What AI teams are actually hiring for

Every skill below is counted across 34,217 live job postings on our board, of which 8,738 are data, AI or software roles. It is a count, not an opinion — and it is recomputed daily as the board changes.

The ranking deliberately leads with specialised skills rather than languages. Python appears in more postings than anything else here, and tells you nothing: everyone writes Python. What separates candidates is the domain work — post-training, inference optimisation, simulation, the infrastructure underneath a model.

Specialised skills

Ranked by how many data / AI / software postings ask for them.

#	Skill	Category	Roles	Share
1	Agents & tool use	AI systems	422	4.8%
2	Robotics & embodied AI	Applied domains	312	3.6%
3	AI infrastructure	AI systems	270	3.1%
4	Streaming & real-time data	Production ML	182	2.1%
5	RAG & retrieval	AI systems	102	1.2%
6	Model serving & MLOps	Production ML	86	1.0%
7	Simulation	Applied domains	85	1.0%
8	Multimodal & VLA	Model work	68	0.8%
9	Computer vision	Applied domains	55	0.6%
10	Fine-tuning	Model work	46	0.5%
11	Reinforcement learning	Model work	44	0.5%
12	NLP	Applied domains	39	0.4%
13	Experimentation & A/B testing	Production ML	34	0.4%
14	Evals & benchmarking	Model work	33	0.4%
15	Post-training (RLHF/DPO/SFT)	Model work	33	0.4%
16	CUDA & GPU kernels	AI systems	33	0.4%
17	AI safety & interpretability	Model work	32	0.4%
18	Inference optimisation	AI systems	19	0.2%
19	Recommenders & ranking	Applied domains	17	0.2%
20	Compilers (XLA/MLIR)	AI systems	8	0.1%
21	Pretraining	Model work	7	0.1%
22	Data curation & synthetic data	Model work	7	0.1%
23	Chip & silicon design	Applied domains	6	0.1%
24	Distributed training	AI systems	5	0.1%

Baseline tools

Table stakes rather than differentiators — shown for scale.

#	Skill	Category	Roles	Share
1	Python	Core tools	1,326	15.2%
2	SQL	Core tools	537	6.1%
3	AWS	Core tools	406	4.6%
4	Kubernetes	Core tools	212	2.4%
5	PyTorch	Core tools	155	1.8%
6	Spark	Core tools	103	1.2%

How this is measured, and where it's weak

Each posting's title, description and tags are scanned for the terms behind each skill. A company never counts towards its own name — Databricks has hundreds of open roles that all say “Databricks”, which measures who is hiring, not what the market wants.

The limit worth knowing: the source truncates every description at 603 characters, so what gets scanned is a posting's opening blurb, not its requirements list. Only about 18% of the data/AI/software postings name any specialised skill at all — the rest are cut off before they get there. So read these as relative demand between skills, not as “82% of teams don't want this”. Absolute counts are floors, and the more specialised the skill, the more it is understated.

A term also has to be written to be counted: a team can want post-training experience without using the phrase.

There is no pay data here on purpose. Fewer than 0.2% of these postings state a salary, and the upstream source carries no salary field, so any “highest paid skills” ranking built from them would be noise dressed up as a finding.

Last computed August 19, 2026.

Browse the roles these come from →