Addressing 2026's Challenges with 2023's Technology

This week, Microsoft introduced the Surface Laptop Ultra, a device that breaks away from the traditional Surface design. It stands as Microsoft's most potent Surface Laptop to date and marks the debut of the NVIDIA RTX Spark in a Windows laptop.

However, it might not be the perfect computer for the Agent era.

According to Microsoft, the Surface Laptop Ultra is tailored for creators, developers, and AI builders, supporting large-scale 3D scenes, lengthy compilation times, and local models and datasets. NVIDIA describes the RTX Spark as a superchip for Windows PCs, targeting personal AI Agents with the following features:

- 6,144 CUDA cores, fifth-generation Tensor Cores, FP4 precision, connected via NVLink-C2C to a 20-core NVIDIA Grace CPU, supporting up to 128GB of unified memory.

As a result, the RTX Spark can locally run large models with 120 billion parameters, handle up to 1 million token contexts, render 3D scenes exceeding 90GB, edit 12K 4:2:2 video, generate 4K AI video, and run AAA games at over 100 FPS in 1440p.

Image Source: NVIDIA

This level of power is almost unprecedented for a traditional laptop.

The strength of the RTX Spark primarily lies in its GPU, driven by CUDA, Blackwell GPU architecture, unified memory, and local model inference capabilities. If the key question for AI PCs is whether they can run larger models locally, the RTX Spark is undoubtedly compelling. However, at Leitech, we believe that if the focus has shifted to whether Agents can autonomously perform tasks over the long term, the answer is not so straightforward.

Agents involve breaking down tasks, invoking applications, searching files, running code, opening web pages, maintaining sandboxes, handling permissions, waiting for external responses, and switching contexts between tasks. These workloads demand robust cloud-based model capabilities and heavily rely on end-side CPU, I/O, system scheduling, security isolation, application interfaces, and cloud-side state management.

Conversely, the performance of end-side models is far from sufficient to support the model capabilities required by Agents. Even a "beast" like the RTX Spark resembles more of a compact AI workstation squeezed into a laptop form factor rather than a personal computer redesigned for the Agent era.

In essence, it's attempting to answer 2026's Agent questions with 2023's vision of local large models.

From NPU to CPU: NVIDIA's Strengthened PC Presence

NVIDIA's interest in PC processors didn't emerge overnight.

From a short-term business perspective, there's little need for the company to develop a PC SoC personally. It's already well-established in the PC market: its discrete GPUs dominate, with strongholds in gaming, creation, professional graphics, and the CUDA ecosystem. Its data center GPUs are in high demand, with customers eagerly waiting in line.

However, AI PCs have altered the workload structure of PCs and NVIDIA's position within them. (Note: AI PCs primarily refer to laptops, excluding DIY desktops.)

In the past, PC workloads revolved mainly around CPU and graphics performance. NVIDIA could simply enhance its GPUs and connect them to PCs via PCIe to maintain its top position in the performance hierarchy.

But in the AI PC era, more workloads are centered around local inference, multimodal understanding, voice and video processing, semantic search, local knowledge bases, and system-level AI functions. NPUs are more aligned with everyday PC usage than GPUs.

The low power consumption, always-on nature, quiet operation, long battery life, and system-level integration of NPUs are precisely what AI PCs need in their early stages. Traditional discrete GPUs require data to be shuttled between system memory and VRAM, leading to high power consumption, noisy fans, and poor battery life.

For laptops, which are inherently constrained by battery life, thermal management, and form factor, the logic of NPUs is highly appealing.

In 2024, Microsoft set a new standard with AI+ PC (initially Copilot+ PC): NPUs must deliver at least 40 TOPS to support next-generation Windows AI features. Qualcomm, AMD, and Intel have all stepped up, providing PC manufacturers with a fresh selling point beyond CPU core counts, screen quality, and battery life.

Image Source: Microsoft

However, if future PC AI workloads are entirely taken over by NPUs, NVIDIA's role in PCs could shift from a platform definer to a high-end graphics accessory.

This is why NVIDIA developed the RTX Spark, integrating CUDA, RTX, TensorRT, OptiX, DLSS, FP4, Blackwell GPU, and unified memory into Windows laptops. The goal is to ensure that PCs not only run AI features but also support local AI workflows, preventing the AI PC era from bypassing GPUs and CUDA.

During a Microsoft Build connection, NVIDIA CEO Jensen Huang mentioned a discussion he had with Microsoft CEO Satya Nadella three years ago about a new kind of personal computer: one suitable for designers and creators as well as AI, with both local processing capabilities and deep integration with Windows, creative software, and AI software stacks.

Image Source: X

The RTX Spark and Surface Laptop Ultra are the results of that conversation. They do answer a question: if the Windows ecosystem wants to build a high-end laptop capable of running local large models, local creation, and local AI development, how aggressive should the hardware be?

But AI has changed significantly in three years.

Three years ago, the industry was still in the early stages of imagining local inference following the ChatGPT boom. Many believed that the key to AI PCs was running models locally, allowing users to avoid sending private data to the cloud, paying token bills for each inference, and achieving lower latency and more stable AI experiences.

This logic still holds true today, but only partially. By 2026, AI has evolved beyond chatbots. From inference to Agents, context, reasoning chains, and KV caches—nearly all requirements have increased, and the capability gap between local devices and cloud models has widened for many tasks.

Image Source: X

In fact, nearly all large models trained for Agents today have hardware requirements that far exceed previous standards. Quantized and compressed models cannot meet the demands for smooth Agent operation and user expectations.

In short, at least for now, end-side devices cannot support a good local Agent experience; cloud-based solutions are inevitable. Therefore, for personal PCs, the CPU becomes even more critical.

Can an "Outdated" ARM CPU Adapt to the Agent Era?

In the Agent era, users don't just want answers—they want AI to complete tasks.

But completing tasks is not the same as generating text locally. An Agent executing tasks often needs to access web pages, invoke software, run code, read files, handle permissions, verify results, and operate continuously in the background. It functions more like an operator than an offline model. The more complex the workflow, the more it relies on CPU, I/O, system scheduling, browser environments, sandboxes, and cloud services.

In contrast, the RTX Spark focuses almost entirely on GPU and AI capabilities: 1 petaflop of AI performance, 6,144 CUDA cores, Blackwell RTX GPU, 128GB of unified memory, local 120B models, and million-token contexts.

For the CPU, NVIDIA chose MediaTek, adopting a 10-core Cortex-X925 and a 10-core Cortex-A725. These cores are based on Arm IP released two years ago and have been widely used in flagship and sub-flagship smartphone SoCs over the past two years, including the Dimensity 9400, Dimensity 8400, Xuanjie O1, and Exynos 2500.

Meanwhile, last year's Dimensity 9500 already adopted Arm's latest flagship architecture, C1-Ultra, and SoCs featuring the next-generation C2 core are expected to launch in the coming months.

Image Source: Arm

Of course, the RTX Spark has significantly more CPU cores, but from planning to execution, the CPU is unlikely to be the core focus of NVIDIA's consumer PC SoC.

This is not surprising, as NVIDIA's strongest moat has always been GPUs and CUDA. However, the Agent era will re-elevate the importance of CPUs. NVIDIA's Vera CPU for data centers essentially acknowledges that as AI evolves from chatbots to Agents, code execution, data processing, sandbox environments, and task orchestration will become critical paths, and CPUs will no longer just support GPUs.

Yet, the RTX Spark allocates too much budget, power consumption, chip area, and system design to GPUs and local inference without fully addressing the most critical aspects of Agents: execution, scheduling, long-term state management, and cross-device collaboration.

Meanwhile, Project Solara, unveiled alongside the Surface Laptop Ultra, represents Microsoft's alternative answer.

According to Steve Bathiche, head of Microsoft's Applied Sciences Group, Project Solara is a "chip-to-cloud" platform designed for agent-first experiences and new device form factors. It doesn't just bring intelligence to PCs, browsers, or phones—it integrates intelligence into workflows, environments, and task contexts. Devices are no longer designed around apps but around Agents.

By the way, Project Solara runs on Android, not Windows.

More importantly, Project Solara's state is not confined to a single device but is managed across a set of dedicated devices via Azure. Microsoft demonstrated portable and desktop form factors and confirmed that Qualcomm and MediaTek will be the first chip partners.

Project Solara currently includes two devices. Image Source: Microsoft

This approach may seem less groundbreaking than the Surface Laptop Ultra and is still in its early stages, but it aligns more closely with the actual needs of Agents. The value of Agents lies not in having a complete local AI factory at every entry point but in appearing at the right time, place, and device, and delegating tasks to cloud-based states and backend intelligence for continuous progress.

In other words, the Surface Laptop Ultra "enlarges" the PC, while Project Solara "streamlines" devices. The Agent era may favor the latter.

Final Thoughts

Local computing power remains important.

Privacy-sensitive data, local files, low-latency interactions, offline scenarios, creative assets, and development environments all require robust end-side capabilities. For professional users needing local models, rendering, video generation, and CUDA workflows, the Surface Laptop Ultra could be an excellent machine.

However, when it comes to what personal computers should look like in the Agent era, the Surface Laptop Ultra may not be the best answer.

Agents are inherently better suited to a cloud-centric model, with multiple lightweight devices serving as entry points. Phones, PCs, badges, desktop screens, earphones, and glasses can all act as Agent touchpoints, but they don't need to become compact AI workstations.

From this perspective, AI PCs like the Surface Laptop Ultra and RTX Spark resemble using 2023's vision of local inference to answer 2026's Agent questions. They are powerful and important but not the starting point for next-generation Agent devices.

NVIDIA, Microsoft, Agents

Source: Leitech

Images in this article are from the 123RF licensed library. Source: Leitech