All 16 niches, one screen

Rank

Six scores, one to five, applied identically. They are judgements built on the evidence on each niche page, not measurements — and the sum is a blunt instrument, because a 1 on room to win should end the conversation whatever the total says. Read the columns.

Niche	PRF	BUD	DEF	CST	RCH	RM	Σ	Read	Who is there
Video, motion and VFX	5	4	4	4	4	5	26	build	—
Mechanical CAD and manufacturing	4	4	4	4	5	5	26	build	—
Defensive security	5	5	4	4	4	3	25	build	Loginsoft — sells 'Security Data for AI Training' and AI model validation at $25–49/hr from a 201–500 person firm founded 2005
Architecture, engineering and construction	4	3	4	4	5	5	25	build	—
Accounting, audit and tax	4	2	3	5	5	5	24	build	—
Clinical medicine	4	5	4	2	5	4	24	watch	Centaur AI, Protege (labelled artefacts only — reasoning traces are unoccupied)
Law	4	3	4	3	5	4	23	build	Vals AI (adjacent — benchmarking, not data supply)
Senior code review	3	5	2	4	5	2	21	watch	Datacurve, AfterQuery, Mechanize, Turing, Scale, Surge, ~20 more
Investment banking and financial modelling	3	5	2	5	4	2	21	crowded	Halluminate, AfterQuery, Vals AI
Offensive security	2	5	3	3	5	2	20	crowded	Gray Swan, Irregular, Dreadnode, Trajectory Labs, 10a Labs, Palisade Research
Design and UI/UX	2	4	2	4	5	1	18	crowded	Contra Labs, Taste Labs, Design Arena, LM Arena, Verita AI
Sales and GTM	2	1	1	3	5	5	17	avoid	—
Game development and 3D art	2	2	2	3	5	3	17	avoid	Worldmodeldata, Good Start Labs (both adjacent, neither selling dev expertise)
Skilled trades and field service	2	1	2	1	5	5	16	avoid	—
Life-science wet lab	2	3	4	1	4	1	15	crowded	Edison Scientific, Mercor (via Sepal AI), Lila Sciences, Periodic Labs
Chip design and EDA	2	2	3	1	3	4	15	avoid	Phinity Labs (unconfirmed)

01

What the columns mean

Higher is always better for the operator.

PRF

Speed to proof

How fast you can be visibly the best provider in this niche — the reference customer, not the first invoice.

BUD

Budget now

Whether a buyer is demonstrably spending on this domain today, evidenced by a posting, a system card or a contract.

DEF

Defensibility

Whether the supply stays scarce and the data keeps needing refreshing three years out.

CST

Cheap to start

What it costs to assemble the first cohort and run the first pilot. Higher is cheaper.

RCH

Reachability

Whether a named register, forum or marketplace puts you in front of the supply this month.

RM

Room to win

How much of the position is still open. A niche with an entrenched specialist scores low however good it looks.

02

The same table in words

Each niche's one-line read, in rank order.

Video, motion and VFX

The most under-priced position on the shortlist. A named lab rate card, an embarrassing benchmark, nobody selling — and a workforce whose expectations of what AI companies pay were set at $4.88.

Mechanical CAD and manufacturing

Everything is already in place except the price. Labs currently buy mechanical engineering as generic STEM reasoning; the whole business is repricing it as a vertical.

Defensive security

There is no SWE-bench for the SOC. Whoever builds it owns the category's vocabulary — and the only way to build it is on public malware, never on customer telemetry.

Architecture, engineering and construction

Recruiting is nearly free at these rates and the vertical is empty. Stamping liability caps you at evaluation, never production — build accordingly and the ceiling stops mattering.

Accounting, audit and tax

Best structural economics of the eight. The buyer is an accounting software company, not a frontier lab, and you must be honest with yourself about that before you start.

Clinical medicine

Everything about the demand says build. The wage floor says you can only afford the specialties the buyer cares least about — solve that and this is the best niche here.

The strongest under-served niche in the set, on the condition that you build the drafting benchmark before Vals AI gets round to it.

Senior code review

Do not enter the general coding lane. The review-and-critique sub-lane is genuinely underserved, and it is a feature of someone else's roadmap, not a company.

Investment banking and financial modelling

The most obvious niche in the report, which is why the buyer has insourced it and a funded competitor is already there. Strong economics, weak defensibility.

Offensive security

Ranked first on evidence and last on availability. The benchmark layer has already been captured by vendors, which makes the usual entry weapon useless here.

Design and UI/UX

The inspiration for the model is the one niche you should probably not enter. Copy the method, not the market.

Rank last. When a frontier lab enumerated the four domains it was surging specialists into, sales was not one of them — and the buyer who does own this data already owns it.

Game development and 3D art

The door was closed before you got here. Unity's July 2026 terms prohibit training AI on Unity data, and no frontier lab is asking for game-development judgement anyway.

Skilled trades and field service

Rank last. The intuition that physical-world expertise must be valuable is strong, widely held, and wrong here — the buyer has not appeared and the cost base is capital, not labour.

Life-science wet lab

The layer you can afford is the layer four funded companies already occupy. The layer where you would be unique needs a bench, and the marginal experiment costs hundreds to thousands of dollars.

Chip design and EDA

Attractive on paper, structurally hostile in practice. The data that matters is telemetry Cadence already collects as a by-product of software its customers pay for.