Two numbers underneath every AI budget belong to someone else: what a provider charges for a token, and how capable the model behind that token is this quarter. Buyers named both — one described a provider raising token prices overnight, another feared an isolated tool drifting two years away from language-model progress. One question sits behind both: which exposure can architecture remove, and which simply has to be priced?
Model Market & Provider Economics
Where Is the AI Model Market Heading,
and Will Provider Pricing Hold?
Two exposures sit under every AI plan: the price somebody else sets, and the capability somebody else ships. Architecture absorbs one. The other has to be measured on your own work.
Built from real buyer questions in our sales meetings
Direct Answer
Plan for two moves you do not control, and make the model a component you can replace. A model provider can change prices or terms on its own schedule, and an open-weight model on your own hardware will trail the newest cloud release by some margin at any given moment. Better negotiation fixes neither. The durable mitigation is architectural: keep the corpus, the workflows and the license portable, so the model underneath becomes a part you swap rather than a platform you are married to.
The limit: nobody prices this reliably, Iternal included. Iternal expects the frontier model makers to raise their rates as the cost of building data centers lands — a prediction, not a measurement. Iternal has also found an incorrect price for one hosted model in the blended-average column of its own cost calculator, and it rations its internal model allowance, giving it to developers first until the limit is hit. Treat any provider-economics comparison you are handed, Iternal's included, as a snapshot with an error bar, and re-run it on your own consumption.
Verify on your own workload and on your own paper. Measure the capability gap against the tasks your people actually perform, because a gap that ends a coding project is invisible in a payroll workflow. Get the pricing and term protection you count on written into the agreement. Then rehearse the swap once: export the prepared corpus, point the application at a different endpoint, and read what the license lets you keep on the day you walk away.
Capability gap and provider dependency are separate problems. The gap decides which tasks a local model carries today; dependency decides what happens to your work when someone else changes a service you rely on. For more information on what a private AI assistant costs, visit the cost page; for how the software is licensed, visit the licensing page.
Where a Local Model Actually Sits Against the Frontier
Iternal positions the models that run on a business laptop as closer to a widely used cloud assistant of 2024 than to a top-tier model inside a large data center — a rough placement offered as guidance, never as a benchmark. Sellers describe the same line from the other side: a laptop today runs what needed a mega cloud a year ago.
Both descriptions hold, because they measure different machines. The medium-sized model that fits a mainstream business device lands near that 2024 mark; a well-specified workstation costing roughly twenty thousand dollars runs an open-weight model close to top-tier cloud parity. Device class decides which sentence is true for you, so trust only the measurement taken on hardware you intend to buy. One gap closes on no schedule at all: a local open-weight model cannot answer recent public questions, because its knowledge of the outside world stops roughly a year before you load it.
When One Provider Stumbles, the Work Stops
Buyers raised this exposure repeatedly, always as something already lived through. A cloud coding tool went out and users could reach nothing. Applications vanished from a productivity suite without warning. Cloud models get withdrawn, here today and gone tomorrow, and buyers raised the prospect of a foreign government switching off a subscribed service.
The continuity answer is a floor, not a second subscription. Iternal has watched cloud outages break its own customer demonstrations, so it builds the fallback into the architecture. AirgapAI runs the model on the device and saves chats into your own file system, keeping prior work reachable when a provider goes dark; where the application points at a larger data-center model, a kill switch drops back to the local model without changing the interface. Make it a setting somebody flips, not a project that starts on the worst morning of the quarter.
Where the Gap Bites, and Where It Does Not
State the objection at full strength, because buyers do: an open-weight model running locally will always sit two or three revisions behind the latest, and if the distance grows far enough the cheaper option stops being worth having. Iternal accepts the premise and rejects the assumption underneath it, that the distance matters equally across tasks.
| Kind of work | Does the gap decide it? | What Iternal says |
|---|---|---|
| Frontier-class coding and long agentic chains | Yes | AirgapAI is not agentic today, because the models it runs sit well behind the state of the art. Iternal claims no equivalence to the leading cloud coding tools. |
| Recent public events, spreadsheets, image generation | Yes, for now | Public knowledge stops roughly a year before you load the model, spreadsheets are handled poorly, and on-device image creation is about two years out. |
| Question and answer over your own governed corpus | No | Answer quality tracks how well the source material was prepared, so the model is rarely the constraint. |
| Drafting, summarizing and back-office workflow | No | Frontier intelligence is almost too smart for back-office automation, and unnecessary for something like automating payroll. |
Split your population before you buy for it. Developers and coders need the newest frontier model; the rest of the office needs broad capability at a bounded cost. For many corporations and mid-sized enterprises, Iternal says plainly, a revision or two behind is good enough, and an older model keeps working on what it already handles. Refresh the application roughly quarterly and models every three to six months.
The Portability Checklist That Makes the Model a Component
Portability is bought at design time rather than at renewal. Four properties decide whether the model under your assistant is a component or a commitment:
- The corpus. Blockify output is designed to be agnostic to the embedding model and the vector database, and Iternal states it drops into any retrieval pipeline rather than only its own. Prepared data no single engine owns outlives every model decision after it.
- The endpoint. AirgapAI calls a model server over an OpenAI-compatible endpoint, accepts models you supply, and loads several while activating one at a time. Iternal states it does not care which models a customer picks.
- The license. Iternal sells AirgapAI as a one-time perpetual license per device, owned for the life of that device, with no token fees. A license nobody can reprice mid-term closes one exposure outright.
- The fallback. A kill switch drops a server model back to the local one, making the floor under your workflow a setting rather than a migration.
Two mechanics land on you, and Iternal says so. AirgapAI does not orchestrate the swapping of server-side models — a customer programs their own container to spin instances up and down — and prompting travels with workflow configuration in one JSON payload. See both demonstrated on your own hardware.
Will Provider Pricing Hold?
The direction is widely agreed; the timing is genuinely open. Token pricing today is heavily subsidized, and subsidies go first when a builder of data centers needs the capital back, which is why Iternal expects the frontier model makers to raise their rates. Buyers pushed back on the calendar rather than the logic: a near-term rise looks unlikely while competition from open-weight releases would cost a provider subscribers. Open-weight providers meet the same gravity eventually, because a model given away has to be monetized. What the evidence supports is a direction rather than a number, so treat any date or percentage you are quoted as a forecast.
Your remedy is contractual and architectural. An open-weight model carries no token billing, turning a recurring inference bill into hardware you already own plus a one-time data-preparation effort. Iternal answers the same exposure in its own terms: a one-time perpetual license per device, published updates included, no metering. Hold the caveat alongside it — Iternal grounds that argument in a predicted price correction rather than in current capability, and concedes that cloud economics beat its own in-sourcing pitch for some workloads today. Take your figures into these questions:
Pin it down: questions for your evaluation
-
On our tasks and our hardware, how far behind the cloud tool our team uses is the model you would ship us?The capability gap measured on your workload, not asserted about a device class.
-
What price and term protection can we hold for the life of the devices we are buying?Whether the pricing exposure the market expects is yours to carry or closed in writing.
-
Can you demonstrate pointing the application at a different model endpoint, and show what we keep if we leave?Portability as a rehearsed operation, and who owns the prepared corpus afterwards.
-
Which figures in any cost comparison you hand us are measured, and which are projected?Which half of the business case is auditable today, and which half carries an error bar.
Answered elsewhere
- Which models are supported, and running your own — see choosing a local model.
- Where a model came from, and adapting it — see model provenance and adaptation.
- How fast a model responds on a given device — see the local speed page.
- Bringing a metered AI bill under control — see the inference-cost page.
- Whether to move now or wait for things to settle — see acting now versus waiting.
- Translation, transcription and voice on the device — see languages and voice.
FAQ
FAQ: The Model Market and What It Costs You
Expect it to. Buyers describe a locally run open-weight model as two or three revisions behind the latest, and Iternal agrees its models sit behind the state of the art. That distance decides frontier-class coding, long agentic chains and recent public facts; it rarely decides question and answer over your own corpus, drafting or routine workflow.
With a cloud-only assistant, work stops: buyers described a coding tool going out and users reaching nothing. A device-resident model gives you a floor. AirgapAI runs the model locally and saves chats into your own file system, and a kill switch drops a server connection back to the local model.
A provider can reprice, and buyers named that exposure plainly. Your answer is contractual and architectural. An open-weight model on hardware you own carries no token billing, turning a metered bill into a fixed asset plus one data-preparation effort. Iternal sells AirgapAI as a one-time perpetual license per device with no metering. Put the term protection you need in writing.
In some segments, yes: capability that needed a large data center a year ago now runs on a well-specified local machine. Two forces cut the other way. The supply of strong open models has narrowed, much of it now from one region, and the largest providers keep their best work on their own platforms.
There is no confident answer available, and a manufactured one would not help. The incentives point the other way: large providers want users captured on their own platforms, and Iternal notes a model company is unlikely to release its best model for on-premise use, because weights on someone else’s hardware can be taken.
Rehearse the Swap Before You Need It
Forecasts about model pricing age badly; a swap you have already performed does not. Load a prepared corpus, point the assistant at one endpoint and then another, and read the license with the day you leave in mind.