NVIDIA RTX Spark Brings AI Petaflop to ARM Laptops
Key insights
- Vera Rubin is in full production with a supply chain described as twice the size of the Grace Blackwell build-out.
- RTX Spark pairs a 20-core Grace CPU with 6,144 Blackwell CUDA cores via NVLink for 1 petaflop of AI in slim laptops and compact desktops.
- Nemotron 3 Ultra, a 550B-parameter mixture-of-experts model, delivers 5x faster inference at roughly 30% lower cost than leading open models.
Why this matters
Vera Rubin entering full production with a supply chain twice Grace Blackwell's size and 150 Taiwan partners already committed means NVIDIA's next data center generation is scaling faster than any prior cycle, compressing the window for cloud competitors. RTX Spark is the first NVIDIA product to combine Grace CPU cores and Blackwell CUDA on a single NVLink package for consumer laptops, which, if successful, brings the full CUDA software stack to ARM-based Windows devices without a discrete GPU and opens a new competitive front against Qualcomm and Apple in AI-on-device. The Vera CPU being designed specifically as 'a CPU for agents,' paired with the NVIDIA Agent Toolkit and OpenShell secure sandboxing, signals that NVIDIA is structuring its hardware roadmap around autonomous AI agent workloads, not just model training or static inference.
Summary
NVIDIA opened Computex 2026 with two headline announcements from CEO Jensen Huang: RTX Spark, combining a 20-core Grace CPU with 6,144 Blackwell CUDA cores over NVLink for 1 petaflop of AI performance in slim laptops and compact desktops, and confirmation that Vera Rubin is now in full production with a supply chain twice the size of Grace Blackwell.
RTX Spark is being built alongside MediaTek and Microsoft for Windows PC integration, putting NVIDIA's full AI stack into consumer-grade ARM form factors for the first time.
Essentially: (NVIDIA, MediaTek, Microsoft) are moving AI inference out of data centers and into laptops.
- Vera Rubin's five-rack system delivers 10x higher inference performance per watt and 10x lower cost per token versus prior generations
- Nemotron 3 Ultra, a 550B-parameter mixture-of-experts model, runs 5x faster and roughly 30% cheaper than leading open models
- DGX Station for Windows handles up to 1 trillion parameter models with 748GB coherent memory
Huang framed the moment as economic infrastructure, declaring AI "now a GDP generator" and the AI factory buildout "the largest infrastructure build out in human history."
Potential risks and opportunities
Risks
- Vera Rubin's 150 Taiwan supply chain partners and 350+ factories across 30 countries create concentrated geopolitical exposure if Taiwan supply chain access is disrupted at any point during the ramp.
- RTX Spark's Windows PC integration depends on a three-way partnership with MediaTek and Microsoft; delays in Windows ARM driver support or MediaTek production timelines could slip the product's availability window.
- Nemotron 3 Ultra's 5x inference speed and roughly 30% cost advantage over leading open models may compress quickly if Meta or Mistral release comparable mixture-of-experts architectures before NVIDIA locks in enterprise deployments.
Opportunities
- Liquid cooling infrastructure vendors are positioned to benefit from Vera Rubin requiring 100% liquid cooling at 45 degrees Celsius inlet temperatures across all five-rack deployments at scale.
- Named cloud partners CoreWeave, Nebius, and NAVER Cloud can differentiate on early Vera Rubin access, capturing inference workloads before hyperscalers complete their own deployments at the same performance-per-watt tier.
- Enterprise buyers evaluating on-premises AI can now target DGX Station for Windows, which runs 1 trillion parameter models with 748GB coherent memory in a deskside form factor without requiring a full data center build-out.
What we don't know yet
- No OEM partners, device names, or pricing were disclosed for RTX Spark in the article, leaving launch timing and retail availability unconfirmed.
- Whether Vera Rubin's supply chain across 350+ factories in 30 countries can sustain the doubled scale relative to Grace Blackwell without production bottlenecks through the second half of 2026.
- Performance comparisons between RTX Spark and competing ARM laptop chips were not provided, making competitive positioning against Qualcomm or Apple unverifiable from this announcement alone.
Originally reported by blogs.nvidia.com
Read the original article →Original headline: Jensen Huang's Computex 2026 Keynote Officially Launches NVIDIA N1X ARM Laptop SoC — Dell, Lenovo, ASUS, MSI Confirm Holiday Devices; Huang Teases Third Unnamed H2 2026 Product