MISA builds an "AI factory" with the NVIDIA platform to automate businesses in Vietnam

MISA, a longtime software enterprise in Vietnam, is building an on-premise "AI factory" based on NVIDIA infrastructure to bring general AI and AI agents deep into accounting, tax, sales, human resources and public administration processes. This is also a practical example of the "5-layer cake" model that CEO Jensen Huang describes, where accelerator hardware, networks, systems, software and AI models are designed as a unified architecture, serving the entire lifecycle from development to deployment of AI in the enterprise.

According to NVIDIA, MISA currently provides SaaS and AaaS services to more than 450,000 government and corporate customers, and 3.5 million individual users. Previously, the company relied heavily on external LLM APIs with high operating costs, high latency and little control over data security, so it developed its own models MISA AI 1.0, MISA AI 1.0 Plus and MISA AI 2.0, and brought "Core AI" to its domestic self-built infrastructure to optimize performance, cost and compliance.

In the two lowest layers of the "5-layer cake" of energy and chips, MISA uses a cluster of 8 NVIDIA DGX H200 systems (a total of 64 H200 GPUs) connected by NVIDIA Quantum InfiniBand network infrastructure. This cluster acts as the "engine" for the AI ​​factory, responsible for training and inferring models from 8 billion to 235 billion parameters, while the NVIDIA A40 GPU takes on specialized workloads such as eKYC or LLM routers. The "infrastructure" layer in the 5-layer architecture described by NVIDIA is the place where land, electricity, cooling, networks and systems gather to operate tens of thousands of processing cores as a single machine, not for data storage but for "intelligence production". MISA deploys AI factory in exactly that spirit, bringing together the DGX H200 infrastructure, InfiniBand network and coordination layer into a unified architecture, in which NVIDIA Run:ai helps virtualize and share GPUs in a "fractional" fashion, allowing to increase the number of parallel AI experiments from 1–2 to 10–20 per week.

The top layer is “applications,” which according to Jensen Huang is where economic value is created, from drug discovery platforms to industrial robots to autonomous vehicles. With MISA, this layer is realized with more than 50 pre-built AI agents and more than 500 customizable agent templates for many businesses: from retail, service to B2B consulting. Some typical agents are MISA AVA (internal search assistant), MISA AMIS OneAI (expert question-and-answer platform), MISA AMIS Support (24/7 support), MISA AVA Mail (email management) and MISA AVA Recruitment (recruitment automation). In finance and accounting, MISA AVA automates tasks such as invoice checking, reconciliation and data entry, helping to create financial reports in minutes instead of hours and significantly reducing human errors. In sales and operations, the system supports answering product information, prices, call summaries, composing emails and extracting customer data, so that the sales team focuses more on interaction and revenue growth. With recruitment - human resources, the entire process from receiving CVs, screening, evaluating to extracting data is automated in a few minutes, instead of needing five HR personnel to manually process hundreds to thousands of records. At the executive level, MISA AVA aggregates financial, sales, and human resources data into real-time reports, analysis, and forecasts, helping leaders query in natural language and immediately receive comparisons, trend analysis, and predictive suggestions. According to NVIDIA, the full-stack architecture including DGX H200 cluster, NVIDIA AI Enterprise and Run:ai helps MISA shorten model fine-tuning time by 4 times and save about 200,000 USD in operating costs per month. The system currently handles about 2.5 million feature uses per day and more than 1.1 million LLM requests, serving more than 450,000 organizations nationwide, and also appearing in public services, such as AI robots supporting people in looking up information at the Public Administration Center. MISA intends to develop AI factory into a comprehensive AI agent ecosystem serving both internal operations and external customers, while mastering the entire LLM stack from pre-training to model optimization to ensure long-term technological autonomy. The company's roadmap also extends beyond text to images, voice and 3D, in line with the vision of "AI is a 5-layer infrastructure" that Jensen Huang describes: from energy, chips, infrastructure, models to applications, all pulling together to build an "intelligence factory" for the AI ​​era.