and shadow ledgers, which is a solved problem until it isn’t, at which point reconciliation breaks publicly. Rebuilding the ledger as event-sourced from the ground up is the foundation underneath the other four properties.
The third is agent configurable operations. Big Three implementations cost what they cost because the work is fragmented across specialists in compliance, payments, deposits, lending, and reporting, sequenced by a program management office that bills hourly. Coding agents can collapse that fragmentation. Agents can ingest a bank’s tenant configuration, reconstruct the rules in plain English, reconcile against integration feeds, and generate a configuration draft in days rather than quarters. The same pattern works for ongoing operations: changing a comp rule, adding a new product, reconfiguring a workflow.
The fourth is a bank owned data layer. The bank’s own data, accessible directly by the bank’s own agents, through a real-time event stream and APIs that are not gated behind separate purchase agreements or token pools. The current architecture treats the customer as a tenant whose data the vendor controls, which made sense when integration was expensive and direct database access was a security risk. It does not make sense in a world where the bank wants its own agents to operate against its own customer data, and where its corporate clients want their agents to settle programmable transactions against the bank’s ledger. The bank owns the data, the vendor operates the platform, and the architecture has to distinguish them.
The fifth is agent grade controls. As banks deploy agents across compliance, operations, treasury, and customer service, "agent" becomes a first class principal type alongside human roles. A KYC agent should see customer onboarding data, not the trading book. A treasury agent should see intraday positions, not employee payroll. The technical pattern, scoped tokens with auditable trails, is an extension of existing IAM rather than a new category, but the data model has to be designed for it from the start. The same control plane has to handle the regulatory side. The compliance surface around banking, BSA/AML, OFAC, fair lending, CRA, the EU AI Act and MiCA for institutions with European operations, plus the new GENIUS Act layer for payment stablecoins, is expanding faster than most compliance team can track manually. An AI native architecture pulls regulatory monitoring into the core release cycle, with an agent that flags what needs to change and routes the update through human approval. The same stack, the same permissioning, one audit trail across the whole system. Difficult to retrofit into a platform built around quarterly release cycles, and natural to build into one designed for continuous deployment.
How do we get there, and why this is different
I think the entry point matters more than the endpoint. The right wedge is probably the one Capital One ran on payments infrastructure. Do not fight the incumbent renewal, rather sell into the adjacent budgets that are not locked. A bank’s core contract is typically locked for seven years, but its real-time payments program, FedNow integration budget, compliance technology spend, and data and analytics budget are not. A scoped project that delivers FedNow native settlement, AI driven compliance monitoring, or real-time GL reconciliation across the existing core can be sold cleanly into one of those budgets. By the time the core renewal opens, the challenger is already inside the bank, integrated with the existing core, and delivering value the CIO can point to. The question stops being “rip and replace your core.” It becomes “expand what is already working into the renewal we were going to negotiate anyway.”
The Big Three are not going to sit still. I expect aggressive bundling, with the core renewal packaged with payments processing, debit network economics, BSA/AML, and digital banking. I expect to see steep multi year discounts on renewals that happen to land in the middle of a challenger’s evaluation. I expect some FUD from the analyst firms and consulting partners with nine figure Big Three practices to protect. Expect contractual friction on data portability when a bank tries to extract its own tenant. None of those moves address the underlying architectural problem, but any one of them can easily slow a design partner deal by a quarter or more.
All of this is why the core is so different from any other software replacement. The core is a regulated balance sheet system. A core banking outage is not a productivity problem, it is a 24 hour news cycle and a regulatory examination. Stability beat modernity, and stability is what the Big Three sold. I think now stability requires modernity. CIOs who renewed for seven years in 2023 are sitting on contracts that expire in 2030, and between now and then their bank has to support real-time payments, programmable money, and AI in production. The architecture they renewed onto cannot do those things, and the vendor’s answer is a consolidation program that pushes them onto the next generation platform anyway.
The sidecar approach definitely buys time, but it does not buy architecture. Banks already run multiple specialty cores in parallel for mortgage servicing, wealth, and treasury, so adding a tokenization stack is not architecturally novel. What makes the tokenization sidecar different is that, unlike mortgage servicing, the deposits being moved through it are the same deposits the core ledger holds. Reconciliation between the two systems remains an operational burden. Intraday liquidity management still requires the core to participate, because the deposit balance lives in the core. As tokenized volume scales, the sidecar becomes a parallel system of record, and the bank ends up running two cores instead of one.
Project Keystone has a second problem that bank consortium models have run into repeatedly. The pattern is not necessarily that consortia always fail, but it’s that they struggle when value depends on broad counterparty adoption before any single participant gets enough standalone utility. Closed networks need every counterparty to join before they reach critical mass, and the previous decade is littered with bank consortium blockchain initiatives that failed to clear that bar: Marco Polo, We.Trade, Vakt, Komgo, Contour, B3i. Partior is the one that worked, and it took five years and a JPMorgan/DBS/Standard Chartered consortium to get there. Five super regional banks plus an unnamed sixth is a real start, but it’s not yet a network. The treasurer who needs to settle a tokenized payment between a Project Keystone bank and a non-Keystone counterparty has to fall back to dollar rails or use a stablecoin issuer with broader reach. The bank's tokenization capability fails to compete for that transaction.
In a community bank somewhere right now, an operations analyst is reconciling a discrepancy between the FedNow settlement file and the overnight GL post by hand, in a spreadsheet. In another bank, a relationship manager is explaining to a corporate client that the bank cannot settle the client's payment in tokenized form. The client routes the payment through their JPMorgan account instead, the one they opened for exactly this kind of thing. That conversation is going to happen more and more every quarter.
Someone is going to build the system that replaces both these pieces of work, and the bank that buys it first will spend the next decade taking corporate relationships from the banks that didn't.
References
Core banking market structure and concentration
Federal Reserve Bank of Kansas City (April 2024). Market Structure of Core Banking Services Providers.
Engage fi (2025). Core Banking Modernization Part 2: The Big Three.
SDK.finance (April 2026). Best Core Banking Software Providers 2026.
Big Three modernization programs
Everest Group (January 2026). The Great Core Banking Shakeup.
CCG Catalyst (December 2025). Earnings Roundup: Fiserv, FIS, and Jack Henry.
Finopotamus (April 2026). GAC 2026: Core Conversations with Fiserv, Jack Henry and Pediment.
Fiserv (January 8, 2026). Fiserv Collaborates with Microsoft to Accelerate AI-Driven Innovation.
FIS (April 29, 2026). FIS Launches New Platform Giving Banks Control Over Digital Money. Lyriq launch.
FIS (April 30, 2026). FIS and Leading Financial Institutions to Build Their Own Digital Tokenized Money Network. Project Keystone.
Programmable money and tokenized settlement
JPMorgan Kinexys. Kinexys Pilots First USD-Denominated Deposit Tokens (June 2025).
The GENIUS Act (2025). S.1582 - GENIUS Act. Effective January 2027.
OCC (December 12, 2025). OCC Conditionally Approves Five National Trust Bank Charter Applications. Conditional approvals for Circle, Ripple, Paxos, BitGo, Fidelity Digital Assets.
Partior. Partior platform overview. Bank-owned consortium platform doing atomic cross-bank tokenized deposit settlement since 2021.
BIS (2024-2026). Project Agorá. Seven-central-bank tokenized cross-border settlement project.
Citi. Citi Token Services. Live in US, UK, Singapore, Hong Kong as of 2025.
Tradeweb (April 2026). 2026 ICD Portal Client Survey.
ABA Banking Journal (March 2026). Tokenized deposits: the future of tokenized money for financial market settlement. Cites McKinsey $2 trillion forecast.
Enterprise modernization comparables
Tessera Labs (2026). Tessera platform overview. Multi-agent AI for ERP modernization.
DataMigration.AI (2026). Global Bank Core Banking Migration Case Study. 18-month migration in 4 months, 8 AI agents.
Backbase (January 2026). AI Core Banking Integration: Strategies That Scale. Industry data: 30% of full migrations succeed.