AI agents begin moving deposits: banks build defenses, Anthropic's high-yield financing and Japan bond-market spillover

Several recent financial threads point to the same shift: AI agents and the interest-rate environment are repricing banks' liabilities, markets' risk appetite and AI companies' cost of capital. Banks are starting to block automated operations by AI agents on deposit accounts; beneath a stable U.S. equity index, ordinary stocks are under pressure; Anthropic is pursuing both high-threshold regulatory lobbying and high-cost debt financing; and rising expectations for Japanese monetary policy normalization are adding another layer of repatriation pressure to global bond markets.

Banks fortify against AI agents: the moat around low-cost deposits

Citigroup, Bank of America and JPMorgan Chase have successively announced restrictions on AI agents' access to customer accounts. Citigroup explicitly said it will not allow third-party AI agents or assistants to log into customer accounts or conduct transactions directly on its digital banking platform, saying the move is intended to protect customer accounts and information security. Bank of America CEO Brian Moynihan told CNBC on Wednesday that the bank is restricting AI agents' access to consumer deposit accounts while studying how to determine who bears responsibility when AI initiates payment instructions. JPMorgan CEO Jamie Dimon said on the third-quarter earnings call that the bank will restrict AI agents' access to customer accounts because 'you don't want rogue agents, and there could be thousands of them.' Behind this collective defense is the banking industry's deep anxiety that trillions of dollars in low-cost deposits could be 'swept away' by AI. ^[1]

That anxiety has a clear interest-rate-spread backdrop. According to the Financial Times, after Meta launched its Muse AI personal assistant in September, industry concerns rose markedly, with the market broadly worried that AI agents will automatically move consumers' idle funds from bank accounts into higher-yielding products. Apollo chief economist Torsten Slok quantified the risk in a report: AI assistants could automatically move households' idle cash from the national average checking account rate of just 0.1% into accounts yielding 3.3% to 5%. The report warned that if every household used AI agents to optimize cash-balance yields, banks could lose a large volume of low-cost deposits used to make loans, posing a threat to the entire financial system. ^[1]

Deposit flight strikes at the root of bank profits because non-interest-bearing deposits are huge. JPMorgan, Bank of America and Wells Fargo currently hold a combined $1.6 trillion in non-interest-bearing deposits, equal to 16% of their total liabilities. If forced to pay 3% on those deposits, the three banks would bear an additional cost of about $47 billion a year, nearly half the combined total of their annual after-tax net profit. For banks, AI agents are not just a technical tool but a competitive entry point that could bypass traditional customer relationships and directly trigger deposit repricing. ^[1]

The defenses differ by bank, but the core logic is the same: until responsibility is clearly assigned, AI agents should not be allowed to autonomously move account funds. Brian Moynihan explicitly said human involvement should be required in authorizing account transfers. Bank of America is communicating with AI agent service providers, warning them of the risk that consumers may later dispute where funds went. Jamie Dimon stressed the risk-control logic, saying no external agent will be allowed into the company's systems before the corresponding tools are built. Citigroup said customers can still authorize limited read-only access to account information through specific open-banking providers. According to people familiar with the matter, the Federal Reserve has held private discussions with bank executives on the compliance boundaries for AI agents in financial transactions, but the Fed declined to comment. Regulators privately support banks' moves to restrict AI agents' access to accounts, with officials saying the relevant authorization and liability determinations may need to be clarified through legislation. ^[1]

Facing deposit competition pressure from AI, banks have begun looking for responses. Christoph Stegmeier, a senior partner at consultancy Simon-Kucher, said banks need to strengthen the stickiness of deposit products by expanding customer loyalty programs and optimizing loan rates, making deposits harder to replace. This AI-driven deposit battle tests whether banks can buy time on the technology defense line until regulatory rules and liability boundaries are in place. ^[1]

Divergence inside U.S. equities: tighter financial conditions land on ordinary stocks

U.S. equities are showing a rare internal rupture: the index is unmoved, but beneath it the picture is messy. In a latest report, Morgan Stanley chief equity strategist Mike Wilson warned that tighter financial conditions are quietly doing part of the Fed's work, but the shock is being absorbed concentratedly by ordinary stocks rather than the few AI-centric giants. The 10-year Treasury yield has broken above 5.3%, the dollar has risen to a new high for this cycle, and oil prices remain elevated. With these three pressures combined, Morgan Stanley's composite financial conditions index has jumped to about 1.4 standard deviations, a level breached only a few times in the past 35 years, most recently in 2022 when the Fed raised rates by 75 basis points at every meeting. Meanwhile, Goldman Sachs has revised its forecast to one more rate hike at the October FOMC. ^[2]

The market's key fracture is already clearly visible. The S&P 500 equal-weight index has lagged the market-cap-weighted index by 6% over the past six months. In September, when the 10-year yield broke above 5%, the equal-weight index fell 5% in a month while the cap-weighted index was nearly flat. Goldman Sachs' Tony Pasquariello, while partly reserving judgment on the 'breadth collapse' thesis, acknowledged that the median S&P 500 constituent has fallen 15% from its 52-week high, the worst reading since the dot-com bust. Wilson's breadth gauge also shows that the share of S&P 500 constituents above their 200-day moving average fell to 43% at the end of September, close to the late-March low of 44%; it has since edged back to 50%, but overall pressure has not cleared. The S&P 500 forward price-to-earnings ratio has also retested about 19 times, near the March low. More notably, the S&P 500 equity risk premium has fallen below zero, a level almost unseen since 2002. ^[2]

Stability at the index level masks structural damage. Financial stocks are another warning signal: the sector fell 7.7% in 30 days, and its relative performance versus the S&P 500 has dropped to the lowest level since 2019. Wilson sees financials as one of the most important near-term guideposts for equity direction; before the large-bank earnings season begins, whether the sector can stabilize will be a key window for judging the market's path. He also gave two trigger conditions: financials stabilizing, and the MOVE index, which measures Treasury volatility, falling back. According to Bloomberg, Goldman Sachs' Rich Privorotsky recently observed that long-end bond implied volatility has fallen about 30% from its recent high, giving the market some breathing room. In the coming weeks, September CPI data, Warsh's remarks at the IMF and the October 28 FOMC meeting will follow in succession. Consensus expectations show that, affected by a 9.6% month-on-month jump in gasoline prices, September CPI may rise 0.6% month on month and 3.6% year on year. ^[2]

How this divergence ends—whether ordinary stocks catch up with the index or the index converges downward toward ordinary stocks—remains the market's central suspense. Wilson believes the shock is still at the valuation level and has not yet become a growth threat, but he also acknowledges that an equity risk premium at zero is historically rare. Pasquariello noted that a wave of lockup expirations in 2027 could push net U.S. equity supply to the highest level since the 2000 dot-com bubble, potentially stressing the market's supply-demand balance. With the equity risk premium at zero, breadth gauges near dot-com levels and a potential supply shock emerging, the market may not be well prepared for the tail risk of the index converging downward. ^[2]

AI regulation and financing: high-threshold legislation and high-cost debt

On AI regulation, Anthropic is quietly promoting an AI safety bill that it largely drafted, seeking to set a very high threshold for federal intervention in frontier AI models while strictly limiting regulatory power to the Commerce Department and federal courts. Reuters, citing Politico on the 17th, reported that Anthropic representatives have privately lobbied members of Congress on the draft law. Under the bill, the government could intervene to block a model's release or use only when the AI model faces a foreseeable, imminent risk—that is, it could cause more than $1 billion in losses or lead to 50 or more deaths or serious injuries. The triggering risk must also stem from one of three specific situations: the AI model being used to develop or spread biological weapons, the model escaping its developer's control, or the model causing a cyberattack. The bill also designs an oversight mechanism centered on independent evaluators, which must register with the Commerce Department and regularly evaluate labs and their frontier models. Once an evaluator determines that a frontier model poses an imminent enumerated risk, it must report to the Commerce Department within 72 hours after completing the evaluation. After finding a risk, the Commerce Department may apply to a court to suspend the model's operation, with a federal judge in the U.S. District Court for the District of Columbia making the final ruling. Any other U.S. government department, agency or official would be explicitly prohibited from suspending the development, deployment or internal use of an AI model. ^[4]

Supporters argue that this design gives the government the ability to stop or contain dangerous deployment of the most powerful models and requires developers to test frontier AI models, publish evaluation results and accept independent third-party assessments. Critics, however, may argue that the triggers are too narrow, effectively providing a considerable legal shield for the continued operation of frontier AI labs such as Anthropic. The bill allows AI labs to pay evaluators but stipulates that compensation must not be linked to evaluation conclusions—an arrangement that could still raise questions about evaluator independence. For investors, if the bill becomes law, it would mean frontier AI models enjoy a high degree of operational certainty in the U.S. market, with regulatory intervention risk strictly constrained by law. ^[4]

Running parallel to the regulatory lobbying is expensive financing. Anthropic is advancing its largest-ever chip financing at a cost far above market benchmarks. According to the Financial Times on the 17th, Anthropic is offering Blackstone a yield above 10% in exchange for the latter's participation in the debt financing code-named Project North Star, with a total size of up to $60 billion. People familiar with the matter said a Blackstone fund will subscribe to $9 billion of that, about half of the $18 billion subordinated loan, and receive a 10% coupon, with the actual total return higher after adding new-issue discount and other fees. The financing is expected to set a record for chip financing and become an important gauge of investors' appetite for AI debt. Notably, just four months earlier, Anthropic completed about $4.4 billion in financing code-named Project Big Sky at an 8.5% rate. The rapid rise in borrowing costs highlights growing investor wariness about AI companies' intensive financing needs. ^[6]

Project North Star aims to fund Anthropic's procurement of about 1.5 gigawatts of custom semiconductors developed by Google, namely tensor processing units. The overall debt structure has two layers: about $42 billion in senior debt, guaranteed by Broadcom, which is involved in chip development; and the remaining $18 billion in subordinated loans, of which Blackstone is subscribing $9 billion. The remaining share will be sold to a broader group of investors over the coming months, possibly at terms lower than those Blackstone received. The debt is expected to enter the institutional 144A bond market after the funds are fully drawn in 2027. According to the ICE BofA BB U.S. High Yield Index, borrowers with credit profiles similar to Anthropic typically pay about 7% in the public bond market. Anthropic's 10% coupon to Blackstone is more than 300 basis points above that benchmark, and the actual yield gap is wider when discounts and fees are included. Rating agency KBRA rated Anthropic one notch below investment grade in June this year, and the subordinated debt is expected to receive a similar rating. ^[6]

Behind the high-cost financing is a clear strategic logic. According to media reports citing people familiar with the matter, Blackstone has told its investment partners that the deal is a rare opportunity to invest in Anthropic-linked debt at a high yield; once Anthropic completes an IPO, it is expected to receive an investment-grade rating, which would sharply lower its financing costs. This is essentially a time arbitrage: before going public, Anthropic trades a higher rate for the necessary computing reserves and bets that a post-IPO credit-rating upgrade will create refinancing room, thereby lowering long-term capital costs. For early investors such as Blackstone, it is a way to lock in above-market returns during the risk window. The financing is both a funding guarantee for Anthropic's expansion and a key stress test of whether the AI sector's debt-financing market can continue to absorb large transactions. ^[6]

Japan rate-hike expectations and capital repatriation from global bond markets

The International Monetary Fund has unusually endorsed Japan's monetary policy normalization and warned that if inflation exceeds expectations, the Bank of Japan may need to accelerate rate hikes. Krishna Srinivasan, director of the IMF's Asia and Pacific Department, told Reuters during the IMF annual meetings in Bangkok this week that the BOJ is doing the right thing and stressed that if inflation rises faster than expected, tightening must be faster because it does not want inflation expectations to become unanchored—this depends on the data. BOJ Governor Kazuo Ueda also reiterated at the IMF meetings that the central bank will focus on anchoring core inflation near its 2% target and will not tolerate any overshoot, further strengthening market expectations for another rate hike this year. The BOJ raised rates in June and September this year, pushing the policy rate to 1.25%, the highest in 31 years, and most analysts expect at least one more hike this year. Srinivasan noted that a weak yen has further raised raw-material import costs and intensified domestic inflation pressure, one important backdrop for the BOJ's continued tightening. The BOJ will hold a key policy meeting later this month; the market widely expects rates to stay unchanged then, but the meeting may send a hawkish signal to pave the way for a December hike. ^[5]

The IMF's remarks come as global bond markets are under pressure. The French 10-year government bond yield briefly approached 5%, a new high since 2002, with borrowing costs now above those of Greece and Italy. The U.S. 10-year Treasury yield climbed to a multi-decade high of 5.28%, and even softer inflation data failed to stem the selloff. In the view of some market analysts, Japan's policy normalization and the resulting repatriation of Japanese capital may be an unignorable structural force behind both bouts of turbulence. Bloomberg columnist Gearoid Reidy wrote in an analysis that the world may not be fully prepared for a Japan with positive rates, and France may be only the first market hit. Data show that as of July this year, Japanese investors held about 23 trillion yen, or roughly $145 billion, of French government bonds—Japan's largest overweight position in the euro area. However, as Japan's 10-year government bond yield rose above 3% last month, a 30-year high, the yield advantage of French 10-year bonds over JGBs after currency hedging has narrowed to about 40 basis points, sharply reducing their appeal. ^[5]

With the yield advantage fading and France's fiscal position deteriorating, Japanese capital has begun to loosen. Data show that Japan's holdings of French government bonds have fallen 2.5% from the end of last year. A global fixed-income fund managed by Shinji Kunibe at Sumitomo Mitsui DS Asset Management has sold all its French bonds due to fiscal concerns. Mizuho's Masayuki Nakajima said that even if valuations look cheaper, Japanese investors' incentive to rebuild positions is declining. As the IMF provides policy support for faster BOJ rate hikes, the continued recovery in the appeal of Japanese domestic assets may extend this trend and create persistent structural pressure on overseas bond markets that rely on Japanese capital. ^[5]

On exchange rates, the IMF endorsed the Japanese authorities' stance. Srinivasan said Japanese authorities are fully committed to a free-floating exchange-rate regime and that he has had many discussions with them and has no doubt about it. In the analytical chapter of its Asia economic report, the IMF noted that the large swings in the yen since mid-2025 are largely unrelated to fundamentals such as yield differentials and instead moved in sync with speculative positions in yen futures. Japan conducted a rare coordinated intervention with the United States on July 31 this year to stop the yen from falling to a near-40-year low. Srinivasan said Japan's foreign-exchange market is deep enough, authorities use exchange-rate volatility as the first line of defense, and intervention is used only in specific circumstances. Beyond monetary policy, Japan's fiscal sustainability also leaves markets with doubts. Markets worry that the government's large spending plans, combined with a two-year consumption-tax cut starting next April, will further expand an already huge debt burden. Srinivasan acknowledged market concerns about fiscal risk but also noted that since the pandemic Japan's fiscal consolidation has actually been stronger than that of any other G7 country, partly because of continued growth in tax revenue. He stressed that Japan needs to clarify funding sources for new spending and tax cuts and establish a medium-term fiscal framework to stabilize market expectations. ^[5]

Conclusion: the financial system's defense lines under AI and rate repricing

Taken together, these threads show that AI is simultaneously changing how funds flow, how markets price risk and where regulatory responsibility lies. Banks restricting AI agents' access to accounts is an access control to protect low-cost deposits at a time when deposit rates are still low and the gap between 0.1% checking accounts and 3.3% to 5% money-market yields is wide; but with liability determination unclear, individual bank blocks may only buy time. In U.S. equities, a financial conditions index at about 1.4 standard deviations and the equal-weight index's 6% lag indicate that the tight-money shock is being borne by ordinary stocks, with index stability masking deteriorating breadth; financials and the MOVE index are key to watching the convergence path. At the AI company level, Anthropic is promoting a high-threshold safety bill while financing at a 10% coupon, showing that regulatory certainty and capital cost are two sides of the same coin: legislation reduces tail regulatory risk, while high-yield debt buys computing power. In Japan, the IMF backs faster rate hikes; Japanese investors hold about 23 trillion yen of French government bonds and have cut holdings by 2.5%; normalization is adding structural pressure to global bond markets through capital repatriation. Going forward, watch the liability rules between banks and AI agent service providers, Anthropic's financing pricing and rating progress, the BOJ meeting and inflation data, and whether U.S. financial stocks and bond-market volatility gauges can stabilize. ^{[1][2][4][5][6]}