For the last three years, enterprise AI happened somewhere else. Models trained in someone’s cloud, inference ran behind an API, and the endpoint stayed where it had always been: on a laptop that opened a browser.
With local AI computing platforms like NVIDIA DGX Spark (opens in a new tab), RTX Spark (opens in a new tab), and DGX Station (opens in a new tab), that framework is ending. Frontier-scale AI compute is now on our desks, providing us with machines capable of running very large models locally and agents that operate on-device against local data.
This means the supercomputer is becoming a workstation, and the workstation is becoming something IT has never had to manage before. So, what does it mean for users, adoption, and Tanium customers?
What does this mean for users?
An NVIDIA DGX system (opens in a new tab) on an engineer’s desk is not a server, nor is it a normal PC. It holds models and inference output that may be deeply sensitive, and it runs AI agents that touch local files, call external services, and consume resources unpredictably. It arrives on operating environments most enterprise management stacks were never built around: all DGX systems run NVIDIA DGX OS (opens in a new tab), a customized version of Ubuntu Linux, while RTX Spark runs Windows on ARM. NVIDIA also announced coming availability of DGX Station (opens in a new tab) with Windows on Arm to support agentic engineering workflows with large models running on the same machine.
This creates a familiar problem in an unfamiliar place. IT cannot answer basic questions about these machines, like if it’s patched, if it’s compliant, what is actually running on it, and if we can do something about it right now. Ultimately, procurement approves the hardware while operations inherit the blind spot.
According to Stratix, only about 20% of organizations say they fully monitor how employees use AI tools and 63% report a data compromise tied to shadow AI. What’s more, IBM found that shadow AI was involved in 43% of the security incidents it studied in 2026, which is more than double the year before. Now put a machine capable of running a trillion-parameter model under the same desk, outside the same controls, and we’re looking at some serious IT and security concerns.
What does adoption actually look like?
Gartner® states that 100% of enterprise PC purchases will be AI PCs by the end of 2026. This doesn’t necessarily mean every buyer wants AI, but it does mean it’s what will be on the truck.
Additionally, research from IDC shows us that 27% of organizations have deployed AI PCs this year, 33% have actively started piloting them, and another 21% intend to deploy them within twelve months.
According to IDC, by 2030, 50% of enterprise AI inference workloads will be processed locally on endpoints or edge nodes, reducing cloud traffic and latency while supporting greater control over sensitive data. Gartner projects “By 2029, over two-thirds of all enterprises globally will deploy edge AI, up from 10% in 2025.”
Furthermore, the software consuming that compute is arriving faster than the infrastructure to govern it. Gartner® forecasts “up to 40% of enterprise applications will include integrated task-specific agents by 2026, up from less than 5% today.” IDC predicts by 2027, G2000 agent use will jump 10x and token/call loads 1,000x, making agent vetting, orchestration, and optimization essential IT responsibilities.
All this points to one uncomfortable conclusion. Within a short amount of time, much of the corporate fleet will be AI-capable, a meaningful share of inference will run on it, and the software consuming it will be agentic autonomous, resource-hungry, and largely invisible to the tools IT uses today.
What does this mean for Tanium customers?
We have always stood by the fact that you cannot manage what you cannot see, and you cannot secure what you cannot act on. This is especially hard to do in real time and across every endpoint you own.
Here’s what we know: If the AI endpoint becomes its own silo with its own tooling, patch cadence, and standing compliance exception, enterprises end up with the least governed and most sensitive machines in the building.
So, what are we doing about it? We’re focusing our efforts on four key areas.
- Coverage first. Tanium clients run on DGX OS/Ubuntu and Windows on ARM, so these systems enroll into the same platform as every other endpoint. Patching, asset inventory, configuration compliance, and process management were validated on NVIDIA hardware in our Durham engineering hub, ensuring they will work day one when deployed in Enterprises, with no custom build required.
- The AI-specific layer. We’re implementing real-time visibility into CPU, memory, GPU, and temperature across the fleet. This includes CIS hardening with CIS- and NIST-mapped compliance reporting, continuous SBOM across system packages, containers, and language runtimes, just-in-time privileged access with full session recording without requiring SSH, and air-gapped deployment support for regulated environments.
- Governance for the agents themselves. This is the part no endpoint management platform does well today. We’re building toward deploying AI agents across a fleet simultaneously, controlling their runtime so an agent can be started, monitored, and automatically stopped when it exceeds a defined time or resource threshold, and monitoring consumption at fleet level to show which models are running and whether machines are operating within expected parameters.
- Staying ahead of the device family. We’re extending coverage as DGX Station, RTX Spark, and DGX Spark (opens in a new tab) reach customers.
In typical platform shifts, the management layer shows up late, making the gap in between where the incidents happen. This time, we’d rather be early.
If you have any questions about the current state of AI compute, or what Tanium’s doing about it, talk to one of our experts.