What 998 event records say about AI communities in Southeast Asia

The busiest community is not necessarily the one drawing the most people. A packed calendar can describe a dependable meetup, a venue that hosts many groups, or a database with better historical coverage. Registration totals can be just as misleading when a few large events account for most of the number.

We reviewed 998 completed, approved events linked to communities in the AISEA database. Of these, 700 are explicitly located in Malaysia, Singapore, Indonesia or Vietnam. We looked at three things together: how often communities meet, how registrations are distributed, and how many recurring communities have gone quiet.

The clearest finding is a monthly rhythm. Among 35 communities with at least two recorded event dates, the median community has a typical gap of 28 days between events. The average of those community-level medians is 40 days, pulled upwards by groups that run larger events every two or three months.

This is a more useful picture than an overall event count. It suggests that a durable community does not need to fill every week. For many groups, one dependable gathering a month is the operating rhythm.

One in five recurring communities has gone quiet

Seven of the 35 recurring communities have no event in the past 90 days and nothing upcoming in the database. That is 20 per cent.

The wider directory makes inactivity look worse than it is. Eighteen of 50 operating community units meet the same 90-day rule, but 11 of those have no meaningful event history or only one old event. We cannot tell whether they stopped, never became recurring communities, or are simply missing from our sources. Calling all 18 inactive would turn a data-coverage problem into a claim about community failure.

The defensible conclusion is narrower: four in five communities that established a repeated calendar are still active or have another event scheduled.

Registrations are concentrated in a small share of events

The database contains 38,170 recorded registration instances across 907 events with a value. These are registrations, not attendance or unique people. Someone who registered for five events appears five times.

The top 10 per cent of measured events account for 66.5 per cent of all recorded registrations. A small number of hackathons, showcases and large programmes therefore carry much of the reach, while recurring meetups carry much of the continuity.

There is also a serious measurement caveat. Of the 907 events with a registration value, 430 are recorded as zero. The database cannot tell us whether every zero means nobody registered or whether the historical listing no longer exposes a recoverable count. We keep those records in the audit, but we do not use zero as proof that an event had no audience.

For that reason, the country comparisons below show both event volume and the median among events with a positive registration count. The positive-only median is useful for understanding known reach, but it is not a substitute for the missing history.

Malaysia has the broadest base of measured participation

Malaysia contributes 232 completed events. We have a positive registration count for 164 of them, totalling 10,871 registrations. The median positive event has 26 registrations.

Large programmes matter, but they do not completely dominate the Malaysian total. The top 10 per cent of measured events contribute 50.9 per cent of registrations. That is still concentrated, but less so than Singapore or Vietnam.

The result is a broad base of smaller rooms with periodic spikes. Malaysia's strength in this dataset is not one blockbuster format. It is the coexistence of frequent, modest gatherings and occasional events that bring in hundreds of people.

Malaysian.AI is a useful example. The database links it to 35 completed events between May and September 2026, with a median four-day gap between event dates. Thirty-three events have a positive registration count, with a median of 15. This is a high-frequency, small-room model: reach accumulates through repetition rather than a single large turnout.

Singapore has the fullest calendar and the weakest registration coverage

Singapore contributes 398 completed events, the largest country total in the dataset. Yet only 87 have a positive registration count. Another 261 are recorded as zero and 50 have no count.

The known registrations total 8,769, and the median positive event has 65 registrations. The top 10 per cent of measured events account for 81.2 per cent of the total.

We can say with confidence that Singapore has a dense calendar. We cannot say with the same confidence how its total reach compares with Malaysia. Too much of its historical registration data is recorded as zero or missing. The events with known positive counts tend to be larger, but they are a selective slice of the calendar.

Lorong AI shows why event count needs interpretation. The database links 283 completed listings to Lorong AI across 232 dates, with a median gap of one day. That describes an active hub where many gatherings happen, not necessarily 283 programmes produced by one organising team. Two hundred and thirty-one of those listings carry a recorded zero, so the calendar is strong evidence of activity but weak evidence of total reach.

Indonesia has fewer events, but participation is more evenly spread

Indonesia contributes 27 completed events. Twenty-two have positive counts, totalling 2,408 registrations. The median positive event has 100.5 registrations, the highest typical positive count among the four focus countries.

Its registration base is also less dependent on a few outliers. The top 10 per cent of measured events account for 31.1 per cent of registrations. Within this dataset, Indonesia looks less like a large meetup calendar and more like a smaller set of consistently substantial workshops and build sessions.

The limitation is coverage. Twenty-seven events are enough to describe the recorded cluster, not a whole country's community activity. The finding should be read as a pattern in our sample rather than a national ranking.

Vietnam combines small recurring events with very large programmes

Vietnam contributes 43 completed events. Thirty-nine have positive counts, totalling 7,037 registrations. The median positive event has only 14 registrations, while the mean across measured events is 164.

That gap is the story. The top 10 per cent of measured events account for 88.5 per cent of registrations. Most recorded gatherings are small, while a handful of large programmes create nearly all the measured reach.

Vietnam therefore has two different kinds of community activity in the same total: intimate recurring sessions and mass-participation build programmes. Reporting only the event count understates the large programmes. Reporting only registrations makes the smaller continuity layer disappear.

What the data says about a healthy community

Event count, registration reach and continuity answer different questions.

• Event count tells us how much activity is visible.
• The gap between dates tells us whether people have a dependable reason to return.
• Registrations tell us recorded interest, not attendance or retention.
• Registration concentration tells us whether reach is spread across the calendar or carried by a few large programmes.

On those measures, the typical recurring community in our database operates monthly. Most established communities are still active. Reach is far more concentrated than activity: smaller gatherings keep the calendar alive, while a small group of large events produces most registrations.

That division of labour is not a weakness. Large programmes bring people in. Recurring communities give them somewhere to return. The practical question is whether the two connect: whether someone who joins a large hackathon can find a local room, another build session and people who remember what happened last time.

Method note

This analysis uses approved AISEA community and event records available on 20 September 2026. Umbrella records with active city or country chapters were excluded from the operating-community count so the same network was not counted twice. Cadence is calculated from distinct event dates. For each recurring community, we calculated its median gap between dates, then compared those community-level medians.

A community is classified as inactive only when it has at least two past event dates, no event in the previous 90 days and no upcoming event in the database. Registration figures use the displayed source-reported count, including approved overrides. They are registration instances, not verified attendance or unique participants. Recorded zeroes are retained but treated cautiously because the current data does not distinguish a true zero from an unrecoverable historical count.