Shopee's Custom LLM Scales 113x in 8 Months, Cuts Fraud Costs by 90%
nvidia · x · 2026-08-21
Shopee published results on Compass, its frontier LLM built for multilingual e-commerce in Southeast Asia. To bridge the "jagged frontier" where general models fail at local tasks, Shopee used the NVIDIA AI Factory stack. Key outcomes include API token scaling from 3 billion to 340 billion monthly (113x growth) and a 50x speedup in fraud detection at 90% lower processing cost via TensorRT-LLM. The tech stack utilizes NVIDIA Megatron-Core for pretraining, NeMo for post-training, and runs on A100, H100, and Blackwell GPUs.
More from Companies & People
- Bundling Era Ends as Agents Force Products to Stand on Merit — matt_slotnick · 2026-08-21
- PyTorchCon NA Keynote: Cohere CAIO Joelle Pineau to Speak — PyTorch · 2026-08-21
- Opinion: OpenAI Growth Surpasses Anthropic Due to Cost and Tone — bindureddy · 2026-08-21
- OpenAI Pledges Zero Data Retention for Enterprise in Bid to Win Customers — eliano · 2026-08-21
- Elevation of Slack is smart; bundling age is over — matt_slotnick · 2026-08-21
- Glen: Org-Wide Memory Layer for Agents Cuts Context by 29% — ycombinator · 2026-08-21