What This Is
RTX Spark is rumored to launch on October 7, featuring a GB10-class chip, 24GB to 128GB of VRAM, priced at $1,800-$2,900 (roughly 13,000-21,000 RMB). What we care about: NVIDIA is packing compute capable of running 70B (70 billion) parameter large models into a ~$20,000 device—a key signal that AI compute is moving from data centers to offices.
It's a stripped-down version of DGX Spark (NVIDIA's flagship compact AI workstation for developers), with the professional-grade high-speed networking interface ConnectX-7 removed, meaning no multi-node clustering—but single-machine local inference (running a trained model to generate answers) is largely unaffected.
Industry View
The LocalLLaMA community on Reddit defines it as "a DGX Spark without the pro interfaces." The optimistic read: roughly $2,900 for a complete 128GB VRAM machine—half the price of equivalent local-LLM-capable hardware. In a market where a used 4090 still costs ~$20,000, this is a clear signal of AI compute democratization for enterprises.
But cautious voices are worth hearing: this remains community rumor—NVIDIA has not officially confirmed it. DGX Spark's recent significant price hikes show NVIDIA holds strong pricing power, meaning RTX Spark's actual shipping price and configuration may diverge meaningfully from these leaks. More critically, while the hardware barrier drops, actually deploying and fine-tuning these models still requires specialized talent—a hidden cost far exceeding the machine itself.
Impact on Regular People
For SMB IT: Companies that don't want to send customer data to the cloud but want on-prem private Q&A systems can now see "local LLMs" shift from "100,000+ RMB data center setups" to "20,000 RMB workstations." But don't rush to order—wait for the official launch and real benchmarks.
For individual professionals: 24GB-128GB of VRAM in a ~$20,000 machine means individual developers and product managers can run open-source models like Llama on their desktops for prototyping—no more dependence on cloud APIs. The tradeoff is heat and noise—these devices are almost certainly not silent.
For consumer markets: In the short term, little impact on average consumers. But if local inference costs continue falling, "AI devices" could extend from phones to home NUCs (mini PCs)—a path running parallel to cloud-based conversational products.