Using AI Remotely

Why This Matters

In a city, the AI tool you pick is about features. Remote, it's also about what your connection can carry. A tool that runs fine on city fibre can stall on a single 4G bar, burn through a prepaid data cap in a week, or fall over the moment 15 people on a venue Wi-Fi try it at the same time.

This page gives you practical numbers for each of the main tools: text chat, voice mode, image generation, file uploads, and video. It then translates those numbers into guidance for the kinds of connections you actually have in remote NT: Telstra 4G, Optus 4G, Starlink, venue Wi-Fi, and tethered hotspots.

How Honest Are These Numbers

An honesty note before the figures. None of the major AI vendors, OpenAI, Anthropic, Microsoft, Google, or Perplexity, publish official "megabytes per message" or "minimum kilobits per second" figures for their consumer chat products. What they publish is general guidance like "requires a stable internet connection".

The numbers below are engineering estimates, built from the underlying audio codec bitrates, typical text chat payload sizes, observed image and audio file sizes, and the Telstra and Ookla speed data cited in the Tennant Creek connectivity research. Treat them as ballparks for planning, not promises. Where a number comes from a published vendor limit (for example Claude's 30 MB file upload cap), it's marked as published. Everything else is estimate, with a range.

Data Use Per Tool, Per Mode

One-line summary in the first column, data cost in the second, minimum comfortable bandwidth in the third, source basis in the fourth. Specifics are estimates unless marked.

The four rules of thumb that come out of this table:

Text chat, on any of the six tools, is cheap. A full hour of heavy back-and-forth text is typically well under 10 MB. You can do this on a weak 4G bar.

Voice mode is roughly 10 to 20 times the data cost of the same conversation typed. That's still modest in absolute terms (around 60 to 120 MB per hour), but it adds up fast if you're on a prepaid SIM or a shared hotspot.

Image generation is the biggest single-click cost in everyday use. A batch of four images can move 5 to 10 MB in a few seconds.

Gemini Live with video on is the worst case. It's the only common consumer mode that genuinely needs broadband-class upload, not just download.

What Remote NT Connections Actually Deliver

These figures come from the Tennant Creek connectivity research already done for this project (see TC deep-research-report.md in the project folder), with the bandwidth research overlay on top.

Telstra 4G in remote NT towns. Present in Tennant Creek and most regional centres. Telstra's own typical speed range is 2 to 500 Mbps on 4G, with 5G running 10 Mbps to 1 Gbps where available. In practice, a useful planning floor is "a few Mbps" on a decent bar, which is more than enough for any text or voice AI use.

Optus and Vodafone. Present in Tennant Creek but with smaller footprints outside town. Broadly comparable speeds to Telstra on a good bar, more variable on the edges.

Northern Territory regional median (Ookla 2025, via WhistleOut). Around 72 Mbps download. That's a regional median, not a remote-community median, so take it as an optimistic ceiling rather than what you'll see in the bush.

Starlink residential in remote Australia. Community reviews commonly report 50 to 200 Mbps down and 10 to 20 Mbps up. That's effectively fixed broadband for every AI tool on the table above, including Gemini Live.

Starlink Mini on the Roam plan. Often lower, commonly 20 to 100 Mbps, variable by time of day and sky obstruction. Still broadband-class for every AI mode if you have line-of-sight.

Public and venue Wi-Fi. Shared, and the most-variable thing on this list. A 50 Mbps venue link that tests fine with one person can collapse when 15 people all hit voice mode at once. Plan around congestion, not the speed test.

Traffic Light Guide for a Tethered 4G Hotspot

If you're tethering to a single phone or using a basic mobile hotspot, here's the shortcut. Green is fine, amber needs care, red is usually a mistake.

The Voice Mode Trap

Voice mode feels magical. It also quietly eats your data. Here's the maths that matters:

One hour of ChatGPT, Copilot, or Gemini voice, by our estimate, moves about 60 to 120 MB in each direction, so 120 to 240 MB total. That's nothing on home broadband, but on a shared venue hotspot with five people doing it at once, you're pushing 500 MB per hour through one phone. That single phone will collapse under the load.

Voice costs roughly 10 to 20 times more data than typing the same conversation. If the job is to ask the AI to summarise a meeting or draft an email, type it. Save voice for situations where you genuinely need hands-free, or where speaking is faster than typing on a small screen.

Images Cost Far More Than Text

A single phone photo uploaded for the AI to look at is typically 100 to 500 times the size of a text question about the same topic. A modern phone photo can easily be 3 to 5 MB. The same question described in words is maybe 100 to 500 bytes.

On a solid connection this doesn't matter. On a weak 4G bar or a constrained hotspot, it's the difference between a snappy chat and a hanging upload.

What to do about it. If the AI doesn't need to see the exact picture, describe it in words. If it does, crop the photo tight to the bit that matters, or resize it before sending. Most phones have a built-in "edit and crop" step that can turn a 4 MB photo into a 400 KB one without losing the information you actually need.

For bulk image generation (say, drafting a flyer), do it on broadband or Starlink. For one-off image generation (a logo, an icon), 4G will cope.

A Month of AI Use, Translated to GB

So what does all this mean for your monthly data bill? Three realistic profiles, estimated at the high end so you don't get caught short.

Typical Australian postpaid mobile plans sit in the 20 to 150 GB per month range. Budget prepaid is often 10 to 40 GB. Even heavy AI use fits inside a normal postpaid plan. The gotcha is the shared heavy use case: one hotspot, five family members, everyone using voice mode, no one checking the counter. That can eat 30 to 50 GB before anyone notices.

What Works Offline or Cached

The honest answer: nothing in the consumer tier is genuinely offline. All of ChatGPT, Claude, Copilot, Gemini, NotebookLM, and Perplexity need a live connection to produce any answer. "It's in Word" does not mean Copilot works without the internet. Every Copilot action in a desktop M365 app still goes to the cloud and back.

What you can pre-cache and use without a connection:

NotebookLM Audio Overviews. Generate on good connection, download the MP3, play back anywhere. This is the single best offline-friendly feature on any of these tools.

ChatGPT and Claude chat transcripts. Save or export as text or PDF so you can re-read without going back online.

Perplexity answers. Export or screenshot the answer and the source list.

Generated images. Save them locally once created.

When Your Connection Drops Mid-Prompt

It happens. Here's what each tool does and what you can do.

Text chat (any tool). The reply usually stops mid-sentence. Wait for signal to come back, refresh or reload the chat, and either scroll up to see what did arrive or just ask again. Your prompt is typically preserved. Nothing is "lost" in a damaging sense.

Voice mode. The session usually ends. You'll need to start it again. Unlike text, there's no transcript by default unless you turned it on, so the content of the conversation is gone. Worth knowing before you rely on voice for anything important.

Image generation. If the generation completed on the server side, the image is usually waiting for you when you reconnect, in the same chat. If it was still running when the connection dropped, you'll need to ask again.

File upload. Interrupted uploads fail. Start again. For big files (over 10 MB) on a flaky connection, try uploading in the strongest part of your signal, or split the file into smaller pieces.

Copilot in M365 apps. You may see a "Copilot is unavailable" banner. The rest of Word, Excel, or Outlook keeps working. The document itself is safe, only the AI features pause.

What the Vendors Don't Tell You

Short version: none of OpenAI, Anthropic, Microsoft, Google, or Perplexity publish a supported minimum bandwidth for their chat products. The closest vendor statements are:

OpenAI describes Voice Mode as needing "a stable internet connection" and recommends Wi-Fi for long sessions. No numeric floor. ChatGPT does not support offline use.

Anthropic requires a live connection for Claude. Published limits are about file size (30 MB per file, 500 MB per project), not bandwidth.

Microsoft publishes network guidance for Teams (audio in the 30 to 130 kbps range depending on quality) which is the closest numeric signal for Copilot voice, but it's Teams data, not Copilot-specific.

Google requires a connection for Gemini and Gemini Live. The Gemini API "Live" documentation describes real-time sessions but doesn't publish a supported client minimum bandwidth. NotebookLM documents source-size limits (200 MB per source, 500,000 words per source), not bandwidth.

Perplexity publishes no bandwidth guidance. Requires a connection.

The practical implication: don't expect a vendor service level agreement on "will this work on my 4G bar". The useful guidance is the rules of thumb above, not vendor policy.

Practical Checklist

Ten things to do, in rough order of usefulness, if you're using AI from a remote NT connection.

1. Default to text chat unless voice is genuinely needed.

2. Crop or resize phone photos before uploading for analysis.

3. For NotebookLM audio, generate on the best connection available, download the MP3, then play offline.

4. Don't put more than five people on the same hotspot running voice or image-heavy sessions at once.

5. Split big PDF uploads into smaller files when on 4G.

6. Save important AI conversations as text or PDF before you lose signal.

7. Turn off video in Gemini Live if you're on mobile data.

8. Watch your monthly data counter if you share a hotspot or a prepaid SIM with others.

9. If a connection is flaky, type one short prompt to anchor the chat before doing anything long.

10. Know the local backup, the nearest public library or service with reliable Wi-Fi, for the moments when you really need it.

References

These are the canonical pages to verify any specific figure against before printing it in a handout or quoting it to a funder. Bitrates and limits change as the products evolve.

Prepared 2026-04-19. Per-tool figures are engineering estimates unless flagged as published. Re-verify against the primary sources above before printing to a participant-facing handout.