# Intel Vietnam Country Manager: "Vietnamese businesses are moving from Edge AI pilots to real deployments"

- Original: https://migovi.com/en/intel-viet-nam-edge-ai-phong-van/
- Author: [Hero Association](https://migovi.com/en/author/migovi/)
- Published: 2026-11-18T03:35:55+07:00
- Categories: AI, Business

---

Edge AI, which means running AI models on the device itself instead of sending data to the cloud, is leaving the lab and entering factories, stores, and urban infrastructure. At Computex 2026, Intel said its Series 3 processors have already won more than 130 design projects for edge devices. Migovi sent six questions to Phùng Việt Thắng, Country Manager of Intel Vietnam, about where domestic businesses should start.

Key takeaways in 60 seconds

1. Vietnamese businesses are moving from experimenting to deploying Edge AI in the real world
2. Main benefits: lower bandwidth and storage costs, lower latency, data kept on-premises
3. There are 4 obstacles: legacy system integration, training data, model optimization, and heterogeneous hardware
4. Intel recommends splitting workloads across CPU, GPU, and NPU on a "right job, right engine" principle
5. OpenVINO lets developers train once and deploy across many types of hardware

Migovi  Intel says Edge AI is entering a new phase with Agentic AI and Physical AI. Where do Vietnamese businesses in manufacturing, retail, and urban infrastructure stand in this process? What are the opportunities and technical obstacles in moving AI processing from the cloud to the edge?

Phùng Việt Thắng Vietnamese businesses are shifting their focus from experimentation to real-world Edge AI deployment, and the most immediate opportunity is cost. Picture a component factory with 500 4K IP cameras monitoring the production line. If it streams raw video to the cloud around the clock, it pays tens of thousands of USD a month for bandwidth and storage. When inference runs on the edge device itself using the OpenVINO toolkit, the camera only sends data when it detects a defect.

Next is latency. Processing on-site allows real-time responses, which is essential for quality monitoring, defect detection on the line, and smart traffic. Sensitive data from manufacturing, healthcare, and infrastructure doesn't have to leave the premises, which makes security requirements easier to meet. The system also keeps running reliably when the connection drops, a common situation in industrial environments or places where telecom infrastructure is uneven.

There are four groups of obstacles. The first is integration, because most companies run on many different platforms. The second is training data: real production environments differ greatly from what was planned for, and collecting consistent data from thousands of sensors is still a hard problem. The third is that models must be heavily optimized before they can run on power-efficient edge hardware. Finally, devices, operating systems, and connectivity protocols are not uniform, which makes it harder to secure and manage AI applications at scale.

![Phung Viet Thang of Intel Vietnam presenting on Edge AI and Physical AI for businesses](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-6.jpg)

Migovi  Intel's edge hardware spans everything from mainstream to high-performance tiers, combining CPU, GPU, and NPU. In terms of investment cost and performance per watt, how should businesses allocate AI workloads across Edge PCs and Edge Servers?

Phùng Việt Thắng The principle we recommend is the right job for the right engine. The CPU orchestrates sequential data, handles latency-sensitive tasks, and runs traditional machine learning models. The GPU has large parallel processing capacity, so it takes on heavy AI work such as multi-stream video analytics or large language models (LLMs). The NPU suits tasks that run continuously in the background, such as voice and face recognition. Moving these to the NPU frees up the CPU and GPU and also lowers the power draw of the whole system.

The goal is to use just enough resources. If the AI capability built into the processor, such as Core Ultra Series 3, already meets the application's needs, a business may not need an additional discrete graphics accelerator. That makes the design simpler, cuts power consumption, and lowers total cost of ownership. For heavy AI workloads, a discrete accelerator may still be necessary.

Ella from Sensory AI is one example. On a single Core Ultra Series 3 chip, the system simultaneously recognizes images, talks with customers, analyzes business data, runs multiple AI agents, and controls a robot in real time. Tasks are divided among the CPU, GPU, and NPU through OpenVINO instead of requiring separate chips for AI and control. That's also why we advise Vietnamese software engineers to use OpenVINO, so workloads are allocated automatically and code doesn't need to be rewritten over and over.

> Right job, right engine: the goal is to use just enough resources.

*Phùng Việt Thắng*

Migovi  Many businesses are wary of depending on a closed architecture when investing in AI. Given Intel's contributions to the Linux kernel and the OpenVINO toolkit, how does Intel help them move AI models between hardware platforms without rewriting code from scratch?

Phùng Việt Thắng When investing in AI, businesses care most about three things: risk control, long-term total cost of ownership, and the ability to carry a project forward. Dependence on a single closed architecture is making many organizations hesitant to scale up their Edge AI systems.

With OpenVINO, a business trains a model once on any framework and then deploys it on many kinds of devices just by changing a setting, with no code changes. It can also run several types of hardware in parallel within the same system. Software development costs drop and deployment time gets shorter. More importantly, the business isn't locked into one vendor, so it is free to negotiate prices and choose hardware for each stage.

On Linux, Intel has been the largest code contributor to the Linux kernel for more than 15 consecutive years, nearly double the second-place vendor. As a result, hardware from Xeon to Core Ultra is fully supported on popular Linux distributions such as Ubuntu LTS and Red Hat Enterprise Linux. Businesses can upgrade their infrastructure without worrying about compatibility, and can scale from edge to cloud or the other way around thanks to cloud and container platforms that are already integrated.

![Phùng Việt Thắng](https://migovi.com/wp-content/uploads/2026/11/thang-vncm-2-1200x800.jpg)

*Businesses aren't locked into a single vendor, so they're free to negotiate prices and choose hardware for each stage. Phùng Việt Thắng Intel Vietnam Country Manager*

Migovi  Real-time processing combined with AI vision is critical for factories going through digital transformation. How do the Robotics Reference Platform and Robotics AI Suite address latency, bandwidth, and reliability for cobots and AGV/AMR autonomous vehicles in factories and warehouses in Vietnam?

Phùng Việt Thắng These two platforms bring AI inference directly onto the device to handle three problems: latency, bandwidth, and reliability.

On latency, inference capability sits right on the robot's processor. Sensors and AI vision fuse their data to make instant decisions. A cobot stops immediately when it detects someone stepping into a restricted zone. An AMR can run at higher speeds to increase the number of goods-transfer cycles.

On bandwidth, the robot processes images on-site and sends only tiny packets back to the management system, such as "At shelf 3" or "Obstacle detected." The network infrastructure carries a lighter load, and scaling up the fleet becomes easier.

On reliability, the power-efficient NPU in Core Ultra Series 3 lets each AMR operate fully on its own, without interruption when the network drops or it enters a wireless dead zone. Robots can therefore work longer and need to charge less often.

[![Cách Intel cùng đối tác đưa Edge AI, Physical AI vào hạ tầng Việt Nam](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-12-768x432.jpg)](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-12.jpg)

[![Cách Intel cùng đối tác đưa Edge AI, Physical AI vào hạ tầng Việt Nam](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-14-768x432.jpg)](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-14.jpg)

[![Cách Intel cùng đối tác đưa Edge AI, Physical AI vào hạ tầng Việt Nam](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-13-768x432.jpg)](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-13.jpg)

[![Cách Intel cùng đối tác đưa Edge AI, Physical AI vào hạ tầng Việt Nam](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-3-768x512.jpg)](https://migovi.com/wp-content/uploads/2026/08/intel-edge-ai-physical-ai-doanh-nghiep-migovi-3.jpg)

Migovi  Edge AI systems often have to run in harsh environments: high temperatures, constantly moving vehicles, and fanless machines with little room for heat dissipation. How does Intel optimize performance per watt to run heavy AI vision workloads while staying reliable 24/7?

Phùng Việt Thắng Core Ultra Series 3 includes versions designed specifically for industrial Edge AI, targeting high performance per watt and 24/7 operation. For AI vision, the architecture combining CPU, GPU, and NPU splits the work up instead of piling everything onto one processor. The system can therefore run algorithms at low power and sustain a continuous workload.

Software plays an important part too: OpenVINO optimizes models through quantization, which makes algorithms leaner while preserving accuracy. Businesses can choose compact, low-power Edge PC models that still meet their needs.

Migovi  Processing AI at the edge helps address privacy concerns. Which hardware security technologies does Intel build in so that a business's visual data and AI models can't be tampered with or leaked right at the point of collection?

Phùng Việt Thắng Visual data from surveillance cameras and product quality inspection, along with AI models such as defect-detection algorithms, are strategic assets for a business. We protect them with hardware security technologies, depending on the scale of the device.

On high-performance Edge Servers using Xeon, Confidential Computing solutions, including Intel SGX and Intel TDX, create isolated encrypted partitions. Data and algorithms stay protected even while the system is processing them.

On Edge AI devices using Core Ultra, a hardware-based security layer is built into the silicon. A secure boot mechanism prevents spoofed firmware. Physical memory is encrypted to protect models and data, even if someone tampers directly with the device. Businesses keep their intellectual property and meet privacy regulations without trading away performance.

**MIGOVI VIEWPOINT**

These answers come from a hardware vendor's perspective, so businesses should verify them with a pilot project using their own data. OpenVINO is open source, but it is optimized first and foremost for Intel hardware, and its engineers are actively working to improve compatibility with other vendors. The "no vendor lock-in" claim should also be tested on the exact devices a business plans to buy or use. The most useful takeaway is the principle of using just enough resources: many machine vision problems on factory floors don't need the power of a discrete GPU at all.
