BankingNewsAI Daily Brief  · 

Anthropic CEO admits huge uncertainty about next year just days before the company's IPO

🏦 3 Banking AI🤖 3 General AI

Banking AI

Financial institutions & fintech technology

3 stories

Fed's Barr outlines potential financial stability implications of artificial intelligence

For the chief credit officer at a regional bank

Seven AI firms are expected to carry over $800 billion of on-balance-sheet debt by year end, and the Fed now compares them to the 1990s telecom buildout. Your data-center loans sit inside that comparison.

Twice a year the Fed publishes a Financial Stability Report. This one covers the AI buildout and how it was paid for. The seven largest AI firms are expected to carry more than $800 billion of on-balance-sheet debt by year end. The simple model is that they borrow now and expect returns later. If returns arrive slower than investors expect, prices reset, investment falls, and the lost wealth cuts spending. Why worry when the borrowers are rich? Because the Fed also names circular financing networks and other complex obligations, which do not all sit where the $800 billion is counted. The honest caveat is that these firms have not yet passed the telecom borrowers of the cycle that began in 1995. The Fed says they may soon.

→ Action

Credit risk: Map exposures to data-center borrowers, lessors and AI-linked counterparties, including leases and guarantees that sit off borrowers' balance sheets.

Full story →Read article →  from Federal Reserve

ECB says new AI capabilities could favour attackers over bank defenders

For the CISO at a mid-size US bank

The ECB says attackers get AI agents cheaply while banks wait on budgets and approvals, and banks lean on a few providers. Most vendor reviews assume the defender is ahead and the supplier is replaceable.

This is a European Central Bank financial-stability study of how AI changes cyber risk. The model is simple. Attackers can run low-cost AI agents with no committee to answer to. Defenders pay for frontier models, work through governance, and make slower decisions bound by regulation. The ECB calls defenders structurally disadvantaged, at least in the near term. The model is imperfect: AI helps defenders too, and euro area banks start from high cybersecurity maturity. The exception that matters is concentration. Banks rely heavily on a few third-party providers, and a systemic event can start when one hard-to-replace supplier fails, even outside the financial system. A bank can lock its own doors well. It shares the locksmith with everyone else.

→ Action

Third-party risk: Map which critical services depend on the same AI model and ICT providers, and test exit plans for each.

Full story →Read article →  from Europa

Eltropy launches Branded Calling for credit unions and community banks

For the chief operating officer of a community bank

In a Twilio test, branded calls were answered 62% of the time and unbranded calls 20%. Your customers on AT&T, which is not yet supported, will still see a bare number. So will the scammers' targets.

Branded Calling puts a bank's verified name and logo on the customer's screen when the bank calls. STIR/SHAKEN authenticates the number, Rich Call Data carries the name and logo, and Eltropy first vets the institution and checks its logo against trademark registries. Why must a bank prove who it is to its own customer? Because customers now ignore unknown numbers, and scammers call from them too. In Twilio's test, branding lifted answer rates from 20% to 62%. Unanswered calls still cost money and redials, and they can earn a spam label. The catch is coverage. It works on T-Mobile and Verizon, about 250 million phones, and AT&T is only planned. A uniform helps only where people check uniforms. Once customers expect the logo, an unbranded call from your bank looks like the fraud.

→ Action

Fraud operations: Pull answer rates on outbound fraud-alert and collections calls, split by carrier, to size the gain before signing.

Full story →Read article →  from Thepaypers

General AI

Large language models & AI infrastructure

3 stories

Anthropic CEO admits 'huge uncertainty' about next year days before IPO

For a vendor-risk officer at a bank that runs Claude

Anthropic has committed about $901 billion, mostly to compute, with $71 billion due by September 2027, and its CEO calls next year a huge uncertainty. Write exit clauses and a fallback model before the Nov. 19 IPO.

Anthropic sells Claude, but first it must buy the chips that run it, committing one to two years before it knows what customers will pay. Those commitments total about $901 billion, with roughly $71 billion due by September 2027. Revenue runs at $81 billion a year. The simple model: buy too little and lose sales, buy too much and overextend. That model is imperfect, since Anthropic buys modestly upfront and tops up on the spot market. The exception is cash. The IPO seeks up to $64.4 billion, plus $26 billion privately. Why raise money while revenue climbs? Because the bills come first, like a restaurant ordering two years of food before booking diners. A vendor insuring against its own forecast is telling customers to insure too.

→ Action

Vendor risk: Add insolvency termination rights, data-export terms and a tested fallback model to every Claude contract before renewal.

Full story →Read article →  from The-decoder

Congress wants to tighten AI chip export rules after Nvidia spy scandal

For a credit officer with data-center loan exposure

Your data-center loans rest on chips worth whatever their renters pay. Who may rent them is now a Senate Banking Committee question, and that committee wants Nvidia's former CEO testifying.

Export rules decide where an AI chip may ship. After the scandal over Jensen Huang, now Nvidia's former CEO, and a suspected Chinese spy, both parties want tighter rules. Senate Banking Chairman Tim Scott wants Huang and Commerce Secretary Howard Lutnick to testify. Speaker Mike Johnson cites a "growing consensus" on limiting exports to China. The simple model: fewer legal buyers, less demand. The imperfection: the Remote Access Security Act, introduced in April 2025, stalled in the Senate and may not become law. Why would a domestic rack care about export law? A server, like a warehouse, is worth what its tenants pay.

→ Action

Credit risk: List GPU and data-center loans whose borrowers rent compute to foreign customers, and stress collateral values under remote-access limits.

Full story →Read article →  from The Verge

Meituan releases a 1.6-trillion-parameter agent model trained without Nvidia GPUs

For the head of model risk at a us bank

A developer can point Claude Code at LongCat-2.5 for a promotional €0.27 per million input tokens. Your policy on Chinese-origin weights should exist before that change shows up in a config file.

LongCat-2.5 is a coding agent from Meituan, the Chinese food delivery company. Its weights are free to download under the MIT license. It has 1.6 trillion parameters, uses about 48 billion per token, and was trained entirely on Chinese hardware. Its API speaks the OpenAI and Anthropic formats, so it plugs into Claude Code and Codex. The simple model: free, near the frontier, easy to swap in. Meituan reports 88.4 on Terminal Bench v2.1 against 89.1 for Claude Opus 5. The catch: on the harder v4.0 it scores 38.5 against 53.9, and Meituan ran the tests. Why care where weights came from if they run in-house? The self-hosting guide assumes 194 Huawei chips, so most developers will call Meituan's API. Governing the download is easy. The endpoint is one line in a config file.

→ Action

Model risk: Add Chinese-origin open weights and LongCat API endpoints to developer-tool allowlist rules and network egress blocks.

Full story →Read article →  from The-decoder

Get this in your inbox every morning

Free · No spam · Unsubscribe anytime