/ techAugust 4, 20268 min read

Home Assistant went from 69,000 installs to 660,000. The board it runs on went from $120 to $305.

Operators who can now write their own software need somewhere to run it that they own, and the demand is measurable. So is the obstacle: the same AI buildout that made the software nearly free is bidding against them for the memory to run it. Raspberry Pi's price increase works out to exactly $5 per gigabyte.

Something changed about who owns the computer that a person's software runs on. It is easier to see in a public JSON file than in any trend piece.

Home Assistant publishes its opt-in telemetry at analytics.home-assistant.io. Pulled on August 4, 2026, it reports 661,887 active installations. Five years ago the same counter read about 69,000.

ACTIVE HOME ASSISTANT INSTALLATIONS / OPT-IN TELEMETRY, AUG 1 EACH YEAR600K400K200K69K660K202120222023202420252026

Values read from the Home Assistant analytics history file, sampled at August 1 of each year: 69,110 / 159,865 / 245,046 / 350,815 / 494,399 / 660,155. Telemetry is opt-in and the project estimates under 25% of users participate, so the level understates reality. The slope is the point, not the level.

Nine and a half times in five years, and still accelerating in absolute terms — the last twelve months added 166,000, more than any year before it. That is not a product growth chart. It is a count of houses that acquired an always-on server.

What is actually on those boxes

The same file reports what the median install is doing. Per installation: 2 users, 29 integrations, 14 automations, 6 add-ons, 382 entities.

Fourteen automations is the number worth sitting with. Nobody sold anyone fourteen automations. There is no marketplace where you bought them, no vendor who ships them, no SKU. They are conditional logic that a specific person wrote about their specific house, and they are worthless to everybody else on earth. That is the definition of personal software, and it now has a median installed base of six figures.

The install-type breakdown says something else again. Of 661,887 installs, 527,462 — 80% — run Home Assistant OS, the appliance image, rather than a container or a hand-configured Linux box. Four out of five people running a private always-on server are not sysadmins. They flashed a card.

Both halves of the cost moved at once

The reason this could not happen in 2015, when the Pi was already cheap, is that owning a machine has two costs and only one of them ever fell.

Writing the software. The 2025 Stack Overflow survey has 51% of professional developers using AI tools daily and 84% using or planning to. The interesting part of that survey is the second finding: trust in the output dropped to 29%, down eleven points. Adoption up, confidence down. That combination does not describe people shipping production systems for strangers. It describes people building throwaway tools for themselves, where being wrong costs an afternoon.

Operating it. This is the half everyone forgets. SaaS did not win because software was hard to write — it won because running a server was a job, and nobody wanted the job. Appliance images, containers, one-command installers and overlay networks that make a box reachable without opening a port took that job and made it a Saturday. The 80% running the appliance image are the measurement of it.

Personal software needed both. You cannot have a tool you wrote for yourself if the price of keeping it alive is being on call for it.

The ladder is denominated in memory

What you can put on your own hardware is set almost entirely by one number, and it is not the processor.

WHAT A RUNG COSTS, AND WHAT IT HOLDSPi 5 · 16GB$305an always-on service stack. an 8B model, slowly.Framework · 32GB$1,099a 30B model at 4-bit. desktop-class everything else.Framework · 128GB$1,999a 120B model at 4-bit, roughly 65GB resident.

Raspberry Pi 5 16GB at its current published price. Framework Desktop at its announced configuration pricing (Ryzen AI Max 385/32GB and Max+ 395/128GB); street prices for 128GB Strix Halo boxes have been reported above list. Model sizes are 4-bit quantised and approximate.

Every rung on that ladder is a memory tier wearing a product name. The processor differences are real and they are second-order; what decides whether a model fits on your desk is how many gigabytes are addressable. Which is unfortunate, because memory is the single most contested commodity in the economy right now.

The increase is exactly five dollars a gigabyte

Raspberry Pi raised prices twice in nine weeks and published the numbers both times. On December 1, 2025 the 16GB Pi 5 went from $120 to $145. On February 2, 2026 every variant moved again — by $10, $15, $30 and $60 respectively.

Plot the total increase against the memory on the board and it lands on a straight line.

RASPBERRY PI 5 PRICE INCREASE vs ONBOARD MEMORY / LAUNCH TO FEB 2026$75$50$25increase = $5 per GB, plus $5+$15+$25+$45+$852GB4GB8GB16GB

Cumulative increase from launch price to the February 2, 2026 price, derived from Raspberry Pi's two published announcements: $50 to $65, $60 to $85, $80 to $125, $120 to $205. The dashed line is not a regression — the four points sit on it exactly.

Each doubling of memory adds exactly twice the previous increment: $10, then $20, then $40. Solve it and the whole repricing is $5 + $5 × GB. Not a fitted trend — the four published prices land on that line without residual.

Which means there is no tariff component, no margin grab, no freight story, nothing about the SoC. Raspberry Pi passed through a memory bill and nothing else, and said so: an unprecedented rise in the cost of LPDDR4 memory, driven by "competition for memory fab capacity from the AI infrastructure roll-out."

The 16GB board has since gone again. It is $305 today154% above the $120 it launched at, in under a year.

It is every rung, not just the cheap one

Apple discontinued the 256GB Mac mini on May 1, 2026, moving the entry price from $599 to $799. Be precise about that one: the storage tier changed at the same time, so it is not a clean like-for-like increase. But Apple named the cause itself — Tim Cook cited a global memory shortage and "significantly higher memory costs" — and then raised the M4 Pro Mac mini by $200 to $1,599 in June with no config change at all.

Underneath both: TrendForce has conventional DRAM contract prices up 93–98% quarter-over-quarter in Q1 2026, with PC DRAM up 105–110%, and server DRAM still climbing 13–18% into Q3.

So the two halves of this story have one cause. The AI buildout wrote the software that made personal hardware worth owning, and the same buildout is outbidding the person who wants to own it for the memory to run it on. Those are not competing trends. They are the same trend arriving at your desk from opposite directions, and they meet on a spec sheet line item.

What would make this wrong

The Raspberry Pi is not evidence for any of this, and it is worth saying plainly. Raspberry Pi Holdings' FY2025 results — revenue $323.2M, up 25%, on 7.6 million units — come roughly 75% from industrial and embedded customers, not hobbyists. Pi's growth is machines in factories. It is the price ladder that tells you something here, not the volume.

Home Assistant telemetry is opt-in, self-selecting toward the engaged, and the project says under a quarter of users report. It is a good series and a bad level. Anyone quoting 661,887 as the size of anything is quoting the wrong number.

Most people will never do this, and the appliance image is doing more work in that 80% than any argument about sovereignty. The moment a personal stack requires a second decision — a certificate, a port, a backup that was never tested — the population that sustains it shrinks hard. Self-hosting's failure mode is not dramatic. It is a service that quietly stopped syncing in March.

And memory is cyclical. Raspberry Pi called the pressure "painful but ultimately temporary," and they have been through this before. If DRAM normalises in 2027, the squeeze in this piece evaporates and the ladder gets cheap again. The demand curve is the durable finding. The price curve is a moment.

What we take from this

TEGO is not a hardware company and has no advice about your homelab.

But we made this exact trade and can price it. TEGO runs a single rented box that hosts our automation runner and our design platform behind a tunnel. It replaced two subscriptions, it costs less than either of them did, and the reason it is worth operating is not ideology — it is that the software running on it was cheap enough to write that owning the machine finally made arithmetic sense.

That is the whole shift, and it is smaller and more boring than the word "sovereignty" suggests. Nothing about control, nothing about privacy, nothing about escaping anybody. Just a line item that used to be a subscription and is now a capital cost, because the thing that used to be expensive stopped being expensive.

The part to watch is that the arithmetic now has a memory price in it. Renting compute versus owning it was a preference in 2024. In 2026 it is a bet on DRAM.

If you are trying to work out which side of that line a system of yours belongs on, come talk to us.