AsiaAI Tech Index Methodology

What the index measures

The AsiaAI Tech Index ranks companies in Japan, China, South Korea, and Taiwan by how much they matter to artificial intelligence. It is not a size ranking, and it is not a stock screen. A company can be very large and rank modestly, and a mid-cap supplier can rank near the top.

That is deliberate. The question the index tries to answer is closer to: if this company stopped shipping tomorrow, how much of the AI supply chain would stop? Scale is part of the answer but not most of it.

The five pillars

Pillar	Weight	What it asks
Ecosystem criticality	30%	How much of the AI supply chain is disrupted if this company stops shipping, and how quickly competitors could absorb the volume.
Technology and IP depth	25%	Technical position against named rivals, patent and IP defensibility, and what a competitor would have to replicate. Half of this pillar is judgement, half is R&D intensity and operating margin.
AI relevance	20%	Direct participation in AI infrastructure: compute, memory, advanced packaging, EDA, networking, power. A company whose AI link runs through its customers rather than its own products scores lower.
Scale and financial strength	15%	Revenue, market capitalisation and three-year revenue growth. Purely mechanical, with no judgement involved.
Momentum and geopolitical risk	10%	Geographic diversification, export-control and sanction exposure, governance, and recent trajectory. A high score means low risk.

Scale carries only 15% on purpose. A weighting that favoured size would produce a list of the region’s biggest technology companies, which is a list that already exists and tells you nothing about AI.

How the scores are produced

Each company has a research profile built from its own filings: annual reports, securities filings, exchange disclosures and results releases, with market data used only for share prices and market capitalisation. Those profiles are published on this site, with their sources, so any figure in the index can be traced back.

The financial pillar is arithmetic on the reported numbers. Revenue is taken in the currency the company reports in and converted at one documented rate per currency, so no company benefits or suffers from a rate a particular source happened to use. Revenue growth is calculated in the local currency, so it measures growth rather than growth plus currency movement.

The four judged pillars are scored by a language model reading four evidence paragraphs written for each company, against a fixed rubric. The model is not shown the financial pillar and is not asked to produce a rank. Its reasoning and the sources behind each judgement are stored alongside every score, so a score can be questioned against the evidence that produced it rather than taken on trust.

Every score is a position, not a grade

Scores are percentile ranks within the universe the index covers, not absolute marks. A company scoring 80 on technology depth is ahead of 80% of the companies covered; it does not mean it scored 80 out of 100 on some fixed scale.

Two consequences follow, and both matter when reading the table:

• Adding companies moves everyone. A company whose composite falls between one run and the next may not have changed at all. Between runs, a change in rank is meaningful; a change in score on its own is not.
• Absolute figures are ranked against the right peers. Revenue and market capitalisation are ranked within each country, because a national universe is the fair comparison for scale. Ratios such as R&D intensity and operating margin are ranked within sub-sector, because a fabless designer’s 21% R&D intensity and a foundry’s 8% are not the same kind of number.

Who is not ranked, and why

Of the 164 companies profiled in this run, 141 are ranked, and 23 are not. Every unranked company is missing the financial pillar, and almost all of them are private companies that publish no revenue, no market capitalisation, and no growth history. Several are significant: national foundry projects, well-funded model developers, humanoid-robotics firms.

Those companies keep their full profiles on this site. They are absent from the ranking, not from the coverage. Where a pillar cannot be evidenced, it is left empty and the company is excluded with the reason recorded, rather than being given a neutral score. A company assigned a placeholder 50 on a pillar carrying 30% of the composite would sit above every genuinely-scored company below the median, which would be worse than saying plainly that it could not be scored.

What this index does not do well

Stating the limits is part of the method.

• Some companies are a year behind. A small number are ranked on their most recent published year, which is one year older than the rest. Their profile page shows the fiscal period used.
• Not every company discloses everything. R&D spending and capital expenditure are missing for some. A missing input is dropped, and the remaining weights are rescaled rather than a number being invented, so those companies are scored on fewer inputs.
• Market capitalisations come from market data, observed close to the research date, and move with the market.
• The judged pillars are judgements. They are evidenced, rubric-anchored and reviewable, and they are still judgements. Reasonable analysts would score some companies differently.
• Data confidence varies. Each profile carries a grade from A to D: A for audited filings, B for filings plus market data, C for analyst reconstruction, D for uncorroborated reports. The grade is shown on the profile and is never used to adjust a score, because multiplying a composite by a confidence factor does not say “we are less certain”, it asserts that a company is worse.

Editorial overrides

Ranks are computed, not chosen. If a published rank ever differs from the computed one, the change, the reason, and the person who approved it are recorded, and the difference is disclosed on this page. No overrides have been applied to the current run.

Coverage and updates

This run covers 164 companies: 56 in mainland China, 42 in Japan, 35 in Taiwan and 31 in South Korea. Each run is a snapshot, identified by its date, and previous runs are retained so a published list can always be reproduced.

Corrections are welcome. If a figure on a company profile is wrong, the profile lists the sources it was built from, which is usually the fastest way to show it.