Turn idle compute into private AI retrieval
AiGrid is a self-hosted appliance that uses machines you already own to read, OCR, and index your document archive, then serves it to any AI assistant as cited passages. No GPU to buy, and no document leaves your network.
From $19,995 per year, with no per-seat or per-query fees.
The desktop window is open. All three pools are converting documents inside your network.
“In effect, a distributed virtual AI supercomputer.”
How one CIO described AiGrid to his peers. It is a fair analogy with one caveat: the grid does a single job, which is converting your documents for retrieval.
Your AI assistants can only work with what they can read
Most organizations hold decades of contracts, reports, manuals, and scanned archives that no AI assistant can use until every page has been read, laid out, OCR'd, and indexed. That processing is heavy compute, and until now there have been two ways to get it.
You can send the documents to a cloud service, which is hard to accept for sensitive material and is billed by the seat, the page, or the query. Or you can buy GPU infrastructure and run the pipeline privately, which is expensive, slow to procure, and becomes a permanent workload for your team.
There is a third source of compute, and you have already paid for it.
The capacity is already on your balance sheet
AiGrid groups the machines you lend into pools, and an administrator sets the schedule for each one. The agent runs at low priority, so it does not compete with other work.
After hours
The desktops your staff leave at night
Hundreds or thousands of workstations sit idle every evening. AiGrid uses them overnight and hands them back before the working day starts, and each one follows its own local clock.
Default hours for a workstation pool: weekday evenings and overnight.
Around the clock
The servers waiting for a bad day
Standby capacity in a DR site can do useful work on every other day. A server pool can run all day, every day, or on whatever schedule you choose.
A server pool can be set to work around the clock.
Spare capacity only
The headroom in the clusters you run
On a cluster, AiGrid runs below every other workload. It scales up from zero when there is work and gives the capacity back the moment the cluster needs it.
Illustration. What a cluster lends depends on what it has spare at that moment.
What makes AiGrid different
- It runs on idle compute
- Cloud RAG services are plentiful, and so are RAG engines that need a GPU fleet. AiGrid does the heavy document conversion on ordinary CPUs across machines you already own, so there is no GPU to procure and no per-page cloud bill.
- Private by architecture
- There is no AiGrid cloud service for your data to go to. The documents, the extracted text, the index, and the models all stay in your environment, and the appliance only ever has read access to your document store.
- It retrieves and never generates
- Every result is a real passage from your own documents with its filename and page number. Your assistant writes the answer and cites the source, so people can check what they are told.
- It fits the AI stack you have
- Any assistant or agent that can call an HTTPS endpoint can use it. You can use your own vector database in place of the built-in one, or keep your whole RAG layer and use AiGrid purely for text extraction.
See it in 87 seconds
Built for the questions your CISO will ask
- Machines in a pool never receive an inbound connection. Every call is outbound HTTPS.
- On a desktop the agent runs under its own service account, at low priority, and not as the user.
- Air-gapped capable. Processing and search call no external service.
From document store to cited answer
Your documents stay where they are
AiGrid scans your S3-compatible storage on a schedule and queues anything new or changed.
Machines ask for work
A machine inside its allowed hours asks the appliance for a document. Work is never pushed to a machine.
The pool does the heavy lifting
The machine works out the layout, OCRs scanned pages, and cuts the text into short passages.
The appliance builds the index
Passages are indexed centrally, inside your network, alongside their filename and page number.
Your assistant asks and cites
The assistant sends a question to the retrieval endpoint, gets the best-matching passages, and answers with sources.
See it index a real corpus
We will walk you through AiGrid ingesting documents across a live pool, and show an AI assistant answering with filename and page citations.
- A live demonstration, followed by your questions.
- A trial scoped to your estate, run on your infrastructure.
- No obligation. A specialist will contact you to arrange a time.