Four ways to make documents usable by AI, and a fifth
The comparison below is by category. Products within a category differ, so check any vendor's current documentation and pricing before you decide.
| Cloud document AI and hosted RAG | Self-hosted RAG on GPUs | Enterprise search suites | Build it yourself | AiGrid | |
|---|---|---|---|---|---|
| Where documents are processed | In the provider's cloud | In your network | Usually in the vendor's environment | In your network | In your network, on machines you own |
| Compute for ingestion | The provider's, billed to you | GPU capacity you procure and run | The vendor's | Whatever you build and operate | Idle CPUs on workstations, DR servers, and clusters |
| Typical pricing shape | Per page, per query, or per seat | Hardware plus software and operations | Per seat | Engineering time and infrastructure | Appliance plus a flat annual fee per machine |
| What you get back | Varies, often generated answers | Varies | Search and generated answers | Whatever you build | Cited passages for your own assistant |
| Who runs it | The provider | Your team | The vendor | Your team, permanently | One appliance on your Kubernetes, operated from a dashboard |
What the same job costs at four scales
An illustration at published list prices, for archives averaging 10 pages per document, converted with their layout and tables intact.
| Option | 850,000 documents 8.5 million pages | 9 million documents 90 million pages | 20 million documents 200 million pages | 100 million documents 1 billion pages |
|---|---|---|---|---|
| Google Document AI, Layout Parser $10 per 1,000 pages | $85,000 | $900,000 | $2,000,000 | $10,000,000 |
| Amazon Textract, text with tables $0.015 per page for the first million in a month, then $0.01 | $90,000 | $905,000 | $2,005,000 | $10,005,000 |
| AiGrid $19,995 appliance, plus $25 per workstation and $295 per server (Kubernetes node), per year |
$25,880 per year with 200 workstations and 3 servers |
$49,420 per year with 1,000 workstations and 15 servers |
$78,845 per year with 2,000 workstations and 30 servers |
$314,245 per year with 10,000 workstations and 150 servers |
Cloud figures are US list prices from the Google Document AI and Amazon Textract pricing pages, read in November 2026, with each job run in one billing month, which gives the lowest cloud price. Amazon's rates are for US West (Oregon). Negotiated discounts, storage, and the cost of a hosted vector database are not included. The AiGrid pools are examples of workstations and servers working together. You can mix the two in any proportion, and as a rough guide one server can do the work of about 100 workstations. The pool you lend sets how long the job takes, and the page count does not change the price. Check current prices before you rely on these figures.
- Where a cloud service costs less
- Plain text recognition, without layout or tables, is far cheaper in the cloud for small and mid-sized archives: about $6,000 to $9,600 for 8.5 million pages. The gap narrows as the archive grows, and at 1 billion pages it comes to about $601,000 to $604,500. Small archives favor the cloud in any case. Against $10 per 1,000 pages, AiGrid with the pool in the first column breaks even at about 2.6 million pages in the first year.
- What the table does not show
- AiGrid's price does not grow with the number of pages, but the time the job takes depends on the size of your pools and the hours you give them. The cloud figures are one-time for each archive, and AiGrid's is an annual fee that also covers every new and changed document.
- The cost that is not on an invoice
- With a cloud service, every page of the archive leaves your network to be processed. For many of the organizations we speak to, that settles the question before price does.
When each one is the right call
- A cloud service
- A reasonable choice when the documents are not sensitive and the volume is modest. It stops being one when the corpus has to stay inside your boundary, or when per-page and per-query charges are applied to decades of archives.
- A GPU-based stack
- The right answer if you are already running GPUs for other AI work and have capacity to spare. If you would be buying GPUs just to read documents, it is worth knowing that document conversion and retrieval do not need them.
- An enterprise search suite
- These products do much more than retrieval, including connectors to SaaS applications and their own assistants. If you need that breadth, AiGrid is not a replacement. If you need your own assistant to read your own archive privately, a suite is more than the job requires.
- Building it yourself
- The open-source parts exist, and AiGrid uses several of them. The work is in scheduling conversion across machines that switch off at random, without opening ports on them. If you have built the rest already, AiGrid can run purely as the text extraction tool and give you that piece.
- AiGrid
- The right call when the documents must stay inside your network, you would rather not procure GPUs for ingestion, and you already own machines that sit idle for part of every day.