Firmus

About Firmus

Firmus provides GPU-first AI cloud infrastructure across Asia-Pacific, combining on-demand and reserved GPU clusters, bare metal, and S3-compatible WEKA storage. Its modular AI Factories deliver high-density performance with lower power and water usage, backed by MLPerf-validated systems and sovereign deployment options. Users select a region, GPU SKU, and capacity, then attach S3-compatible storage to launch training or fine-tuning jobs. Clusters scale horizontally by adding nodes or vertically by reserving larger GPU pools. Telemetry tracks performance, energy, and thermal metrics to optimize runs and control costs. The platform supports LLM training, LoRA fine-tuning, vision and multimodal pretraining, recommender systems, and high-performance inference serving with predictable latency. It is designed for AI-native companies, enterprises consolidating fragmented GPU estates, research labs, and public-sector teams requiring data residency. Teams across multiple sites benefit from consistent hardware, easing reproducibility, compliance reviews, and long-term capacity planning. Performance and energy telemetry help tune batch sizes, parallelism, and placement, while support assists with benchmarking, runbooks, and cost planning. Implementation includes security reviews, workload migration planning, and SLAs for capacity, availability, and response, with optional multi-site designs for resilience and data residency.

Key features

• ✓Modular AI Factories with liquid-cooled, high-density GPU clusters
• ✓GPU options: Blackwell GB300, H200 HGX, and L40S for training, fine-tuning, and inference
• ✓NVSwitch intra-node bandwidth and full-rail InfiniBand for cluster-scale networking
• ✓S3-compatible WEKA storage for high-throughput data pipelines and artifact management
• ✓On-demand instances and reserved capacity with predictable scale
• ✓MLPerf-validated systems for efficiency and throughput benchmarks
• ✓Sovereign deployment options and data residency across Singapore and Australia
• ✓Real-time telemetry for performance, energy, and thermal monitoring
• ✓BlueField DPUs for I/O offload and isolation
• ✓Support for LLM training, LoRA fine-tuning, vision pretraining, and inference serving

Use cases

• ✓Training and fine-tuning large language models with predictable throughput
• ✓Consolidating fragmented GPU estates for enterprises with multi-site operations
• ✓Running high-performance inference serving with low and predictable latency

Frequently asked questions about Firmus