COMPUTEX 2026 Edge AI Hardware Roundup: From Palm-Sized Supercomputers to Five-Layer AI Ecosystems

Over 6,000 exhibitors descended on the Taipei Nangang Exhibition Center from June 2 to 5, 2026, for COMPUTEX 2026. The official theme, "AI Together," was less a marketing slogan and more a precise description of a fundamental industry shift playing out across the show floor: the migration of powerful AI workloads from centralized cloud data centers to the edge . While major platform keynotes from Intel and NVIDIA set the stage, the most tangible evidence of this trend came from four companies—MSI, Innodisk, Longsys, and Karrie—each attacking a different layer of the edge AI hardware stack.

MSI EdgeMesa N AI+: A Petaflop in Your Palm

The most visually striking embodiment of local AI was the MSI EdgeMesa N AI+ mini PC. Announced as one of the first devices built on NVIDIA's new RTX Spark superchip, this compact system packs a 20-core Grace CPU, a Blackwell GPU with 6,144 CUDA cores, and up to 128 GB of unified LPDDR5X memory into a palm-sized chassis . The result is roughly 1 petaflop of AI performance, a figure that until recently was the domain of data center racks, not a device you can place on a desk .

MSI designed the EdgeMesa N AI+ as a workstation for AI developers, data scientists, and creators who need to run large language models (LLMs), real-time inference, and generative AI without the latency and privacy trade-offs of cloud services . A 10GbE LAN port and multi-display support underscore its role as a self-contained productivity node, not a toy . MSI demonstrated working applications in healthcare, retail, finance, robotics, and smart-city management at its booth .

Innodisk: A Five-Layer Blueprint for Enterprise Edge AI

If MSI showcased a product, Innodisk showcased a system. The company's five-layer edge AI ecosystem spans compute, memory, storage, sensing and communication, and software—the essential architectural layers any enterprise needs to deploy AI outside a data center .

Live demonstrations brought the concept into sharp focus. Visitors saw on-premises LLM fine-tuning running on the AccelTune platform (a no-code tool hosted on the APEX-S100 server with Intel Xeon 6700-series processors and dual NVIDIA RTX PRO GPUs), heavy machinery intelligent security systems processing sensor data locally, autonomous mobile robots (AMRs) navigating factory floors, and OCR container recognition systems that never phoned home to a cloud . The core pitch was data sovereignty: enterprises can keep sensitive information within their own infrastructure while still running modern AI models .

Longsys: AI-Native Memory for the Edge

Longsys (江波龙) addressed a less glamorous but critical bottleneck: memory designed for AI. The company debuted two purpose-built storage modules targeting the specific demands of edge inference .

The AIDIMM is a socketed module built on LPDDR5X with a 256-bit wide bus, running at 9600 MT/s. A single module supports up to 128 GB of capacity with 307.2 GB/s of bandwidth, which Longsys says is enough to run a 70-billion-parameter LLM locally without choking . Its four DRAM chips are arranged in a same-side layout compatible with mainstream AI host motherboards, with a tool-less high-pin-count connector .

For more compact, embedded scenarios, the AILPBGA is a soldered BGA1764 package at 22mm × 22mm, offering capacities from 24 GB to 64 GB while maintaining the same 256-bit base architecture . Both modules implement dynamic voltage scaling (0.9V-1.05V) and an intelligent efficiency mechanism called FDVFS .

Beyond the modules, Longsys also revealed a deeper architectural play: its Storage Processing Unit (SPU) and Intelligence Storage Agent (iSA). The SPU is a dedicated processing chip fabricated on a 5nm process, purpose-built to offload storage intelligence from the main CPU, with a maximum single-drive capacity of 128 TB. The iSA is the scheduling software that forms a closed-loop hardware-software system for edge AI storage . Under its consumer brand Lexar, Longsys framed its entire portfolio around the theme "Edge AI Storage • Integrated Implementation" .

Karrie: The Invisible Infrastructure of AI Factories

Karrie (嘉利) operates a layer beneath the end-user devices, but its presence at COMPUTEX 2026 was no less significant. The global server mechanical engineering provider demonstrated AI server chassis, rack-level structural solutions, and precision manufacturing capabilities that physically support the next generation of AI infrastructure .

Karrie's work was visible across multiple partner booths at the event. The company highlighted its inclusion on NVIDIA's approved vendor list for server chassis and rack components—a designation earned in September 2025—and showcased products built on the NVIDIA MGX modular reference architecture, spanning 1U, 2U, and 6U form factors . Its official statement, "We are honored to support NVIDIA MGX and be an important part in shaping the next-generation AI Factory ecosystem," captured the company's strategic positioning as a behind-the-scenes enabler of the AI buildout .

The Bigger Picture: Physical AI, Robotics, and the Local-First Wave

The four companies above operated within a broader narrative that dominated COMPUTEX 2026. Physical AI and robotics were everywhere. ADLINK demonstrated unified hardware deployment matrices spanning controllers, smart displays, and robotics . AAEON showcased edge Physical AI for humanoid robotics, smart transportation, and healthcare . Inventec turned its booth into a mini "AI Factory," demonstrating smart manufacturing and robotics applications .

NVIDIA's ecosystem anchored much of the hardware: RTX Spark powered MSI's mini PC; MGX provided the blueprint for Karrie's chassis; and the Jetson platform drove demos across AAEON, ADLINK, and other robotics exhibitors . Simultaneously, Intel emphasized rack-scale AI infrastructure and a new disaggregated inference model called "neocloud" aimed at cloud-to-edge scalability .

The consistent, cross-vendor signal was that AI inference is moving out of the cloud. Latency-sensitive industrial applications, privacy regulations, and the sheer cost of continuous cloud inference are pushing enterprises to run models locally . COMPUTEX 2026 wasn't a show about AI possibilities; it was an exhibition of AI hardware that companies can buy, install, and run today.