Local AI, Big Tech Wants AI on Your Computer

The next sales pitch for artificial intelligence may come with a power cord.

Microsoft and Nvidia are expected to present the Surface Laptop Ultra at a San Francisco event on October 7, Reuters reported Wednesday morning. The machine is part of a broader effort to move demanding AI work onto personal computers, giving users more computing power at home and at work.

For buyers, the appeal is understandable: greater control over sensitive information and less dependence on a distant service. For the companies selling the technology, there is another attraction. Customers who purchase powerful computers also pay for some of the infrastructure needed to serve them. The question is how much of that bargain benefits the person buying the machine.

Local AI runs a model on the device in front of you. Once a suitable model has been downloaded, software such as LM Studio can answer questions and work with documents without an internet connection. The model was developed elsewhere; the computation needed to produce its answers happens on your computer.

Microsoft’s Windows plans include a platform for agents that can carry out tasks locally, with operating-system controls governing their access. Nvidia’s RTX Spark platform supports configurations with up to 128 gigabytes of shared memory, intended to give those applications considerably more room to operate.

“Our goal is to deliver unmetered intelligence to every home and every desk with Windows,” Microsoft chief executive Satya Nadella said in the companies’ May announcement.

Unmetered is an appealing word in an industry increasingly experimenting with charges tied to consumption. It also leaves open a substantial question about where the costs go.

Microsoft reported $41 billion in capital expenditures in the quarter ended June 30, including finance leases. Roughly two-thirds went toward relatively short-lived assets, primarily processors. During the same earnings call, the company described expanding software pricing beyond a charge for each user to include charges based on consumption. It is spending heavily to meet demand while developing more ways to bill for that demand.

Moving a suitable task onto a customer’s computer can reduce the remote computing that task requires. The customer supplies the hardware and electricity. Microsoft can retain the operating system and the software relationship. Nvidia gains another place to sell its chips. That gives both companies a reason to encourage local AI alongside further investment in data centers.

Microsoft’s investments point toward a business that expects substantial demand for both local and remote computing. The company said it added 31 data centers in its June quarter. Moving some work onto personal computers is compatible with building an enormous business around the work that remains in the cloud.

For computer makers, the commercial logic is equally clear. Useful AI software could give a customer a reason to replace a machine that still handles email and spreadsheets perfectly well. Buyers will have to judge whether the extra capability justifies an earlier upgrade, or a more expensive one.

There are credible reasons to want the capability. Consider a business with a large collection of internal documents and a recurring need to search or summarize them. Keeping that work on its own equipment could reduce the amount of information sent to outside services. A downloaded model could also remain useful during a network outage. LM Studio already documents offline chat with local files, so that basic use does not depend on the arrival of a new laptop.

The economics become less attractive when expensive equipment spends most of its time waiting. A buyer who runs a few requests a day has fewer opportunities to recover the purchase price than a business processing a steady workload. Someone replacing an aging computer also faces a different calculation from someone buying a second machine solely for AI.

At the more specialized end of the market, the upfront cost is rising. Reuters reported that Nvidia recently raised the price of its 128-gigabyte DGX Spark desktop to $6,950, citing higher memory costs. That is a substantial investment before a business has measured a single hour saved.

Capacity brings its own complications. A model needs memory, and the material it considers while answering consumes additional memory. Longer documents and simultaneous requests can increase those demands. Ollama, another tool for running models locally, documents how concurrent requests increase memory requirements and how work can queue when resources are insufficient.

Those constraints help explain why a successful demonstration can be a poor purchasing guide. A machine may handle one carefully chosen task comfortably, then struggle with the workload its owner actually needs. A useful test would include the same files, competing applications and turnaround times the buyer expects on an ordinary working day.

There is also the labor that an ownership calculation can overlook. Choosing models, applying updates, investigating failures and checking answers all take time. A business may decide that responsibility is worthwhile. A household looking for help with routine administration may place a higher value on a service somebody else maintains.

Privacy deserves the same close accounting. LM Studio’s policy says messages and documents remain on the device when models run locally. Its cloud features, including web search, involve external processing. A locally installed application can therefore contain both private, offline functions and features that send information elsewhere. The relevant question is what happens during the particular task being performed.

Access matters as well. Apple’s documentation says an application granted Full Disk Access can reach files belonging to other applications, including Mail and Messages. Running an assistant on your own computer does little to limit that reach if you also grant it broad permission to inspect your digital life.

The distinction becomes especially consequential when an assistant can act. Reading a folder to prepare a summary and having permission to change its contents are different grants of authority. Microsoft’s plans for Windows agents explicitly include containment and access controls. Those controls will influence whether people feel comfortable allowing an assistant to do more than answer questions.

The products themselves suggest that many users will end up mixing local and cloud services. Nvidia describes software that can route requests according to privacy policies. Apple’s June announcement says some AI features rely on server models and links increased access to iCloud+ plans. Owning capable hardware can coexist with a continuing subscription.

That gives buyers a more practical standard than the promise of unlimited intelligence. They can ask how much of their actual work the computer completes, how much information still leaves it and how much assistance it requires from its owner. A fast answer that needs extensive correction has little economic value, wherever it was generated.

Microsoft and Nvidia have a persuasive reason to bring more AI onto the desktop. The purchase will make sense for customers when the machine reliably saves enough work to justify its cost. For a business owner considering another expensive computer, the most revealing question may be a familiar one: who gets the call when it stops working?