Microsoft × NVIDIA Start Selling Surface Laptop Ultra: Runs 120B Models Locally, Top Config ~$7,000

What Happened

The numbers first: the event took place October 7 at 1 PM ET in San Francisco, with Microsoft CEO Satya Nadella, NVIDIA CEO Jensen Huang and Windows & Surface chief Pavan Davuluri on stage. The two companies announced their partnership on May 31. RTX Spark shrinks NVIDIA's Grace Blackwell datacenter superchip down to a 3nm consumer SoC — a 20-core Arm Grace CPU and a 6,144-core Blackwell GPU on one die, linked by NVLink-C2C, sharing up to 128GB of LPDDR5X unified memory (~300GB/s bandwidth), with 1 petaflop of AI compute at FP4 sparse precision. Microsoft's claim: this 15-inch laptop runs 120-billion-parameter models locally with 1M-token context. Two SKUs: RTX Spark S2 (18-core CPU / 5,120-core GPU / 24–32GB memory) and S3 (20-core CPU / 6,144-core GPU / 32–128GB memory). The display is a 15-inch mini-LED PixelSense Ultra (2,000 nits peak HDR), the haptic trackpad is ~30% larger than the previous generation, and ports include HDMI / USB-C / USB-A / SD card reader / headphone jack; under 18mm thick, under 4.5 lbs (~2 kg).

Two layers to the launch details: multiple outlets report pre-orders opened on October 7 with first shipments in mid-October. But Microsoft has not announced official pricing — leaker Roland Quandt puts it at "under ~$3,000" to "over ~$7,000"; Windows Central says the entry model starts "well above $2,000." Also announced: Snapdragon X2 Plus versions of the Surface Pro 12-inch ($1,149.99) and Surface Laptop 13-inch ($1,199), on sale October 13 in select markets. Microsoft also announced Windows on Arm has crossed 7,000 verified apps, and dropped the Copilot+ PC branding for the Ultra. By Microsoft's own account, the machine is still in "pre-release" status — the launch isn't final until regulatory approvals clear.

Key Facts

1. "1 petaflop" is an FP4 sparse figure — real-world speed is untested: Microsoft's 1 petaflop uses FP4 low precision plus sparsity (skipping unnecessary math). As gagadget puts it: "Real-world speeds are untested." RTX Spark's unified memory bandwidth is ~300GB/s.
2. The 24GB entry model can't run 120B: The entry S2 has only 24GB of unified memory — by wccftech's math, it can only run 4-bit quantized Qwen3.8-27B-class models; running 120B requires the 64GB / 128GB tiers. The leaked top-config price of ~$7,000 is in MacBook Pro M5 Max territory; Windows on Arm app compatibility is "not perfect yet" by Windows Latest's own testing.
3. Announced May 31, on sale October 7 — four and a half months later: The machine was already shown at Computex in June, while Microsoft withheld pricing and the ship date. After the event, official pricing is still unannounced. Pre-orders opening October 7 with mid-October shipments is currently per "multiple outlets."
4. Microsoft, NVIDIA and memory costs: Microsoft wants to move part of the AI compute currently running in Azure datacenters onto machines the customer pays for; NVIDIA enters the Windows PC market long held by Intel / AMD. NVIDIA's DGX Spark recently rose ~75% to $6,950 (128GB). Moor Insights' Anshel Sag told Reuters that when Microsoft pitched "local AI on PC" two years ago the hardware was ready but the software wasn't — now the software is here and the hardware is too expensive: "only people with budgets can afford local AI now."

Context

RTX Spark launched May 31 — the first time NVIDIA's Grace Blackwell architecture came down to consumer silicon. Back then Microsoft and NVIDIA promised to "reinvent the Windows PC for the age of personal AI"; at Build 2026, Microsoft customized Windows 11 for it with Workload Profile Scheduling (task scheduling across the 20 cores) and a tuned Prism x86 emulation layer. Until today, both the Surface Laptop Ultra and the Surface RTX Spark Dev Box (a developer mini-PC) were "pre-release." Today's event was Microsoft's first major Windows / Surface showcase in two years. Lenovo's Yoga Pro 9n / Yoga 9n 2-in-1 and ASUS ProArt P14 / P16 with RTX Spark are also on the way.