Meta's AI Push Sparks Job Anxiety

The Keystroke-Tracking Tool That Changed Everything

Meghan Bobrowsky's page-A1 feature reports that Meta has deployed an internal software tool that records employees' keystrokes — not for security or compliance, but to train AI on the basic computer tasks employees do every day. There is no opt-out. Internal sentiment, measured by the workplace-anonymous app Blind, is at record lows. Meta has framed the tool as “productivity research,” but the function is unambiguous: capture human procedural knowledge in machine-readable form so that the AI can reproduce it.

This is the cleanest articulation yet of what AI productivity actually means at the corporate level: the AI is not augmenting the worker; it is studying the worker so that the next-version AI can replace the worker. Bobrowsky reports that Meta's Reality Labs has already absorbed 1,500 layoffs this year; a new AI-engineering organization launched in March with a 50:1 employee-to-manager ratio (versus a more typical 8:1 in tech orgs). Mark Zuckerberg has reportedly begun developing a so-called “CEO agent” to assist him in his own role.

Block, Amazon, Oracle — The Pattern Generalizes

The article quotes Block CEO Jack Dorsey: “Within a year, the majority of companies will reach the same conclusion.” Block has restructured around AI-augmented small teams. Amazon has signaled its AI investments will materially reduce headcount. Oracle has taken a similar position. The pattern is structural: the companies pushing AI hardest have realized that the productivity gain comes through headcount compression, not headcount addition.

For an investor, this matters in a precise way. The bull case for AI infrastructure (NVIDIA, Broadcom, the hyperscalers) was that AI represents a new revenue layer on top of existing economic activity. The reality, increasingly, is that AI represents a margin substitution — output stays roughly constant while labor costs drop. That is positive for corporate margins in the medium term and for the productivity statistics in the long term, but it is negative for the consumer-spending base that underwrites discretionary categories. It is also negative for the “AI productivity will lift everyone’s wages” political narrative that the Bessent interview in this same paper carefully tries to manage.

$135 Billion of Capex, Not Yet a Revenue Story

Meta has guided to $135 billion of infrastructure capex this year. Microsoft, Amazon, Alphabet, and Meta combined will likely spend over $670 billion in 2026, the third consecutive year of greater than 60% combined capex growth. The depreciation tail of this spend — the non-cash expense that flows through the income statement over 4-6 years — will compress reported earnings in 2027-2028 even if cash generation stays strong.

The market is currently pricing the AI infrastructure trade as if 2026 capex equals 2027 revenue. It does not. The revenue contribution from AI products, on a fundamental rather than promotional basis, is still single-digit billions across the hyperscaler complex. The right framing for now is: NVIDIA and Broadcom are direct beneficiaries because they sell the picks and shovels in the period of capex; the application-layer companies (Meta, Amazon application businesses, Google Search ads) are net costs in the period of capex and net beneficiaries only in 2027-2028 if and when the revenue layer materializes.

What This Means for the Book

(1) Trim AI/Cyber/Data from 20% to 18%. The headcount-displacement narrative is now in the price for the application-layer (META, AMZN, SQ) names. The capex-beneficiary names (NVDA, AVGO, GOOGL, MSFT) stay as core. Redeploy the freed 2% to Energy.

(2) Cyber stays at 5%. The Anthropic Mythos breach narrative from earlier this week and the keystroke-tracking story together signal that AI runtime security is a board-level priority. CRWD, PANW, ZS, S, OKTA stay as a higher-conviction segment within the trimmed AI sleeve.

(3) Watch consumer discretionary. If the Block/Meta/Amazon AI-displacement pattern accelerates, the consumer-spending base softens through Q3-Q4. UAL is the exception (Kirby playbook is a moat story), but DAL/AAL/LUV and the broader discretionary complex face a real headwind.