AI 'leaves the cloud' and heads to the desk.

The emergence of miniaturized computing systems like the Nvidia DGX Spark at Computex 2026 is concretizing the trend of operating large-scale AI models at the edge without relying on a central server system.

At the global technology event Computex 2026 held in Taipei, Taiwan, the focus of discussions in the semiconductor and computing industries shifted significantly from "Cloud AI" to "Edge AI." The hardware industry is undergoing a restructuring as data center-level computing power is integrated into personal devices placed directly in the workspace.

The shift from responsive AI to autonomous agents (Agent AI)

In the early stages of the artificial intelligence wave, the common operational process involved users submitting data requests to cloud servers such as OpenAI, Google, or Microsoft and receiving responses. However, this architecture revealed many limitations regarding transmission latency, bandwidth costs, and the security of the source data.

Nvidia DGX Spark is a line of personal computers specifically designed for AI and will be distributed in Vietnam.

Photo: Anh Quân

The development of Agentic AI—a generation of autonomous software agents capable of planning, reasoning, and interacting directly with local file systems—is placing new demands on hardware infrastructure. Instead of passively responding, these agents act as digital human resources, processing a continuous stream of information in real time. To ensure data integrity and security, bringing AI models to operate offline on users' devices has become an essential technical solution.

A prime example of this trend is the DGX Spark AI personal computer, introduced at Computex 2026. The device boasts a compact desktop design but delivers the performance of a miniature supercomputing system thanks to its single Nvidia GB10 Grace Blackwell Superchip.

You may also like

Leveraging the power to push Vietnam's digital economy to reach 30% of GDP.Amazon Web Services (AWS) has just launched Local Zone in Hanoi, as Vietnam aims for the digital economy to contribute 30% of its GDP by 2030.

MyTV Sport Hub: An "AI assistant" for World Cup 2026 fans.For football fans, the World Cup is never just about 90-minute matches. The excitement often begins hours before kick-off, as viewers eagerly research the starting lineups, the form of the stars, and predict their favorite team's chances of progressing. And after the final whistle, that journey continues with discussions about tactics, memorable moments, and fascinating side stories.

Renova Cloud has achieved AWS Premier Tier status in the Amazon Web Services partner network.Renova Cloud, a leading strategic partner in cloud computing consulting services in Southeast Asia, has become an AWS Premier Tier Services Partner – the highest rank in the Amazon Web Services Partner Network (AWS Partner Network – APN).

The device's independent operation relies on a 128GB LPDDR5X Unified Memory system with high-speed bandwidth. In AI architecture, memory capacity and speed determine the ability to process large language models (LLMs). This allows data engineers to directly run models with up to 200 billion parameters on the device itself, rather than deploying them on cloud servers.

In terms of specifications, the Blackwell architecture GPU integrates 5th generation Tensor cores (FP4 precision format) providing 1 petaFLOP of computing power. The 20-core ARM CPU is responsible for coordinating data between the local file system and the AI ​​model.

The workstations serving the AI ​​needs at the enterprise edge now come in compact sizes, making them easy to deploy at various scales.

Photo: Anh Quân

At the exhibition booths, infrastructure solutions for this trend were clearly differentiated through synchronized systems from original manufacturers and specialized hardware integration solution providers. A prime example is Leadtek, showcasing a range of workstations and servers from its Nvidia-Certified Systems. Targeting the on-premises (internal) operational needs of small and medium-sized enterprises, the WinFast WS950 AI workstation supports multi-GPU configurations with two professional Nvidia RTX PRO 6000 Blackwell Workstation Edition graphics cards, providing a total of up to 192 GB of GDDR7 GPU memory. On a larger scale, their WinFast GS5855T server system allows for the integration of up to eight RTX PRO Blackwell architecture GPUs to meet the demands of intensive AI inference and training tasks.

Optimizing security and operating costs.

Operating AI at the edge through a local hardware system addresses three core challenges of today's technology infrastructure. First is data security. All business information, internal source code, and personal data are stored and processed in a sandbox environment isolated from the internet, limiting the risk of data leakage to third parties.

United States - Vietnam: Don't miss

Strengthening the friendship between Vietnam and the United States.On July 3rd, as part of the Pacific Partnership - Friends of the Pacific 2026 program, the U.S. Army Pacific delegation, led by Lieutenant General Joel Vowell, Deputy Commander of the U.S. Army Pacific, paid a courtesy visit to the Quang Tri Provincial Military Command.

The U.S. Army Pacific delegation visits the Quang Tri Provincial Military Command.2026 marks the first time the Pacific Partnership-Friends of the Pacific program will be jointly organized in Quang Tri, creating a large-scale cooperation framework with 25 activities.

The Ministry of Foreign Affairs received a copy of the Letter of Credence for the U.S. Ambassador to Vietnam.On the afternoon of July 2nd, at the Ministry of Foreign Affairs headquarters, Mr. Le Cong Dung, Director of the State Protocol and Foreign Interpretation Department, received a copy of the Letter of Credence from Ms. Jennifer Wicks, Ambassador of the United States of America to Vietnam.

New Edge AI solutions showcased at Computex 2026

Next is the issue of fixed computing costs. Renting cloud infrastructure, which is charged based on token amounts, incurs significant variable costs as it scales. Operating on offline hardware transforms these costs into a fixed asset investment, optimizing long-term operations. Finally, there's the issue of local scalability: Through high-speed connectivity protocols, users can link edge computing systems to share resources, scaling edge modeling processing capabilities to massive sizes.

Source: https://thanhnien.vn/ai-roi-dam-may-ve-ban-lam-viec-185260605224532968.htm