English
Computex 2026 officially kicked off in Taipei, where NVIDIA CEO Jensen Huang delivered a keynote covering the company’s complete hardware and software roadmap. Most datacenter-focused content centered on mass production of Vera Rubin GPUs and AI agent-driven computing trends, which we skip for conciseness. The most noteworthy part tailored for consumer users was a brand-new Windows PC ecosystem co-launched with Microsoft, marking NVIDIA’s largest client hardware update this year. Besides PC products, the revised LPX rack for data centers also attracted our attention with two clear hardware tweaks.
Updated LPX Rack
NVIDIA’s LPX rack has received two visible upgrades: the chassis size was upgraded from 1OU to 2OU, and the four built-in Groq C2C high-speed interfaces adopted a redesigned physical layout. No other exterior changes can be observed, leaving the function of the newly added second internal layer unconfirmed and worthy of further follow-up.
Three-Tier RTX Spark PC Product Family
NVIDIA rolled out a complete three-tier PC lineup including laptops, small-form-factor desktops and professional workstations. All devices are built on Spark and Blackwell architectures and fully optimized for Windows systems. Jensen Huang described this lineup as the first fully reimagined PC product family in 40 years, covering mainstream consumer gaming, content creation and high-end professional workstation scenarios respectively.
RTX Spark Laptops
Co-developed with MediaTek, the RTX Spark superchip (codename N1X) is manufactured on TSMC 3nm process with 70 billion transistors via dual-chiplet packaging. Its integrated Blackwell GPU carries 6,144 CUDA cores and fifth-generation Tensor Cores, delivering 1 FP4 petaflop local AI performance. It pairs with a 20-core Grace CPU, connected via 600GB/s NVLink-C2C interconnect — five times faster than PCIe 5.0 while consuming far less power. The chip supports up to 128GB unified memory with 300GB/s memory bandwidth.
This superchip shares almost identical core silicon with the data-center-oriented DGX Spark, but targets Windows rather than Linux systems, with complete RTX gaming and creative software stacks adapted for consumer scenarios. NVIDIA claims its overall performance is comparable to mainstream RTX 5070 laptop discrete GPUs, while its unified memory architecture brings obvious performance differences compared to traditional discrete graphics cards working through PCIe buses.
Different from highly unified DGX Spark servers with identical reference designs, RTX Spark laptops will see obvious performance gaps among different OEM brands. The product lineup covers ultra-thin 14mm lightweight laptops and standard gaming laptops, with configurable power ranges from single-digit watts up to 80W. Diversified cooling solutions and fan curve tuning from different manufacturers will lead to distinct sustained performance, even for devices equipped with the exact same chip. NVIDIA will launch multiple Spark SKUs with memory ranging from 16GB to 128GB. The first batch includes over 30 laptops and more than 10 desktops from mainstream vendors such as Dell, HP and Microsoft Surface.
NVIDIA made two practical hardware adjustments for consumer laptops: removing onboard ConnectX-7 200GbE networking module to save power and board space for faster NVMe storage, and allowing OEMs to customize I/O ports instead of unified fixed interfaces. For battery performance, NVIDIA promises all-day battery life for daily office work and web browsing, while heavy gaming and full-load local AI inference will drain the battery within 45 to 60 minutes. We will launch targeted professional benchmarks to test real battery consumption under actual AI workloads, instead of only referring to official marketing data.
RTX Spark Desktops
Compact RTX Spark desktops will launch this fall alongside new laptops, and MSI showcased its prototype model during the keynote. The overall design is similar to DGX Spark but adopts a revised industrial appearance. Two key questions remain to be verified after actual testing: whether desktop models deliver much better sustained performance thanks to sufficient thermal headroom compared with thin laptops, and whether they cancel high-speed CX7 networking modules just like laptops to lower consumer retail prices.
DGX Station for Windows
As the top-tier workstation in the new PC family, DGX Station for Windows adopts GB300 Grace Blackwell Ultra desktop superchip, combining a high-performance Blackwell Ultra GPU and a 72-core Grace CPU. It supports up to 748GB coherent unified memory and 20 FP4 petaflops AI computing power, capable of running trillion-parameter large language models and hundreds of AI agents locally. It retains 800Gb/s ConnectX-8 SuperNIC for high-speed networking, and supports optional RTX PRO 6000 professional workstation discrete GPU.
Core Concern: Microsoft’s Windows on Arm Adaptation
NVIDIA’s new PC hardware has been fully verified by DGX Spark in terms of silicon stability and performance, yet we still hold concerns about Microsoft’s software adaptation capability. Recent Windows updates have triggered widespread user complaints, including forced Copilot integration across the whole system, mandatory OneDrive bundling, and immature optimization for Windows on Arm devices. NVIDIA solves compatibility issues via Prism emulator optimization and cooperation with global anti-cheat vendors, but similar optimization attempts in the past failed to achieve ideal results.
If emulation and driver adaptation work perfectly, RTX Spark integrated chips can become power-efficient and cost-effective alternatives to traditional discrete GPU gaming laptops. Besides, the full Linux compatibility inherited from DGX Spark may boost the popularity of Linux desktop systems if Windows fails to give full play to the hardware performance. Furthermore, launching Windows-version DGX Station is controversial: the B300 chip has no dedicated ray tracing cores, and this product misses the best early iteration window for cutting-edge AI hardware development.
Long-Term Product Roadmap & Physical AI Layout
NVIDIA confirmed that this three-tier RTX Spark PC family will become a fixed iterative product line, updating synchronously with every new GPU architecture generation with full support from mainstream PC manufacturers. The company also unveiled a compact 400 TFLOPS N1 Grace Blackwell Spark chip, which will be applied to unreleased small-size terminal devices.
The second half of the keynote focused entirely on physical AI layout. NVIDIA open-sourced a full set of physical AI agent development tools for industrial partners including TSMC, Foxconn and SK hynix, which have already been deployed on real factory production lines. Four core physical AI platforms are released as follows:
Isaac GR00T Humanoid Robot
Built based on Unitree H2 Plus chassis, this humanoid robot is nearly 6 feet tall and weighs 150 pounds, with 75 degrees of freedom and dual tactile five-finger hands for precise manipulation. It supports 7kg standard payload and 15kg peak payload with powerful joint torque. Powered by Jetson AGX Thor T5000, it carries a Blackwell GPU with 2070 FP4 teraflops, and supports around 3 hours of continuous operation with a built-in 1kWh battery. The whole development stack is fully open-source, with top global research institutions joining early tests. Unitree will launch official commercial hardware in late 2026.
Cosmos 3 Multimodal Foundation Model
Cosmos 3 is NVIDIA’s flagship open foundation model for physical AI, supporting unified processing of text, images, videos and ambient audio. It has three versions targeting different scenarios: Cosmos 3 Super for high-precision robot training, Cosmos 3 Nano for fast reasoning, and upcoming Cosmos 3 Edge for local real-time inference. It greatly shortens physical AI training cycles from months to days, leads mainstream open-source benchmarks, and has been adopted by enterprises including Samsung and Li Auto.
Alpamayo 2 Super Autonomous Driving Model
Based on Cosmos foundation models, this 32-billion-parameter vision-language-action model supports full 360-degree surround perception for L4 autonomous driving. It can generate traceable reasoning for every driving decision and realize automatic data labeling to cut annotation time significantly. It can be distilled into lightweight models for vehicle-end deployment, won Computex Best Choice award, and will be open-sourced on GitHub and Hugging Face this summer.
DRIVE Hyperion Autonomous Driving Platform
DRIVE Hyperion is NVIDIA’s mature L4 autonomous driving reference platform equipped with Halos safety system and dedicated DriveOS. It has secured abundant global automotive partners: Uber will launch robotaxi fleets in Munich, while Foxconn, VinFast and other manufacturers will promote L4 self-driving vehicles across Asia and the Middle East in the next few years.
Closing Thoughts
NVIDIA has completed a comprehensive hardware layout covering thin-and-light laptops, desktop workstations and robot terminal chips, with mature and reliable underlying silicon and ecosystem. Nevertheless, two critical uncertain factors remain untested: Microsoft’s actual optimization level for Windows on Arm, and real performance gaps caused by different cooling designs among various RTX Spark laptops. We are optimistic about NVIDIA’s hardware strength but remain cautious about Windows software adaptation, and we will release complete lab benchmark results once retail hardware arrives for actual testing.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!