The Clarity Act cleared committee last Thursday with the yield exemption for non-bank stablecoin issuers intact. Bank CEOs were on the phone with senators beforehand at what the ABA called its highest lobbying intensity in years — and it didn't matter.
That's the same week Robinhood hit 4 million Gold subscribers averaging $60,000 in balances, approaching JPMorgan's typical customer profile, built in three years from a standing start.
Bank of America's Erica has fielded 3.2 billion conversations, and most banks still can't tell when a customer is about to leave until the account's already empty.
The incumbents have the data, the relationships, the regulatory relationships, the branch networks — and they're watching a brokerage app eat their depositor base while Congress writes the rules that allow them to continue to do so.
It’s not the lobbying doesn’t work. We’re just finding out that (maybe?) regulation was never the moat in the first place.
The Bank Is Still Standing. But the Building Has Changed.
Five separate conversations about AI and banking infrastructure landed on the same fundamental observation from different angles: the bottleneck isn't capability, it's architecture. Backbase CEO Duke Plietta put the sharpest point on it — 60% of frontline bank work happens in the white space between systems, not inside them. That's not a technology problem. It's a design choice that became load-bearing over decades, and AI can't fix it by sitting on top.
The FIS-Anthropic partnership announced earlier this year, with BMO and Amalgamated Bank among the first adopters, looks at first like a straightforward enterprise AI distribution deal. It probably isn't. As Oban MacTavish noted on Fintech Insider, payments data is one of the last major data categories that large language models don't yet have access to. The 5% jump in FIS stock on announcement day was the market pricing in something other than a financial crimes chatbot. Data access — proprietary, transactional, at scale — is the actual asset in play.
Meanwhile, Customers Bank, a $26 billion Pennsylvania institution, signed a multi-year deal to embed OpenAI engineers directly inside the bank to automate lending and payments workflows. Their CEO used an AI clone to deliver the first 30 minutes of an earnings call before disclosing the substitution. That either reads as a confident demonstration or a warning, depending on your priors. Ron Shevlin's framing is more useful than either reaction: the strategic value of agents isn't headcount replacement, it's cycle-time compression. The banks that understand that distinction will deploy differently than the ones chasing a labor-cost narrative.
The most counterintuitive data point across all five sources belongs to Trade Republic, which scrapped its AI customer service deployment entirely and hired 1,000 human agents after customer satisfaction collapsed in its core German and Austrian markets. At the same moment, Starling Bank was launching what it describes as the UK's first agentic financial assistant. Google Cloud's Georgina Buckley argues that banks — precisely because they hold identity, liquidity, and trust — are the indispensable foundation that agentic commerce will run on, not the layer it routes around. That argument is self-serving coming from someone selling cloud infrastructure to banks. It also happens to be correct.
The Plumbing Is Being Replaced While the Water Is Still Running
The MoneyGram float number is the cleanest way to see what's actually at stake. Luke Tuttle's team has integrated Fireblocks directly into their ERP, partnered with Kraken to bridge stablecoins to physical cash agents, and launched USD-denominated consumer wallets in Colombia and El Salvador — none of which require the end user to understand any of it. That last part is the product. The stablecoin is infrastructure, invisible by design, and the float savings fund the whole bet.
Meanwhile, the ECB is quietly doing something structurally similar at the wholesale end. Its shift from Pontos to Apia — a fully distributed ledger technology-native settlement layer for simultaneous securities and payment clearing — isn't bridging old rails to new ones. It's replacing the old rails entirely. Europe's WERO instant-payment network is making the same move at the consumer layer, explicitly framed as a sovereign alternative to Visa and Mastercard. The official explanation is efficiency and innovation. The real reason is that European institutions have decided they don't want American card networks owning the rails under their economy.
The liability question is where the optimism runs out. Alipay's single-use credential model for AI agent transactions — a scoped authorization that expires on execution — is elegant. But Chargebacks 911's warning is harder to wave away: the entire fraud detection stack was built on human behavioral signals, typing speed, device angle, session duration. An AI agent executing a transaction produces none of those signals. It's instantaneous and flawless, which is exactly what fraud detection is designed to flag as suspicious — and exactly what it can't catch when the agent is legitimate.
Camilla Bullock's (Emerging Payments Association Asia) arithmetic on quantum computing is the one the industry is most aggressively ignoring. Migration to quantum-safe encryption takes roughly seven years. Google's quantum readiness target is 2029. Less than 20% of senior payments leaders are actively tracking the threat. Those numbers don't resolve neatly. The payments infrastructure being rebuilt right now — the stablecoin rails, the DLT settlement layers, the agentic credentialing systems — is being built on encryption that has a known expiration date, and almost nobody responsible for that infrastructure is treating it as urgent. That's not a knowledge gap. That's a choice.
The Collateral Problem Is Getting Solved From Every Direction at Once
Thomas Galbraith's (Barkr’s CEO) description of the traditional appraisal industry is worth sitting with: Five firms value the same private jet, you get five different numbers, and not one of them is contractually liable for any of them. Lenders have been pricing risk on top of that foundation for decades — not because it was acceptable, but because there was nothing else. Barkr's $2 billion in covered valuations across luxury assets and GPUs is still small against a $1.5–2 trillion collateral market, but the NVIDIA partnership, which is routing Apollo and similar names directly to Barkr's door, suggests the volume problem may resolve faster than the incumbents expect.
The Bitcoin lending story is structurally similar, and the people building it know it. Both Arch's Himanshu Sahay and Mezo's Matt Luongo are making the same foundational argument: holders of appreciating non-cash assets have always been able to borrow against them cheaply and roll the loans indefinitely — if those assets were the right kind. Sahay says plainly that he started Arch in 2022 because you couldn't do that with Bitcoin. The CLO structure and Anchorage's qualified custody are attempts to build the same institutional scaffolding that makes a private bank comfortable lending against a Picasso. The 60% loan-to-value ceiling and on-chain transparency are the concessions to the fact that Bitcoin isn't a Picasso yet.
Wayflyer's Aiden Corbett makes the point that survives all of this: in lending, your dollar is as good as anyone else's dollar, so underwriting advantage isn't optional — it's the whole game. Barkr's answer is a warranties product that makes collateral risk contractually enforceable. Arch's answer is a CLO that lets it offer rates the bilateral market can't match. Wayflyer's answer is Shopify and Facebook Ads data that no bank was reading in 2019. Different instruments, identical logic.
Scott Bass, who spent time at Capital One before Monzo, adds the uncomfortable corollary: bad underwriting decisions in lending don't announce themselves. The business can look healthy for six months after the decision that kills it. That lag is the reason the current wave of infrastructure investment — warranties, on-chain custody, real-time data feeds — matters beyond the marketing. Better collateral valuation and better forward-looking data don't just help lenders win deals. They shorten the feedback loop before the losses are already baked in.
The Regulatory Floor Is Setting. The Yield Problem Isn't.
Hannah Meakin of Norton Rose Fulbright, covering the April regulatory wave across five jurisdictions in a single month, put it plainly: "April feels like one of those months where crypto regulation stops being theoretical and starts to feel very real. There's more certainty now with much less patience." That's not hyperbole. The UAE issued a new federal framework. Dubai published detailed rules on exchange-traded derivatives. The shift from policy signaling to binding deadlines happened fast, and firms that treated regulatory prep as a future problem are now the ones scrambling.
The commercial beneficiaries of this clarity are already positioning. Nexo — which retreated from the US retail market in 2022 under regulatory pressure — is back, sponsoring Formula One and golf tournaments, and preparing a crypto-backed credit card pending US approvals. Neil Steinhardt of Nexo frames the FIT21 Clarity Act as the last piece needed. That's a reasonable read. It's also the explanation of a company that timed its re-entry to regulatory tailwinds and is now dressing the timing in principle.
What the regulatory maturation doesn't resolve is the risk-pricing failure that Mauricio Di Bartolomeo of Ledn identified with uncomfortable precision. Shortly after the Kelp/Aave exploit in April, USDT on Aave was yielding 3.92% — 32 basis points above the Federal Reserve's rate — despite unresolved smart contract exposure. Senior secured private credit from Apollo or Blackstone pays 12–18%. The market is pricing DeFi stablecoin risk like it's a Treasury with a footnote.
That mispricing matters more now, not less. As regulatory frameworks harden globally, institutional capital will increasingly treat licensed crypto platforms as a legitimate asset class. More capital chasing the same on-chain yields won't solve the compensation gap — it'll compress it further. Ledn's Bitcoin asset-backed security, rated BBB-minus by S&P and more than two times oversubscribed at 6.84%, points at where the arbitrage actually lives: off-chain credit structures that borrow crypto's collateral logic and price the risk honestly.
The Clarity Act's Quiet Carve-Out Is the Only Fight That Matters
Treasury Secretary Bessant has been candid about why the White House won't kill the yield language: stablecoins, in his framing, are a demand mechanism for U.S. Treasuries that lowers government borrowing costs. That's the fiscal interest underneath the official financial-innovation explanation. When the administration says it wants a "competitive stablecoin framework," it means it wants a buyer for its debt. The yield carve-out isn't an oversight — it's the point.
J.P. Morgan appears to have drawn the same conclusion. The bank filed for a second tokenized money market fund on Ethereum this month, explicitly targeting stablecoin issuers as reserve management infrastructure. Two filings in the same product line, before the Clarity Act has even reached the floor. That's not hedging. That's a bank deciding the bill passes substantially as written and positioning accordingly — while the ABA's member CEOs are still calling senators.
Meanwhile, the compliance architecture around everything else is being rebuilt simultaneously. The CFPB's revised examination manual dropped May 15, resetting procedures across the full Consumer Protection Statute inventory, effective immediately. NYDFS issued an industry letter the following day requiring disparate impact analysis under state law — directly contradicting the Trump administration's rollback of disparate impact obligations at the federal level under the revised Regulation B. Institutions with material New York consumer lending activity now run two parallel fair lending compliance programs. That's not deregulation. That's jurisdictional arbitrage creating more paperwork, not less.
The one genuinely clarifying data point in all of this came from an FDIC staff study on the 2023 bank failures: depositor sophistication, not insurance status, predicted who ran first. Large uninsured depositors executed near-complete withdrawals. The policy implication — that deposit insurance reform matters less than the composition of your deposit base — has been sitting in the data for three years. The CAMELS overhaul announced by Fed Vice Chair Bowman, shifting examination focus toward capital adequacy and liquidity over procedural documentation, suggests at least one regulator read it.
Banks Have the Data. They're Still Losing the Customer.
Rex Salisbury put the business-model asymmetry plainly: Robinhood earns roughly 2% return on assets for its clients, which means it can afford to offer a 3% cash-back card that every bank has already quietly declined to match. Banks haven't done the math wrong — they've done it exactly right, for themselves. The product that would retain the customer would also compress the margin. So the customer goes.
The deposit franchise is the longer-term problem. Approximately $20 trillion in U.S. bank deposits currently earns near-zero yield, held in place largely by friction — the mild inconvenience of moving money. Agentic AI is about to make that friction disappear. When software can automatically seek yield on idle cash in real time, the economic logic that kept deposits parked at Chase or Wells Fargo for decades stops working. Banks have roughly a decade, by Salisbury's estimate. That's not long if your core system migration is already three years behind schedule.
BoA’s Erica data point cuts differently. Bank of America built something genuinely impressive — 50 million users, eight years of behavioral data, a platform now evolving toward autonomous action on the customer's behalf. Jorge Camargo describes a third stage beyond answers and actions: autonomy, where the AI executes tasks the customer didn't even know to ask about. That's a real capability gap most regional banks can't close organically.
But here's the irony that both episodes circle without quite colliding: the institution best positioned on AI-driven customer intelligence is still structurally constrained by the same margin logic that's letting Robinhood eat its lunch. Having the best assistant doesn't help if the underlying product is still designed around extracting yield from customer inertia rather than competing for it. Capital One paying $5.15 billion for Brex suggests at least one major bank has concluded that the build-versus-buy calculation has already been settled — just not in their favor.