Your AI should work on a plane, a train and in a basement
UK train signal was poor in up to 83% of Ofcom’s tests, and a car park can swallow it whole. Which of your AI tools still answers?
On this page
The seatbelt sign goes off, you open the laptop, and the assistant you use forty times a day shows a spinner. The plane has Wi-Fi, technically. It costs money, it crawls, and the reply you were halfway through reading never finishes. Nothing is wrong with the model. It just lives somewhere else, and you are no longer connected to somewhere else.
That moment is a useful test, and three ordinary places run it for you: a plane, a train and a basement. Each one breaks the network in a different way. If an AI tool stops working when the network does, it is a service you are borrowing, not a tool you own. This post walks through what fails in each place, with measured numbers, and what a setup that passes all three looks like on an ordinary laptop.
Three places show whether your AI is yours
We have argued before that privacy is something you can test by switching the Wi-Fi off; that is the airplane-mode test. This post extends the same switch from privacy to dependence. The question is no longer who can read your prompt. It is whether you can still work when the link is weak, flickering or gone.
The three places are not exotic. A plane gives you a link that exists but is slow and rationed. A train gives you a link that comes and goes every few minutes. A basement, an underground car park or a hospital lower floor gives you no link at all. Most people pass through at least one of these every week, and each failure looks different on screen: a stall, a half-finished answer, or an error.
The idea is older than generative AI. In April 2019, Martin Kleppmann and three colleagues at Ink & Switch published Local-first software, a list of seven ideals for apps people can rely on. The third reads simply: the network is optional. Their target was documents and spreadsheets. Seven years later the same standard fits AI, because assistants have become the place where documents get written.
On a plane, the link is there but slow and rationed
Airlines are improving fast, and the improvement shows how bad the baseline was. When United announced in February 2026 that it had fitted Starlink to most of its regional jets, the press release put the share of its daily departures with the new system at just over 25%, about 1,200 flights. It also said live sports, online games and home cameras had historically not been possible on traditional satellite Wi-Fi. The airline expects more than 800 equipped planes by the end of 2026. Today, just under three in four of its own departures still fly on older systems.
Older systems are slow for a reason no software update can fix. Many relay through satellites in geostationary orbit, which ESA puts at 35,786 km above the equator. A request goes up and down, and the reply goes up and down again: at least 143,000 km at the speed of light, or roughly 0.48 seconds before any server has done any work. Streaming chat replies arrive in many small pieces, so that delay compounds, and a large upload such as a photo or a PDF can time out altogether.
A local model has no round trip at all. The prompt goes from the keyboard to the laptop’s memory and back. Speed depends on the machine rather than the sky, and the cost of a seven-hour flight of drafting is some battery. That last part is the honest trade, and we come back to it below.
On a train, a flickering signal is worse than none
Trains look connected and are not. In June 2026 the UK regulator Ofcom published a study of connectivity on trains covering 24 stretches of main line. Depending on the network, between 58% and 83% of tests were poor. Ofcom’s bar for “good” is modest: 5 Mbit/s down, 1.5 Mbit/s up, and no more than 50 milliseconds of delay.
The on-board Wi-Fi, which many passengers fall back on, did well about 1% of the time. Ofcom blamed old equipment and speed caps, and said some carriage types block signal on their own.
A dropping signal is worse for AI than a missing one, because it fails halfway. You send a long request, the train enters a cutting, and the reply stops at paragraph three. Some tools retry and some do not. A cloud image or video job may already be running, and billed, when the connection drops, and you return to an error with no result. With no signal at all, at least you know not to start.
Good cloud tools soften this. They keep the partial answer, queue the request, or let you resend it with one click. That helps, but it does not change the shape of the problem: the work still waits for the next mast. A model on the laptop has no next mast to wait for, so the reply that starts in the station is the reply that finishes in the tunnel.
Underground, the signal is simply gone
The third place is the bluntest. Concrete and earth absorb radio waves quickly, and the loss compounds with depth. In 2016, a team from Aalborg University and Nokia Bell Labs drove a measuring rig through Denmark’s two deepest car parks and recorded live 4G and 3G signal level by level. Their average: about 5.2 dB lost for every metre down.
At 12 metres the worst case passed 60 dB of extra loss. Because each 10 dB is a tenfold cut, that is roughly a millionth of the power available at the ramp. The study also found that the lower-frequency 800 MHz band pushed through concrete walls, while 2,000 MHz needed openings such as doors to get in at all. Which band your phone happens to use decides whether you get one bar or none.
People work in these places every day. Editors and musicians rent basement studios because they are quiet. Archives, labs and hospital records rooms sit below ground because the space is cheap and stable. Field teams work in plant rooms, tunnels and remote sites. For all of them, a cloud assistant is a tool that only works upstairs.
Full bars do not protect you from someone else’s outage
There is a fourth place the test reaches, and it is your own desk. The network can be perfect on your side and still fail on someone else’s. On November 18, 2025, a configuration change inside Cloudflare made its proxies crash. Its own account puts the start at 11:20 UTC, core traffic back by 14:30 and every system normal by 17:06. The company called it its worst outage since 2019. Any service behind it, AI assistants included, was out for people sitting on excellent home fibre.
Outages end. Retirements do not. On April 22, 2026, OpenAI announced it would shut down a batch of older models, and its deprecations page lists October 23, 2026 as the date: four days after this post. The list includes gpt-4, gpt-4-turbo, o1, o3-mini, o4-mini and gpt-image-1. Six months of notice is fair warning, and it is still a date someone else picked. If a workflow depends on how one of those models behaved, the workflow changes on that day.
Even installed software can be borrowed. Microsoft’s licensing documentation says Microsoft 365 Apps must reach its licensing service at least once every 30 days. Miss that, and Word and Excel drop to view-and-print until you reconnect. The same page says plainly that the apps will not work on a computer completely cut off from the internet. An AI tool with a sign-in check has the same shape, even when the model runs on your machine.
None of this is a reason to distrust any one company. Cloudflare published a minute-by-minute account the same day, and OpenAI gave six months of notice. The point is structural. A service has dependencies you cannot see and cannot fix from your seat, and every one of them is a way for the tool to stop working while your laptop is fine.
A kit that passes all three fits in about 12 GB
The setup that passes is not exotic either. It is a handful of open model files on the laptop you already carry, plus an app that runs them without phoning home. The models are smaller than the frontier, and that is the honest cost. For drafting, summarising, rewriting, answering questions about your own files and transcribing recordings, they are good enough to get real work done. Our offline AI post goes through those everyday jobs one by one.
On the Ollama library, the 4-bit build of Gemma 4 12B is an 8.0 GB download and Qwen3.5 4B is 3.3 GB. The whisper.cpp table lists its small speech model at 466 MiB. Together that is just under 12 GB, downloaded once on a good connection and then kept. None of those files check a licence server, and none of them can be retired out from under you.
Setting it up takes three steps. Install an app that runs open models on your machine, either a runtime such as Ollama on its own or an app that bundles one. Download the models at home, on the fastest connection you have. Then turn the Wi-Fi off and ask each model a real question before you rely on it, because a model that is listed is not always a model that loads. That last step takes two minutes and catches most surprises, such as an app that wants you to sign in before it shows you anything.
The limits are real, and we would rather name them. Local generation uses the processor hard, so a long session on battery drains it faster than browsing does. Video generation and the largest reasoning models still want a big machine or the cloud. The point is not to give up cloud models. Use them when you are connected and the job is worth it, and keep a local floor under everything else. Our serious-software-is-local essay makes the longer case that this is how every other kind of professional software already works.
In CSuite, the bundled local runtime runs the same chat and writing actions as the cloud providers, so moving a task offline is a model switch rather than a new app. Whichever app you use, check two things: that the models live in a folder you can see, and that the app opens and answers with the network off.
Run the test at home, before the trip
Do it tonight, while you still have Wi-Fi to fix whatever fails. Turn on airplane mode and open every AI tool you rely on for work. Ask each one for something real: a summary of a file, a rewrite, a transcript. Write down which ones answer, which ones spin, and which ones refuse to open.
The ones that answer are yours. The ones that spin are borrowed, which is fine for work that can wait for a connection. The ones that refuse to open are borrowed on terms you did not read. For anything you need on a plane, a train or below ground, swap in a local model now, while the download is fast. If you are deciding how much machine that takes, our personal compute post covers the hardware floor. Your most important tools should still work when the signal goes.
