GMI Cloud AI-Native GPU Cloud and Inference Engine

GMI Cloud is an AI-native cloud provider headquartered in Mountain View, California, founded and led by CEO Alex Yeh, with more than 120 staff and a strong Asia-Pacific footprint. It supplies GPU capacity and inference services to AI companies and enterprises that cannot easily secure high-end NVIDIA hardware.

The platform offers dedicated bare-metal GPU instances with root access, a cluster engine that orchestrates multi-node training clusters, and a serverless inference engine with automatic scaling, request batching, scale-to-zero and production APIs for LLM and multimodal models. Newer products include GMI Router and the Agentbox agent runtime. It is an NVIDIA Preferred Partner built on NVIDIA reference architectures, with Tier-4 data centers in Silicon Valley, Colorado, Taiwan, Thailand and Malaysia. Customers include Fireworks, Higgsfield, HeyGen, OpenRouter, Reflection and Trend Micro.

• Bare-metal GPU compute: dedicated H100 and H200 servers with root access and custom stacks.
• Cluster engine: orchestrates multi-node clusters for large-scale training.
• Inference engine: serverless model serving with autoscaling, batching and scale-to-zero.
• Model APIs: production endpoints for LLM and multimodal models.
• GMI Router: optimizes performance and token consumption across models.
• Agentbox: runtime infrastructure for AI agents.

Vietnamese AI startups can rent H100 or H200 clusters in nearby regions to train Vietnamese language or video generation models. Software firms can serve open models through the serverless inference engine without managing GPUs. Enterprises can reserve dedicated capacity for private model training.

Published on-demand prices of 2.60 USD per GPU-hour for H100 and 3.20 USD for H200 give predictable cost. The company reports 3.7x higher throughput, 5.1x faster inference and 30% lower cost on production traffic. Customer Higgsfield cut p95 inference latency by 65% and compute cost by 45%. Data centers in Thailand, Malaysia and Taiwan provide capacity close to Southeast Asian users.

Shortages of high-end GPUs and rising inference demand are fueling specialized AI clouds and national AI factories across Asia. Southeast Asian enterprises and AI startups want regional capacity with lower latency and data locality, and providers with facilities in Thailand and Malaysia are positioned to serve them.

Headquartered at 278 Castro St, Mountain View, California, with 120+ staff globally as of May 2025. Raised an 82 million USD Series A in October 2024 led by Headline Asia with Banpu Next and Wistron, and 668 million USD in September 2026 led by ARCHIV with NVIDIA participation. Announced a Taiwan AI Factory in 2025 and a 1 GW sovereign AI project in Japan in March 2026.