NVIDIA RTX Spark AI PCs Start Shipping This Month

Starting this month, you can walk into a store and buy a desktop PC with the kind of AI horsepower that needed a server rack just two years ago. NVIDIA's RTX Spark machines — built around a new chip called GB10 — are now shipping from Lenovo and Acer, with ASUS, Dell, HP, MSI and Microsoft's own Surface line expected to follow through October.

What's actually inside the box

The GB10 is what NVIDIA calls a "superchip": it fuses a 20-core Arm-based Grace processor with a Blackwell-generation GPU on one package, sharing 128GB of unified memory between the two. In plain terms, that means the CPU and GPU aren't fighting over separate pools of RAM the way they do in a normal laptop — the whole chip acts like one big, fast workspace for AI models. NVIDIA rates it at roughly one petaflop of AI performance, a measure of how many trillion calculations it can do per second, which is the kind of number that used to belong to data-centre hardware, not something sitting under your desk.

That's enough memory and horsepower to run AI models with up to around 200 billion parameters — the dial-settings that make a language model smarter but also hungrier for memory — directly on the machine, no cloud server required.

Why "local" is the actual headline

The appeal isn't just raw speed. It's where the computation happens. Running an AI model locally instead of sending every query to a cloud API can cut response times dramatically and removes the recurring per-query cost that adds up fast for anyone building AI products. NVIDIA is pairing the hardware with NIM — its library of pre-packaged AI models — so developers can get a chatbot or image model running on the machine in minutes rather than wrestling with setup.

Local inference keeps data on the device, cuts out per-token API bills, and can respond many times faster than a round trip to a cloud server.

For anyone who's felt the sting of a ballooning OpenAI or Anthropic API bill while prototyping, that pitch lands differently than the usual "faster GPU" sales pitch.

Where India fits in

This matters more for Indian developers than the press releases let on. A big chunk of India's AI startup ecosystem — from small language-model teams in Bengaluru to healthcare-AI founders building diagnostic tools — currently runs on rented cloud GPUs billed in dollars, which is brutal on the rupee when the exchange rate moves against you. A desktop that can run a 200-billion-parameter model locally, one-time hardware cost instead of a metered cloud bill, changes the math for a founder bootstrapping a product. It also has a privacy angle that's increasingly relevant under India's Digital Personal Data Protection (DPDP) Act: keeping sensitive user data — medical records, financial documents, whatever a startup is processing — on a machine in the office rather than shipping it to a server in another country sidesteps a whole category of compliance headaches.

The catch is price. Early estimates put RTX Spark machines somewhere between $2,500 and $3,000 before any India-specific pricing or import duties are added, which likely pushes the Indian retail price well past the ₹2.5 lakh mark. That's a serious outlay for an individual developer, though it's a fraction of what a cloud GPU cluster costs over a year, and the kind of number an AI-focused startup or a university research lab could justify more easily than a solo coder.

A crowded, fast-moving shelf

RTX Spark isn't arriving in a vacuum. It follows NVIDIA's DGX Spark, a similar but more workstation-oriented machine the company put out roughly a year earlier, and it's landing right as Intel and AMD push their own "AI PC" chips with on-device neural processing units. The difference with RTX Spark is scale: this is aimed squarely at people who want to run genuinely large models, not just handle Windows' built-in AI features a little faster.

• Lenovo and Acer are first to ship, in October 2026
• ASUS, Dell, HP, MSI and Microsoft Surface are expected to follow within the same window
• 128GB of unified memory supports local models up to roughly 200 billion parameters
• Pricing is expected to start around $2,500–$3,000, with no official India pricing confirmed yet

Whether this becomes a mainstream category or stays a niche tool for AI builders depends less on the chip's specs and more on whether software catches up — most people don't have an obvious everyday use for 128GB of unified memory yet. But for the growing number of Indian teams trying to build AI products without a Silicon Valley cloud budget, having this much compute available as a one-time purchase, sitting on a desk rather than metered by the hour, is the more interesting story than the petaflop number on the box.