Azure Document Intelligence on-premise: what the disconnected container actually costs
TL;DR
Running Azure Document Intelligence on your own hardware means buying a disconnected container commitment, priced annually and charged in full up front.
At one million pages a year, pre-built extraction costs €8,586.60 through the cloud S0 endpoint and €61,823.80 as a disconnected commitment, a 7.2 times premium, verified against the Azure Retail Prices API on 10 September 2026.
Connected containers are not an air-gap option, because Microsoft documents that they stop serving queries when they cannot reach the Azure billing endpoint.
Container model coverage is narrower than the cloud service: v4.0 containers cover Read and Layout only, while Invoice, Receipt and ID Document remain on v3.1.
Below roughly 250,000 pages a year, paying Microsoft beats any self-hosted build once hardware, engineering and internal ownership are counted honestly.
Questions people ask
How much does Azure Document Intelligence cost on-premise?
Between €4.17 and €210 per 1,000 pages, depending on the model family and commitment size. Pulled from the Azure Retail Prices API on 10 September 2026 for West Europe in EUR: the disconnected Read commitment works out at €4.17 per 1,000 pages at the 16 million page tier, pre-built extraction at €61.82 per 1,000 pages at the one million page tier, and custom extraction at €157.65. Commitments are bought annually and charged in full at purchase.
Can an Azure Document Intelligence container run fully offline or air-gapped?
No, not in connected mode. Microsoft states that containers are not licensed to run without connecting to Azure for metering. A connected container reports usage every 10 to 15 minutes, and if the billing endpoint stays unreachable it continues running but stops serving queries, then stops entirely after ten failed attempts. Only the disconnected commitment, which uses an offline licence file and no outbound connection, will pass a genuine air-gap review.
What is the premium for disconnected container pricing versus the cloud API?
For one million pages of pre-built extraction in West Europe, the cloud S0 endpoint costs €8,586.60 a year and the disconnected commitment costs €61,823.80. That is a 7.2 times premium. For custom extraction the same volume is €25,759.90 in the cloud against €157,650.70 disconnected, a 6.1 times premium. Plain Read is far closer, at roughly three to four times cloud pricing.
Do the containers support every Azure Document Intelligence model?
No. The v4.0 containers support Read and Layout only. The Invoice, Receipt and ID Document containers remain on version 3.1, and custom generative extraction has no container release at all. If a demo showed you custom generative extraction on a cloud endpoint, that capability does not currently transfer to an on-premise deployment, so price the gap rather than assuming the catalogues match.
At what page volume does self-hosting stop being worth it?
Below roughly 250,000 pages a year, pay Microsoft. At 100,000 pages the disconnected pre-built commitment is €7,418.86 a year, and no combination of GPU hardware, a consulting engagement and internal ownership beats that. Between 250,000 and one million pages, model it properly, because you break even in year two rather than year one. Above one million pages a year the case for self-hosting is clear.
What does an Ækora engagement actually involve?
It is a consulting and deployment engagement, not a product. There is no free tier, no trial and no self-serve sign-up. You get a measured accuracy baseline on your own documents, a sized hardware plan and a cost model with sources attached. You also own the infrastructure afterwards, which means no vendor SLA or status page, and a slower time to first result than an Azure API key.
Want this worked out on your documents?
We will price a real sample against OpenAI or Anthropic and tell you whether Bulk document processing or a local model on your hardware is the cheaper first step.
Running Azure Document Intelligence on hardware you control costs between roughly €4.17 and €210 per 1,000 pages, depending on which model family you need and how large an annual commitment you buy. Pre-built invoice extraction at one million pages a year is €61,823.80 as a disconnected container in West Europe, against €8,586.60 for the same volume through the cloud S0 endpoint. Same models, your own metal, 7.2 times the price.
Every figure on this page was pulled on 10 September 2026 from the Azure Retail Prices API, filtered to region West Europe, currency EUR, product name containing "Document Intelligence". You can run the same query yourself in a browser. Prices move, so check the date before you quote anything here in a business case.
Microsoft sells two things called containers, and only one is on-premise
This is the distinction that wastes the most time in procurement, so deal with it first. Azure Document Intelligence ships as a Docker container you can pull into your own datacentre. There are two licensing modes for that container, and they are not interchangeable.
The connected container runs on your infrastructure but bills through Azure. Microsoft's own documentation is blunt about it: containers "aren't licensed to run without being connected to Azure for metering", they report usage "about every 10 to 15 minutes", and if the billing endpoint stays unreachable the container "continues to run, but doesn't serve queries". After ten failed attempts it stops serving requests altogether. Read that again if your DPIA turned on network egress. A connected container is an Azure service that happens to execute on your rack.
Definition · Disconnected container
A licensing mode in which the Document Intelligence container runs with no outbound connection to Azure at all. You buy an annual commitment up front, download a licence file, and the container serves requests offline for the term. This is the only Azure Document Intelligence deployment that survives a genuine air-gap review.
If your legal team has already said the documents cannot leave the building, the connected container does not solve your problem. The disconnected commitment does. It also costs considerably more, and it is sold on terms that are unfamiliar to teams used to pay-as-you-go.
What the disconnected container costs per page
Microsoft sells disconnected commitments in three model families: Read (OCR and text extraction), Pre-Built (invoice, receipt, identity document), and Custom (models you train on your own layouts). Each family has tiers sized by annual page volume. Divide the commitment price by the tier size and you get an effective per-page rate.
Azure Document Intelligence: cloud S0 versus disconnected container rates, West Europe, EUR, pulled 10 September 2026
Model family
Cloud S0, per 1,000 pages
Disconnected, per 1,000 pages
Premium
Read
€1.288
€5.56 at the 2M tier, €4.17 at the 16M tier
3.2x to 4.3x
Pre-built (invoice, receipt, ID)
€8.5866
€74.19 at the 100K tier, €61.82 at the 1M tier
7.2x to 8.6x
Custom extraction
€25.7599
€210.20 at the 100K tier, €157.65 at the 1M tier
6.1x to 8.2x
Azure Document Intelligence: cloud S0 versus disconnected container rates, West Europe, EUR, pulled 10 September 2026
Two things stand out. The premium falls as you commit to more volume, which is normal. And the premium is much heavier on structured extraction than on plain OCR. If all you need is text off a page, the disconnected Read container is only three to four times cloud pricing. If you need invoice fields, you are paying seven or eight times.
The same numbers as an annual invoice
Per-page rates are easy to wave away. Here is what finance will actually see, for pre-built extraction, at three volumes.
Annual cost of pre-built extraction, cloud S0 versus disconnected commitment, West Europe, EUR
Annual volume
Cloud S0
Disconnected commitment
Difference
100,000 pages
€858.66
€7,418.86
+€6,560.20
500,000 pages
€4,293.30
€32,972.69
+€28,679.39
1,000,000 pages
€8,586.60
€61,823.80
+€53,237.20
Annual cost of pre-built extraction, cloud S0 versus disconnected commitment, West Europe, EUR
The container catalogue is narrower than the cloud catalogue
There is a second cost that does not appear on the invoice. Container model coverage lags the cloud service. As of Microsoft's current install-and-run guidance, the v4.0 containers support Read and Layout only. Invoice, Receipt and ID Document containers remain on v3.1. Custom generative extraction, the newest and most capable offering, has no container at all.
So the honest comparison is not "the same product, on-premise, for more money". It is an older product, for more money, with a smaller catalogue. That may still be the right purchase. But price the gap rather than assuming parity, especially if a solution architect demoed custom generative extraction to your team on a cloud endpoint.
The alternative: open-weight models on your own GPU
The reason this question is worth asking in 2026 and was not worth asking in 2023 is that small open-weight document models have become genuinely good. On Roboflow's comparison of OmniDocBench v1.6 results, PaddleOCR-VL-1.6 scores 96.34% at 0.9B parameters under Apache 2.0, MinerU2.5-Pro scores 95.75% at 1.2B, and GLM-OCR scores 95.22% at 0.9B under MIT. These are models that fit on one mid-range accelerator, not a cluster. Check the licence terms individually before you commit, because they are not all equally permissive.
On cost, Spheron's self-hosting guide works a concrete example: 10,000 pages on an L40S at roughly $7.27 on-demand, against $15.00 for the same pages through AWS Textract. That is about half. It is also a figure worth treating carefully, and the same guide says why: quoted throughput numbers are peak decode rates, while sustained end-to-end throughput including PDF rasterisation and I/O is roughly half of that. Halve the pages per hour and you double the GPU hours. Any vendor who quotes you peak decode as a capacity plan is quoting you a number they cannot hold.
One thing we will not tell you: that these models are more accurate than Azure Document Intelligence. Microsoft does not publish an OmniDocBench score, so no head-to-head exists that you could open and read. Anyone claiming a percentage-point win over Azure is extrapolating. The only benchmark that settles it is your own documents, which is why a serious evaluation starts with a few hundred of your real pages and a measured field-level error rate. Our review of local OCR models goes through the current field in more detail.
What you actually need in place
A self-hosted stack is not a licence purchase. It is four separate things, and skipping any of them is how these projects stall in month four.
Hardware sized to sustained throughput, not peak. One L40S-class card handles a surprising amount of steady document volume, but the sizing has to start from your pages per hour at the busiest week of the month, not the annual average. We wrote up the method in GPU sizing for document processing.
A measured accuracy baseline on your documents. Field-level precision and recall on a labelled sample, per document type, before anything goes live. Without it you have no way to know whether a model change helped.
An exception path. No extraction stack is right on every page. You need a confidence threshold and a queue where low-confidence pages go to a human, plus reporting on what fraction lands there.
Somebody who owns it on Tuesday morning. Patching, model updates, GPU driver upgrades, capacity. This is real internal cost and it belongs in the business case alongside the hardware.
The volume below which you should just pay Microsoft
Here is the part most vendors leave out. Run the arithmetic against the disconnected commitment, because that is your real alternative once cloud is off the table.
At 100,000 pages a year, Microsoft's disconnected pre-built commitment is €7,418.86. No consulting engagement plus GPU hardware plus a year of somebody's attention comes in under €7,418.86. Not ours, not anyone's. At that volume you should buy the container, or reopen the cloud conversation, and we will tell you so on the first call rather than at the end of a paid scoping exercise.
At 500,000 pages a year the commitment is €32,972.69 and it recurs every year. A first-year engagement plus hardware lands in roughly the same range, so you break even in year two rather than year one. That is a real decision, not an obvious one, and it turns on how long you expect to be doing this.
At 1,000,000 pages a year the commitment is €61,823.80 annually and the case flips clearly, because year two of a self-hosted stack is hardware you already own plus support, not another €61,823.80. Spheron's independent break-even estimate against Textract sits at 50,000 to 100,000 pages per month for raw compute alone, which is 600,000 to 1.2 million a year. Two different methods, same neighbourhood.
So: below roughly a quarter of a million pages a year, pay Microsoft. Between there and a million, model it properly. Above a million, the arithmetic is not close. The same shape of calculation, run against per-token API pricing rather than a container licence, is in our fixed cost versus per-token comparison.
Where we lose
Ækora is a consultancy and a deployment engagement. There is no product to sign up for, no free tier, no trial. That has consequences you should weigh honestly.
Time to first result. An Azure API key works this afternoon. We scope, pilot, then roll out. If you need extraction running before the quarter closes, Azure wins and we are not going to pretend otherwise.
No SLA, no status page, no vendor-side SOC 2. You run the infrastructure, so you own the uptime. Some risk committees will not accept that, and that is a reasonable position.
Microsoft shops. If your data already lives in Azure, your identity is Entra, and your volume is moderate, Azure Document Intelligence is genuinely the right answer. We have said this to prospects and we will say it to you.
On compliance, one clarification worth making plainly. Deploying locally does not make you GDPR compliant. It removes an international transfer and a processor from your record of processing activities, which makes the compliance work smaller and the DPIA easier to write. The obligations remain yours. We cover the distinction in GDPR-compliant document AI.
How to decide this week
Pull your actual page volume for the last twelve months, split by document type. Multiply by the rates in the first table. If the disconnected commitment for your volume is under about €10,000 a year, buy it and move on to a more interesting problem. If it is over €50,000 a year and recurring, the build-versus-buy question is worth a fortnight of proper analysis.
If you want that analysis done with your documents rather than a vendor's sample set, that is what our bulk document processing engagement starts with: a measured accuracy baseline on your pages, a sized hardware plan, and a cost model you can take to your CFO with the sources attached. If the numbers say pay Microsoft, the report will say pay Microsoft. You can also read our wider comparison of Azure Document Intelligence and Textract alternatives in Europe first, or just get in touch with your volumes and we will tell you which side of the line you are on.