Training in the Cloud, Running at Home
Your next computer could change what you are willing to pay for an AI subscription.
The Economics of Personal AIWhen you buy a powerful computer, you expect it to do the work. You paid for the processor, the memory, and the electricity. Yet when you open a cloud AI assistant, much of that hardware waits while another computer generates the answer.
There are good reasons for this. Large models can need more resources than a personal device can provide. But the balance can change as computers improve and models become more efficient.
That raises a business question as much as a technical one: what happens to the AI subscription when the customer already owns a computer capable of doing the job?
Apple offers a particularly interesting case. It sells the hardware, controls the operating system, and can put an assistant directly into the applications people already use.
My thesis is that local AI could become another reason to buy a better computer — and, for some users, a reason to cancel a separate AI subscription.
Apple has already started
This is not a proposal for Apple to enter a field it has ignored.
Its 2026 Apple Foundation Models family includes both local and server-based models. Apple describes AFM 3 Core Advanced as a 20-billion-parameter model designed for its more capable silicon, with only part of the model active for each request. [1]
What does “20 billion parameters” mean?Parameters are numerical settings adjusted during training. Together, they shape how a model processes information and generates answers. Twenty billion parameters means twenty billion of these learned values — not twenty billion facts or stored answers.*More parameters generally require more storage, but the number alone does not tell you how capable a model is. In a sparse model like this one, only selected parts participate in each computation. That reduces the work needed to generate an answer, although the other parameters still need to be stored.*Apple and Google also announced a multi-year collaboration using Gemini models and Google’s cloud technology as a foundation for the next generation of Apple Foundation Models. [2]
So the ingredients already exist: external technology and computing resources, models adapted to Apple’s products, and a combination of local and cloud execution.
The open question is how much everyday work can move onto the device — and what that does to the economics.
Training and serving customers are different bills
Developing a capable model is expensive. Training requires computing resources, data, experimentation, and people who know what they are doing. Renting infrastructure changes how a company pays for that work; it does not make the work free.
Then comes another bill: running the model whenever customers use it.
A cloud assistant needs resources to answer requests throughout the day. More usage means more work, although the cost of each request depends on the model, hardware, and how efficiently the service operates.
Local inference changes who supplies those resources. The customer’s computer generates the answer. The provider still pays for development, distribution, updates, and support, but it does not have to supply server inference for that local request.
Apple could therefore sell a device that performs useful AI work while also reducing the amount of that work Apple must serve centrally.
That is the business opportunity. The computation still happens, but the customer has already bought the machine that performs it.
Deliver the model, then let the device work
The basic approach is straightforward: develop models centrally, optimise them for supported hardware, and distribute them to devices. Updated models could arrive alongside operating-system updates or as separate downloads.
Making a model smaller takes more than putting it in a ZIP file. Quantisation reduces the precision used to store its numerical values. Distillation trains a smaller model using guidance from a larger one. These approaches can reduce resource requirements, with tradeoffs that need testing.
Apple has already documented aggressive quantisation in its on-device models. [3]
The user does not need to understand any of that. The useful experience is receiving an update and finding that the computer can now do more.
Over time, capable local assistance could become something buyers expect from the operating system, much as they expect search, dictation, and a web browser.
Local does not have to mean open
Open-weight models already let people download model weights and run them on suitable hardware. Mistral, for example, publishes models for this purpose. The permissions attached to those releases depend on the individual licence. [4]
Apple does not have to adopt that distribution model for its own proprietary technology. It could deliver an optimised model as a controlled component of the operating system, accessible through Siri and supported application interfaces, without offering a freely reusable model download.
Apple already provides developer access to its on-device model through the Foundation Models framework. Access to a model’s capabilities and publication of its weights are different things. [5]
Compression and protection also serve different purposes. Compression makes a model practical to store and run. Access controls and protected packaging can make unauthorised reuse harder; they do not guarantee that weights delivered to a user-controlled device can never be extracted.
The commercial point remains: Apple could distribute local capability as part of a paid device and software ecosystem, then replace or improve the model through subsequent updates. It would not need to give everyone an unrestricted copy to run on competing hardware.
Open-weight models make local AI available to people willing to choose and install them. Apple’s opportunity is to make comparable everyday usefulness part of the computer’s normal experience, with deployment, integration, and updates handled for the user.
The next Mac versus the next subscription renewal
Consider a future Mac generation — an “M7,” purely as a scenario, not a claim about an announced product.
Suppose its local assistant handles the tasks for which you currently pay a cloud service: rewriting text, explaining code, summarising documents, translating messages, and answering everyday questions.
You might still find a cloud assistant better at difficult research or complex programming. But if you rarely need those capabilities, would you keep paying every month?
Apple would not have to reproduce every feature of ChatGPT to change that decision.
Apple can earn money without charging for every answer
A cloud service can sell access to its models. A hardware company can also sell the device that makes those models useful.
If better local AI encourages buyers to choose more memory, upgrade an older machine, or stay within the platform, Apple has potential ways to benefit beyond a standalone assistant subscription.
This is my strategic interpretation, not a disclosed Apple financial forecast.
It also creates a tension for customers. AI might help them avoid a subscription while encouraging a more expensive hardware purchase. And local execution does not prevent Apple or another developer from charging for the software itself.
“Runs on your computer” and “included in the purchase price” are separate promises.
The operating system may matter as much as the model
A personal assistant becomes useful when it can work with your documents and applications, subject to the permissions you grant it.
That gives an operating-system vendor an important opportunity. It can make AI available where the task already happens, without requiring the user to move everything into a separate chat window.
For many everyday jobs, convenient integration and reliable execution may matter more than having the highest-scoring model.
Local processing also offers a practical privacy benefit: tasks completed entirely on the device need not send their contents to a remote model. But an assistant can still use online search, connected services, or cloud fallback. The application’s actual behavior matters more than a “local AI” label.
Why the cloud will still have a role
A model fitting in memory does not mean it runs comfortably alongside everything else. Response speed, long documents, power consumption, and heat all affect the experience. A desktop with substantial memory and a thin laptop running on battery face different limits.
Cloud systems keep improving too. They can offer capabilities that remain beyond personal hardware, while serving users who prefer inexpensive devices.
Apple’s own architecture includes server models for demanding work, which makes its present approach hybrid. [1]
My forecast is that more routine requests will become practical locally, while remote services retain value for tasks that benefit from greater resources. That does not guarantee lower total data-center demand: wider adoption and more ambitious tasks could outweigh the work moved onto devices.
The subscription will need a stronger reason to exist
The important change would be in what customers consider worth paying for.
If a capable assistant comes with the computer and handles their daily needs, a separate subscription becomes a purchase for additional capabilities. Providers would have to earn it through better results, specialised tools, collaboration, or access to services the local assistant cannot replace.
For Apple, that could turn AI capability into hardware value. For cloud providers, it could make everyday assistance harder to sell on its own.
We are used to asking whether a personal computer can run a large model. The more consequential question may soon be:
Can it do enough that you stop paying someone else to run one for you?
Sources and further reading
Current product descriptions are distinguished from the author’s forecasts. The hypothetical M7 scenario does not imply announced specifications, pricing, or capabilities.
Top comments (0)