Microsoft has unveiled an on-device AI coding model, a security container to keep AI agents in check and an Nvidia-powered Surface Laptop Ultra, challenging Apple’s push for local AI while facing steep hardware prices.
Â
Microsoft has introduced an AI coding model that can run directly on personal computers, along with new security technology designed to stop AI agents from accessing data without permission, as it competes with Apple’s drive to bring more artificial intelligence onto personal devices.
At an event in San Francisco, the company also showed off the Surface Laptop Ultra, a high-powered machine built around Nvidia’s RTX Spark chips. The Redmond, Washington-based firm wants Windows to become a platform where AI agents handle tasks such as writing code or running complex business projects on desktops and laptops.
The strategy is a bet that some workloads now handled in Microsoft’s expensive Azure cloud data centres can move to powerful Windows machines owned by businesses and consumers, who pay for the hardware. Apple is chasing the same opportunity with new Mac computers. For Nvidia, a foothold in Windows PCs could open one of the last big markets dominated by Intel and AMD.
Guardrails for autonomous agents
To address security risks posed by increasingly autonomous agents, Microsoft released Microsoft Execution Containers (MXC), a tool that blocks agents from reaching data or performing unauthorised tasks on a laptop or desktop. Pavan Davuluri, executive vice president for Windows and devices, said Anthropic, OpenAI and Nvidia will use it. IT departments will be able to set rules for agents, which Windows will enforce on each employee’s machine.
“We needed to make the desktop the most secure place for agents to execute,” CEO Satya Nadella said. Nvidia chief Jensen Huang, who appeared on stage, said the technology “is going to revolutionize how agents are built and deployed”, adding that without it agents are “a complete nonstarter”.
Microsoft also announced that models once dependent on the cloud can now run locally, including Nvidia’s open-source Nemotron. Davuluri said a version of China’s DeepSeek V4 can run on machines with at least 60 gigabytes of memory and beat OpenAI’s GPT-5 on some coding and reasoning tasks. Meta’s Muse assistant will arrive on Windows as a native app, and the open-source OpenClaw agent system can use the security tools. Copilot will split work between cloud and device, with chief Jacob Andreou saying local models can handle cases where cost or privacy matter more.
The price problem
Cost remains a major hurdle after a memory-chip shortage pushed up prices. The Surface Laptop Ultra starts at $2,599 and reaches $5,899 for a version with a 20-core processor, 128GB of memory and 1TB of storage. Apple’s top MacBook Pro configuration costs $6,700, though the machines are not directly comparable. Microsoft’s entry-level Ultra exceeds Apple’s $1,999 base MacBook Pro but offers 24GB of memory against 16GB.
Nvidia has raised the price of its DGX Spark AI desktop by about 75% since launch, to $6,950. Analyst Anshel Sag of Moor Insights & Strategy said the software is now ready but the hardware is too expensive, so “only the people who have the budget” can run AI locally.
