Microsoft and NVIDIA have put local AI on the PC roadmap in unusually concrete form. At Build 2026, Microsoft highlighted the Surface RTX Spark Dev Box, a developer computer built for sustained AI work and local model fine-tuning, with enough memory and performance to run large language models locally without separate cloud GPU instances. In its Windows announcements, the company said RTX Spark will power laptops and small desktop PCs including Surface, ASUS, Dell, HP, Lenovo and MSI beginning this autumn.

NVIDIA supplied the hardware signal. Its RTX Spark announcement described a superchip for slim Windows laptops and small desktops, with up to 128GB of memory and the ability to run 120-billion-parameter language models locally. The numbers point to a clearer direction. The PC is being redesigned so AI work can happen on the machine in front of the user.

Local AI has existed for years. People already run models on personal machines with tools such as LM Studio, Ollama and other open model software. The new part is commercial packaging: better hardware, operating system support and familiar developer tools are being lined up around the same idea. Local models are becoming a normal option in the workplace toolkit.

The business argument starts with privacy. When an employee sends sensitive text, code or documents to a cloud model, that information leaves the company environment. Running the model on the device reduces that movement. LM Studio says downloaded models can be used offline and that chats and documents stay on the machine; Ollama says it runs locally and does not see prompts or data when used locally.

Cost is the second reason. Cloud AI is usage-based, and the bill grows with repeated use across teams, documents and coding sessions. OpenAI’s API pricing page tells customers to monitor spend and set monthly budgets; Anthropic’s pricing docs describe token usage being rated into dollar charges. Local models give companies another place to send work when the strongest cloud model is unnecessary.

Every machine has limits. Local models still trail the leading cloud systems on raw capability, especially on complex tasks that need strong reasoning or large amounts of context. The useful framing is workload choice. Some tasks deserve the best external model; many routine drafting, search, coding and internal assistance tasks may be good enough on a local model, especially when privacy or cost is the constraint.

The starting point is already available. LM Studio’s docs tell users to install the app, download a model, load it, and chat with it. Ollama’s docs point users to download versions for macOS, Windows or Linux and describe getting started with models such as Gemma, Qwen and others. The test is practical: install one tool, download a few models, and see what works on the hardware already in use.

Integration is now part of the story. GitHub says Copilot CLI can connect to local models such as Ollama, vLLM and Foundry Local, and can run in offline mode for restricted environments. Every AI assistant will still have its own limits and setup. Local models become useful when they sit inside the tools people already use.

Companies can begin before the next buying cycle. The right first move is a small internal trial with low-risk tasks, a short list of approved local tools and clear rules for what stays local. The hardware coming later will make the experience faster and broader. The operating habit can start now.

Sources

Microsoft, Microsoft Build 2026: Be yourself at work — blogs.microsoft.com
Windows Experience Blog, Introducing a powerful new chapter for Windows PCs, accelerated by NVIDIA RTX Spark — blogs.windows.com
NVIDIA Newsroom, NVIDIA and Microsoft Reinvent Windows PCs for the Age of Personal AI — nvidianews.nvidia.com
LM Studio, Offline Operation — lmstudio.ai
Ollama, FAQ — docs.ollama.com
OpenAI, API Pricing — openai.com
Anthropic, Pricing — platform.claude.com
LM Studio, Get started with LM Studio — lmstudio.ai
Ollama, Ollama’s documentation — docs.ollama.com
GitHub Changelog, Copilot CLI now supports BYOK and local models — github.blog