Why this matters
AI tools are built and marketed by people on fast city internet. Out bush, the connection is the constraint. This page gives practical numbers for each main way of using AI (text, voice, image generation, file uploads, video), then translates them into guidance for the connections you actually meet: Telstra and Optus 4G, Starlink, venue Wi-Fi, and tethered hotspots.
How honest these numbers are
None of the major vendors (OpenAI, Anthropic, Microsoft, Google, Perplexity) publish official "megabytes per message" or minimum-speed figures for their chat products. The numbers below are engineering estimates, built from audio codec bitrates, typical text payloads, observed image and audio file sizes, and published Telstra and Ookla speed data. Treat them as ballparks for planning, not promises. Where a figure is a published vendor limit (such as Claude's 30 MB file cap) it is marked published; everything else is estimate.
If you plan to print a figure on a handout or quote it to a funder, check it against the vendor's current documentation first (links at the foot of the page). Bitrates and limits change as products evolve.
Data use per tool, per mode
Specifics are estimates unless marked published.
| Tool, mode | Typical data per use | Min usable bandwidth |
|---|---|---|
| Text chat (any tool) | ~2 to 80 KB per exchange | ~50 to 100 kbps |
| Voice mode (ChatGPT, Copilot, Gemini) | ~1 to 2 MB per minute | ~64 to 128 kbps sustained |
| Image generation | ~0.5 to 3 MB per image | Slow 4G works; painful on weak signal |
| Image upload (phone photo) | ~150 to 500 KB after resize; original 1 to 5 MB | Upload headroom ~256 kbps+ |
| Claude file upload | Equal to file size; 30 MB/file, 500 MB/project (published) | Upload headroom matters most |
| Gemini Live (video plus voice) | ~3 to 10 MB per minute; video uplink dominates | ~500 kbps floor, ~1 Mbps up comfortable |
| NotebookLM source upload | Equal to file size; 200 MB or 500,000 words/source (published) | Upload-limited |
| NotebookLM Audio Overview | ~10 to 25 MB for a 10 to 20 minute piece | Download once, play offline |
| Perplexity text answer | ~20 to 100 KB per answer with sources | ~100 kbps |
Four rules of thumb from this table:
- Text chat is cheap. A full hour of heavy back-and-forth is typically well under 10 MB. It works on a weak 4G bar.
- Voice costs 10 to 20 times more than typing the same conversation (around 60 to 120 MB per hour). Modest alone, but it adds up on a prepaid SIM or shared hotspot.
- Image generation is the biggest everyday single-click cost. A batch of four can move 5 to 10 MB in seconds.
- Gemini Live with video is the worst case, the only common mode that genuinely needs broadband-class upload.
What remote connections actually deliver
- Telstra 4G in regional centres: a useful planning floor of "a few Mbps" on a decent bar, plenty for text or voice AI.
- Optus and Vodafone: comparable on a good bar in town, more variable at the edges.
- Starlink (residential): commonly 50 to 200 Mbps down, 10 to 20 Mbps up; effectively fixed broadband for every tool here, including Gemini Live. Starlink Mini on Roam is lower but still broadband-class with line of sight.
- Public and venue Wi-Fi: shared and the most variable; a link that tests fine with one person can collapse when fifteen people hit voice mode at once. Plan around congestion, not the speed test.
One signal bar is not no signal. Text tools usually work on a single stable 4G bar. What kills the experience is jitter and packet loss, not peak speed; if your bar flickers between one and zero, text chat feels frozen even if a speed test passes.
Traffic-light guide for a tethered 4G hotspot
Green (fine on 4G): text chat on any tool; short bursts of voice; document uploads under 5 MB; reading Perplexity answers.
Amber (with care): long voice sessions (budget the data); phone-photo uploads (crop or resize first); image generation in small batches; PDFs of 5 to 20 MB; more than one person on the hotspot.
Red (prefer broadband): Gemini Live with video on; ten or more image generations; streaming NotebookLM Audio Overviews live; PDFs over 20 MB; any group activity with five or more people on the same hotspot at once.
The voice-mode trap
Voice feels magical and quietly eats data. One hour of voice moves roughly 60 to 120 MB each way, so 120 to 240 MB total. Nothing on home broadband, but with five people on a shared venue hotspot at once you are pushing 500 MB per hour through one phone, and it will collapse. Voice costs 10 to 20 times more than typing the same thing. If the job is to summarise a meeting or draft an email, type it; save voice for genuine hands-free moments.
Images cost far more than text
A phone photo uploaded for the AI to look at is often 100 to 500 times the size of a text question about the same thing. On a solid connection this does not matter; on a weak bar it is the difference between a snappy chat and a hanging upload. If the AI does not need to see the exact picture, describe it. If it does, crop tight or resize first; most phones can turn a 4 MB photo into a 400 KB one without losing what matters. Do bulk image generation on broadband or Starlink.
A month of AI use, in GB
Estimated at the high end so you are not caught short.
| Use profile | What it looks like | Monthly data |
|---|---|---|
| Light | A few text chats a day, the odd upload, no voice | ~0.2 to 0.5 GB |
| Moderate | Daily text, ~30 min voice most days, occasional images, a weekly PDF | ~2 to 5 GB |
| Heavy | Lots of text, regular voice, Gemini Live video, bulk images, large uploads | ~10 to 20 GB |
Even heavy individual use fits inside a normal postpaid plan (20 to 150 GB). The real gotcha is shared heavy use: one hotspot, several people, everyone on voice, nobody watching the counter, which can eat 30 to 50 GB before anyone notices.
What works offline or cached
Honestly, nothing in the consumer tier is truly offline; ChatGPT, Claude, Copilot, Gemini, NotebookLM and Perplexity all need a live connection. "It is in Word" does not mean Copilot works without internet; every Copilot action still goes to the cloud and back.
What you can pre-cache and use without a connection: NotebookLM Audio Overviews (generate on a good connection, download the MP3, play anywhere, the single best offline-friendly feature); chat transcripts saved or exported as text or PDF; Perplexity answers exported or screenshotted; and generated images saved locally.
For genuinely offline AI, local-model tools (Ollama, LM Studio) run a small model on a laptop with no internet. Less capable than the big models and they need a reasonably modern machine, but they work on the plane, in the bush, or when the tower is down. Worth knowing about, not a first stop.
When the connection drops mid-prompt
- Text chat: the reply stops mid-sentence; reconnect, reload, and scroll up or ask again. Your prompt is usually preserved.
- Voice mode: the session ends and, unless you turned on a transcript, the content is gone. Worth knowing before relying on voice for anything important.
- Image generation: if it finished server-side, the image is waiting when you reconnect; if not, ask again.
- File upload: interrupted uploads fail; restart, and for big files try the strongest part of your signal or split the file.
- Copilot in Office apps: you may see "Copilot is unavailable", but Word, Excel and Outlook keep working and the document is safe.
On a flaky connection, type one short prompt first to anchor the chat. If you drop out, reconnect and say "continue", and the AI will usually pick up where it left off.
A practical checklist
- Default to text chat unless voice is genuinely needed.
- Crop or resize phone photos before uploading.
- For NotebookLM audio, generate on the best connection, download, then play offline.
- Do not put more than five people on one hotspot running voice or image-heavy sessions.
- Split big PDF uploads into smaller files on 4G.
- Save important conversations as text or PDF before you lose signal.
- Turn off video in Gemini Live on mobile data.
- Watch the monthly counter if you share a hotspot or prepaid SIM.
- On a flaky connection, anchor the chat with one short prompt first.
- Know the local backup with reliable Wi-Fi (library or service) for when you really need it.
References
Verify any specific figure against these before printing or quoting it.
- OpenAI, Voice Mode FAQ: https://help.openai.com/en/articles/8400625-voice-mode-faq
- OpenAI, ChatGPT requirements: https://help.openai.com/en/collections/3742473-chatgpt
- Anthropic, Claude file limits: https://support.anthropic.com/en/articles/7996608-getting-started-with-claude-ai
- Anthropic, vision / image input: https://docs.anthropic.com/en/docs/build-with-claude/vision
- Microsoft, Copilot for Microsoft 365 requirements: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-requirements
- Microsoft, Teams media bitrate (closest analogue for Copilot voice): https://learn.microsoft.com/en-us/microsoftteams/prepare-network
- Google, NotebookLM help (source limits): https://support.google.com/notebooklm
- Google, Gemini API "Live" docs: https://ai.google.dev/gemini-api/docs/live
- Perplexity help centre: https://www.perplexity.ai/hub/faq
- Telstra, typical 4G/5G speeds: https://www.telstra.com.au/mobile-phones/coverage-networks
- WhistleOut, NT regional median (Ookla 2025 via): https://www.whistleout.com.au/
- Starlink Australia, plans and pricing: https://www.starlink.com/au
- Opus audio codec (underlies real-time voice): https://opus-codec.org/
Per-tool figures are engineering estimates unless flagged as published. Re-verify against the primary sources above before printing to a participant-facing handout.
Glossary
- Bandwidth
- how much data a connection can carry per second; measured in kbps or Mbps.
- kbps / Mbps
- kilobits and megabits per second; 1 Mbps is 1000 kbps.
- Tethering
- sharing one phone's mobile data with other devices.
- Latency / jitter
- delay and the variation in delay; high jitter makes a connection feel frozen even at decent speed.
- Codec
- the method used to compress audio or video; it sets how much data voice and video use.
- Cached / offline
- saved so it can be used without a live connection; only some AI outputs can be.
- Local model
- an AI model that runs on your own device with no internet, such as via Ollama or LM Studio.
Take it with you: a one-page data-cost and traffic-light cheat sheet is available to download from the button at the top of this page.