The Board Room

Why a loan mark bites harder than a lower multiple

A falling share price hurts your shareholders. A loan markdown reaches your operating plan, because it changes what the next lender will offer. The Bear Cave's reading is that AI disruption has crossed from a multiple-compression story, painful but survivable, into a question of cost of capital and access to capital that lands with the CFO at refinancing. In our view, a markdown at a bank of JPMorgan's weight also works as a template. Other lenders now have a precedent for wider spreads, tighter covenants and explicit AI-disruption diligence. Bank stocks show the strain: the largest US banks have fallen 11% to 25%, with Capital One down 25% and Goldman Sachs down 24%. Those drawdown figures come from social posts, so treat them as directional until filings confirm them.

Who gets asked first

The Bear Cave ranks sponsor-backed, seat-priced software with near-dated maturities as the most exposed borrowers. The seat detail matters because the substitution evidence has reached macro data. a16z charts show call-center employment flipping from roughly 4% annual growth over 15 years to a roughly 4% annual decline. That makes it among the first visible job categories where AI substitution shows up in aggregate numbers rather than vendor case studies. A vendor pricing per seat into support organizations is selling into a shrinking seat base, and lenders can now see it.

The Information's analysis of Amazon shows the same logic at the top of the market. Amazon trades at 29x forward earnings against Walmart's 37x, despite growing about three times faster. One of the two risks the market appears to be charging for is agents replacing human shoppers. If that happens, Amazon's $80 billion ad business, mostly merchant promotion aimed at people, could be wiped out. Seats and ads rest on one assumption: a human on the other end. Credit and equity markets are now both pricing what happens when it fails.

Why an AI slowdown would not rescue you

A skeptic would argue that if AI adoption stalls, disrupted software borrowers recover. The Bear Cave disagrees, and so do we. The disruption thesis is already embedded in the underwriting, and a lender does not reverse a markdown on a hunch. It reverses on evidence: cohort retention and expansion in AI-exposed segments, revenue tied to AI usage, and realized pricing on AI features. A company that cannot show it is where the AI budget lands gets priced as a casualty in both scenarios.

The smart move

Treat your next refinancing as a diligence process that has already started. Map every facility, maturity and covenant, and model terms at materially wider spreads. Then decide whether terming out debt or adding lenders is worth the cost while windows are open. Build the evidence file in parallel, because credit committees weigh data over narrative. The same work pays on the buy side. If credit keeps tightening, AI-exposed software assets will come to market at reset valuations. A corp dev team already holding a ranked watchlist with price triggers can move at those prices.

Once AI disruption sits in the loan file, the only argument a lender accepts is data showing the AI budget lands with you.

The arithmetic behind the bonds

Data-center debt is repaid in dollars, not tokens. For borrowers to service it, total token spend has to keep rising, which means usage must grow faster than unit prices fall. The evidence shows unit prices falling from two directions at once. The first is list price. Anthropic cut its small-model rate 90% with Claude Haiku 5.5, a move TheSequence reads as a play for volume share, not margin. The second is compute per task. Microsoft Research Asia's Agent Lightning, which trains agents inside the tools they run in production, lifted a 9-billion-parameter open model from 41.8% to 56.4% on the SWE-bench Verified coding benchmark with roughly 6,000 training samples. Results like that let small, cheap models do work that previously needed frontier ones.

The Bear Cave names these exact forces, slowing token spend or efficiency gains that cut compute per task, as what breaks the infrastructure thesis. In our view the efficiency half is already visible. The open question is whether volume growth outruns it.

Where the concentration sits

The exposure is not spread evenly. TheSequence reports that about 70% of Lambda's $50 billion backlog is a single Anthropic commitment, while Lambda raises up to $4 billion at a $14.5 billion pre-money valuation plus $1 billion of debt. A supplier whose backlog rests on one lab carries that lab's strategy as counterparty risk for every other customer. At the top of the market, The Information notes Amazon is spending $220 billion on data centers this year, with Alphabet and Meta making the same bet, and that the payback remains uncertain.

The public window is closing, not the money

Capital has not dried up. It has become private, sovereign and selective. The Information reports that SoftBank's Masayoshi Son is seeking up to $100 billion from Gulf investors and that Valar Atomics closed a $1 billion Sequoia-led round. Nvidia-backed Firmus Grid pulled its IPO citing volatility and will seek private funding. Sources disagree on that deal's scale. The Information describes a raise of nearly $5 billion at a valuation above $30 billion. Social posts flagged by The Bear Cave cite a $44 billion target against $51 million of revenue and 46 MW of installed capacity. Neither figure is confirmed by a filing. For a customer, the pattern matters more than the number: providers that planned on a public exit now depend on private rounds that can reprice or disappear.

The smart move

The efficiency curve is good news for your AI unit costs and a warning about whom you buy capacity from. We would bank the savings and shorten the exposure. Credit-review every compute and data-center counterparty without an investment-grade balance sheet, cap prepayments to them, and negotiate portability, termination or step-in rights. Keep fixed-price inference commitments short, with repricing clauses, because cheaper tiers keep arriving. Then read the Q3 hyperscaler reports against a baseline you wrote down beforehand. If capex guidance holds while AI demand commentary softens, the debt side of this trade is getting riskier.

The cheaper each AI task becomes, the more your compute supplier's debt depends on the world buying far more of them.

A compliance finding became a lost customer

Patch hygiene used to arrive as an audit finding with a remediation deadline. The Accenture case suggests it can now end a flagship customer relationship. The PeopleSoft breach exposed data on almost all FBI special agents, and the bureau responded by changing contractors. It did not demand a corrective plan. Accenture's side is not in the reporting we reviewed, so the causal link is reported and unconfirmed. We expect patch SLAs and termination triggers to tighten at the next renewal cycle for any company selling into government or regulated buyers.

The German ruling applies the same logic to individuals. The court treated the manager's recent anti-scam training as proof he should have known better. Training exists to reduce risk, and it still does. It also creates a record that raises the standard of care for whoever received it. Risky Bulletin reads the ruling as weakening both “we were deceived” and “the tool did it” as defences.

Error budgets cannot see where a change came from

The metrics leaders review do not register the newest failure modes. Paul LaPosta, in a piece SRE Weekly featured, argues that a change depending solely on an AI model, on one person's knowledge, or on a single point of failure is riskier than a comparable change the team can explain and operate unaided. An error budget, the allowance of unreliability a team may spend, scores the two identically until one fails. Alex Ewerlöf puts the human side plainly: “AI can explain it to you but it cannot understand it for you.”

Volume makes this pressing. The Institute for Ethical AI & ML reports that agent-authored pull requests grew 9x in eight months and now outnumber human ones, and one engineer it quotes calls human review “theater.” Much of that adoption data comes from vendors and AI-heavy firms, so it is directional. It warns that an agent can be 100% available, 100% within latency, and 100% wrong while every dashboard shows green. Ungated change paths already cause public outages. Railway's September 30 disruption came from code shipped before the database change it depended on. Another firm ran key business logic as scripts analysts edited live in production, outside version control.

A named-human rule

Two measures most dashboards omit belong in board review: patch latency on internet-facing systems against contractual commitments, and the share of material changes only a model can explain. The companion rule is that a named engineer must be able to explain and operate an AI-assisted change without the model before it ships. The rule costs speed. Railway shows what skipping it costs. Vendors warrant the same evidence standard. incident.io published how it measures the accuracy of its AI SRE, a tool that assists with incident response, and SRE Weekly's editor called that disclosure rare. That disclosure is the right procurement bar, along with testing any AI operations tool against the buyer's own recent incident history before signing.

The FBI dropped a contractor and a German court held a manager liable, both over failures no dashboard flagged.