"From Apple PCC to NVIDIA DGX"... The Business Landscape to be Changed by Enterprise 'Private AI'

In 2027, the core of AI technology is expected to move beyond simple chatbots to 'AI agents' that collaborate with humans. Analysis suggests that corporate organizational structures will be reorganized into organizations that collaborate with agents, organizations that work with chatbots, and organizations that work with people. According to a T Times TV video, as companies enter the era of 'Private AI' by building their own AI servers due to security and cost issues, the roles of companies related to GPUs and data centers are becoming increasingly important. In particular, if companies come to possess their own GPUs, the landscape of Big Tech is also expected to change.

Apple's Layered AI Strategy, from On-device to Private Cloud

Apple is recently strengthening the AI functions of Siri and building its own independent AI ecosystem. In the video, Park Jong-cheon, CAIO of JiranJigyo Soft, explained that Apple is taking a strategy of operating models in layers. First, in the on-device environment, Apple utilizes its latest chipsets such as the A20 to process most tasks inside the device without leaking personal information through lightweight models like the 'AFM3 Core'. In this stage, personal data such as photos or text can be processed without being sent outside the device.

However, on-device alone has limits in processing complex calculations. To solve this, Apple utilizes 'PCC'. PCC is a data center directly managed by Apple, characterized by maximizing security as it is built based on high-performance silicon such as the M5 Max. It is expected that data centers combining CPU, GPU, and NPU, which are advanced server-grade silicon, will be built within PCC. The speaker noted, "It is important for Apple to take the lead in gaining experience by dividing models into layers and mixing hardware and cloud," suggesting that this technology stack could expand into business-oriented Siri AI in the future.

This strategy of Apple also provides implications for the enterprise market. This is because cases may emerge where companies adopt devices such as NVIDIA's DGX Spark (in the 10 million won range) or use Mac mini running 24 hours a day as 'token factories' to reduce API costs. Ultimately, the trend of personal AI moving from on-device to Apple PCC, and then to cloud AI such as Google Cloud, is expected to align with the Private AI construction strategies of corporations.

NVIDIA's Entry into the Model Market and the Rise of China's Cost-Effective AI

The moves of NVIDIA, a powerhouse in the AI infrastructure market, are also changing. NVIDIA, which previously focused on GPU supply, has recently stepped into the model market by directly presenting AI models such as 'Nemotron 3 Ultra'. This is interpreted as a strategy to directly provide models that can deliver optimal performance on specific hardware when companies possess their own GPUs and operate Private AI. NVIDIA's direct creation of models is intended to provide a powerful ecosystem to companies using GPUs.

In this flow, the 'cost-effectiveness' strategy of Chinese AI models is also drawing attention. According to the video, China's 'QN 3.8(27B)' model delivers performance comparable to high-performance models from the United States while maximizing cost efficiency by drastically lowering hardware requirements. For example, if a machine worth 1 billion won is required to run Nemotron, it is analyzed that the QN model can run on a machine worth 50 million won. The analysis suggests that US GPU sanctions have instead prompted Chinese companies to develop high-efficiency models that run well even on low-specification hardware, triggering a cost-effectiveness war that maintains performance while lowering prices to one-twentieth of the level.

The Core Challenge of Enterprise AI: 'Data Security Grades' and Permission Management

When the era arrives where companies grant all work permissions to AI agents, security issues are expected to be the biggest obstacle. The speaker emphasized the need for a system to classify the security grades of internal corporate data. Citing the government's N2SF (Data Security System), they explained that data should be managed by dividing it into Classified, Sensitive, and Open grades.

For example, a CEO can access data of all grades, a team leader can see up to the Sensitive (S) grade, and a general employee can only see up to the Open (O) grade, controlling permissions in this manner. In this case, the AI must support work while complying with the user's permissions. The speaker predicted, "When a user asks for information they do not have permission for, an intelligent security engine that provides 'hints' helpful for work, even if it does not disclose the direct content, will become the core of future corporate security." This refers to advanced security technology where the AI identifies correlations between data and provides an appropriate level of response without infringing on the user's permissions.