Don't Buy a RTX Spark Laptop For Local AI Until You See This

Summary

The video examines whether the $5,899 Surface Laptop Ultra, an RTX Spark laptop with 128 GB of unified memory, is a good choice for running local AI—especially coding agents. The presenter argues that impressive memory capacity and output speed don’t tell the whole story: for long prompts, the time to process input may matter more than the time to generate a response.

Hardware and model fit

• The laptop’s Nvidia RTX Spark N1X combines a 20-core CPU and a 6,144-core Blackwell GPU, with CPU and GPU sharing unified memory.
• Microsoft’s MAI Code 1.1 Flash is described as a 137-billion-parameter mixture-of-experts model, with about 6.8 billion parameters active per generated token. Compressed to roughly 53 GB, it can fit in the laptop’s memory.
• The laptop’s advertised memory bandwidth is around 300 GB/s. The presenter says this is relevant to decode—generating output—but does not determine how quickly the model reads a long prompt.

The key distinction: prefill vs. decode

• Prefill is the model processing the prompt, conversation history, files, and tool results before it begins responding. The video characterizes this as compute-intensive.
• Decode is generating the response. It depends more on memory bandwidth because the model repeatedly reads its active weights.
• For short prompts, decode can dominate the wait. For long prompts and agent workflows, prefill may dominate: the agent repeatedly sends code, test results, and other context back to the model.

Reported performance

According to Microsoft’s developer-blog figures cited in the video, a Surface Laptop Ultra running MAI Code 1.1 Flash with a Windows on ARM build of llama.cpp, CUDA, and DFlash2 speculative decoding achieved:

• 923 tokens/second prefill on a 64K-token prompt—about 71 seconds before the first response.
• 769 tokens/second on a 128K-token prompt—about 170 seconds before the first response.
• About 63 tokens/second decode on short prompts.

The presenter compares this with an Nvidia developer-forum report of 2,000–2,477 tokens/second prefill on a DGX Spark. Because the report used a different model and engine, it is not a direct benchmark comparison; the presenter treats it as an indication that the desktop may process long prompts roughly 2.3 times faster.

The proposed explanation is power: the DGX Spark is cited at 140 W, while Nvidia’s laptop-chip specifications list 45–80 W. The video says the Surface’s final sustained wattage had not yet been confirmed, despite earlier reporting that Microsoft had discussed a 110 W design figure and subsequently clarified that it was not a target for sustained TDP.

What this means for coding agents

The video focuses on Microsoft’s pitch of running local coding agents. Agents can resend a growing context—including repository files and tool results—on successive turns, making prefill important. Cached context can make later “warm” turns much faster, but a new session, repository, large log, or cache miss may require processing the prompt from scratch.

Price, alternatives, and limitations

• The presenter says Nvidia raised the 128 GB DGX Spark price to $6,950, making it about $1,050 more expensive than the laptop. The DGX still has a higher stated power budget, a 200-gigabit networking port for linking units, and Nvidia Linux; the laptop runs Windows 11 and includes a screen, battery, and keyboard.
• Microsoft’s Surface RTX Spark dev box is cited at $5,999, with 128 GB of memory, 2 TB of storage, and a 100 W thermal envelope. The video says it was expected in November.
• A 128 GB ROG Flow Z13 with AMD Ryzen AI Max+ 395 is cited at around $3,200. A separate result for an AMD-based mini PC reportedly showed about 1,320 tokens/second on a similar long-context test, or roughly 99 seconds for a 128K prompt. The presenter cautions that the model and engine differed, so this is directional rather than a definitive comparison.
• The video notes that the GPU can address 110 GB rather than the full 128 GB, with Windows and other apps also consuming memory. At 256K context, the model reportedly used 75.5 GB, leaving less room for applications such as VS Code, a browser, and Docker.
• The presenter also mentions reported delivery delays and that the laptop was listed as out of stock at the time of recording.

Who may or may not benefit

The long-prompt delay may matter less for short chats or small code questions, where decode speed is more significant. The laptop may also suit people who need local, portable AI for sensitive data or who use a long-running agent that can reuse cached context. Its all-in-one Windows form factor may appeal to people who also need a general-purpose or gaming laptop.

The video recommends waiting for independent reviews, particularly for a user’s real workload, before deciding.

Review checklist highlighted in the video

1. Check cold prefill time at the context length and with the model you actually intend to use—ideally 64K or 128K, not a short chat prompt.
2. Repeat the test after 30 minutes of sustained load and on battery, using a large model.
3. Confirm the laptop’s final sustained wattage.
4. Verify whether the 128 GB configuration ships on the expected schedule.

The presenter also notes Microsoft’s stated holiday return window, which would allow buyers to test the machine with their own workflows.

Main speaker and sources

• Main speaker: Kai, the video’s narrator and analyst.
• Sources cited: Microsoft’s developer blog and product/store information; Nvidia’s specifications and developer forum; Tom’s Hardware reporting on Microsoft’s earlier power-design comments; and user-reported DGX Spark and AMD-system benchmark results.