Azure Document Intelligence On Premise: Container Pricing

Short answer: Yes, Azure Document Intelligence runs on premise, in Docker containers, and it comes in two very different forms. A connected container runs on your hardware but phones home to Azure every 10 to 15 minutes to report usage, and a prebuilt invoice commitment starts at $152 a month for 20,000 pages. A disconnected container runs fully offline, requires Microsoft's approval, and starts at $8,640 a year for 100,000 pages a month. The prebuilt invoice model in either container is the 2023-07-31 (v3.1) release, one major API version behind the cloud service.

Every price on this page comes from Microsoft's Retail Prices API for East US, which carries the container meters that the public pricing page does not display. The requirements come from Microsoft Learn's container documentation and the disconnected container FAQ. If you only need invoices read accurately and do not have a hard offline requirement, you can test that right now with the extractor at the top of this page, without a Docker host or an approval form.

Can Azure Document Intelligence run on premise?

Yes. Microsoft ships Document Intelligence as Linux Docker containers you pull from the Microsoft Container Registry and run on any x64 host, VM or edge device. The models run on CPU, not GPU. What changes between the two options is billing and connectivity: connected containers must reach an Azure billing endpoint, and disconnected containers need a license file and an approved annual commitment.

Not every model is available on premise, and the container versions lag the cloud. According to Microsoft's install guide, the v4.0 (2024-11-30) containers cover only Read and Layout. The Invoice, Receipt and ID Document containers are on v3.1 (2023-07-31). Custom and General Document are on v3.0 (2022-08-31). If your evaluation was done against the current cloud invoice model, the model you would run on premise is not the same release, so test on the container before you commit.

Azure Document Intelligence container pricing

Containers are not billed per page at the pay as you go rate. They are billed through commitment tiers, and each tier has three prices: Azure hosted (the cloud service on commitment), connected container and disconnected container. Here is the prebuilt tier, which is the one that covers the invoice model.

For reference, the cloud pay as you go rate for the prebuilt invoice model is $10.00 per 1,000 pages with no commitment. So a connected container at the 20,000 page tier is 24% cheaper per page than calling the cloud API, and it is 20% cheaper than the same commitment bought as a hosted service. That 20% gap holds at every prebuilt tier.

The number most people do not expect is in the last column. A disconnected container costs exactly the same as a connected one once you divide the annual price by twelve. $8,640 over 12 months is $720, which is the connected 100,000 page price. The same is true at 500,000 ($38,400 a year, $3,200 a month) and at 1,000,000 ($72,000 a year, $6,000 a month). Microsoft charges no premium for running fully air-gapped. What it charges instead is access: you have to qualify, and the smallest tier is five times larger than the smallest connected tier.

Custom extraction and Read tiers

If your documents need a custom model, or you only need raw text, the tier ladders look like this. The discount for moving into a connected container is smaller on these meters: 15% on custom and about 10% on Read.

The disconnected column again equals twelve months of the connected column on every row. The entry points for offline use are therefore $8,640 a year for prebuilt, $12,960 for Read and $24,480 for custom.

How much does Azure Document Intelligence cost when you exceed a commitment?

On connected containers and hosted commitments, pages above your tier are billed at that tier's effective rate. The connected prebuilt overage meters are $7.60 per 1,000 at the 20,000 tier, $7.20 at 100,000, $6.40 at 500,000 and $6.00 at 1,000,000. Going over is not punished, which also means a connected 20,000 tier behaves like a flat $7.60 rate with a $152 monthly floor.

That floor sets the break-even. At $10.00 pay as you go, the $152 commitment pays for itself at 15,200 pages a month. Below that, pay as you go is cheaper even though the per-page rate is higher. Disconnected containers have no overage meter at all, because no usage leaves the building. Microsoft's FAQ says you can add units at any point in the year, charged pro-rated for the days left in the billing year.

What is a disconnected container in Azure?

A disconnected container is the same Docker image as the connected one, started once with DownloadLicense=True while it can still reach Azure. That writes a license file with an expiration date. From then on the container runs with no internet connection, writes usage records to a mounted output folder, and exposes a /records/usage-logs endpoint so you can report consumption against the commitment you bought.

The connected container is where the "on premise" label gets misleading. Microsoft's install guide says the container reports usage about every 10 to 15 minutes. If it cannot reach the billing endpoint, it keeps running but stops serving queries, retries 10 times at the same interval, and then stops serving requests entirely. Documents stay on your hardware, but the service still depends on outbound connectivity to Azure. If your security team signs off on "on premise" because they assume no internet dependency, show them that paragraph first.

Who qualifies for disconnected containers?

Access is gated, and the bar is high. Microsoft's disconnected container FAQ lists these requirements, and you submit them through a request form before you can buy anything.

• Your organization must have an Enterprise Agreement or an equivalent agreement and be identified as a strategic customer or partner of Microsoft.
• You need a valid offline scenario: zero connectivity, a remote site with occasional internet, or regulation that prohibits sending any data back to the cloud.
• The Azure subscription ID must belong to the applying organization. Applications from personal email addresses are denied.
• Microsoft reviews and replies within 10 business days, then sends purchasing instructions for the commitment tier.

Two more lines from the same FAQ matter to anyone in a regulated industry. Disconnected containers carry no SLA, because you control the hosting. And Microsoft states that Azure AI containers don't have any compliance certifications. The certifications you rely on for the cloud service do not transfer to a container on your server, so the host, the storage and the access controls become your audit scope. Teams that already run formal programs usually need to map the container host into their SOC 2 and HIPAA control set, and a compliance automation platform is the practical way to keep that evidence current. The FAQ also notes that disconnected resources are created from the public cloud, not from Azure Government.

What hardware does the invoice container need?

The invoice container does not run alone. Microsoft lists Layout as a required supporting container for Invoice, so a working invoice deployment is two containers. Each has its own sizing.

Plan for at least 16 cores and 32 GB of memory for one invoice stack, and 48 GB to match Microsoft's recommendation, before redundancy. Since there is no SLA, a second node for failover is on you as well.

Is on-premise Azure Document Intelligence worth it for accounts payable?

Run the numbers on a typical US AP department. At 5,000 invoices a month with an average of two pages each, you process 10,000 pages. Pay as you go in the cloud costs $100 a month. The cheapest connected container commitment costs $152 a month and needs a 16-core host you have to buy, patch and monitor. The disconnected option is not available to you at all, because the smallest tier is 100,000 pages a month and you likely do not have strategic-partner status with Microsoft.

On premise pays off when three things are true at once: you process well above 15,000 pages a month, you have a documented rule that documents cannot leave your network, and you have engineers to build the rest of the pipeline. The container returns JSON. Getting vendor, line items, tax and totals into an ERP still requires mapping, validation and an import step, and our Azure Document Intelligence pricing breakdown covers what that build looks like on the cloud side.

For everyone else, the question is really about data handling, not deployment. If the concern is where invoices are stored and who can see them, a hosted extractor with a clear retention policy answers it without the hardware. That is what we built: upload PDFs or scans, get vendor, dates, totals, tax and every line item back as Excel, CSV or JSON, with plans from $49 a month, or $24 a month billed annually. If you are comparing engines on accuracy rather than deployment, our Mistral OCR vs Azure Document Intelligence test and the Azure Document Intelligence alternative page show where each one struggles with line items.

Frequently asked questions

Does Azure Document Intelligence have containers?

Yes. Microsoft publishes Read, Layout, Invoice, Receipt, ID Document, Business Card, General Document and Custom Template containers on the Microsoft Container Registry. Only Read and Layout are on the current v4.0 API. Invoice, Receipt and ID Document are on v3.1 (2023-07-31), and Custom is on v3.0.

How much does an Azure Document Intelligence disconnected container cost?

The smallest prebuilt tier is $8,640 a year for 100,000 pages a month, which works out to $7.20 per 1,000 pages. Custom starts at $24,480 a year and Read at $12,960 a year. Commitments are billed annually from the day the resource is created.

Is a connected container really on premise?

The processing is. Documents are analyzed on your hardware, but the container sends usage metadata to Azure every 10 to 15 minutes, and after 10 failed attempts to reach the billing endpoint it stops serving requests. It needs outbound internet access to keep working.

Can I use Azure Document Intelligence containers with Azure Government?

Microsoft's FAQ says disconnected container resources are currently created from the public cloud. Sovereign clouds, including Azure Government, are not listed as a source for them, so check with your Microsoft account team before planning a GCC High deployment around containers.

Is the container invoice model as accurate as the cloud version?

It is a different release. The invoice container runs the 2023-07-31 model, while the cloud service offers the 2024-11-30 API. Microsoft does not publish a side by side accuracy comparison, so run your own invoices through both before you sign an annual commitment.

Which is cheaper, an Azure container or the cloud API?

Below 15,200 pages a month, the cloud pay as you go rate of $10 per 1,000 pages is cheaper than the smallest connected commitment. Above that, a connected container commitment is 20% cheaper than the hosted commitment at the same tier, before you count the cost of the hardware you run it on.