Washington piles into crypto while Airwallex bets on agents and banks insure against hallucination
7 min read · 11 sources
- SEC proposed crypto exemptions up to $75 million annually with a conditional safe harbor as the CLARITY Act fight intensifies in Washington.
- JPMorgan ended its banking relationship with Polymarket in October 2025 amid regulator scrutiny of prediction markets.
- Lloyd's-backed Testudo began underwriting up to $10 million in AI-hallucination liability for banks, with premiums of $10,000 to $20,000 a year.
- Airwallex raised $320 million in a Series H at an $11 billion valuation and launched T:0 and Airi for "autonomous finance."
- Citadel shed more than 80% of the risk it took on from the Situational Awareness fund via over 100 block trades worth $4 billion in market value.
Washington is moving on crypto from three directions at once, and the rest of this week’s fintech news sits in the wake of that. Trump hosted Coinbase’s Brian Armstrong, a16z’s Marc Andreessen and Ben Horowitz, Ripple’s Brad Garlinghouse, and Kraken’s Arjun Sethi at the White House to push for a “fair version” of the CLARITY Act, even as he objected to ethics provisions from Sens. Thom Tillis and Ruben Gallego aimed at officials’ personal crypto holdings. The bill is the unfinished business of the previous push: who actually regulates digital assets, the SEC or the CFTC, and on what terms.
CFTC Chairman Mike Selig made the jurisdictional threat explicit. If CLARITY stalls, the agency will build its own crypto regime using authority it already has, which would put derivatives, spot markets, and token oversight on parallel tracks and force issuers and exchanges to read two rulebooks at once. The SEC, meanwhile, formally proposed Regulation Crypto Assets: offerings up to $5 million over four years, or $75 million annually, could move without full registration, and a conditional safe harbor would let issuers iterate inside a defined perimeter. For engineers shipping token infrastructure, the compliance thresholds and the SEC-versus-CFTC question are now the two numbers that actually change the build.
Citadel ran more than 100 block trades totaling over $4 billion in market value to unwind more than 80% of the risk it absorbed from the Situational Awareness fund.
JPMorgan walked away from Polymarket in October 2025
JPMorgan ended a banking relationship with Polymarket in October 2025 as state regulators began treating prediction markets less like exchanges and more like gambling operations. Polymarket’s read is different: the company says it still works with JPMorgan across multiple entities and customer fund flows, suggesting a partial rather than full cut-off. Either way, the episode is the first hard data point on what “debanking” looks like for a high-volume, consumer-facing crypto venue at a top-tier US bank, and Kalshi is the obvious next name to watch. If a Tier 1 bank can drop a customer because a state AG files a letter, the operational risk model for every prediction-market platform just got worse.
Banks are now buying insurance against AI hallucination
Testudo, a Lloyd’s-backed startup, is underwriting up to $10 million in liability for harms caused by generative AI, with annual premiums between $10,000 and $20,000. The product is aimed squarely at financial institutions that want to put models in front of customers or in the loop on transactions, where a confident wrong answer about a balance, a wire, or a credit decision is a lawsuit waiting to happen. Some vendors still prefer to self-insure or eat the risk as part of their SLAs; the skeptics’ argument is that a dedicated hallucination policy is mostly indemnity theatre until case law catches up. For the people shipping the systems, the practical effect is a new column in the risk register: a line item for “what happens when the model is confidently wrong on a regulated workflow,” and a reason to keep humans in the loop one step longer than the demo suggests.
Affirm's slow grind to profitability
Max Levchin’s Affirm has finally turned profitable after more than $2 billion in cumulative losses, putting the company near a $25 billion valuation. The thesis Levchin has held since the PayPal days is that fraud and credit risk are the moat, so most of the underwriting, collections, and decisioning stack was built in-house rather than licensed from the usual banking partners. The corollary is that Affirm ships features slowly and is ungainly to integrate with, but the cost of bad debt stays predictable. For engineers outside the company, the read is that vertical control of risk infrastructure paid off, at the price of a decade and a few billion in losses, and the next crop of “we’ll just use Stripe Underwriting” BNPL plays should probably reread the loss history before raising.
The application layer is where the AI money actually has to land
Tidemark Capital’s System of Action essay makes the contrarian case that apps are not dead, they are the prize. The argument is that foundation models do not book revenue, chase invoices, file compliance forms, or wrangle the per-tenant permissions and semantics that vertical workflows actually require. That work is done by systems that have to run with infrastructure-grade reliability, because they are now being driven by autonomous agents rather than humans clicking buttons. The implication for engineers is uncomfortable: the differentiator in 2026 is rarely the model, it is the scaffolding underneath it, the connectors, the idempotency, the audit trail, the escalation path. Vertical SaaS incumbents and AI-native challengers are both being told to harden up or be replaced by something that can be trusted to act on its own.
PayPal and Venmo take the campus tuition rails
PayPal and Venmo are now integrated with Illumia, Nelnet Campus Commerce, and TouchNet, the three back-office systems most US universities use to process tuition. Students and parents can pay through the school’s existing payment portal with a Venmo or PayPal balance or linked funding source, no separate registration, no new merchant agreement at the school. For PayPal, this is a wedge into a category it has been circling for years, peer-to-peer in the dorm, ticketing for athletics, and now the actual tuition payment. For anyone who has tried to wire money to a US university, the question is what the fee schedule looks like and whether Venmo’s standard 1.5% to 1.9% card surcharge actually shows up on a tuition transaction.
Airwallex bets the next $320 million on agentic finance
Airwallex closed a $320 million Series H at an $11 billion valuation, up from $8 billion in December, led by Addition with Baillie Gifford, T. Rowe Price, Amex Ventures, and Washington University in St. Louis. The company now serves 675,000-plus businesses and is past $1 billion in annualized revenue and EBITDA positive. The interesting part is what the money is for: T:0, an automated bookkeeping product president Lucy Liu calls “assisted driving” for finance, and Airi, an agentic consumer wallet built around one-click checkout. The pivot is from being a cheaper SWIFT alternative to being the rails and the agent on top of them, and the Series H valuation implies investors are willing to underwrite that as a payments-plus-AI story rather than a pure FX margin story. An IPO still hinges on market conditions, but the company has the cash and the metrics to choose its window.
Citadel unwinds $4 billion of Situational Awareness risk
Citadel has shed more than 80% of the risk it took on from Leopold Aschenbrenner’s Situational Awareness fund, Ken Griffin told clients, executing more than 100 block trades totaling over $4 billion in market value. The purchase talks started July 29, one day before CNBC reported Situational Awareness was forced to liquidate public positions after margin calls driven by losses on Sandisk and Bloom Energy, each down more than 50%, and a short book in software names like Adobe that rallied. Griffin’s note also confirmed Citadel’s flagship Wellington fund returned 5.94% in July, its best month since 2022. For anyone tracking AI-thematic flows, the lesson is that concentrated, leveraged bets on a narrow AI hardware basket can break a fund at the worst possible moment, and the unwind itself is large enough to move small-cap names.
Wall Street tech banking keeps reshuffling
JPMorgan is hiring Bank of America’s David Fishman to run North America technology M&A, elevating Vineet Seth to vice chair of investment banking and forming a new advisory group for top tech clients. The move is part of a wider talent drain from BofA, which has also lost senior dealmakers to Barclays, Deutsche Bank, and Citi. For anyone selling a company, expect a more competitive bid process on US tech M&A, and for anyone working at a target, expect the buyer to be a JPMorgan banker you’ve met twice rather than a BofA banker you’ve known for ten years.
In brief
Francisco Partners agreed to acquire healthcare payments and communications software vendor Weave for $650 million and take it private, with the plan to pour more into AI features. Ripple closed a $275 million private placement of senior unsecured notes out of its Ripple Prime unit, rated BBB by KBRA, which puts it in the top 1% by size of the 53 debt rounds by US blockchain companies in the last 48 months. The use of investment-grade debt rather than equity is the tell: Ripple is now borrowing like a brokerage, because that is what it is becoming.
You May Also Like
CFTC to Regulate Crypto With or Without Congress; Bitcoin Posts Largest Weekly Dollar Gain Ever
Bitcoin surged $14,264 in a single week to close at $77,387 — its largest-ever weekly dollar gain — while spot ETFs absorbed $1.92 billion in net inflows and …
Anthropic's hardware spec, Gemini's video update, and $105B in combined AI revenue
Anthropic opened a research preview of the Model Hardware Standard, a model-agnostic spec for letting AI agents safely operate lab and manufacturing equipment …
Nvidia reportedly buys Hugging Face for $12.9B as GPU demand stays supply-constrained through 2028
Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, the same week Amazon tripled its Nvidia GPU order to 3 million chips and Nvidia told …




