Follow ZDNET: Add us as your preferred source Google.
Important points of ZDNET
- AI runs on Linux. period. There are no substitutes.
- Canonical and Red Hat are building Linux distributions specific to Nvidia Vera Rubin.
- The Linux kernel is tailored for AI and ML workloads.
Modern AI started with open source and ran on Linux. Today, Linux is not only important for artificial intelligence; This is the foundation upon which today’s entire modern AI stacks run. Everything is Linux from top to bottom, from hyperscale training clusters to edge inference boxes.
Also: Things went wrong when I tried a Linux distribution that promised free built-in AI
The magic tricks of AI are actually the collective output of very mundane infrastructure, such as supercomputers, GPU farms, and cloud clusters that run almost all Linux. All core machine learning frameworks (TensorFlow, PyTorch, scikit-learn, etc.) were originally developed and tuned on Linux. From Jupyter and Anaconda to Docker and Kubernetes, the tools around these tools are similarly optimized for Linux.
Why IT jobs will survive and die on Linux
why? Because researchers and production engineers are actually deploying AI on Linux. The future of IT jobs will live and die on Linux.
AI runs on Linux because it is the most flexible, powerful, and scalable environment for the GPU-intensive distributed workloads that modern AI requires. Additionally, the entire tooling and cloud ecosystem is standardized on Linux.
Yes, all AI platforms, whether it’s OpenAI, Copilot, Perplexity, Anthropic, or your favorite AI chatbot, are built on Linux, and the drivers, libraries, and orchestration are all tied together in different ways. The unique parts may gain brand power, but without Linux, they are nowhere to be found.
That leads to more Linux jobs.
As noted in the Linux Foundation’s 2025 State of Tech Talent Report, AI is driving net growth in technology jobs, especially Linux jobs. What this will look like will be determined by the “AI”. [is] The report says that reshaping roles, rather than eliminating them, “will lead to changes in skill demand and new opportunities for workforce growth.”
In addition to the rise in Linux systems and network administration jobs, the site Linux Careers sees “a burgeoning trend involving professionals who combine Linux expertise with operational artificial intelligence and machine learning.” These new AI/Linux jobs include AI Operations Specialist, MLOps Engineer, ML Engineer, and DevOps/AI Engineer.
Of course, Linux distributors know all this, so when a new Linux distribution is released, its manufacturers emphasize AI features.
Canonical and Red Hat, for example, are racing to plant the Linux flag on Nvidia’s new Vera Rubin AI supercomputer platform. The race continues over who will own the operating system layer of the “Gigascale AI Factory.”
Red Hat is introducing Red Hat Enterprise Linux (RHEL) for Nvidia. This selected RHEL edition is specifically optimized for Nvidia’s Rubin platform, including the Vera Rubin NVL72 rack-scale system.
Also: How AI and Rust are rewriting Linux and Windows programming
According to the company, this variant ships with Day 0 support for Vera CPUs, Rubin GPUs, Nvidia’s CUDA X stack, and verified OpenRM drivers and toolkits are provided directly through Red Hat repositories.
Linux kernel and AI
Canonical is also rolling out official Ubuntu support for the Nvidia Rubin platform, also targeting Vera Rubin NVL72. The London-based company’s story is to make x86-equivalent custom Arm-based Vera CPUs “first-class citizens” with its upcoming Ubuntu 26.04 release.
So unlike Red Hat, which has RHEL specifically for Nvidia processors, the new Ubuntu supports Nvidia. This version also upstreams features such as Nested Virtualization and ARM Memory Partitioning and Monitoring (MPAM) to better partition memory bandwidth and cache for multi-tenant AI workloads.
Also: Linus Torvalds also does vibe coding now
Running all of this is the Linux kernel, which has been steadily modified to meet AI’s voracious appetite for hardware acceleration. Modern kernels make good use of GPUs and specialized accelerator drivers, advanced memory management for moving tensors quickly, and schedulers tuned for massively parallel batch jobs.
In other words, the kernel has been rewired as an operating system for AI hardware accelerators over the past decade.
Memory: Put your data where the GPU is
Specifically, one of the most important enablers is heterogeneous memory management. This allows device memory such as graphics processing units/video random access memory (GPU VRAM) to be integrated into the Linux virtual memory subsystem.
This, combined with Direct Memory Access Buffering (DMA-BUF) and Non-Uniform Memory Access (NUMA) optimizations, allows the AI runtime to keep tensors close to accelerators and reduce data copies that tend to degrade performance.
Also: I’ll show you how I finally solved a sticky Linux problem using this AI terminal app.
Modern kernels also treat advanced CPU-GPU combinations, such as tightly coupled NUMA-style CPU/GPU nodes, as first-class citizens. This allows on-demand migration of memory between CPU-attached RAM and high-bandwidth GPU memory.
As Nvidia explains, this “enables the CPU and GPU to share one page table per process, giving all CPU and GPU threads access to all memory allocated to the system.”
Accelerators: actual subsystems, not add-ons
Linux now includes dedicated compute accelerator subsystems designed to expose GPUs, Tensor Processing Units (TPUs), and custom AI application-specific integrated circuits (ASICs) to AI and machine learning (ML) programs.
On top of that, GPU support has matured from graphics-first to compute-focused through Direct Rendering Manager (DRM), open stacks like ROCm and OpenCL, and Nvidia’s Compute Unified Device Architecture (CUDA) drivers.
Also: AI is already part of Linux – whether developers like it or not.
The kernel work has been expanded to cover new AI accelerators such as Intel’s Habana Gaudi, Google’s Edge TPU, and FPGA/ASIC boards with driver and bus abstractions. This allows AI programs like PyTorch and TensorFlow to see and use them as just another device. So anyone creating new AI silicon today assumes for granted that Linux will run on top of it.
Scheduling: Feeding the Hungry Accelerator
Linux’s default scheduler, Earliest Eligible Virtual Deadline First (EEVDF), real-time scheduler, and NUMA balancing are all tuned to help AI workloads lock onto the CPU, isolate noisy neighbors, and feed accelerators without jitter. Efforts to increase the default kernel timer frequency from 250 Hz to 1000 Hz have already shown measurable improvements in large-scale language model (LLM) acceleration at negligible power cost.
Also: We found 7 must-have Linux apps for students, including local AI
Although not the default setting on Linux, some distributions such as the Ubuntu low-latency kernel come with this as a standard setting.
Direct Path: Eliminate the CPU middleman
The latest kernels use technologies like Nvidia’s GPUDirect and peer-to-peer DMA to give the GPU direct access to memory, storage, and even peer devices. combine with Computing Express Link (CXL)) and input/output memory management unit (IOMMU) processing to allow accelerators to bypass the CPU when moving data. This eliminates the bottleneck that previously halted ML training runs. This invisible plumbing allows AI clusters to scale out without collapsing on their own I/O.
Also: My two favorite AI apps on Linux – how to use them to get more done
Put all this together: When executives talk about “AI strategy,” what they’re not talking about is the sober reality that AI strategy depends on managing Linux at scale. The key is to patch your kernel, harden your containers, and secure your opaque workloads. AI may be making headlines, but the operating system that does the real work is still Linux.
