‍#分享美股交易赢英伟达股票

The Business Opportunities Behind Nvidia's "Intelligent Agent Computer": Data Never Leaves the Machine, Tokens Are Completely Free

Earlier this month, Nvidia founder Jensen Huang announced the world's first Windows PC (personal computer) super chip designed specifically for personal AI agents—RTX Spark. This release is seen as the most disruptive transformation in the PC industry since the widespread adoption of Windows graphical interfaces in the 1990s. Huang even called it "the first complete redesign and reconstruction of the personal computer in 40 years."

Perhaps it is premature to declare that the PC has been completely overturned, but it is foreseeable that the definition of the PC is loosening. A senior industry insider told reporters that hardware manufacturers are fully capturing AI opportunities, with new designs emerging constantly, and the edge ecosystem showing vigorous vitality.

Changing the Architecture of Computers for 40 Years

RTX Spark is not Nvidia's "solo effort," but the result of assembling the strongest global tech ecosystem alliance. The chip was co-designed by Nvidia and MediaTek, manufactured using TSMC's 3-nanometer process, with deep system-level optimization involving Microsoft.

This chip is not a traditional standalone CPU or GPU but a highly integrated system-on-chip (SoC). Its core architecture follows the "heterogeneous fusion" concept, integrating multiple core components that were originally dispersed on the PC motherboard into a single chip. Its most revolutionary technological innovation is the unified memory architecture—traditional PCs use separate system memory for the CPU (central processing unit) and VRAM for the GPU (graphics processing unit), requiring data to be transferred back and forth, causing latency and power consumption. RTX Spark allows the CPU and GPU to share a maximum of 128GB of memory pool, with intra-chip bandwidth up to 600GB/sec enabled by Nvidia's second-generation interconnect technology. This design completely eliminates data transfer bottlenecks, allowing the GPU to directly access large-capacity system memory, providing critical support for running ultra-large AI models.

Nvidia officially provided performance metrics for RTX Spark across multiple core scenarios: locally running a 120 billion parameter large language model (LLM) supporting 1 million tokens (word units) of context; running AAA games at 1440p/100FPS; smoothly editing 12K 4:2:2 videos. Among these, the ability to locally run a 120 billion parameter model is particularly crucial—users can run a GPT-3.5 level large language model on a laptop without network connection or cloud fees, supporting entire novels or hundreds of pages of professional documents in context.

From "Operating Software" to "Expressing Needs"

Huang mentioned in his speech that the core difference between AI computers and traditional computers is: "Over the past 40 years, everyone has opened applications, clicked, and input commands to complete work. With RTX Spark, users only need to express their needs, and the PC will handle the rest."

This signifies a shift in computer design philosophy. In the past, computers were always designed around "people," but in the future, people will design computers for AI intelligent agents. Digital economist Liu Xingliang said that people need to shift from "knowing how to operate" to "knowing how to express," and that the most important skills for individuals in the future may become asking questions, judgment, creation, and decision-making.

Recalling the nearly simultaneous "Lobster" (OpenClaw) installation and uninstallation wave not long ago, the underlying reason was uncontrollable token bills and security concerns. Local deployment of large models perfectly avoids these two pain points—running a 120 billion parameter model locally, data never leaves the machine, with unlimited runs and no cloud fees.

In fact, since 2023, AI PC products have already appeared on the market, but why is this year regarded as the dawn of the AI PC era? The key lies in consumers' willingness to pay. Early AI PCs had only 10–40 TOPS (trillions of operations per second) of computing power, capable of running small models below 7 billion parameters, with limited practicality; subsequent products increased to 40–60 TOPS, capable of running models with 13–30 billion parameters, but AI was still like a plugin. RTX Spark, from the chip architecture level, is specially designed for AI agents, with a local AI computing power of 1 PetaFLOP (equivalent to 500 TOPS), supporting 120 billion parameter large models— for ordinary consumers, using large models on PCs now feels almost as good as in the cloud, with better security and no token fees, significantly boosting purchase intent.

The market responded quickly: on the day RTX Spark was released, Nvidia's stock price rose by 6.26%, ARM surged by 15.73%, while Intel and Qualcomm fell by 4.67% and 8.78%, respectively.

Running "Full-Strength" Large Models Locally

This release excited a Chinese chip company called Houmo Intelligent. Previously, they struggled to persuade the market to accept bringing large models to the terminal. The design concept of RTX Spark aligns with their M50 chip—whoever can run a more "full-blooded" large model on a computer will have the ability to define the entire AI PC category's rules.

Nvidia's approach is to integrate CPU and GPU on a single chip, while Houmo Intelligent's M50 is a compute-in-memory architecture chip. Both share the common point that through architectural innovation, AI computing power is upgraded from an auxiliary function to the core dominant capability of the chip.

Lenovo Launches AI Host P7

Last month, Lenovo announced the AI host P7 equipped with Houmo Intelligent's M50 chip, capable of smoothly running a local large model with up to 122 billion parameters within a palm-sized body. Houmo Intelligent's VP of PC sales, Ni Xiaolin, said that the M50 is an efficient chip designed specifically for large model inference, capable of achieving 160 TOPS at just 10 watts of low power consumption, combined with large memory and high bandwidth, perfectly meeting the dual needs of low power and high performance for local deployment.

Not long ago, Shanghai-based "one-person company" Liaoqi Intelligent partnered with Houmo Intelligent to create a holographic interactive personal computing center, which can organize schedules, summarize notes, and serve as a conversation partner. The product has been deployed in a nursing home in Huangpu District, allowing elderly residents to interact with distant family members through holographic digital images.

A NAS (Network Attached Storage) company also hopes to become the next-generation home hub. They aim to leverage the large model capabilities integrated into the M50 to mobilize personal data stored in NAS devices, turning it into a more powerful personal data center.

According to China Academy of Information and Communications Technology data, in Q1 2026, the penetration rate of AI PCs in China will reach 42%, a year-on-year increase of 380%. We can either wait for the "rule-changing" AI PC to launch this fall or look forward to a blooming variety of smart terminal products.