NVIDIA Blackwell is a GPU architecture and platform designed for large-scale AI training and inference, succeeding the Hopper architecture (H100/H200). It operates at 1kW of power per GPU and requires direct liquid cooling, representing a new era for data center design. Blackwell powers AI factories for extreme-scale inference and is available across major cloud providers including AWS, Azure, and Oracle OCI.
United States Updated 2026-09-12· researched 2026-09-12· next review 2026-10-12· listed 2026-09-12 ⚠ research overdue — re-review queued
AI Revenue Exposure
NVIDIA Blackwell directly benefits from surging AI infrastructure spending as organizations deploy large-scale AI training and inference workloads; it is the core platform powering AI factories and is offered across major cloud providers (AWS, Azure, Oracle OCI) as GPU-accelerated compute instances.
Why it matters to AI
NVIDIA Blackwell operates within the AI industrial ecosystem across
semiconductors, compute, data centers, ai infrastructure software, cooling, networking. Its role sits
and upstream of 14 dependencies.
Oracle offers orderability for NVIDIA GB300 NVL72 and HGX B300 NVL16 with Blackwell Ultra GPUs across public cloud, OCI Dedicated Region, and Oracle Alloy.
AWS offers P6-B200 instances accelerated by NVIDIA Blackwell GPUs and P6e-GB200 UltraServers accelerated by NVIDIA Grace Blackwell Superchips for training and deploying the largest AI models.
NVIDIA announced a $5 billion investment in Intel, which will give Nvidia an immediate holding of about 4% in Intel and allow Nvidia to offer GB300 data center servers based on Blackwell GPUs on Intel's X86 architecture.
Institutional investors collectively hold approximately 64-68% of NVIDIA's shares, with Vanguard Group (~9.33%), BlackRock (~7.92%), and Jensen Huang (~3.50%) as top shareholders.
Saitech delivers custom-configured NVIDIA Blackwell AI infrastructure including HGX systems, high-speed networking, storage, and integration services for enterprise AI, large language model training, and inference workloads.