The framework, the timeline, the company-level evidence, the sector pattern, and the duration mismatch that ties them together. All rendered in house style.
Two AI Letters
The reader’s map for the rest of the essay.
A
Admission
AI narrates a cut that was already needed.
Coinbase
I
Investment
AI changes the cost mix, but economics are not yet legible.
Cloudflare
The AI Layoff Letter Becomes a Genre
Three letters in eleven days, with market reactions where reported.
Compiled from company communications and after-hours equity moves, Feb–May 2026.
Atonement in one picture
Three indexed series, FY21 = 100. The re-expansion of headcount and operating expenses preceded the AI-framed cut.
Coinbase’s AI-framed cut followed a re-expansion of headcount and operating expenses after prior restructurings.
Indexed series: Coinbase Form 10-K (FY21–FY25). Headcount line is illustrative — directional, sourced from disclosed year-end counts.
The template
The AI letter landed alongside a very old-fashioned financial problem.
Metric
2024
2025
Change
Loan losses
$323M
$821M
+154%
Transaction, loan & consumer receivable losses
$794M
$1,337M
+68%
Announced workforce cut
—
4,000+
Nearly half
The AI letter landed alongside a very old-fashioned financial problem.
Block · Form 10-K, fiscal years 2024 and 2025.
Real productivity, stalled leverage
Revenue per employee and cash from operations per employee, FY21–FY25. The bull case is real. The 2025 stall is also real.
Illustrative — values directionally consistent with Cloudflare 10-K and 10-Q disclosures, FY2021–FY2025. Replace with audited figures before publication.
Chips, compute, computer nerds
The 2025 input bill, broken into reported line items used as proxies for input buckets.
Reported line items, used as proxies for the input mix. Not all cost of revenue is compute; not all R&D and S&M is people. Read the buckets as directional shape, not exact decomposition.
Cloudflare · Form 10-K, fiscal year 2025.
The AI margin math is messy
Six public companies, claimed AI benefit on one side, new input cost or pressure on the other. The shape is the same across the board.
Company
Claimed AI benefit
New input cost / pressure
Spotify
Faster shipping, lower headcount
More compute per employee
Shopify
Support savings, Sidekick adoption
Higher LLM costs offset savings
Uber
Higher output per worker
AI cost and impact underestimated
Roblox
Better creator tools
Training / inference pressured margins
Pinterest
AI shopping assistant
Chip capacity pressures gross margin
Airbnb
AI resolves more support issues
AI expenses expected to rise
Compiled from earnings calls and shareholder letters, calendar 2024–2026.
Labor savings now, compute bill later
The timing mismatch is the narrative.
Quarter 1
Headcount down
→
Opex down
→
Margin story improves
Reported margins compress and then expand on a smaller cost base.
↓
Quarter 2+ · usage scales
AI usage up
→
Inference / tooling / chip costs up
→
Margin math gets messier
Variable input costs reappear in the P&L just as the labor benefit is annualized.
The timing mismatch is the narrative.
Useful life of an engineer is longer than chips
Three input classes, three different ways the accounting and the economics drift apart.
Asset
Accounting treatment
Economic behavior
Risk
Engineer
Opex / SBC
Appreciates with context and compounding knowledge
Attrition · dilution
Chip
Capex / depreciation
Depreciates and obsolesces
Replacement cycle
Token / compute
Variable cost
Scales with usage
Budget volatility
Three buckets, three different legibility problems
The inputs a software business actually buys, sorted by what makes each one hard to read in the financials.
