AI labor economics · tracker
Capability is here. Viability is the question.
The industry has spent hundreds of billions proving AI can do the work. Whether it pays — to actually run, role by role — is a different number. This page is that number on open methodology: the macro proof first, then every role costed.
Role numbers are the matrix-derived read at a representative cell (Tier 2 US city × mid experience). Your mix will differ — compute your own Wagecard for the precise read.
The industry proof
Hundreds of billions bought the capability. The payback isn't showing up.
The capital pouring into AI infrastructure is on the record. What AI earns back mostly isn't. Same pattern as every role below — capability is the easy part; viability is the question.
The industry proof
The capability is bought. What it earns back, the market mostly can't see yet.
Amazon, Alphabet, Microsoft, Meta — filed FY2025
Two of the four disclose none at all
On $6B of its own capex — the supplier wins
Public hyperscalers — capex vs disclosed AI revenue
Amazon
FY2025 capex (company-wide)
AWS AI revenue run-rate, Q1 2026
Amazon FY2025 capex & AWS AI run-rate (Yahoo Finance) (2026-04-09)
Alphabet (Google)
FY2025 capex
No AI-specific revenue line
Alphabet Q4 & FY2025 earnings release (2026-02-04)
Microsoft
Only one to disclose AI revenue and AI capexFY2025 AI data-center capex
AI business run-rate, Q3 FY2026
Microsoft FY2025 AI data-center capex (CNBC) (2025-01-03) · Microsoft Q3 FY2026 newsroom (AI run-rate $37B) (2026-04-29)
Meta
FY2025 capex
No AI-specific revenue line
Meta Q4 2025 results (8-K, exhibit 99.1) (2026-01-28)
Nvidia
Sells the shovelsFY2026 capex (own property & equipment)
FY2026 data-center revenue
Nvidia Q4 & FY2026 earnings release (SEC) (2026-02-25)
Private frontier labs — capital raised vs revenue run-rate
OpenAI
Capital raised, Mar 2026 round
Revenue run-rate, Feb 2026
OpenAI $122B round (Bloomberg via Yahoo Finance) (2026-04-02) · OpenAI ~$25B run-rate (The Information via Yahoo Finance) (2026-02-15)
Anthropic
Series H, May 2026
Revenue run-rate, May 2026 (self-reported)
Anthropic Series H announcement ($47B run-rate) (2026-05-28)
xAI
Series E, Jan 2026
FY2025 revenue (consolidated, filed)
xAI Series E announcement (2026-01-06) · xAI FY2025 revenue, SpaceX S-1 (via Yahoo Finance) (2026-05-21)
How these numbers work
Figures are compiled from company SEC filings, official earnings releases, and named financial reporting as of June 2026 — each cited in the row above. They are not like-for-like: capex is cash committed in a fiscal year; an AI-revenue figure is an annualized run-rate a company chose to disclose, not booked annual revenue; capital raised is balance-sheet, not spend. We do not net them — one “profit” number across these bases would be wrong. Where a company discloses no AI-specific revenue, we show “Not disclosed” rather than estimate it. Two reasons the company-level picture stays murky: most AI revenue isn't broken out, and the AI economy is circular — the same dollars cycle between cloud providers, chipmakers, and labs. Figures tagged est. rest on press or self-reported estimates, not a filing. Macro framing inspired by isaiprofitable.com; figures and sources are our own.
At the company level the payback is mostly undisclosed — and where it shows, the spend dwarfs it. At the role level it's calculable. That's the rest of this page.
Why it resolves at the role level
Capability is cheap. Viability is the question.
They say
AI runs a knowledge-work role for the price of a few API calls — the token bill is rounding error against a salary.
The math says
Across these 15 roles, raw token spend is about 45% of what the AI substitute actually costs to operate — oversight, retries, orchestration, and integration carry the rest. Mean substitution exposure is 41/100, and 57% of the hours stay human-led or human-critical.
What the math says, in aggregate
Most knowledge work is augmentation territory, not replacement
Across 15 roles, the average economic substitution exposure is 41 / 100. By hours, 6% of the typical week lands in fully-replaceable territory — the rest stays human-led, human-critical, or AI-augmented.
Roles costed
Each at a representative Tier 2 / mid cell.
Average exposure
Lower = more human-leveraged. 57% of hours stay human-led or human-critical.
Fully-replaceable share
Of the average week, by hours, across all roles.
Where the hours sit, on average
Role by role
What AI actually costs to run, per role
Sort by exposure, by what the market pays, or by the gap between pay and AI run-cost. Every row opens the full methodology read.
Exposure runs 0–100; lower means more of the role's hours stay human-leveraged. AI run-cost is what it costs to run AI across the same task-week — tokens, oversight, retries, integration — and excludes human-critical hours the model doesn't hand to AI. It is a run-cost, not a savings figure.
Computed against capability matrix v1 · model v1-mvp · representative cell (Tier 2 / mid). Median pay from BLS OEWS blended with market compensation data. Numbers refresh when the matrix refreshes. Open methodology at /methodology.
FAQ
Questions the number raises
Is this an AI-risk score?
Where do the numbers come from?
Why is the AI run-cost so much lower than the pay?
Why these roles, and why one representative cell?
Your week isn't the representative cell. Get the precise read.
Pick your geo, experience, salary, and the task mix that fits your actual role. Anonymous preview before sign-in; full Wagecard with a free account.