Perfect internet is not a prerequisite. Here is the arithmetic.

One sentence turns up in nearly every first conversation, whether the business is in Nouméa, Port Vila or a shed outside Whangārei: we will look at this once we have a better connection. It is offered in good faith, and it rests on a measurement error. A business automation does not move video, it moves text — and the gap between those two is a factor of a thousand.

What an automation actually weighs

Take the most ordinary task there is: a system that drafts a payment reminder, sends it, and records what it did. What crosses the network is the request sent to the model and the answer that comes back — say 3,000 words of context and 700 words of reply. In characters, roughly twenty kilobytes for the round trip. No video, no images, no attachments.

Now set that against what your line already carries without complaint. One minute of high-definition video calling moves about 15 megabytes. So a single minute on a video call shifts as much data as 750 payment reminders. A 4K film, at fifteen gigabytes, is worth seven hundred and fifty thousand of them. A system handling two hundred operations a day consumes roughly four megabytes daily — a quarter of one minute of video, spread across eight hours.

The island constraint is latency, not bandwidth

Response time is the genuinely island-specific question. Gondwana-1, the first submarine cable serving Kanaky (New Caledonia), runs 2,080 kilometres from Nouméa to Sydney and entered service in September 2008. Light in fibre travels about 200,000 kilometres per second, which puts the physical round trip to Sydney at roughly twenty milliseconds — thirty to forty in practice, once the equipment along the way is counted. To a data centre on the US west coast, budget 150 to 200 milliseconds.

Two hundred milliseconds on an operation nobody is watching has no consequence whatsoever. A quote formatted in eleven seconds rather than nine changes nobody's day. Latency only bites when a human is waiting at a screen — and most of what gets automated in a small business happens with nobody at a screen at all. That is the property worth hunting for: asynchronous work is indifferent to geography.

The state of the network, in published figures

It also helps to argue from figures rather than impressions. According to OPT-NC, the territory's operator, 63,327 fibre accesses had been created by the end of 2024, more than 8,600 of them during that year alone, for a 81 per cent connection rate across the fixed-line base. Fibre was available in 20 of 33 communes in the first quarter of 2025, and those twenty account for 97 per cent of the households and businesses using fixed services. On mobile: 537 active sites as of April 2025, with 4G coverage reported at 99 per cent of the population.

Which means the business waiting for fibre usually already has it, and the one that does not has 4G that comfortably clears the twenty-kilobyte bar described above. The region has also stopped hanging by a single thread. As reported by Outremers360, Gondwana-2 — linking Nouméa to Fiji — and the domestic Picot-2 cable entered service on 25 August 2022 at a final cost of 4.675 billion XPF. Before that date, one break cut the territory off. That is no longer true.

The real risk is the outage, not the slowness

One serious objection remains: networks go down. They do here, as they do everywhere. But that risk is not managed by waiting for more bandwidth — it is managed at design time, and it costs nothing extra.

A fragile system is one that insists on being online at the exact moment you use it. A sound system queues the work, retries by itself, and picks up where it stopped when the line comes back. That is a configuration decision, not a capital expense. We build automations that way for a plain reason: the question is never whether the connection will drop, only when — and what happens in the meantime.

The three cases where the objection holds

It would be dishonest to claim bandwidth never matters. Three uses depend on it. Video calling, which needs a sustained stream and forgives jitter badly. Syncing large files — drawings, site photography, video rushes. And sending audio or images to a remote model: an hour of meeting audio to transcribe weighs about thirty megabytes, fifteen hundred times a payment reminder. If your project lives there, the line is a real constraint.

There is also an answer that removes the question entirely. A model installed on a machine sitting inside the business uses no bandwidth at all once downloaded: it keeps working through the outage, and the documents never leave the building. It is not the right answer everywhere — the reasoning power available locally still trails the largest remote models. It is the right answer when the line is genuinely poor, or the data genuinely sensitive.

What to check before starting

1. Measure instead of guessing. Run a speed test at the hour your team actually works, three days running. Below a stable 4 Mbps some uses get uncomfortable; above it, bandwidth is no longer your problem.
2. List the candidate tasks and mark, for each one, whether a human is waiting on the answer. The ones where nobody waits are insensitive to the network. Start there, without exception.
3. Insist on the queue. Every automation must be able to fail, wait and resume unattended. A supplier who cannot answer that question has answered a different one about themselves.
4. Separate the sensitive data. Accounts, medical records, identity documents: whatever must not leave can be processed on the premises, regardless of what the network is doing.

Common questions