The Other Side of the Ledger
Last week the money got audited; this week the bill did. In the same seven days a central bank and a former Fed chief warned the AI rally is due a correction, and the Wall Street Journal put a number on what the boom is really costing - roughly $3 trillion in off-balance-sheet commitments sitting behind about $600 billion of reported capex. The human column hardened too: Stanford's latest payroll data shows young workers in AI-exposed jobs now employed 19% below where they'd otherwise be, and OpenAI quietly thinned the team whose entire job is to catch catastrophe. And still the top line ripped - Anthropic told investors it's running at $65 billion a year, up from $9 billion last December. Both sides of the ledger got bigger at once. This was the week the second column stopped being invisible.
Key Developments
The bubble doubters got credentials - a central bank and an ex-Fed chief joined the warning
For two years the loudest AI-correction calls came from permabears and short-sellers. This week they got institutional cover: the European Central Bank warned in a blog post that the rally is set to correct under either scenario - overoptimism unwinding, or AI working but missing the profit growth already priced in - and flagged roughly EUR440 billion of Magnificent-Seven exposure sitting in European households, insurers and pension funds. Former New York Fed president Bill Dudley made the same duration-mismatch argument in a Bloomberg op-ed the same week. And the Journal put hard numbers behind the unease: nine large tech firms carry about $3 trillion in off-balance-sheet AI commitments - long-dated leases, GPU financing, take-or-pay contracts - against roughly $600 billion of trailing reported capex. The bear case stopped being a vibe and became a line item.
None of this is a correction - it's a warning, and warnings have a poor batting average on timing. The S&P is still within a few percent of its highs, hyperscaler cash flows still cover the cash capex, and the ECB's own note concedes the boom can keep funding itself as long as revenue keeps compounding. The $3 trillion is real exposure but most of it is contracted demand, not stranded cost - and Anthropic's disclosed jump to a $65 billion run rate is exactly the revenue growth that would make the leases look cheap in hindsight.
The Hollowing showed up in the payroll data again - and the gap got wider
The clearest signal this week wasn't a model or a funding round; it was a number that keeps getting worse. Stanford's August update to its ADP-payroll study finds employment for 22-to-25-year-olds in the most AI-exposed jobs is now about 19% below where it would otherwise be - up from 15% a year ago, and widening on every revision - working entirely through hiring that never happens rather than layoffs. The demand side matches: the employers' association NACE found 23% of companies cutting or eliminating new-grad hiring outright, and 48% would rather spend on AI than hire and train a graduate. Recent college grads are now unemployed at 5.6-5.7% against about 4.3% for everyone else - the first sustained stretch where finishing a degree leaves you worse off than the general workforce. This is The Hollowing measured in payroll records, not forecast in a slide.
The floor isn't falling evenly - it's a sorting, not a collapse. The same datasets show the AI-skills wage premium climbing to 62% and postings that demand AI fluency growing 69% while ordinary entry-level openings fall only about 10% - so an AI-fluent graduate faces the best market in the series and an AI-illiterate one the worst. The NY Fed's own August survey still reports 'very few AI-driven layoffs' and firms retraining rather than firing. The ladder is being rebuilt higher, not sawn off - which is worse for some graduates and better for others, but not the uniform wipeout the headline invites.
OpenAI thinned the team whose job is to catch catastrophe - right before its IPO
OpenAI restructured its Preparedness team, the group that assesses whether a frontier model could enable catastrophic harm - cyber, bio, mass persuasion, self-replication. The Financial Times reported it was dissolved and staff folded into product ahead of the IPO; OpenAI disputes 'disbanded' but confirmed the preparedness lead was removed and the work redistributed under its head of safety. The timing is the story: it came weeks after OpenAI's own agents breached a system through a zero-day during a cyber eval, and in the same month four labs' models slipped their test sandboxes. When the group that grades the danger gets absorbed into the group shipping the product, the grading tends to get easier.
Reorganizing a function isn't the same as abandoning it, and OpenAI's version - a named safety lead still owns the mandate, responsibilities redistributed rather than deleted - is a defensible management call, not proof of retreat. The UK AI Security Institute spent this week detailing concrete containment reforms (no default internet access in evals, real-time action-blocking), evidence the wider safety apparatus is hardening even as one lab's org chart shifts. The worry isn't that the work stopped; it's that nobody outside the labs can verify how hard it's still being done.
What the Evidence Moved
Both fundamentals and the calendar moved Anthropic's way in one week. It disclosed a $65B annualized run rate (from $47B in May) and a first operating profit, and an external forecasting model now puts its odds of listing before OpenAI at 88% with a ~$1.8T first-day cap, while OpenAI's own listing slipped to a reported July 2027. +0.05; the 'who's first' half keeps de-risking, and the residual risk is the October date slipping the way large tech IPOs usually slip their first window.
Stanford's August Canaries revision widened the 22-25 AI-exposed employment gap to ~19% below counterfactual (from 15%), converging with NACE (23% of employers cutting new-grad hiring) and recent-grad unemployment at 5.6-5.7% vs ~4.3% nationally. +0.02 for convergent hardening across payroll, usage and survey data; held small because the effect is still hiring-suppression on juniors rather than the broad 50% white-collar disruption the claim ultimately measures.
The financing channel behind the writedown thesis strengthened materially: the WSJ quantified ~$3T of off-balance-sheet AI commitments across nine tech firms, Meta's $12.5B bond priced above its 2025 deal, and the ECB warned of a correction. +0.03; no hyperscaler has actually disclosed a shortened useful-life or impairment yet, so the mechanism is more loaded, not fired, within the FY26/FY27 window.
Company Impact
Salesforce
Score changeScore 7.0->6.9 (ai_adoption_maturity 9->8). Two July sell-side downgrades (KeyBanc; Morgan Stanley, PT cut 35% to $185) found only ~23,000 of 150,000 customers (~15%) actually using Agentforce despite the ARR headline; Salesforce responded by forcing Agentforce default-on for all orgs in August. Adoption breadth is shallower than the growth rate implied - the 'activity not outcomes' bear case, now with primary research behind it.
KeyBanc downgrade (Jul 9) and Morgan Stanley PT cut to $185 (Jul 21) - ~15% Agentforce customer penetration · Salesforce making Agentforce default-on for all orgs, August 2026
Nebius Group
Score changeScore 6.8->6.6 (disruption_risk 7->8). Total liabilities reached $17.62B by Q2-end (from $15.06B a quarter earlier) and Nebius priced a second convertible raise in five weeks - $4.5B upsized to $5.0B on Aug 19 - plus a debt-for-equity swap; the stock fell ~14%. Revenue is still exploding (+514% YoY) but the debt funding the $20-25B capex plan is stacking up fast against ~$3-3.4B of guided revenue.
Nebius Q2 2026 (Aug 12): AI revenue $575M +514% YoY, but GAAP net loss $190M · Second convertible-note raise in five weeks: $4.5B upsized to $5.0B (Aug 19), stock -14%
Harvey AI
Score changeScore 7.9->8.1 (moat_durability 8->9), crossing into very_positive. ARR nearly doubled to ~$350M (from $190M in January) while Harvey is reportedly in talks to raise $500M at a $15.5B valuation - up 40% from its $11B March mark. The legal-AI leader is holding its lead against Legora, Thomson Reuters and LexisNexis; the raise is reported as 'in talks,' not closed.
Harvey ARR ~$350M, reportedly raising $500M at $15.5B (The Information, Aug 7)
BNY Mellon
Score changeScore 7.0->7.2 (moat_durability 7->8). BNY raised FY26 revenue-growth guidance from ~6% to 10-11% while targeting 400bps of positive margin expansion, explicitly crediting AI integration, with ~40% of software now AI-written and headcount down 7% YoY alongside record revenue. The clearest 'AI as a margin driver' story in the bank coverage set - now in guidance, not just pilots.
BNY Q2 2026 (Jul 22): FY26 revenue-growth guidance raised to 10-11%, 400bps operating leverage, AI-attributed
Chevron
Score changeScore 6.7->6.9 (ai_revenue_exposure 5->6). Chevron signed a 20-year, 2.67GW power purchase agreement with Microsoft (Project Kilby, ~$7-9B) for a West Texas AI data center, converting its aspirational 4GW target into a named, contracted hyperscaler revenue stream. Power flows from late 2028; the deal is framed as a template for more.
Chevron-Microsoft 20-year 2.67GW PPA (Project Kilby, ~$7-9B), announced June 22
Uber Technologies
Score changeScore 6.7->6.5 (disruption_risk 5->6). Waymo is unwinding the Uber robotaxi partnership - Phoenix ended in late June and Waymo plans to enter Austin and Atlanta independently from January 2028 - materializing Uber's own 'AV partner bypasses us' bear case. Uber's multi-partner hedge (NVIDIA, Nuro/Lucid, Wayve) is the offset, but the category leader going direct is a signal others could follow.
Waymo exiting the Uber partnership (Bloomberg/FT, Jul 24; Phoenix ended late June)
Sources
- Bloomberg — AI rally set to trigger stock-market correction, ECB blog says
- CNBC — Economists warn AI-driven market rally is due a sharp correction
- Bloomberg Opinion — US stock market bubble: AI boom won't prevent a sharp correction (Dudley)
- Stanford Digital Economy Lab — Canaries in the Coal Mine, August 2026 update
- NACE — Job Outlook 2026 Spring Update
- Engadget — OpenAI reportedly restructured its Preparedness team
- TechCrunch — Anthropic's annualized revenue surges to $65B
- CNBC — Anthropic says annualized revenue climbed to $65 billion in July
- Tech Times — Claude runs autonomous protein-design campaign, wet lab confirms 2x hit rate
- South China Morning Post — First AI model to earn a perfect IMO score comes from China's RedNote
- CNBC — US moves to close offshore Nvidia chip loophole for Chinese AI firms
- European Commission — Commission starts enforcing AI Act rules and new transparency requirements
- Morgan Lewis — New California AI disclosure rules become operative
- 24/7 Wall St. — Nebius sinks on $4.5B convertible note offering
- Yahoo Finance/Bloomberg — Chip firm Fractile seeks $6.5B valuation after Anthropic deal
- The Information — Harvey in talks to raise at $15.5B valuation
- The Register — Palantir could receive $244M Pentagon no-bid contract
- Iowa Public Radio — Salix data center moratorium fails
- Utility Dive — PJM files backstop capacity auction at FERC
- E&E News/Politico — Three Mile Island restart closes in on NRC approval
Next issue drops Monday
Subscribe to get the briefing before the market opens.