What Belongs on the Device?

2026-11-18

By Paul DeSalvo

12 min read

architecture

You rent an apartment in one of those buildings that has everything. There's a doorman, a gym on the third floor, a pool on the roof and a party room you can book for birthdays. It came furnished, and when the dishwasher breaks, someone comes to fix it.

Then you add up what it costs each month. Rent, an amenity fee, a parking space. You've used the pool twice and never booked the party room. It's all due on the first, whether or not you were home that month. Mostly you needed a place to sleep.

A cloud bill reads the same way, and if you've never seen one, it's easy to miss how much of it is rent and amenities. The first one I read line by line had a BigQuery charge of about $5,000, as I remember it. A lot of what I was being asked were questions I could have answered on my Mac with the file in front of me. Building apps of my own, the same design would have put that kind of bill in my name: a hosted database, a server and a model call for every summary, each costing something before anyone showed up and more with every person who did.

With an apartment, the alternative costs money too. With software, it's often already paid for. Picture a small house you already own, down the street from the building. A laptop can run an analytical database inside a browser tab. A phone can transcribe a voice note and summarize it with no signal. The people who'd use what I build own that hardware today.

What the rent buys is real. The building is open and staffed around the clock, and there's a common room where people can meet without arranging anything. In software, that's a service that is always running and data anyone can reach with a link.

The question is how often a job needs that. A count of last month's orders doesn't have to be running at three in the morning. A voice note doesn't have to sit anywhere other people can reach. Keeping them always on mostly saves the small effort of sending someone a file, and that's a lot of rent for a little less friction. Some jobs do need the building: a number the whole team expects at six every morning, a record that two people both have to trust. So before I sign a lease for something I'm building, I ask of each part whether it needs the building or can run on hardware its user already owns.

Why everyone moved in

Cloud services: no servers to buy, and a working database the same afternoon.

Worth renting

Why everyone moved in

The building was a good deal when you arrived with two suitcases. Living on your own would have meant buying a place, furnishing it and fixing whatever broke. The building wanted a deposit and a monthly payment, and you could move in that week.

Software faced the same choice. Running an application once meant buying servers, finding somewhere to keep them and employing people to look after them. Cloud services turned that into a monthly bill that started small. A team could have a working database the same afternoon.

The machines people carried couldn't do much of the work either. A phone sent your voice to a server to be transcribed. A laptop ran out of memory on a dataset a server handled easily. The work went to the server because that was the machine that could do it.

And once everything lived in one building, sharing took no effort. Showing a colleague a number meant sending a link. Changing the software meant changing it once. Nobody had to ask which copy was current.

Two of those reasons have weakened. You're no longer arriving with two suitcases: the laptop and the phone can now do much of the work themselves. And the bill that started small grows with every user, every stored record and every hour the service stays on. Sharing is the reason that has held up, so it's the one to price. When a job involves one person and their own data, you're paying the full rent for a common room nobody else walks into.

The machine you already own

Take one of those questions from work: which department had the most orders last month? You can answer that in your browser now, for free. I wanted to see how far that goes, so I built a small tool, Data Explorer, to try. A cut-down version of it runs below, on 5,000 sample orders from last month. The department chart is the answer.

Starting DuckDB in this tab…

Open the full Data Explorer in a new tab

Your device did that work. The page arrived from a web server, like any other page, and your browser fetched DuckDB along with it, an analytical database small enough to run inside a tab. DuckDB made up the 5,000 orders and counted them by department on your processor. The orders never left your browser. The other eight columns are profiled the same way.

So there's no need to send a file to a company when you don't know how long they'll keep it.

And experimenting is free, which changes what you can do. You don't experiment in the cloud. It's too expensive. Every exploratory query in a warehouse goes on the bill, so you learn to ask only the questions you already know are worth asking. You write the one query, read the one answer and stop.

On your own machine a query costs nothing, so there's no reason to stop. You can slice and dice the orders however you like: by department, then by region, then by something that probably leads nowhere. The wrong guesses are free too.

That's the idea I wanted Data Explorer to push on. If queries are free, the tool doesn't have to wait for you to pick one. It profiles every column as soon as the data loads. Further down, under What stands out, it compares the columns with each other, scores what it finds and puts the strongest pairings at the top. Click a bar and everything else filters to it. You can see what's in the data and find something worth a closer look without choosing where to start. I wouldn't build that on a warehouse. The line above the charts counts the queries: about two dozen to open the sample, and nearly as many again on every click. A warehouse would bill each one before anyone had asked a question.

DuckDB handles far more than 5,000 rows. When its team started designing it in 2018, they were betting that a laptop could already handle much of a data scientist's work. In 2025, they loaded 20 GB of Dutch railway records from CSV files in about ten seconds on a laptop with 128 GB of memory. DuckDB's laptop experiment

A warehouse would return the same answer. First someone would have to load the orders into it, keep them up to date and sort out who can see them. When a team needs to agree on one set of numbers, that work buys something. For one person with a file and a question, it's a year's lease for one night's sleep.

Sharing works without the building too. Open a CSV reads a file from your own disk and profiles it the same way, without the file leaving your machine. If I send a colleague the link, their laptop does the counting and their orders stay with them. I'm not running anything on their behalf, and I can shut my laptop. Trying an analysis this way before building a pipeline around it is the subject of Your Laptop Might Be the Best Data Warehouse You Own.

A room for every guest

That count was for me. The sums change when you build something for other people.

Host an application and you're renting a room in the building for each person who uses it. Their records sit in your database. Each summary they ask for is a model call on your account. A database instance bills for every hour it's running, so one sized for a busy day charges the same on a quiet one. Give the app away, and every new user adds to a bill with your name on it.

I met that arrangement as a customer first. I was paying for a journaling app, and the AI features I wanted sat in a premium tier. I remember weighing the price against my whole ChatGPT subscription. I don't know how that app runs those features. But when each summary is a call to a hosted model, someone has to cover it, and a shop passes its rent into its prices.

That made me curious about how much of a journaling app a phone could run by itself, so I built one, Desandar, to find out. I chose iOS for two reasons. The first is that I can record, edit and save everything on the phone. The second is that the AI runs there too: the phone's own speech transcriber turns a recording into text, and a local model suggests a summary for me to review. The original words are still there if the summary misses something. That's the part I find exciting. The whole app can run on the phone I already carry, and nobody pays for a model call however many notes get recorded. The recording and search animation on the Work page

The landlord has a key

Rent is one cost of the building. Another is what you leave out.

Say I've had a difficult conversation with my manager. I want to remember what they said and why it bothered me. If I'm wondering where those words might end up, I can picture myself deleting the specifics and writing "rough day at work." A month later, that tells me almost nothing.

A landlord has a key to every apartment. Most never use it, and you still think twice about what you leave on the table. A journal only works if you write the true version. Keeping the recording, the transcript and the summary on the phone means that question doesn't come up.

What's worth renting

The cloud isn't always bad. What's hard is choosing the parts that are cost effective. Storage can cost very little, and so can a serverless function that only bills while it runs. It's the compute that gets expensive fast. The building has services you can't have on your own, and these are the ones I'd rent.

A locker off site: your other devices

Sharing across devices is the first, and it's one you need. Later I might want to read that journal entry on a second device. Its processor is perfectly capable of summarizing a note. What it lacks is my copy. From the phone, iCloud is there as an option: when it's on, the saved entry and its attachments sync through my private database to the other devices on my account. What I like is whose storage that is. It counts toward my own iCloud quota, so I'm not paying to store anyone else's journal. Each person rents their own locker, and I don't hold the lease on any of them.

A hall for one evening: a remote model

Private Cloud Compute is the other option, and it extends what my hardware can do. On a supported device I can opt in and have a larger remote model summarize the transcript. Apple says those requests are handled privately. I'm taking their word for that, and the transcript text does leave the phone for that request, so it's a trade-off to know about before switching it on. It isn't an extra bill either. Requests count against a daily allowance on my own iCloud account, and if you already pay for iCloud+, a higher allowance comes with it. It's hired by the evening. I can leave it off and still read the finished entry on both devices.

One booking sheet: when someone has to be told no

Sometimes a second copy is exactly what you want. Other times, two copies are the problem. I found that edge with a shared gift list. Imagine two people looking at one from different stores. Both phones are offline. Each shopper sees the same gift available and taps Reserve. Each phone can remember the tap. Neither can promise the gift is still available to its shopper. That promise needs one place both phones can reach, where the first claim takes the gift off the list for the next person.

The party room works the same way. There's one sheet at the front desk, and the first name on Saturday's line has the room.

It's one promise, though. Most of what an app like that does never needs it.

The front desk: numbers a team counts on

I'd make a similar distinction back at work. If the warehouse already holds the current orders and everyone uses the same definition of a sale, I'd run the query there. Downloading copies would give us one more thing to keep in sync.

And if the team wants yesterday's count waiting at six every morning, somebody has to collect the new orders and run that query while our laptops are shut. A front desk is staffed at three in the morning. My laptop is closed.

What owning asks of you

Owning has its own chores, and nobody is on call for them.

Until something syncs or exports it, the only copy of a journal entry is on the phone. Lose the phone before then and the entry is gone. That's a good reason to rent the locker.

The hardware varies. Transcribing on the phone depends on the model of phone and the language, and the local model is only there on devices that support Apple Intelligence. An app has to decide what everyone else gets.

A model that fits on a phone is smaller than one in a data center. It's well suited to turning a rambling voice note into a few lines I can read later. I wouldn't hand it every job.

Fixes travel more slowly. On a server I change the code once and everyone has it. With an installed app, the fix arrives when each person updates.

And sharing takes a step. I have to send the file, or a link to the tool. That's the friction the building removes, and for some work it's worth the rent.

Before you sign the lease

Before I add a server to something I'm building, I ask six questions:

  • Does anyone else need this exact copy? If it's the same person on another device, sync it through their own account. If it's nobody, leave it where it is.
  • Does someone have to be told no? The last gift, the last seat, the last table: one place has to decide who got it.
  • Does it have to happen while the device is asleep? A count at six, a nightly import: that needs something that's always on.
  • Does a team have to agree on the answer? Keep the shared records and definitions in one place and run the query there.
  • Is the job too big for the device? Rent the hall for the evening, and be clear about what leaves the device when you do.
  • Who pays each time it runs? If it's you, for every user, check what their own hardware could do first.

The parts that do move into the building take on the responsibilities in From Prototype to Production.

I'd still pay for a number the whole team can count on at six. For the person who already has the file and wants an answer this afternoon, I'd send the tool and let their laptop do the work. The building has a front desk, a locker and a booking sheet worth paying for. You can rent those without living there.