What are AI labs training for?
The AI Training Demand Index tracks the expertise frontier AI labs are paying for, measured across 270 roles listed in the last 30 days.
Most in demand, last 30 days
Share of roles naming each specialism. Roles usually name more than one, so shares total over 100%.
- Legal16%
- Music11%
- Finance11%
- Financial Modeling7.8%
- Lyrics Evaluation7.8%
- AI Safety6.7%
- Business Strategy6.3%
- Medicine5.6%
- Contract Review5.2%
- Investment Banking4.1%
- Engineering Content Review3.7%
- Mechanical Engineering3.3%
- Python2.6%
- Audio Engineering2.6%
Rising and cooling
How each specialism's share of the last 30 days compares with its share of all 1,000 tracked roles. Measured in percentage points, so it reflects a change in mix rather than in how many roles we collected.
- Music+8.1
- Legal+6.8
- Lyrics Evaluation+5.7
- Contract Review+2.6
- Financial Modeling+2.6
- Engineering Content Review+2.4
- Investment Banking+1.8
- Mandarin+1.5
- Physics-3.2
- Software Engineering-2.9
- Data Science-2.6
- Python-2.0
- English-1.9
What each specialism pays
Median advertised hourly rate. Specialisms with fewer than five priced roles are left out rather than shown on thin evidence.
- Investment Banking$150
- Contract Review$118
- Financial Modeling$100
- Legal$95
- Medicine$95
- Finance$88
- Mechanical Engineering$85
- Engineering Content Review$78
- Business Strategy$75
- Python$75
Who AI labs buy expert training data from
The companies supplying frontier labs with expert-generated training and evaluation data.
Next: once enough history has been collected, all ten vendors will be tracked in the index directly, so their demand mix can be compared side by side.
- 01MercorExpert marketplace matching domain specialists to frontier labs
- 02Rise Data LabsExpert AI training data and RL envs from a pool of half a million domain experts
- 03Surge AIRLHF and human preference data at scale
- 04Scale AIAnnotation and evaluation, with Outlier as its expert platform
- 05Invisible TechnologiesExpert operations, running the Meridial marketplace
- 06Micro1Vetted expert network for data operations and evaluation
- 07AfterQueryExpert-generated datasets and evaluation environments
- 08TolokaCrowd and expert data, running the Mindrift platform
- 09SuperAnnotateAnnotation tooling plus managed expert workforces
- 10AppenLong-running provider of language and search relevance data
Method
The index reads 270 remote AI training roles listed in the 30 days to August 22, 2026, drawn from a leading AI training marketplace, of which 251 publish an hourly rate. It is one consistent sample rather than every listing on AlignList: mixing sources would make the figures jump whenever a new one is added, which would look like a market shift but would only be a change in what we collect.
Specialisms are assigned per role by a language model reading the title and description, and a specialism needs at least three roles to appear. Roles usually name several, so shares total over 100%. Rates are the mid-point of each advertised range, reported as a median, and shown only where at least five priced roles support the figure.
On the comparison. Rising and cooling compare each specialism's share of the last 30 days against its share of all 1,000 tracked roles. Shares are used rather than counts on purpose: how many roles we record in a period depends on how often the source is polled, so a count comparison would report changes in our own collection as changes in demand. A shift in mix is meaningful; a shift in volume, here, would not be.
Figures describe advertised demand, meaning what labs are paying to have trained, rather than hours worked or positions filled.