In August 2026, JH Semiconductor's product center underwent a structural expansion: two new product lines — AI compute servers and Junhangclaw AI inference workstations — went live at /products/ai-server and /products/ai-workstation. The site's product portfolio now extends beyond the enterprise SSD guide and the server memory line to a full map spanning compute systems, desktop inference, and storage plus memory.
AI Compute Servers: Five Platform Classes, from Standard Server Rooms to Liquid-Cooled Clusters
The new AI server line covers edge inference, enterprise inference, training clusters, and high-density nodes, with delivery options spanning air cooling, liquid cooling, rack-scale liquid-cooled clusters, and full-system delivery:
- 8-GPU RTX 6000D platform: eight dual-width GPUs (up to 600W each), 11 PCIe 5.0 slots, 12 front LFF drive bays, separate CPU/GPU air ducts, 4U/5U rackmount — aimed at large-model inference, AIGC, and vision computing;
- 10-GPU Thor T5000 platform: up to ten dual-die Thor T5000 cards, ten PCIe 5.0 x16 double-width slots, four CRPS power supplies in 2+2/3+1 redundancy — high-density dual-socket inference emphasizing GPU density and low latency;
- 8-GPU air-cooled platform: Intel/AMD platforms, eight dual/quad-width blower cards (up to 600W each), 12 PCIe 5.0 slots and 12 LFF bays — rapid deployment in standard server rooms without liquid-cooling retrofits;
- 8-10 GPU cold-plate liquid-cooled platform: full CPU/GPU cold-plate coverage, connects to CDUs and facility water loops with cluster-level supply/return monitoring — built for high power density, low PUE, and scale-out clusters;
- PC-Farm 4U4H/6U6H: four or six hot-swappable PC nodes with fully independent power control per node, 750W peak per node at the backplane, and four CRPS 1100W redundant power supplies.
Junhangclaw Workstations: Bringing Inference Back to the Desktop
Opposite the machine room, the Junhangclaw line addresses the other end of the problem: local AI inference with private data kept on-premises. The four tiers are the JH-W50 (domestic chip, 50 TOPS, for individual users), JH-W120 (120 TOPS, for small teams and lightweight model deployment), JH-W280 (MetaX domestic chip, 280 TOPS INT8, for government and enterprise xinchuang scenarios), and the flagship JH-W2000 — NVIDIA Blackwell GPU, 2,070 TFLOPS (FP4 sparse), 128GB LPDDR5x unified memory, a 140mm-radiator liquid cooler, in a chassis of just 199.3×199.3×100.3mm; the extended W2000N adds four 2.5-inch drive bays. With both lines in place, the company's compute spectrum now runs from 50 TOPS on a personal desktop to rack-scale training clusters.
How the New Lines Fit the Existing Business: One-Stop Compute-Plus-Storage Supply
Since its founding in 2012, the company has run two core businesses: in-house enterprise storage development and server/GPU/memory supply-chain services. Until now, the website's product story centered on storage: a self-developed enterprise SSD portfolio planned across 10 series and 30 models (currently in product definition and engineering validation), and server memory spanning DDR5 RDIMM 4800/6400 and 3DS ECC RDIMM 256GB up to CXL 2.0 memory expansion. With the two compute lines live, an AI server's GPU and memory configuration can be completed within a single supply system; the in-house enterprise SSDs remain in product definition and engineering validation, but series such as TITAN 9 (for AI servers and cloud platforms) and X9 (for KV Cache) were aligned with these system-level scenarios from the planning stage, reserving the storage slot in the full-system offering for when they arrive.
Both pages are now open for browsing; selection specs and solution materials will be updated as the product lines evolve.