Nothing here is for sale.
This is what the parts cost you.
Every figure on this page is money you would pay a supplier — a shop, a distributor, a solar installer — for equipment you own outright. None of it is a licence, a subscription, or a payment for the Ark itself. The software, the specification and the corpus are free and open and stay that way. This page exists so you can plan your own build, not so anyone can quote you.
Everything so far has described one node. There is no one node. There are four sizes, on two independent power axes, and — this is the part that changes the arithmetic — the library is identical in all of them.
The corpus does not care what runs it. Wikipedia in two languages, the medical references, the mathematics textbooks, the planet's maps and the literature are the same bytes on a $580 Raspberry Pi as on a $5,800 Studio, and they read the same way from a phone over local WiFi.
What changes as you go down the ladder is how fast and how well the machine can reason over that library. The librarian gets slower. The collection does not get smaller.
That is what makes this a choice of size rather than a purchase. Most infrastructure decisions are a single irreversible bet on one specification. This one is a set of options in the strict sense — the right, without the obligation, to move — and the expensive half is already paid for at every price point.
Four rungs on two power axes is eight configurations, one library. The decision that normally gates a project like this — can we afford the good machine — stops being a gate. A site that can only fund $580 gets the same archive as one that can fund $8,500.
The corpus is portable. A Reader bought this year becomes a Community node's archive next year by carrying the drives across. Nothing is re-downloaded, re-verified or stranded. You are never buying the wrong rung — only the current one.
Every other line in this budget can wait. Hardware gets cheaper, models get better, solar gets cheaper. The corpus cannot wait, because assembling it requires the exact thing the system exists to survive the loss of. It is the one component whose acquisition window closes.
Every dependency deleted is a choice that stays yours: to keep operating, to not renegotiate, to not be repriced by an organisation you cannot reach or appeal to. That is what the air gap actually buys — and it costs the same at every rung.
Count the links.
Then count how many are yours.
A chain fails at its weakest link, and most procurement decisions never count the links at all. Below is the same capability — ask a question, get an informed answer — delivered two ways. The endpoints are identical. Everything between them is the choice.
This is not a prediction that any of those four counterparties will fail. It is the observation that a deployment in Chocó cannot audit any of them, cannot negotiate with any of them, and cannot appeal a price change, a policy change, a sanctions regime or a severed cable.
The node does not make the chain more reliable. It makes it shorter — and every link removed is a decision that stays with the people who have to live with it.
Because it is the part a price comparison cannot show. Two options that cost the same are not equivalent if one of them requires four organisations to keep making the same decision for ten years. The right-hand column costs between $580 and $8,500 and has the same two links at either end of that range.
Four sizes. Same library in all of them.
Most offline-AI proposals specify one machine and stand or fall on it. Every rung below is stated with what it gives up, because a range presented without its costs is a sales brochure.
Raspberry Pi 5, 16 GB
2 TB working drive plus one cold copy, USB microphone, powered speaker. Draws 10–15 W — a wall socket, or a small power station and a 100 W panel. Serves the full archive and maps over WiFi to every phone in range.
Gives upNo tier-1 model. Answers arrive in minutes rather than seconds. Speech recognition is the small Whisper, noticeably worse in noise. Regional maps rather than the planet. Cannot build its own index.
Mac Mini M4, 32 GB
4 TB working drive plus one cold copy. 5 W idle, around 50 W while thinking. The cheapest rung that runs the full 27-billion-parameter model — tier 1 at 4-bit with vision, planet-scale maps, the good Whisper, conversational-speed answers.
Gives up4-bit only, so quantization damage is unmitigated. One model resident at a time, which means the two-family cross-check costs a reload each time you want it.
Mac Mini M4 Pro, 64 GB
Three 8 TB drives, two kept cold and rotated. A Raspberry Pi as secondary node so archive and map serving never queue behind inference. Named in the build spec as fully adequate.
Gives upMulti-model residency under sustained group load. 8-bit is possible but tight. Index building is slow enough that you would rather be sent one.
Mac Studio, 128 GB
The reference build. Two model families resident simultaneously for genuine cross-checking, 8-bit weights, headroom for the annual re-evaluation. The only rung that manufactures — it builds the retrieval index and re-quantizes weights, then clones both down to every rung beneath it.
Gives upNothing in capability. Costs ten times Rung 1 and needs three times the solar, which is the entire reason the other three rungs exist.
Rung 4 is a factory. Rungs 1 to 3 are what it makes. Fund one hub per region and readers everywhere else.
Solar is the largest line item,
and most sites do not need it.
Compute is one decision; power is a separate one. A school with mains electricity, a clinic in a town, a household on the grid — none of them need a panel to get every benefit described on this site.
Grid power does not weaken the argument, because the argument was never about electricity. It is about the counterparties, and a wall socket removes exactly as many of them as a solar array does.
| Configuration | On the grid | Off-grid | Power adds | Draw |
|---|---|---|---|---|
| Rung 1 — Reader | $580 | $1,080 | +$500 | 10–15 W |
| Rung 2 — Classroom | $1,740 | $2,840 | +$1,100 | 5 / 50 W |
| Rung 3 — Community | $3,800 | $5,600 | +$1,800 | 5 / 65 W |
| Rung 4 — Hub | $5,800 | $8,500 | +$2,700 | 5 / 90 W |
| For comparison — the satellite dish alone | — | +$1,000 of panel | continuous | 20–75 W |
The fact that reverses the intuition
A satellite terminal is not a passive object. It draws power around the clock — connected or idle, used or not — roughly 1.2 kWh a day. A Community node doing real work draws about 0.4. A Reader draws about 0.3.
Off-grid, then, the pipe costs three to four times more to keep alive than the library does. Under a design constraint where watts are the scarcest resource, the offline system is not the austere compromise. It is the one that fits inside the power budget.
Staying connected is the luxury.
Treating power as part of the product is what makes offline AI look unaffordable. It is a site condition, not a feature: the same Rung 2 node costs $1,740 in a town and $2,840 on a hillside. Sites that already have electricity should not be quoted the hillside price, and a funder comparing against a connected stack on mains power should not be shown one either.
Subscriptions price per person.
A node prices per place.
Every AI subscription ever built charges by the seat. However cheap the seat gets, the bill scales with the number of people you are trying to serve — which is the one variable a development programme exists to maximise. A node serves everyone inside its WiFi radius, and the two-hundredth reader costs nothing.
Three of the four configurations cost less than the connection alone — before a single AI seat is bought. Their break-even is zero seats. Only the top rung has to justify itself, and it does so at two.
Sixteen months of satellite service costs more than an entire Reader node, bought outright, with no month seventeen.
The full working, with the opposition given every advantage.
The connected column is built as cheaply as it can honestly be built: the discounted regional satellite tier, the $8 ad-supported chatbot tier, the cheap hardware kit, and grid power so it carries no solar cost either. If the node wins against that, it wins against anything.
| Ten years, one site | Satellite + chatbot | R1 Reader | R2 Classroom | R3 Community | R4 Hub |
|---|---|---|---|---|---|
| Compute hardware | $200 | $200 | $1,200 | $2,000 | $4,000 |
| Storage, working + cold | — | $280 | $440 | $1,200 | $1,200 |
| Secondary node, audio, spares | $150 | $100 | $100 | $600 | $600 |
| Capital, grid-powered | $350 | $580 | $1,740 | $3,800 | $5,800 |
| Add solar and battery, optional | +$1,000 | +$500 | +$1,100 | +$1,800 | +$2,700 |
| Connectivity, 10 yr at $37/mo | $4,440 | $0 | $0 | $0 | $0 |
| AI seats, 10 yr at $8/mo each | $960 ea. | $0 | $0 | $0 | $0 |
| Total — 1 person | $5,750 | $580 | $1,740 | $3,800 | $5,800 |
| Total — 25 people | $28,790 | $580 | $1,740 | $3,800 | $5,800 |
| Total — 300 people | $292,790 | $580 | $1,740 | $3,800 | $5,800 |
| Break-even against satellite | — | 0 seats | 0 seats | 0 seats | 2 seats |
| Value remaining at year 10 | $0 | The archive, on at least two drives, at every rung | |||
| Survives a change of vendor terms | No | Yes — the licences are Apache 2.0 and the weights are on the disk | |||
A 300-seat connected deployment is of course not how anyone would actually do it — they would share one login, or use the free tier. That is precisely the point: the connected model forces rationing at exactly the moment a programme is trying to scale. No rung of the node has a seat to ration.
Two to twenty-eight dollars per person, for a decade.
Per person, for ten years, one node serving a school of 300. Low end is the Reader on mains power; high end is the full off-grid Hub. Halve either figure if the node also serves the clinic and the market — which it does.
| Comparable intervention | Per person, decade | Recurring? |
|---|---|---|
| One-laptop-per-child style programme | $200+ | device replacement |
| School library construction and stocking | $40–150 | restocking |
| Connectivity subsidy, 10 years | $15–50 | evaporates when funding stops |
| Ark node — Hub, off-grid | $28 | $0 |
| Ark node — Community, grid | $12.67 | $0 |
| Ark node — Reader, grid | $1.93 | $0 |
And unlike every row above them, all three arrive bilingual on day one, carry medicine and mathematics and bookkeeping and literature in the same box, and answer questions rather than waiting to be searched.
The expensive part is done once.
The copy costs a drive.
Most hardware philanthropy has flat unit economics — the hundredth school costs what the first one did. This does not, for two compounding reasons. The valuable asset is non-rival: the curated corpus, the retrieval index, the verified manifest and the bilingual recovery procedure took months to assemble and four hours to clone. And the fleet does not have to be uniform.
sites, mixed rungs
70 Readers at $580. 25 Classrooms at $1,740. 5 regional Hubs at $5,800, each building indexes and cloning drives for its region. $113,100 in hardware.
engineering, amortised
Corpus curation, checksum verification, index building, failure documentation, the bilingual recovery procedure, the drill. Sunk once across every unit ever deployed — including units built in 2031.
per person, per decade
$173,100 in total, roughly 30,000 people reached. Under sixty cents a person a year, with no recurring exposure and no counterparty who can raise the price.
For a funder this is the rare shape where the second cheque buys more than the first, and where a pilot is genuinely cheap — twenty Reader nodes is $11,600, small enough to sit inside a discretionary budget and large enough to produce real evidence.
Four things it does better.
One of them is life and death.
Stated plainly, because it is checkable anyway — and because a comparison that only runs one way is not a comparison.
Satellite is two-way. No rung of this node is. It will not summon a helicopter, reach a relative, or report a collapse. In the first seventy-two hours after a disaster the terminal is the more valuable object, and it is not close.
A 27-billion-parameter model quantized to four bits is a competent generalist, not the best model available — and at Rung 1 you are running something far smaller than that. The hosted model improves every quarter while this one is frozen.
Prices, weather, news, who to contact, what is open. The node's knowledge stops at its last refresh, and it is deliberately not in the business of anything volatile.
A subscription needs no named recovery operator, no quarterly drive rotation, no printed procedure, no annual drill. This system asks a community for real custody, and custody is work — the same work at every rung.
So do not fund this instead of connectivity. Fund it beneath connectivity. It is a floor, not a ceiling.
The honest architecture is layered, and the ladder is what makes layering affordable. Satellite for the live, two-way, urgent seventy-two hours. A Hub where a region can support one. Readers everywhere else, for the ten years on either side of the emergency — the ordinary Tuesdays in a place reachable only by river, where the constraint was never the disaster.
The one thing here that cannot wait.
Hardware gets cheaper. Models get better. Solar gets cheaper. Every line in this budget can be deferred except one — and it is the line that populates every rung of the ladder forever.
The corpus can only be assembled while the network still exists. That is not a scarcity tactic; it is the same fact the whole project is built around, pointed at the purchase decision instead of the failure mode.
There is a second, quieter deadline. Every model in this build is permissively licensed and freely downloadable today. That is a policy environment, not a law of nature, and it has moved more than once in the last year. A weights file on a drive in Chocó is not subject to a licence change. A weights file you have not downloaded yet is.
Which is the whole optionality argument in one line: the cheap thing to do now is acquire the corpus and pick the smallest rung that fits the budget. Everything above it can be bought later, from anyone, at prices that will only have fallen.
All figures USD, August 2026, design estimates rather than quotations. The connected stack is priced at Colombian retail and given grid power so it carries no solar cost in the like-for-like comparison. Break-even excludes the value of live connectivity, which is real and is addressed above rather than discounted away. Ten years assumes no hardware refresh; drives are rotated quarterly and re-verified.