Microsoft’s Surface RTX Spark Dev Box Targets Local Large-Model AI Development

Microsoft has introduced the Surface RTX Spark Dev Box, a compact desktop aimed at software developers who want to run large AI models locally instead of relying on cloud infrastructure and per-token billing.

Unveiled on Monday at Microsoft Build 2026, the machine is positioned as a desk-friendly alternative to cloud-based AI experimentation and deployment, particularly for teams that are pushing into very large models and long-context workloads.

The Surface RTX Spark Dev Box centres on Nvidia’s new RTX Spark processor, built on the company’s Blackwell architecture, and pairs it with 128GB of unified memory inside a small-form-factor chassis. According to Nvidia’s own rating, the configuration delivers about one petaflop of AI compute.

In practice, Microsoft says that performance level is meant to let developers load, run, and interact with AI models exceeding 120 billion parameters directly on the device, without issuing any API calls to the cloud. That shifts the cost model from metered usage to a fixed, hardware-based investment for those workloads that can fit on the box.

Pavan Davuluri, Microsoft’s executive vice president of Windows and Devices, framed the design around both model size and context length during a pre-event press briefing. He said Microsoft expects “these class of devices” to be able to run models of around 100 billion parameters. But he stressed that parameter count alone is not the main constraint.

According to Davuluri, a model’s usefulness is tightly tied to how much context it can handle. As context windows grow, the memory footprint of key-value caches grows with them. At a 100,000-token context length, he noted, the key-value cache alone can occupy roughly 40–50GB of memory. That requirement informed the decision to engineer the box around a 128GB unified memory pool that the CPU and GPU share dynamically, rather than splitting memory between components.

Microsoft plans to sell the Surface RTX Spark Dev Box later this year in the United States, exclusively through Microsoft.com. The company has not disclosed pricing.

The system’s focus on running large models and long-context workloads locally is presented as a direct challenge to the per-token pricing that has dominated AI services since the launch of ChatGPT. By enabling developers to execute substantial AI workloads entirely on a desktop device, Microsoft is showing a stronger push toward local, fixed-cost AI infrastructure alongside its existing cloud offerings.