# NVIDIA RTX Spark Laptops and DGX Spark Systems Bring Private, Zero-Subscription AI to the Desk

NVIDIA has rolled out its complete lineup of local AI hardware at GTC Berlin—led by the new **RTX Spark** Windows laptops and **DGX Spark** compact Linux desktop systems—designed for professionals, developers, and small businesses who want to run powerful AI models and personal agents on their own desk without paying monthly cloud token bills or uploading private company files to the internet.

Until now, anyone who wanted to run a capable coding or reasoning AI model locally on a Windows PC ran into a brick wall: graphics card memory (VRAM). Even a high-end gaming laptop typically topped out at 12 GB to 16 GB of video memory, which is only enough to run small, watered-down AI models that struggle with long documents or complex tasks.

NVIDIA's new **RTX Spark** laptop platform breaks that ceiling by packing up to **128 GB of unified memory** and **1 petaflop of AI compute** into portable laptops shipping now from ASUS, Dell, HP, Lenovo, Microsoft, and MSI (while still playing games at 1440p and 100 frames per second). With 128 GB of memory available right on the laptop, a lawyer, financial analyst, or software developer can run 70-billion-parameter models like NVIDIA's Nemotron 3.5 Bolt 74B completely offline on an airplane.

For small teams and engineering offices that want a shared "private ChatGPT box" sitting on a desk, NVIDIA also began shipping new **64 GB DGX Spark** mini-systems on October 23, alongside a one-click tool called **NVIDIA Sync Model Launcher** that downloads a model and hooks it up to coding tools like OpenCode in a few clicks—no command-line Linux headaches required.

## Fast Facts: NVIDIA's New Local AI Hardware Lineup (RTX Spark, DGX Spark & DGX Station)

- **Announcement & Rollout**: Unveiled at GTC Berlin and detailed in [NVIDIA's official local AI hardware announcement](https://blogs.nvidia.com/blog/local-ai-gtc-rtx-dgx-spark-station-sync/).
- **RTX Spark (Windows AI Laptops)**: Up to **128 GB unified memory** and **1 PFLOP** AI performance (+ 1440p 100 FPS gaming); available now from ASUS, Dell Technologies, HP, Lenovo, Microsoft, and MSI (with Acer arriving soon), featuring unlocked bootloaders on select models for custom local agent builds.
- **Underdog Private Desktop Agent**: A new 100% local desktop agent app powered by the Husky inference engine is available now for download on RTX Spark, running personal workflows without user data ever leaving the PC.
- **DGX Spark (Compact Linux Appliance for Devs & Small Teams)**: New **64 GB** systems started shipping October 23 from Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, MSI, and NVIDIA; runs Nemotron 3.5 Bolt 74B (24GB), Qwen3.8 27B, and Gemma 4 26B-A4B out of the box.
- **Scaling Up to 1 TB for Small Teams**: Up to **8 DGX Spark** units can be linked together via built-in ConnectX-7 networking and the NVIDIA Sync Cluster Assistant to create **1 TB of unified memory**, running frontier-class models like GLM 5.3, DeepSeek v4 Flash 0731, and Qwen 3.8 Flash Next for a whole team.
- **DGX Station (Office-Scale Server)**: Packs **748 GB of coherent memory** and **20 PFLOPs** to serve **30+ employees simultaneously** at cloud-like speeds on Linux (with a Windows edition arriving in December 2026) and integrates with Microsoft Intune, Splunk, and Tanium for standard IT management.

## Hardware Buyer's Guide: Which Local AI Machine Fits Your Work?

If you are trying to decide whether to buy an AI laptop, a desktop mini-appliance, or an office server to cut cloud AI bills and keep client data private, the table below compares NVIDIA's three local hardware tiers.

| Hardware Tier | Memory & Compute | Models It Can Run Offline | Who Should Buy It |
|---|---|---|---|
| **RTX Spark Laptop** (Windows PC) | Up to **128 GB unified RAM** / 1 PFLOP | **Nemotron 3.5 Bolt 74B**, Underdog local desktop agent, local voice/video editors | **Individual Power Users, Consultants, Creators & Devs** who need 100% private AI on the go + 1440p gaming |
| **Single DGX Spark** (Compact Linux Box) | **64 GB** (or 128 GB) unified memory | **Nemotron 3.5 Bolt 74B (24GB)**, **Qwen3.8 27B**, **Gemma 4 26B-A4B** | **Solo Developers & Small Studios** wanting a quiet, dedicated local AI box on their desk |
| **Clustered DGX Spark (2 to 8 Units)** | Up to **1 TB unified memory** (via ConnectX-7) | **GLM 5.3**, **DeepSeek v4 Flash 0731**, **Qwen 3.8 Flash Next** | **5-to-15 Person Engineering or Legal Teams** sharing one local AI backend for concurrent coding/document sessions |
| **DGX Station** (Linux now; Windows Dec 2026) | **748 GB coherent memory** / 20 PFLOPs | Full frontier open models (**GLM 5.3**) serving **30+ concurrent users** | **Mid-Sized Companies, Hospitals & Law Firms** replacing $5,000+/month cloud API bills with on-premise hardware |

## Why Local AI Hardware Suddenly Makes Practical Sense for Everyday Work

For the past three years, running AI on your own computer had two fatal flaws: the hardware didn't have enough memory to run smart models, and the software required hours of wrestling with Python environments and Docker containers. NVIDIA's October rollout fixes both problems at once.

First, look at the **software convenience**. Most people don't want to spend their weekend configuring Linux drivers just to summarize PDFs or get coding help. On Windows **RTX Spark** laptops, users can download **Underdog**, a ready-to-use personal desktop agent powered by the Husky engine that organizes files, drafts documents, and automates tasks completely offline. On **DGX Spark**, the new **NVIDIA Sync Model Launcher** (rolling out this month alongside a new Wi-Fi Assistant for appliance mode) lets you click a button on your regular laptop, automatically downloads the right AI model and container onto the DGX Spark, and hooks it straight into **OpenCode** with built-in troubleshooting skills.

Second, look at the **team math**. Suppose a 10-person software agency or boutique law firm spends $100 to $200 per person every month on premium cloud AI plans—or worse, isn't allowed to paste confidential client contracts into cloud chatbots at all. By connecting two to eight **DGX Spark** boxes on a shelf in the office using their built-in ConnectX-7 cables and the **NVIDIA Sync Cluster Assistant**, that team gets up to **1 TB of shared memory**. Every person in the office can connect their laptop to that local box at the same time, running heavy-duty models like **Qwen 3.8 Flash Next** or **GLM 5.3** with zero per-token fees and zero data ever leaving the building.

For larger departments of 30 or more people, the **748 GB DGX Station** plugs right into the exact same IT security tools companies already use for employee laptops—including **Microsoft Intune**, **Cisco Splunk**, **Tanium**, **Puppet**, and **Chef**—and gets a native **Windows version in December 2026**.

## Practical Steps to Switch From Cloud Subscriptions to Local AI

Whether you are upgrading your personal work laptop this fall or setting up a private AI server for your office, follow this four-step checklist.

### The 4-Step Enterprise Implementation Playbook

1. **Prioritize Unified Memory (64 GB–128 GB) Over Raw GPU Clock Speed on Your Next Laptop**: If you want your next Windows laptop to run serious AI agents locally for the next three years, skip standard 16 GB VRAM laptops and pick an **RTX Spark** configuration with **64 GB to 128 GB of unified memory** (available from Dell, Lenovo, HP, ASUS, Microsoft, or MSI). Memory capacity determines whether a 70B-class model fits on your machine at all.
2. **Use Underdog and NVIDIA Sync Model Launcher Instead of Manual Setup**: Don't waste time building custom local inference servers from scratch. Install the **Underdog** desktop agent on RTX Spark for private personal workflows, or use **NVIDIA Sync Model Launcher** on DGX Spark to connect **Nemotron 3.5 Bolt 74B** or **Qwen3.8 27B** to **OpenCode** in three clicks.
3. **Start With 1 or 2 DGX Spark Boxes and Scale as Needed**: Small businesses don't need to buy a massive data-center rack upfront. Start with a single **64 GB DGX Spark** appliance on your office Wi-Fi; if your team's usage grows or you want to run giant models like **GLM 5.3**, daisy-chain additional units (up to 8 boxes / 1 TB total memory) using the built-in **ConnectX-7** ports and Sync Cluster Assistant.
4. **Enroll Office AI Appliances in Microsoft Intune**: If you deploy a **DGX Station** (748 GB) for a 30-person department, have your IT admin register it through the new **Microsoft Intune**, **Tanium**, or **Splunk** integrations so your local AI server gets the same automatic security compliance checks as every other computer in the office.

## Next Steps

- **Add Up Your Team's Current Monthly AI Bills**: Total what your team spends each month on ChatGPT Pro/Dots ($100–$200/seat), Claude, and API credits; if 10+ employees are burning over $1,500/month or holding back sensitive client work due to cloud privacy rules, a shared **DGX Spark** setup will pay for itself quickly.
- **Test the Underdog Local Agent on Windows**: If you pick up an RTX Spark laptop from Lenovo, Dell, HP, or ASUS this month, download the free **Underdog** desktop agent app first and test it on private local spreadsheets and PDFs with your Wi-Fi turned off.
- **Watch for the December Windows DGX Station Release**: If your office IT team runs an all-Windows shop and wants to serve 30+ users from a single 748 GB box without touching Linux, earmark the **December 2026 Windows release of DGX Station** for your Q1 hardware refresh.