Cutting OCR Costs for 64M Pages

A case from a nonprofit that needs to process 64 million pages was posted, and the comments converged on the view that "rather than just running a large VLM blindly, use OCR suited to the document type first."

The key suggestions are as follows.

• Prioritize traditional OCR: Try tools like Tesseract, PaddleOCR, Surya, MarkItDown, and Marker first
• Document classification pipeline: Extract text from digital PDFs directly, and only send scanned copies to OCR to reduce costs
• Small vision model + LoRA: If the document type is consistent, fine-tune a small VLM to balance accuracy and cost
• Dedicated OCR models: Nvidia Nemotron OCR v1, LightOnOCR-2-1B, DeepSeek OCR, OLM OCR, and others were mentioned

Figures from a practical-use perspective were also shared. One comment noted that LightOnOCR-2-1B can process about 5.71 pages per second on a single H100, offering an estimated efficiency of < $0.01 / 1,000 pages. Another comment calculated rough costs and estimated that combining 4090/3090-class GPUs with spot instances could significantly lower costs.

On the funding side, nonprofit/research credit tracks such as Google TPU Research Cloud, AWS Research Credits, and Oracle for Research were mentioned as real alternatives. However, the overall conclusion across the comments was to first measure quality and throughput using sample documents, and then choose the cheapest path based on those results.