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

Microsoft’s own benchmark shows the Surface Laptop Ultra, the flagship NVIDIA RTX Spark laptop, waiting 170 seconds before the first word on a 128K prompt. Before you spend $5,899 on a 128GB AI laptop for local AI, here’s where that number comes from and who it actually affects. Every RTX Spark vs Mac comparison argues about memory bandwidth: about 300 GB/s on the RTX Spark N1X versus 614 GB/s on a MacBook Pro M5 Max. That only describes decode, how fast the model writes back to you. Reading your prompt (prefill) is compute, and compute needs watts. The N1X runs at 45–80 W in a laptop. The same chip family gets 140 W in the DGX Spark. Bandwidth survived the move to a laptop (about 10% faster on paper). Power didn’t. Microsoft published the numbers itself: MAI Code 1.1 Flash (137B total, 6.8B active, 53 GB) on the Windows on Arm build of llama.cpp with CUDA and DFlash2 speculative decoding. Prefill came in at 923.5 tok/s at 64K (about 71 seconds cold) and 769.8 tok/s at 128K (about 170 seconds). Decode was about 63 tok/s on short prompts. A DGX Spark owner on NVIDIA’s forum reports 2,000–2,477 tok/s of cold prefill on long prompts. That’s a different model and engine, but it reads roughly 2.3x faster. For local LLM coding agents that gap is the whole story. An agent re-sends the repo, the failing test and every tool result, so most of the tokens go in, not out. Warm, cached turns are fast. Cold turns (a new session, a new repo, a big log file) pay the full 170 seconds. IN THIS VIDEO Why prefill, not memory bandwidth, decides how long a long prompt takes on a local AI laptop RTX Spark laptop vs DGX Spark: same chip family, 45–80 W vs 140 W Microsoft’s withdrawn 110 W figure, and why final power still isn’t public Microsoft’s own numbers: 71 s at 64K and 170 s at 128K before the first word The DGX Spark jumped to $6,950, so the 128GB Surface is now $1,050 cheaper (and out of stock) The Surface RTX Spark Dev Box: $5,999.99 with a 100 W thermal envelope Why the GPU sees 110 GB of the 128GB, not all of it Who should buy one anyway, plus the AMD Ryzen AI Max+ 395 alternative (ROG Flow Z13, Bosgame M5) Four numbers to check when the reviews land on October 16 CHAPTERS 0:00 170 seconds before the first word 1:49 The case for the RTX Spark laptop 3:38 Prefill vs decode: the two jobs 5:31 Same chip, 80 watts instead of 140 6:57 Microsoft’s own benchmark 8:23 Why coding agents live on prefill 9:39 The DGX Spark price flipped 11:11 128GB that shows up as 110 11:44 Who should buy one anyway 12:44 The AMD alternative 13:38 Four things to check in the reviews 14:53 Verdict SOURCES Microsoft’s local model benchmark (prefill, decode chart, Task Manager): https://commandline.microsoft.com/loc... RTX Spark N1X TDP by SKU, 300.8 GB/s (StorageReview): https://www.storagereview.com/news/nv... NVIDIA RTX Spark: https://www.nvidia.com/en-us/products... Microsoft’s 110 W statement and correction (Tom’s Hardware): https://www.tomshardware.com/laptops/... DGX Spark cold prefill, single-Spark tests (NVIDIA developer forum): https://forums.developer.nvidia.com/t... DGX Spark raised to $6,950 (The Register): https://www.theregister.com/systems/2... DGX Spark $3,999 to $4,699 (NVIDIA forum): https://forums.developer.nvidia.com/t... Surface Laptop Ultra price, stock and holiday returns (Microsoft Store): https://www.microsoft.com/en-us/store... Surface Laptop Ultra footnote 5, GPU memory “less than the total”: https://www.microsoft.com/en-us/surfa... Surface RTX Spark Dev Box: https://www.microsoft.com/en-us/surfa... MacBook Pro M5 Max, 614 GB/s (Apple): https://www.apple.com/macbook-pro/specs/ “99% GPU performance on battery” claim (Gizmodo): https://gizmodo.com/surface-laptop-ul... Bosgame M5 128GB, $3,099: https://www.bosgamepc.com/products/bo... Strata v0.1.40, Strix Halo prefill: https://github.com/Niko1221/Strata/re... New MacBook Pros around October 27 (MacRumors, Bloomberg report): https://www.macrumors.com/2026/10/08/... Every speed in this video is someone else’s published number. Microsoft’s figures are Microsoft’s test of Microsoft’s model. Prices as of October 11, 2026. Already ordered one, or running a Ryzen AI Max+ 395 instead? Tell me in the comments. #RTXSpark #LocalAI #SurfaceLaptopUltra