An AI agent on the laptop: pay once instead of a monthly subscription
Nvidia's RTX Spark brings 128GB of memory and large language models to the desk, and the subscription model of AI providers gets a rival.
Since 16 October, Asus, Dell, HP, Lenovo, Microsoft and MSI have been selling laptops with Nvidia's new RTX Spark chip, which offers up to 128GB of shared memory and, according to Nvidia, up to 1 petaflop of AI computing power. Language models with 20 to more than 100 billion parameters and AI agents can run directly on the device, without data going to a cloud.
What happened
Until now, capable AI has been a service: data centres compute, customers pay per month or per token. RTX Spark moves part of that onto the device. In tests by StorageReview, a locally run coding model with 35 billion parameters produced 89 tokens per second, fast enough for daily work. Windows Central had an agent sort files without using a single cloud token. At GTC in Berlin on 21 October, Nvidia showed the full range, from the laptop to the DGX Spark mini computer and the DGX Station with 748GB of memory, which can run very large open models. For companies, data protection matters most: code, contracts and customer data stay on their own device. Assessment: this is not yet meant to be fully self-sufficient. Microsoft explicitly calls the approach hybrid, local when it makes sense and cloud when needed. The strongest models of the large providers still run only in their data centres.
What follows
The subscription does not disappear, but it gains an alternative with different economics. An RTX Spark laptop costs from $2,599 in the US; the 128GB configuration tested by StorageReview costs about $5,900. In return, usage caps and running costs per request fall away. Assessment: the first winners are Nvidia and the device makers, along with providers of open models that may run locally, many of them from China, such as Qwen, DeepSeek and GLM. Business models that live only on reselling computing power come under pressure. Companies face a new task: if every employee has an agent on their laptop, rules are needed on what that agent may do. Nvidia and Microsoft supply sandboxed execution environments for this.
What changes in practice
- Companies
- Code, contracts and customer data can be processed locally without going to a cloud provider.
- Freelancers
- Entry from $2,599 in the US instead of running subscription costs. The strongest models of the large providers remain in the cloud, however.
- IT departments
- Every laptop with its own agent needs rules on which files and accounts the agent may access.
What to watch
- Prices and availability in euros at German retailers
- Whether the first companies adopt local agents instead of cloud subscriptions for sensitive data
- How the gap between local open models and the cloud models of the large providers develops
Sources
- NVIDIA Newsroom: RTX Spark PCs available today
- NVIDIA Blog: Local intelligence powers next wave of agents (GTC Berlin)
- StorageReview: Surface Laptop Ultra mit RTX Spark im Test
- Windows Central: RTX Spark review and benchmarks