Is Edge AI Booming? Google "Copies" Apple's PCC Private Cloud, and Domestic Mobile Phone Brands Are Also Learning Fiercely

Recently, despite the ongoing frenzy of the AI wave, the "AI bubble theory" has emerged from the periphery into mainstream discussions, and it has been endorsed by many analysts, institutions, media outlets, and even Sam Altman, the founder and CEO of OpenAI, and Jeff Bezos, the founder of Amazon. The core issue is that the capital and computing power investment in AI has skyrocketed, but the returns are still a long way off, which has made many people start to see the shadows of past bubbles again.

Amid these voices of reflection, Apple's name has begun to be mentioned more frequently. In the past two years, Apple was often criticized for being "lagging behind in AI," but now it has attracted much attention for its restraint and stability. Even Huawei, OPPO, Honor, and its old rival Google have partially recognized Apple's AI strategy to some extent.

On November 11th local time, Google released the November Pixel feature update as expected. At the same time, it officially announced its Private AI Compute platform (PAC for short), which has been first implemented on the Pixel 10 phone. For example, the Magic Cue function, which provides context-aware suggestions across apps, is already running on the private AI compute platform.

Pixel 10, Image source: Google

According to Google, the core goal of the PAC private AI compute platform is to "build private and useful AI." This is also the goal of Huawei's HPIC Personal Intelligence Computing, OPPO's PPC Private Computing Cloud, and Honor's HPPC.

However, all these easily make people think of Apple's PCC (Private Cloud Compute) proposed at WWDC24 last year - a new model that combines edge-side privacy with cloud performance.

WWDC24, Image source: Apple

In fact, it is indeed the case. According to the disclosure on the official blog, Google has also adopted a similar strategy for PAC. The difference is that in Apple's case, the AI running in the cloud is not that useful. So recently, Bloomberg reported that Apple plans to obtain the technical usage rights of Google's Gemini model at a price of $1 billion per year to run on the PCC platform to support the Apple Intelligence experience.

Google doesn't have such troubles.

True integration of edge and cloud, going one step further than Apple

To understand Google's PAC platform, we have to start with the problems it wants to solve.

In the era of generative AI, the most powerful models often can only run in the cloud, but the cloud also means risks. Once the data leaves the device, privacy may be exposed. In the past year, major manufacturers have deployed lightweight models on the edge side and hosted large models on the cloud side. But the real challenge is - how to make the two work together without letting the data cross the boundary.

PCC Private Cloud Computing is the answer Apple gave at WWDC24. According to the edge-cloud hybrid architecture of Apple Intelligence designed by Apple, before the device sends data, it first verifies whether the cloud is running a publicly audited system version. Then, the calculation is carried out in a protected "dedicated environment." After all tasks are completed, the data is immediately destroyed, and even Apple itself has no right to access it. The so-called "dedicated environment" is based on its self-developed chips and operating system in its own data center.

Google's PAC private AI compute platform has obviously made a bit of an "acceleration" on this line of thinking.

Google also emphasizes privacy, but it adopts a more cloud-native approach: PAC is built on Google's self-developed TPU (AI chip) and confidential computing infrastructure, combined with AMD's SEV - SNP hardware isolation technology and its self-developed Titanium security architecture. Every time a model is called, remote attestation is required to verify the node identity and system integrity.

Google Cloud data center equipped with TPU, Image source: Google

Only nodes that pass the verification will be allowed by the device to process user requests. At the same time, the input and output channels are encrypted throughout the process, and the intermediate inference data only exists in the memory and will not be stored on the disk.

To put it simply, it's like adding a "peep-proof shell" to the large model in the cloud - when a Pixel device sends a request to PAC, the data will be encrypted and sent to a "zero - privilege sandbox." Even Google's operations engineers cannot view it. After the task is completed, the data and cache are automatically destroyed, leaving no trace throughout the process.

More importantly, Google has truly embedded this mechanism into the system experience.

On the Pixel 10, when users use the Magic Cue function, the system will automatically call PAC to extract context from information such as text messages, calendars, and maps to generate next - step suggestions. Thanks to the capabilities of PAC, the built - in Recorder on the Pixel 10 also supports more voice functions, more accurate multi - language transcription, and better summarization.

Magic Cue function, Image source: Google

Different from the traditional "upload to the cloud and then calculate," PAC performs hardware verification, temporary encryption, and task isolation every time it is called. While ensuring the response speed, it also maintains the privacy boundary. The call logs of PAC can even be directly viewed in the "Settings - Developer Options."

The significance of this strategy is not only about security. When large models become more and more personalized and need to understand your living habits, "private AI" will become the prerequisite for the quality of the experience. PAC provides a new path for this - it can make the model smart enough without sacrificing privacy.

Before Google, Apple had already proposed this concept and framework. But from the TPU and Gemini in the data center (cloud) to the Tensor chip and Android on the Pixel terminal, Google has certain advantages in the entire AI experience chain that Apple can hardly match, especially in the cloud.

OPPO and Huawei are also making efforts to protect edge - side AI privacy

In the AI era, the boundary between the device and the cloud is being redefined. No matter how powerful the computing power of a mobile phone is, it cannot independently carry a model with tens of billions of parameters. And purely relying on the cloud is bound to cross the red lines of privacy and latency. Therefore, "edge - cloud collaboration" has become an inevitable path for all manufacturers: the edge side is responsible for response, understanding, and control, while the cloud side is responsible for inference, generation, and learning.

A truly intelligent experience must find a balance between these two ends. Domestic manufacturers realized this very early. Leading brands represented by OPPO, Huawei, and Honor have almost simultaneously completed the design and construction of their own edge - cloud architectures in the past two years.

In particular, OPPO can be said to be the most active mobile phone manufacturer in exploring AI. In 2023, it launched its self - trained Andes large model and adopted a three - level large - model deployment strategy, including the Tiny lightweight model deployed on the edge side, the Turbo large model deployed jointly on the edge and cloud sides, and the Titan super - large model deployed on the cloud side.

At the 2025 OPPO Developer Conference held last month, OPPO also announced an upcoming "update" - it will jointly build a Private Computing Cloud (also abbreviated as PCC) system with public clouds such as Volcengine, which also proves that Apple's PCC direction is becoming an industry trend.

Image source: OPPO

Specifically, OPPO has introduced the capabilities of multiple public clouds, including Volcengine and Google Cloud (overseas), and integrated them with its own data center to build a secure and trustworthy cloud infrastructure. In addition, OPPO PCC will conduct two - way trusted authentication and encrypted communication with Volcengine's Jeddak AICC. The user data is encrypted throughout the entire link, thus achieving the goal of "data is available but invisible":

No one, including OPPO and Volcengine, or any system can obtain user data.

However, relatively speaking, there are still certain privacy and security concerns and disadvantages in the communication between the third - party cloud and the terminal. This is a challenge for OPPO, Honor, and other manufacturers. Among domestic mobile phones, perhaps only Huawei has the ability to solve it like Google.

In June this year, Huawei released HarmonyOS 6 at the HDC 2025 Huawei Developer Conference. One of its core upgrades is the new Harmony Smart Agent framework. But it wasn't until October when HarmonyOS 6 was officially launched that we found that Huawei also brought an important upgrade - the HPIC Personal Intelligence Computing platform. The official introduction is as follows:

"Establish a dedicated personal computing space in the cloud and extend the privacy protection on the device to the cloud."

Moreover, after further checking the official information, Lei Technology learned that Huawei's HPIC also deploys a confidential computing area in the cloud. When the task complexity or data volume exceeds the device's computing power (such as the Xiaoyi call summary), the system will send the request to the HPI for processing. All data is only used in the calculation process of the current task and will be immediately deleted after the processing is completed, leaving no original content.

Throughout the process, Huawei neither stores nor accesses the user's original data.

Although it hasn't announced a verifiable image like Apple, it is undoubtedly also in the direction of PCC in terms of design. And because Huawei also has its own large - scale data center, AI chip, terminal operating system, and chip, it is actually closer to Google's PAC.

Conclusion

Now it seems that Apple has indeed changed the tide again. At least Huawei, OPPO, Honor, and Google are all learning from it.

Although it's unlikely that other manufacturers can truly achieve full - chain control of software and hardware from the cloud to the edge like Google, Huawei, and Apple, the direction of balancing large - model performance and privacy security is certain.

This article is from the WeChat public account "Lei Technology," author: Lei Technology. It is published by 36Kr with authorization.