Nvidia Open-Sources Its AI Factory OS — 40% More GPUs Per Megawatt
Abhishek Gautam11 min readQuick summary
DSX OS components — NVSentinel, KAI Scheduler, MaxLPS, Dynamo — go open source on GitHub. CoreWeave, Lambda, Red Hat, and Supermicro already run them in production.
Read next
- IEA Just Released 400 Million Barrels of Oil. It Did Not Work. Here Is What That Means for Tech.
- 66% of GenAI Inference Now Runs on Kubernetes — DRA, llm-d, Gang Scheduling
Nvidia released DSX OS as open source this week — the modular software stack it built to run its own DGX Cloud, now published on GitHub for anyone operating GPU fleets. The headline number: DSX MaxLPS power software recovers stranded capacity so operators can run up to 40% more GPUs at peak energy efficiency inside the same megawatt budget, with minimal impact on inference performance.
For an industry where power is the binding constraint — not chips — that is the most consequential infrastructure release of June.
What Is in DSX OS
DSX OS bundles open-source, modular components purpose-built for multi-tenant AI factories at gigawatt scale:
| Component | What it does |
|---|---|
| NVSentinel | Kubernetes-native GPU fault detection + automated remediation — cordons unhealthy nodes and drains workloads in seconds, not hours |
| DSX MaxLPS | Dynamic power management at GPU, rack, and workload level — up to 40% more GPUs per fixed power budget |
| KAI Scheduler + Run:ai | GPU-aware placement, fractional GPU allocation, hierarchical quotas |
| Dynamo + Grove | Distributed inference with disaggregated prefill/decode and per-stage autoscaling |
| NICo | API-driven lifecycle management |
| NVCF | Unified APIs for inference, fine-tuning, batch with native multitenancy |
| Fleet Intelligence | Fleet-wide visibility, integrity verification, health monitoring |
Already running these in production: CoreWeave, Lambda, Mirantis, Red Hat, Supermicro, Crusoe, IREN, Vultr, Nebius, Spectro Cloud, Rafay.
Why Nvidia Gave This Away
Nvidia sells GPUs, not operations software. Every month a neocloud spends rebuilding scheduling, fault handling, and power management from scratch is a month of delayed GPU orders. Open-sourcing DSX OS removes the deployment bottleneck for the entire ecosystem — the same logic as the Vera Rubin DSX reference design and the broader DSX platform push that includes a deal with IREN for up to five gigawatts of AI infrastructure.
It also locks in the stack: DSX OS is optimized for Nvidia silicon end to end. Free software, paid hardware.
Our Analysis: Power Is the Product Now
1. Tokens-per-watt is replacing TFLOPS
The industry metric that matters in 2026 is cost per token within a power envelope. MaxLPS treating grid behavior as part of the platform — not a facilities problem — confirms the shift. If you evaluate providers, ask about tokens per megawatt, not peak FLOPS.
2. Automated GPU fault handling is now table stakes
In large fleets, hardware degradation is a daily event. NVSentinel's seconds-level cordon-and-drain sets the bar; if your provider still pages a human to handle a flaky HBM stack, you are paying for that latency. This matters at home-lab scale too — the 16-GPU residential XFRA build community hit exactly these failure-management walls.
3. The 40% claim reframes the data center backlash
Projects like Kevin O'Leary's halved Utah data center show siting new power is politically hard. Software that extracts 40% more compute from already-permitted megawatts is worth more than new land. Expect every operator to adopt or clone this.
4. Self-hosters get enterprise-grade plumbing free
KAI Scheduler (now a CNCF Sandbox project), Dynamo, and NVSentinel are on GitHub. A 4–8 GPU self-hosted stack running DeepSeek or Qwen weights can now use the same scheduling and fault tooling as CoreWeave. For readers running GPUs behind restricted-API borders, this is directly usable today.
5. Watch the lock-in trade
Everything is tuned for Nvidia GPUs (NVFP4 kernels, NVLink awareness). Adopting DSX OS deepens dependence on the Nvidia supply chain — fine if that is already your reality, a strategic decision if you hold AMD or in-house silicon options.
Key Takeaways
- Nvidia open-sourced DSX OS — the AI-factory software behind DGX Cloud — on GitHub, June 2026
- DSX MaxLPS: up to 40% more GPUs at peak efficiency within a fixed power budget
- NVSentinel: Kubernetes-native GPU fault detection, cordon-and-drain in seconds
- KAI Scheduler, Run:ai, Dynamo, Grove, NVCF cover scheduling, fractional GPUs, and disaggregated inference
- In production already at CoreWeave, Lambda, Red Hat, Supermicro, Crusoe, Vultr, and more
- For developers: judge infrastructure by tokens per watt; self-hosters can adopt the components incrementally
- What to watch: AMD/alternative-silicon responses, CNCF governance of KAI, whether neoclouds differentiate on anything but price once ops software is commoditized
Sources
FAQ
Frequently Asked Questions
What is Nvidia DSX OS?
DSX OS is the open-source, modular software stack Nvidia built to operate its own DGX Cloud AI infrastructure, released publicly in June 2026. It covers GPU fault detection (NVSentinel), power optimization (MaxLPS), GPU-aware scheduling (KAI Scheduler, Run:ai), and distributed inference (Dynamo, Grove).
How does DSX OS let operators run 40% more GPUs?
The DSX MaxLPS component dynamically manages power at the GPU, rack, and workload level, recovering stranded power capacity. Nvidia says this lets AI factories run up to 40% more GPUs at peak energy efficiency within the same fixed megawatt budget, with minimal impact on inference performance.
Is Nvidia DSX OS free and open source?
Yes. The DSX OS components — including NVSentinel, KAI Scheduler, Dynamo, and NICo — are released as open source on GitHub and designed for incremental adoption. KAI Scheduler is also a CNCF Sandbox project. The software is optimized for Nvidia GPU architectures.
Who is already using DSX OS in production?
Nvidia ecosystem partners including CoreWeave, Lambda, Mirantis, Red Hat, Supermicro, Crusoe, IREN, Vultr, Nebius, Spectro Cloud, and Rafay are running DSX OS components in production for AI cloud services.
Why does tokens-per-watt matter more than TFLOPS in 2026?
Power, not chip supply, is the binding constraint on AI data centers in 2026. Operators are measured on how many tokens they can serve per megawatt of permitted power, so software that raises GPU density per watt — like DSX MaxLPS — directly lowers cost per token more than raw FLOPS comparisons.
Free Weekly Briefing
The AI & Dev Briefing
One honest email a week — what actually matters in AI and software engineering. No noise, no sponsored content. Read by developers across 30+ countries.
No spam. Unsubscribe anytime.
More on Infrastructure
All posts →IEA Just Released 400 Million Barrels of Oil. It Did Not Work. Here Is What That Means for Tech.
The IEA approved the largest emergency oil release in history after the Strait of Hormuz closed. Brent crude is still above $90. AWS data centers in UAE and Bahrain were hit by drones. Qatar's helium supply is offline, threatening chip fabs globally. Here is the full developer and infrastructure impact.
66% of GenAI Inference Now Runs on Kubernetes — DRA, llm-d, Gang Scheduling
CNCF 2026 survey: 66% of orgs run generative AI inference on Kubernetes. DRA went GA, Nvidia donated its GPU driver to CNCF, llm-d entered Sandbox, and v1.36 shipped native gang scheduling.
Nvidia NemoClaw: Open-Source Enterprise AI Agent Platform Explained
Nvidia is launching NemoClaw, an open-source AI agent platform for enterprise workforces. It's hardware-agnostic, not CUDA-locked, with Salesforce, Cisco, Google, and CrowdStrike already on board.
Nvidia, Amazon, and Apple Just Closed Their Dubai Offices Because of Iran
Nvidia, Amazon, Apple, and Snap shut Dubai offices as US-Iran tensions ground Gulf flights. Google employees are stranded. Big Tech $50B Middle East AI hub is on pause.
Free Tool
Will AI replace your job?
4 questions. Get a personalised developer risk score based on your stack, role, and what you actually build day to day.
Check Your AI Risk Score →Written by
Software Engineer based in Delhi, India. Writes about AI models, semiconductor supply chains, and tech geopolitics — covering the intersection of infrastructure and global events. 846+ posts cited by ChatGPT, Perplexity, and Gemini. Read in 164 countries.