Satya Nadella, Windows and Surface chief Pavan Davuluri, and NVIDIA CEO Jensen Huang are expected to attend.
Surface Laptop Ultra
Microsoft is expected to reveal more details about the Surface Laptop Ultra, reportedly built around NVIDIA’s RTX Spark platform.
Expected specifications:
- NVIDIA RTX Spark
- 20-core CPU
- Up to 128GB unified memory
- CUDA support
- Local inference for models up to 120B parameters
- Windows 11 optimized for the platform
- Improved Prism x86/x64 emulation
The combination of CUDA and 128GB of unified memory is the key change. It would allow developers to run much larger models locally than on current Copilot+ Surface hardware.
Surface Laptop Ultra vs Surface Laptop 7
| Specification | Surface Laptop 7 | Surface Laptop Ultra* |
| Platform | Snapdragon X Plus / X Elite | NVIDIA RTX Spark |
| CPU | Up to 12 cores | 20 cores |
| AI hardware | Qualcomm Hexagon NPU | NVIDIA GPU |
| NPU performance | 45 TOPS | — |
| Maximum memory | 64GB | 128GB unified |
| GPU | Qualcomm Adreno | NVIDIA |
| CUDA | No | Yes |
| Local AI | Copilot+ workloads | Up to 120B modelsx |
| 86/x64 apps | Prism | Improved Prism |
*Surface Laptop Ultra specifications are not yet official.
The previous Surface generation concentrated on efficient NPU inference. RTX Spark shifts the focus toward GPU-based AI development and inference.
Why 128GB matters
Model size quickly becomes a memory problem.
A 120B-parameter model requires approximately:
- FP16: 240GB for weights
- INT8: 120GB
- INT4: 60GB
These are theoretical weight sizes and exclude KV cache and other runtime overhead.
A 128GB unified-memory system therefore wouldn't run a 120B model at FP16, but it could potentially accommodate heavily quantized versions without the VRAM limits of conventional laptop GPUs. That opens the door to larger local LLMs, multimodal models, RAG pipelines and AI agents using CUDA-compatible frameworks.
The important metric isn't whether the machine can load a 120B model. It's tokens per second, memory bandwidth and sustained performance. Those numbers have not yet been disclosed.
RTX Spark + Windows 11
Microsoft has reportedly been optimizing Windows 11 for RTX Spark. One area is workload scheduling across the platform's 20 CPU cores. Another is Prism, Microsoft's translation layer for running x86/x64 Windows software on Arm-based systems.
This matters for existing development stacks that still depend on x86 binaries, libraries or tools.
The practical test will be compatibility with:
Python → CUDA → PyTorch → containers → IDEs → native dependencies.
Microsoft may also provide details about a Surface RTX Spark Dev Box, potentially positioning it as a desktop system for local AI development.
2024 Copilot+ PC vs 2026 RTX Spark
Microsoft's 2024 Copilot+ PC specification required an NPU capable of at least 40 TOPS. Surface Laptop 7's Snapdragon X platform delivers 45 TOPS.
The architecture was primarily designed for efficient on-device AI features.
The RTX Spark approach changes the target:
- Copilot+ PC: 45-TOPS NPU → Windows AI features and smaller inference workloads.
- RTX Spark Surface: CUDA GPU + up to 128GB unified memory → large-model inference and AI development.
That's a substantially different workload profile.
What to watch on October 7
The most important missing specifications are GPU configuration, memory bandwidth, power limits, inference speed, thermals and price. Microsoft is not expected to introduce Windows 12. The October 7 event is instead centered on Windows 11, RTX Spark, Surface hardware and local AI.
The primary standard for software engineers is how many units of data the Surface Laptop Ultra processes every second on models containing 70 billion to 120 billion parameters. To calculate this, engineers measure the specific level of data compression and the amount of electricity the hardware uses. It is necessary to evaluate those specific metrics to understand the performance of the machine. Under the conditions, the speed of the processor is the most important factor for technical users.
