GMI Cloud’s AI Factory datacenter in Taiwan will use VAST Data storage to feed data to 7,000 Blackwell GPUs.
The forthcoming datacenter is a $500 million investment by Silicon Valley neocloud startup GMI Cloud, in collaboration with Nvidia. The GPUs will be accompanied by NVLink, InfiniBand, and Spectrum-X Ethernet networking. They will be housed in 96 racks, deliver around 2 million tokens per second of performance, and draw 16 megawatts. GMI Cloud was founded in 2021, originally as a Bitcoin miner, and pivoted to pursue the AI datacenter opportunity, following CoreWeave. It raised $82 million ($15 million in equity and $67 million in debt) in a Series A funding round a year ago. Altogether, it’s raised over $93 million in total funding across three rounds. GMI Cloud is the only Nvidia-certified cloud provider in Taiwan under the GPU giant’s NCP/NPN program.
GMI Cloud founder and CEO Alex Yeh said: “This datacenter will become the heart of Asia’s AI future. With thousands of next-generation Nvidia GPUs running in synchrony, we intend for this AI Factory to turn the world’s most ambitious AI visions into reality.
“We are excited to unlock new, heightened AI ambitions for VAST customers with our instant, efficient GPU cloud. VAST’s data platform offers impressive data management capabilities, and we look forward to growing with them.”
VAST’s exabyte-scale data infrastructure, with its support for Nvidia’s Magnum IO GPU Direct, native NFS and object capabilities, will be used for model training, inference, and real-time data processing. GMI Cloud says customers “can manage and scale AI workloads seamlessly with GMI Cloud’s deep Kubernetes integration, from control plane to management APIs.”
VAST co-founder Jeff Denworth said: “This is where the world stops experimenting with AI and starts industrializing it. What Nvidia, VAST Data, and GMI Cloud are building in Taiwan represents the new blueprint for AI infrastructure, where performance, scalability, security, and simplicity converge.”
VAST Data first started supplying GMI Cloud in December last year. Back then, Andy Chen, GMI’s VP of Global Business and Product, said: “We use VAST Data as the shared storage inside a GPU cloud service. Customers can utilize this, either for shared storage or file storage. VAST Data can cover object… VAST actually provides the best solution in the market.”
Chen said GMI uses OpenStack, which VAST supports. GMI tested different vendors, including two major ones, and VAST came out on top. One of them couldn’t support the necessary multi-tenancy; VAST could.
Trend Micro, Wistron (a GMI Cloud investor), Chunghwa System Integration (CSI), and TECO are also supplying technology as part of GMI Cloud’s Taiwan AI Factory business collaboration and partnerships:
- Trend Micro digital twin technology to simulate cyber threats and validate defenses without exposing production systems to risk. Customers will be able to continuously stress-test their infrastructure to learn and adapt in real time, helping to keep defences up to date against evolving malware threats.
- Wistron is building AI-enabled factory systems for computer vision, predictive maintenance, and digital-twin simulation. It can train and deploy models directly on live production lines, reducing downtime and helping smart manufacturing development.
- Chunghwa System Integration (CSI) will provide – AI-enabled integration systems and services, including 5G and AI-enabled IoT tech, datacenter operation and maintenance.
- TECO Electric & Machinery provides the power and system integration backbone for the AI factory with its Energy-as-a-Service offering.
GMI Cloud has datacenters in Taiwan, Thailand, and Malaysia, plus one in Colorado.
Bootnote
VAST Data benchmark performance tests with its Event Broker showed it can handle 136 million streamed messages per second, at 99 percent efficiency across 88 nodes. It says that its Event Broker can sustain 604 percent more throughput than Apache Kafka and one Event Broker node is 156 percent faster than Redpanda. Read a lot more detailed information about this in a VAST Data blog.
