Jensen Huang Bets on AI PCs, but the Best Host for Agents Remains Disputed

Image source: unsplash

"The PC needs to be reinvented."

On October 21 local time, Nvidia CEO Jensen Huang declared at the GTC conference in Berlin that the old paradigm of "input instructions, computer executes" has come to an end, and that only by becoming a personal Agent that understands user intent and proactively executes tasks can the PC meet the demands of the AI era.

Founded in 1993 on the eve of the PC revolution, Nvidia, after more than 30 years of ups and downs, has become the most important driver of the AI revolution, and has also transformed from a startup team that "worked out of an apartment" and "never knew where the next meal would come from" into the world's most valuable listed company.

English text is being prepared.

At the Berlin GTC conference, Huang also showcased Nvidia's "PC redesigned for the AI era," the RTX Spark, and said that for this device the company secretly collaborated with Microsoft for four years, investing thousands of "engineer-years."

However, can the PC really be the best host for an Agent? As to whether this "old bottle" can truly make the "new wine" taste better, there are still many differing views.

Agents Lead the Wave of AI Localization

The way humans and computers interact is changing rapidly—so rapidly that most people cannot yet adapt.

Since the beginning of this year, from the "crayfish" OpenClaw to OpenAI's Codex, and then to Tencent's WorkBuddy and Meta's Muse, new Agents have emerged one after another, setting off wave after wave of excitement. But in terms of deployment and use, in addition to the threshold that still exists for ordinary consumers, the more prominent problems are privacy, security, and cost.

Against this backdrop, models and devices are also making adjustments in tandem, with the more notable manifestations being the emergence of small models, local models, AI PCs, and AI phones. At present, the core logic considered by the industry is still to use the devices consumers are most familiar with, equipped with models that perform better on security, privacy, and cost, so that more personalized Agents can run locally.

Judging from the information Nvidia has released so far, its product logic is consistent with this.

Relevant information shows that in the collaboration between Nvidia and Microsoft, the former provides GPUs, the AI software stack, and local model running capabilities, while the latter attempts to provide execution, isolation, and permission management mechanisms for Agents at the Windows operating system level.

In May this year, Microsoft announced Windows security capabilities for Agents; Nvidia, through tools such as OpenShell, is trying to enable local Agents to run within controllable permission boundaries. An article on Nvidia's official website said that what the two sides hope to solve is not only whether AI can run on a PC, but whether it can continuously execute tasks in real work environments. Microsoft is also advancing mechanisms such as execution containers to limit the scope of Agents' access to files, networks, and system resources.

According to information released by Nvidia, the RTX Spark adopts a chip design that integrates CPU and GPU, provides up to 128GB of unified memory and up to 1 petaflop of FP4 AI computing power, allowing users to run larger-scale local models and, compared with cloud models, use fewer tokens to enable Agents to complete more complex scenarios such as development, design, and game production. At the conference venue, Huang also brought onto the stage the Hugging Face MicroDuck "little yellow duck" robot that he is quite fond of, and introduced that the "brains" of these robots were built on the RTX Spark and that Agents drive them to walk and respond to instructions, thereby proving the powerful AI capabilities of its new-generation PC.

Huang also envisioned that in ten years people's understanding of the PC will be completely different, just as today people still use the term "phone" but rarely use the device to "make calls." In his view, every household may in the future be equipped with a super AI computer that runs Agent tasks around the clock.

However, in discussions about this GTC, there are still many voices mentioning that "the PC is only an expedient," arguing that the Agent boom may not necessarily catalyze AI PCs into the mainstream, and that what will truly enable Agents to land on personal devices and in local scenarios in the future may be other models.

Finding a "Host" for Agents, or Letting Agents "Take Charge"?

At present, the end-side and local intelligent applications that have had longer time and more validation are vertical application scenarios such as smart cockpits and industrial sensors. In terms of consumer adoption, phones and PCs remain the mainstay, while AI glasses, smartwatches, smart bands, rings, and even hats and other wearable devices are also frequently "launched."

Since the beginning of this year, the biggest change has been the rise of box products. In both home and work scenarios, quite a few new box products have emerged. The main feature of these boxes is that they run Agents locally 24/7, which in fact caters to the idea of changing the mainstream human-computer interaction method of "looking at a screen—manual input." In addition, the explosion of embodied intelligence has brought the vision of robots truly entering human life one step closer to reality, but that is still more regarded as the next stage after Agentic AI, namely the era of physical AI.

All the aforementioned devices have their own adherents, and all are promoting the Agent concept, but at the same time each faces its own bottlenecks. Notably, in the current discussion of the "best host" for Agents, more and more voices mention that even in local scenarios, what matters more is not finding a single carrier for the Agent, but how to let the Agent connect to as many trusted devices as possible. In other words, it is about letting the Agent take the lead within an authorized scope.

The box model mentioned above, including workstations, edge servers, and the like, is also one implementation of this approach.

Its core is to put the Agent in the background to achieve "seamless" and "always-on" operation, ensuring the Agent is present and forms memory and intelligence, while phones, PCs, or other devices serve more as the entry layer and presentation layer. Even if devices are replaced, or even if some category of device is replaced by another in the future, the Agent itself is unaffected.

A practitioner in the edge-side chip industry said that the chips his company mass-produced this year were originally designed for AI PCs, but in terms of shipments, purchase orders for edge servers and box-type products instead surged, growing far faster than AI PCs. He believes this reflects both the factor of weakening sales in mainstream consumer electronics due to memory price increases and the fact that new product categories unleash imagination and iterate faster than traditional categories.

In addition, it is said that with the explosion of personal agents triggered by Muse, more and more box products are also exploring paths for better collaboration between local Agents and edge-side Agents networked within devices.

Besides box-type products, another widely discussed approach is the "virtual host" model.

In the past, one of the core advantages of the rise of PCs was that the CPU, memory, storage, and software were all concentrated in one machine, making it easier to use across various scenarios once hardware and software capabilities were up to par. But Agents can precisely break through this centralized structure: users' old computers, game consoles, home servers, and even other devices with computing power can all become distributed compute nodes: one device runs a local model, another handles vision tasks, and the main computer continues to handle daily work. At present, more and more vendors focused on local scenarios are offering solutions that connect multiple PCs in a personal network and allocate AI inference work according to task requirements.

Of course, it is even more likely that the better model for hosting Agents has not yet truly been thought of or seen.

The aforementioned edge-side chip practitioner said that people are accustomed to looking at problems and making predictions along the extension line during technological change, seeking the next outlet or solution, but historical experience has already proven that the true "hit products" and "answers" are always beyond the extension line. (This article was first published on TMTPost APP, written by Feixiang Tai Kong)