Capacity planning for an on-premise deployment: what drives throughput, what a GPU is worth in pages per hour, worked sizing from 50,000 to 5 million pages a month, and how to measure all of it on your own hardware before you commit to anything.
Reference figures measured on an NVIDIA L40S 48 GB worker with 16 vCPU and 128 GB RAM, running analyse, identify and classify over a mixed set of scanned invoices, contracts and claim forms in Arabic and English. Your document mix will move these numbers in both directions.
Sizing conversations usually start with GPU model numbers. They should start with what is on the page, because that is the variable with the widest range.
A native PDF invoice and a fourth-generation fax of a handwritten Arabic claim form are not the same unit of work, even though both count as one page. Image quality recovery, dewarping and dense-script recognition are where the time goes.
Analyse is compulsory — everything else is built on it. Each additional capability you enable adds work, and question answering adds the most because it involves generative inference rather than recognition.
Memory decides how large a batch you can hold; compute decides how fast you clear it. The reference worker is an L40S 48 GB because it fits the full model set with room for a 32-page batch, but the platform is not fussy about silicon.
The same cluster serves both, but they optimise for opposite things. Bulk work batches aggressively for throughput; interactive work runs small batches to keep tail latency down. Separating them is what stops one from ruining the other.
Four realistic volumes, sized for a mixed scanned document set running analyse, identify and classify, with enough headroom to clear a full monthly backlog inside a working week.
GPU workers
Total vCPU
Total RAM
Local NVMe
Object storage, year one
Sustained throughput
Time to clear one month
Matching topology
Licence impact of volume
Sizing assumes a mixed scanned set at roughly 2.0 seconds of GPU time per page with the analyse, identify and classify capability set. Enabling Ask indexing across the same corpus adds about 50% to GPU time — plan an extra worker per five, or accept a longer bulk window. These are engineering starting points to be validated on your hardware.
Capacity is a property of your cluster, not of a contract tier. A GPU worker is a stateless pod that pulls page tasks off a shared queue, so adding one is a node label and a replica count — no re-architecture, no migration, no renegotiation.
Because the licence is sized by deployment footprint rather than pages, adding a worker inside your licensed environment is a hardware decision your infrastructure team can make on its own.
What to promise your users, and what the platform does when you ask it for more than it has.
Latency figures are measured inside the cluster and exclude your own network round trip. All of them are emitted as Prometheus histograms, so you can hold us to the same numbers on your own dashboards.
Every deployment ships a benchmark harness. During the Proof of Value we run it with you, on your documents, on the hardware you would actually buy.
Two to five thousand pages that look like your real workload, including the difficult tail — phone photos, handwritten Arabic, stamped contracts, multi-generation faxes and multi-document packets that need splitting. A clean sample produces a flattering number you cannot plan against.
One GPU, one worker, the capability mix you intend to run in production. This produces pages per hour, GPU utilisation, p50 and p95 latency and a per-stage time breakdown showing exactly where the seconds go.
Repeat with four workers. If throughput does not scale close to four times, the harness tells you which tier saturated first — rasterisation, object storage read bandwidth, or the database write path — before you commit to a hardware order.
Drive a bulk backfill and an interactive load at the same time, which is what a real Monday morning looks like. Confirm that interactive p95 stays inside your service target while the backlog drains behind it.
Throughput without accuracy is not capacity. The same run produces field-level accuracy against a human-adjudicated ground truth, so you can see the cost of a faster configuration before you choose it.
The harness output becomes a written sizing recommendation: GPU count, node shapes, storage classes and expected pages per hour at your volume. You keep that document whether or not you proceed.
On a metered platform, every improvement to the models is an invoice you have to justify before you can benefit from it. That is why archives sit unprocessed at organisations that own perfectly good extraction tooling.
Competitor figures reflect publicly documented list pricing for metered document-parsing platforms as of 2026. Verify against current vendor documentation before making a decision — we would rather you checked.
What platform teams need to know before they size a cluster, answered with numbers.
On the reference worker — an NVIDIA L40S with 48 GB — plan on roughly 1,800 pages per hour for a mixed scanned business document set running analyse, identify and classify. Clean native PDFs run closer to 4,000 pages per hour, and a heavy mix with question answering over dense Arabic scans drops to around 900. The spread is real, which is why we benchmark on your documents rather than quoting a single number.
A Proof of Value gives you measured pages per hour, p50 and p95 latency and field-level accuracy on your own hardware and your own documents, with the threshold agreed in writing before we begin.
