A finance leader rarely kills a cost comparison because the arithmetic is wrong. The comparison dies because someone ran it once, on one figure, already knowing which answer they wanted. Owned AI capacity against a rented cloud assistant invites exactly that failure: one side is a one-time purchase attached to a machine, the other a recurring per-seat fee, and the one-time figure changes with the route you bought through.
Own It Or Rent It
AI PCs Versus Cloud AI Subscriptions:
Which Is Actually Cheaper?
The breakeven arithmetic run across every one-time license figure on record, the device-price pressure that belongs in the same table, and when the subscription wins.
Built from real buyer questions in our sales meetings
Direct Answer
Run the comparison across the whole license spread, never on a single figure. Set the one-time device premium plus a perpetual license against the recurring per-seat fee over the life of the machine. Against a cloud assistant at roughly $30 per user per month, a one-time $96 license pays for itself in about three months — the arithmetic buyers actually ran in front of us. The $695 or $697 price on Iternal’s website reaches breakeven near month 23, the $799 packaged deployment near month 27, and the $1,499 Enterprise + Blockify tier near month 50 — the first three inside the four-year device life the table below assumes, the last one past it. One end of the spread is obvious; the other is a genuine decision.
The limit: permanence is an assumption. A four-year breakeven assumes the one-time option is still sold in year four on the terms you were quoted in year one. Iternal’s own record notes that its current licensing model does not scale for Iternal indefinitely. Put the renewal assumption in the model as a line the CFO can see, rather than leaving it inside the arithmetic.
Verify the figure and the unit on your own written quote. Each one-time figure belongs to a different purchase route — do-it-yourself and channel, direct from Iternal’s website, and the packaged deployment through the channel — and prices have moved. Iternal describes AirgapAI as licensed per physical device for the life of that device, covering everyone who uses it, and the same headline figure also appears quoted per user. Those are one total on a one-laptop-per-person fleet and a different total on shared machines. For more information visit the cost page.
Device prices are the counter-pressure, and they belong in the same table. Memory and component prices have climbed, and buyers told us the AI PC premium is getting hard to justify. Run the comparison against your actual refresh cycle rather than a new fleet purchase. For more information on how the license itself is structured, visit the licensing page.
The Breakeven Table, Run Across the Whole License Spread
The rented side is the easy side. Buyers described paying thirty to sixty dollars per user per month for cloud assistants; the table uses $30, the bottom of what they described, so the owned side has to win the hardest version of the argument. That per-seat figure belongs to the incumbent assistant page. The owned side is where evaluations go wrong, because no single number belongs in the cell. Run the row that matches your quote:
| One-time license | Route it was recorded on | Unit recorded with it | Breakeven at $30 per user per month | Inside a four-year device life? |
|---|---|---|---|---|
| $96 | Do-it-yourself and channel MSRP | Per device in some records, per user in others | About 3 months | Yes, with years of headroom |
| $695 or $697 | Bought directly on Iternal’s website | Per device in some records, per user in others | About 23 months | Yes, at roughly half the life |
| $799 | AI Assist, the packaged deployment, through the channel | One-time perpetual MSRP; device or user not stated | About 27 months | Yes, with roughly 21 months of the life left |
| $1,499 | Enterprise + Blockify, bought directly on Iternal’s website | Per seat, one-time | About 50 months | No — it lands past a four-year life |
Every figure above is stated on the cost page, each with the product and route it belongs to. Each is a real recorded price on a different purchase route, not a point on a range — averaging them produces a number nobody was ever quoted.
The $96 row against $30 per user per month is the comparison buyers ran in front of us, and it is why the owned model gets a hearing at all. Treat it as their arithmetic, not as your answer. Move down the table and the margin narrows sharply: at $799 the payback arrives in the third year of a four-year machine, and at $1,499 it arrives after the machine is due for replacement, which makes device life — not the license figure — the variable that decides the case.
Device Prices Are Moving Underneath the Model
A breakeven table assumes a stable device price. Buyers have spent the past year telling us otherwise: a global memory and storage shortage is pushing hardware prices up, with roughly $300 per unit attributed to memory alone on AI-capable client machines. Carry it as an order of magnitude, not a fixed adder: buyers cited it in a market repricing every week or two.
The behavior it produces is easy to verify inside your own procurement data. Buyers described buying the cheapest laptops available, pulling volumes forward ahead of further increases, and refusing to deal in four-thousand-dollar machines; channel partners described struggling to justify device costs that keep climbing. On the server side the pressure bites harder, with a base server carrying no GPUs put at fifty to a hundred thousand dollars. The reading: a premium that cleared the bar last year has to clear a higher bar this year, and a one-time license does not repair that. It changes which budget line the premium has to survive.
Where the Premium Lands: The Machines You Were Replacing Anyway
Buyers replace client machines on a rhythm, typically every three to four years, and buy the new license at that moment alongside the new laptop. The premium therefore lands on a purchase that already had a budget, an approver and a date — a categorically different argument from asking finance to fund a fleet purchase nobody planned. Iternal states the license runs for the life of the device, so an organization that sweats an asset four or five years pays that price once and keeps the upgrades. Amortized across the device life, a one-time fee becomes a monthly equivalent a budget owner can carry.
Where the framing fails, stated plainly. If the current fleet sits below the memory bar the software needs, the refresh stops being a line you were paying anyway and becomes a hidden precondition — buyers said asking them to acquire minimum hardware may simply be too big an ask. And if the plan is to stretch existing devices to seven years, or the machines sit on a ten-year defense refresh, the premium has no near-term purchase to ride.
The Monument Problem: Hardware Bought With No Workload Behind It
One buyer put it directly: a GPU purchase becomes an expensive monument to failure, because most buyers have no other GPU workload to put behind it. The pattern came up again and again in different clothes — idle servers on the data-center floor with no solution on them, a hospital group that bought all its hardware and GPUs and runs no AI projects, large stacks bought before anyone could say what to do with them. Others refuse to spend a hundred thousand dollars on a server before trying anything at all.
Already-bought hardware inverts the comparison. The capital is gone, so the live question is which workload to put on the metal, not whether to buy more. Iternal answers that unglamorously: run Blockify jobs overnight while the GPUs would otherwise sit idle, and note that querying idea blocks is CPU-driven, so server CPUs can carry the work where that is cheaper. For more information visit the sizing page.
The same risk follows the endpoint, because buyers of AI PCs are probably not fully using them. An unused one-time license is a monument at smaller scale, so adoption belongs in the model as an explicit assumption. For more information visit recovering stalled AI spend.
What Has to Be True for the Subscription to Win
A comparison that only ever produces one winner is marketing, not analysis. The subscription wins real deals for real reasons, and a model that cannot name them will not survive anyone who has bought both. Four conditions decide it:
- Demand is bursty rather than steady. Cloud ends up cheaper than owned capacity when elasticity is genuinely used and idle capacity is never paid for.
- Few people sit behind shared capacity. On-premises economics depend on how many users hit the box; thin usage against a large machine loses to a per-seat fee.
- Devices turn over faster than the breakeven month. Some organizations do not keep a laptop past two or three years, so a license paying back near month 23, or later still, buys device life the fleet never delivers.
- Nothing has to run on the device. Without a device-side reason to buy, commercial buyers told us they choose lower-cost PCs and rent the intelligence.
One buyer stated the strong form: there is no scenario today in which a financial business case can be made for running this on on-premises infrastructure. That judgment is about funding dedicated server capacity, a different purchase from a license riding a client device you were replacing regardless. Keep the two on separate rows. For more information on the metered side of the cloud bill, visit cutting token and inference cost.
The same four conditions decide where the work runs, not only who invoices for it: steady demand and many users behind one machine push the arithmetic toward the device and the local server, while bursty demand pushes it back to rented capacity. For more information visit the edge AI versus cloud economics page.
Per Device Or Per User: Take the Unit From Your Own Quote
Multiply the wrong unit by a fleet and the model is wrong by thousands of dollars before anyone questions the price. Iternal describes AirgapAI as licensed per physical device, good for the life of that device and covering everyone who uses it, with licenses that can be retracted and reassigned. The same headline figure also appears quoted per user — an identical total where each person carries one laptop, a different total the moment a machine is shared across shifts. The $799 packaged figure is recorded as a one-time perpetual MSRP with no minimum purchase quantity, and without the unit spelled out. Four questions settle it inside one call:
Pin it down: questions for your evaluation
-
Does the license on this quote attach to the device or to the named person, in the words of the quote itself?Which number your fleet gets multiplied by, and whether headcount growth changes the total.
-
On a machine worked by three shifts, how many licenses does this quote require?Whether shared and multi-user machines multiply the cost or ride on one license.
-
When a device is retired or a person moves on, can the license be reassigned, and on what terms?Whether the four-year model is a fleet cost or a rolling headcount cost.
-
Which purchase route is this quote on, and does that price hold at the fleet size we intend to buy?Which row of the breakeven table above is genuinely yours.
Iternal states each figure with the route it came from, and the unit travels with the figure rather than being assumed on your behalf. For more information visit the cost page.
Answered elsewhere
- How the license itself is structured, seat by seat against one-time ownership — see how AI software is licensed.
- What the software costs figure by figure, with the route behind each one — see the cost page.
- Turning the comparison into a funding case finance will sign — see building the AI business case.
- Sizing servers and choosing hardware for a shared deployment — see the sizing page.
- Paying for any of it when nothing has been set aside — see when there is no budget at all.
FAQ
FAQ: Owning AI Capacity Against Renting It
Decide it on the life of the machine, not the purchase order. Against a $30 per user per month subscription, a one-time $96 license pays back in about three months; the $695 or $697 price on Iternal’s website near month 23; the $799 packaged deployment near month 27, still inside a four-year device life; and the $1,499 Enterprise + Blockify tier near month 50, past it. The cheap PC wins when the fleet turns over fast, when use is thin or bursty, and when nothing you need has to run on the device itself.
It depends on whether the machine was already being replaced. Buyers told us the premium is getting hard to justify, with a memory and storage shortage pushing prices up and roughly $300 per unit attributed to memory alone. On a planned three-to-four-year refresh, the premium rides a budget that already exists. On a fleet you meant to stretch to seven years, or a ten-year defense refresh, it has nothing to ride on.
Divide the one-time figure on your written quote by the monthly per-seat fee it replaces. Against $30 per user per month, the figures on record pay back in about 3 months at $96, about 23 months at $695 or $697, about 27 months at $799, and about 50 months at $1,499. Those are different purchase routes rather than points on a range, so averaging them produces a number nobody was quoted. Write one limit into the model: a four-year payback assumes the one-time option is still sold in year four on year-one terms.
Where each person carries one laptop, the two counts produce the same total. They separate on shared and shift-worked machines. Iternal describes AirgapAI as licensed per physical device for the life of that device, covering everyone who uses it, and the same headline figure also appears quoted per user. Take the unit from your own quote, ask how many licenses a three-shift machine requires, and get that answer in writing before you buy.
It changes the question from whether to buy to what to run. Buyers described idle servers on the data-center floor with no solution on them, large stacks bought before anyone knew what to do with them, and GPUs running below capacity. With the capital already spent, the comparison is between a rented subscription and the marginal cost of a workload on metal you own. Iternal runs Blockify jobs overnight while GPUs would otherwise sit idle, and querying idea blocks is CPU-driven, so server CPUs can carry the work.
When demand is bursty and elasticity is genuinely used, so idle capacity is never paid for. When few people sit behind a shared machine, which is where on-premises economics stop working. When devices turn over faster than the payback month, so the license buys life the fleet never delivers. And when nothing has to run on the device, which is why buyers with no device-side reason to buy choose cheaper PCs and rent the intelligence.
Run the Table on Your Own Quote
Take the one-time figure and the unit from the quote in front of you, divide by the per-seat fee it replaces, and put the renewal assumption on its own visible line. A model built that way answers the challenge before it is made.