How an extension to Apple’s iCloud could enable personal AI data centers

Last update: Wednesday 5/28/25

   Personal AI Data Center

Yogi Berra, the NY Yankees’ Hall of Fame catcher, once said,”It’s deja-vu all over again.” The following discussion can best be characterized by a slight paraphrase of Yogi’s recollection, “It’s deja-Nvidia all over again.”

Here's the longer more specific subtitle of this note: "How personal AI data centers in our homes and offices running open source small language models could use iCloud routers for authenticated remote connections."

... From Recreation Stations to Personal AI Data Centers …

A. Nvidia | B. Apple History | C. Framework | D. Current Options | E. iCloud Router | F. Components

Introduction
Personal AI data centers, as described in this note, would have immediate appeal to Apple’s savvy GenAI users (20%) who engage intensively with chatbots and agents every day. But personal data centers will probably require substantial evolution before they provide comparable attractions for Apple’s broader base of recreational users (80%).

As will be detailed in the final section of this note, Apple’s intensive AI users would probably double or triple their investments in Apple products and services. They would gladly do so in order to enjoy the substantial benefits to be gained from redirecting most of their generative AI activities from the large language models in the cloud to the small language models (SLMs) in their personal data centers.  

A. Nvidia’s serendipitous path to colossal success 
Here is a concise review of Nvidia‘s path to becoming the provider of the most powerful chips used to train and to run generative AI’s large language models (LLMs).

    Figure 1.

• Nvidia started out as a company that made chips for video games. Gamers were constantly clamoring for more complex games, games that required faster and faster chips. 
  
  
• Suppose a square object on a video screen moved from the lower left side of the screen to the upper right side. The position of each of the square’s four corners would have to be moved the same distance to the right and the same distance upwards. 
  
  
• One could change the position of each corner one at a time. So a classic central processing unit (CPU) would take four cycles to loop through the four identical changes.
  
  
• But suppose you could change all four positions at the same time, change them “in parallel”; then you could move the square four times as fast.
  
  
• In reality, the simultaneous changes couldn’t be made four times as fast, more likely two or three times as fast. But you would still be moving the square much faster than before.
  
  
• So Nvidia began to produce chips that could process many instructions simultaneously, chips that became known as "graphics processing units" (GPUs).
  
  
• As it happens, George Hinton, one of the so-called “godfathers of AI”, realized that he could run his neural network models much faster on these new GPUs because the mathematics was similar. 
  
  
• Recognizing that AI models were likely to quickly evolve into a much larger and far more profitable market than video games, Jensen Huang, CEO of Nvidia, redirected most of his company’s efforts into producing faster and faster GPUs for AI. 

Indeed, the AI models became bigger and bigger LLMs that required faster and faster GPUs. A few years later, Nvidia has become one of the most valuable corporations in the world, whose market capitalization exceeds three trillion dollars. Note that Nvidia does not create LLMs; it only makes the chips, i.e., the platforms on which other companies’ LLMs are trained and run.

So what does all of this have to do with Apple? Apple’s recent history and current opportunities are strikingly similar to Nvidia’s.

B. Apple’s recent history 
Under Tim Cook’s leadership, Apple became a $3 trillion corporation by focusing its efforts on the 80 percent of its user base who use their iPhones and desktops primarily for recreation (and communication). They use the cameras on their iPhones to capture photos and videos that they edit and store on their desktops, then share with family, friends, and associates on WhatsApp, TikTok, and YouTube. And they have had an insatiable craving for higher and higher photo and video resolution. 

Figure 2.

• Photo and video processing require far more computing power than text processing; so Apple’s relentless quest for higher and higher resolution required it to design evermore powerful CPU and GPU chips for its phones and desktops. Its phones and desktops have thereby become the market’s most powerful “recreation stations” … 
  
  
• Last year it become more difficult for Apple to persuade the 80 percent recreation segment of its user base to purchase its latest, most expensive phones and desktops. How much, they rightly asked, will these upgrades enhance the quality of our recreation? 
  
  
• At its WWDC 2024 conference Apple responded with pre-recorded videos and simulations of a new package called "Apple Intelligence" that would only be accessible on upgraded iPhones and desktops. Many recreation users then purchased the upgrades … only to be informed months later that the most important Apple Intelligence features would be indefinitely delayed, delays that angered its recreation base and triggered lawsuits.

By contrast, the insatiable genAI curiosity of the other 20 percent of Apple's user base, i.e., its savvy computer users, will make them eager early adopters of expensive upgrades.

In March 2023, Google declared a “Code Red” because it recognized that OpenAI’s ChatGPT running on its GPT-4 model posed an existential threat to its dominance of Internet search, its most profitable service. Google immediately invested tens of billions into the development of its own powerful chatbots and models in a frantic effort to retain this dominance. And so did Amazon. And so did Meta. But Apple barely took notice. The time to play catchup was back then. Now it’s probably too late. Apple is unlikely to catch up. 

• Why is Apple unlikely catch up? Because today’s “Apple Intelligence” is way behind the powerful chatbots and agents produced by the AI leaders, e.g., OpenAI, Anthropic, and Google. Tomorrow’s “Apple Intelligence” will probably be even further behind because the expensive high pace of generative AI innovation is accelerating, propelled by far bigger $80 billion to $100 billion annual investments by each of the leading AI producers and their financial backers, investment levels that Apple has given no signs that it will match or exceed in order to make up for lost ground.

Ironically, it really doesn’t matter whether Apple’s chatbots and models catch up to its biggest competitors. As Section F will argue, Apple doesn’t need to catch up — it already holds another unbeatable winning hand, much like the one Nvidia has played so successfully.

An Apple version of Nvidia's strategy

C. Framework for an Apple version of Nvidia’s strategies 

Recommended strategies will be derived from from the following important similarities and differences between Nvidia and Apple.

1. Nvidia is also a $3 trillion corporation and grew to this high market capitalization faster than any other company, including Apple.

2. Apple is under antitrust scrutiny in the US, in the UK, and in the EU; whereas Nvidia is not under investigation anywhere. 

3. Apple’s ability to drive growth primarily through upgrades to its iPhones has peaked, as can be seen from its desperate invocation of an untested, nonexistent “Apple Intelligence”  feature last year, so it needs a new growth strategy; whereas Nvidia’s strategies still have strong future growth potential.

4. Nvidia has always seen its new chips as a platform whose use enables other companies to earn more revenue; whereas Apple has primarily regarded its new chips as inducements for its recreation users to upgrade their iPhones and desktops.

5. To be specific, Nvidia’s powerful chips are platforms that enable the development of data centers in which other companies’ generative AI models can be trained and operated.

Apple’s powerful chips could also become platforms that transform its iPhones and desktops into powerful AI workstations in personal AI data centers wherein savvy computer users run open source small language models (SLMs) using iCloud routers for secure private remote connections. Personal data centers would only be accessible to their owners, and to their owner's family, friends, and close associates.

Figure 3. 

6. As a platform, Nvidia does not make special partnerships with specific companies. Indeed, the EU expressly forbids platforms from forming limited partnerships; platforms must be accessible to all providers of services that could use the platforms. 

As suggested by Figure 3, Nvidia's chips are used by a large community of AI providers. By contrast, Apple gave special access to OpenAI, but indicated that it might also partner with Google and other selected providers.

Figure 4.

 

7. Most importantly, Nvidia has strived to create a “road map” for developing new features for its chips through active collaboration with the companies it enables. In the starkest possible contrast, Apple has followed a top down strategy, developing new features for its chips in secret, based solely on its own judgments, then challenging other companies to figure out how to make the best use of the new features.

Unfortunately, Apple does not seem to possess enough in-house expertise to pursue this strategy successfully with regards to new features that are related to generative AI models.

Figure 5.    

  D. Current options … What a savvy computer user can do right now  

• Set up an inexpensive but powerful Mac mini to run open source small language  models (SLM) locally.

• The goal is a personal AI data center that’s easy to configure and manage via simple, graphical apps.

• Avoid recurring subscription fees by using open source SLMs locally, rather than in a corporate cloud. There is no charge for running open source models on one's own computer.

• Create a “library” of domain-specialized SLMs to handle complex questions within narrow areas of expertise. But bear in mind that small open source models can’t do everything:

  • Most cannot search the Internet; so their knowledge stops at the date of their last update.
  • Most cannot engage in chain of thought reasoning; so their capacity to handle complex problems is limited.
• Ideally, a savvy user would only pay subscription fees for access to LLMs in a cloud that could answer complex questions that require powerful chain-of-reason capabilities  … or require knowledge outside of the limits of the user’s library.

• Use LM Studio (or an equally user-friendly app) to provide a local chatbot interface to the SLMs  — no configuration required.

• Today, the SLMs on the Mac mini are only accessible from the devices on the user’s  local network; smartphones or offsite laptops can’t connect without cumbersome software setups.

• What’s missing is a router that provides end-to-end encrypted connections to the data center to authorized devices in remote locations.

In Section F, we will propose that Apple extend its iCloud services to include an “iCloud Router” plus a companion “Local iCloud Router” app (running on a second Mac mini within the personal data center). This setup will ensure complete privacy and local control — no third-party servers that can see any data. 

E. iCloud-Router as an iCloud extension 

The purpose of this short section is to underscore our previous assertion that the iCloud router is an extension of the existing services provided by iCloud. This suggestion does not alter iCloud’s existing functions. Notes, files, calendars, and other synced data services would continue as before. 

What we’re suggesting is a new capability that enables requests from iPhones, laptops, and other authorized devices to gain remote access to personal AI servers within a home or office, such as a Mac mini, that host open source models. iCloud router is merely an extension to iCloud, not a replacement.

Reminder …

Our readers are reminded that the kind of personal data center described in the next section can be constructed in their homes or offices today. But their center will only be accessible to iPhones, laptops, and desktop on the local internet in their homes or offices. Remote access won't be possible without Apple's development of a router extension of its iCloud services for the iClaud router and its development of a no-brainer app that would allow users to configure one of their Mac mini's as a Local iCloud router. 

Remote access would allow them to share their data center with their family, friends, and associates. If remote users were willing to share some of the expenses for purchasing more devices in their center, they could afford to configure a larger library of specialized open source SLMs running on more Mac minis with faster chips and more storage.

F. Components

                       Figure 7.

Four Design Principles ...

          

1. Easy to use, easy to configure, and easy to maintain 

2. Inexpensive components

3. Local and remote access for its owner, and its owner's (authorized) family, friends, and associates

4. Maximum privacy: No one except its authorized users should have access to its data, not even Apple

Figure 7 displays the basic components of an initial version of a Personal Data Center. In keeping with our design principles, we add the following specifications: 

• The iCloud Router is the name we have given to the extension of Apple's iCloud services that Apple must develop. It receives messages from authorized remote devices and routes them to the Local iCloud Router in the Personal Data Center. It has no knowledge of the content of the end-to-end encrypted messages it is transmitting (privacy) to the Local iCloud Router.
  
  
• The Local iCloud Router directs messages it receives from the iCloud router to the appropriate device within the personal data center. 
  
  It should be a Mac mini (inexpensive), preferably with the fastest chip, but a modest amount of storage. The app that will be used to configure its router functions will be written by Apple, so it should be easy to configure and require minimum maintenance.
  
  
• The master storage device is a temporary place holder for a Mac mini that would house the photos and videos of Apple's recreation users (80%). This server would require a more substantial extension of Apple's iCloud services to enable it to interact with remote iPhones, laptops, and desktops. It should have lots of storage, but not necessarily have the fastest chips, and will be fully specified in a subsequent note. 
  
  
• GenAI users (20%) should select the “LM Server” app to host their library of open source SLMs because it provides its users with a companion chatbot app for their iPhones, laptops, and desktops (similar to Claude, Gemini, or ChatGPT); so it should be easy to use. It also has a reputation for being easy to configure and easy to maintain.
  
  GenAI users should acquire a second Mac mini (inexpensive) to host their LM Server app. This mini should have the fastest chips and lots of storage.
  
  

Additional devices

Anything with a client app only needs the usual network routing and authentication to connect over the Internet. We recommended the LM Server app as the “best” host for open source SLMs because it comes with a client app that knows how to “talk” to the server app from local or remote locations. 

For the same reason we could add any other device to our personal data center that comes with a network client app. An important class of devices immediately comes to mind: printers. 

• Here are a few printers that have client apps —  HP, Canon, Epson, and Samsung. Sometimes savvy GenAI users might find it useful to print files on their devices using printers that are in their personal data centers when they are using their devices from remote locations.

... Enhanced privacy ...

More than any other big tech corporation, Apple has maintained a deep, steadfast commitment to safeguarding the privacy of its users. But times are changing. Governments all over the world, including the U.S., are demanding more and more access to user information for natural security and other purposes. Sooner or later, Apple will have to compromise. 

The architecture our iCloud router ensures that Apple has no understanding of what devices are in a personal data center and for what purposes. This alleviates Apple’s burden as a privacy gatekeeper. No matter how hard it is pressured, Apple cannot provide governments with information that it never had. 

But the big tech corporations that run the large language models in the cloud pose far bigger threats to a user’s privacy. While the photos and videos on TikTok, Instagram, and YouTube convey superficial impressions of one’s lifestyle and personality, these visuals can and usually are staged and edited. 

By contrast, the day-to-day spontaneous conversations that genAI users have with chatbots running on cloud-based large language models provide voluminous samples of their thinking, samples that enable the construction of digital portraits of the innermost workings of their minds. These portraits are good enough to facilitate accurate predictions of their most likely responses in a wide variety of circumstances. 

Unfortunately, these insights belong to profit-oriented Big Tech corporations who monetize them by selling these insights to advertisers. What’s worse, these Big Tech corporations offer far less resistance than Apple to government’s demands for access.

The most obvious way to substantially reduce this monetization is to substantially reduce one’s conversations with chatbots running on cloud-based large language models and to redirect one’s conversations to chatbots running on well constructed, open source, small language models in a personal AI data center.