AI agents ignite a frenzy for Mac! Apple (AAPL.US) joins the battle for AI computing power budget, advancing into the AI infrastructure sector

According to Zhitong Finance APP, as American consumer electronics giant Apple (AAPL.US) shifts its growth strategy fully to AI and the expensive foldable iPhone, on Tuesday Eastern Time, when the company's new desktop computers began shipping, Apple's executives started pitching an unusual and highly cost-effective selling point to large enterprise buyers: purchasing these high-performance computers is cheaper than renting a large AI data center.

It is reported that Apple’s newly upgraded Mac mini and Mac Studio desktop computers, priced up to $20,000, can handle some complex AI tasks/inference workloads locally. They will directly compete with Nvidia’s (NVDA.US) high-performance workstation GPUs and new desktops from American PC makers such as Microsoft (MSFT.US), HP (HPQ.US), and Dell (DELL.US), which are launching desktop PCs capable of locally processing AI tasks. These new machines are expected to become a highlight at Microsoft’s Windows event in San Francisco next month.

The new market for Macs unlocked by AI agents comes from the need for local computing power to continuously execute tasks. The agent workflow represented by OpenClaw repeatedly involves model calls, file reading, code execution, browser operations, and result verification, turning computers from waiting-for-user-action terminals into platforms capable of continual task execution. Recently, the globally popular Muse and Astra have undoubtedly further expanded the application range of complex task automation. From engineering and business logic perspectives, this is expected to increase demand for locally-run computing devices that are always-on, energy-efficient, and easy to deploy. Especially, Apple’s Mac series products can handle local tool execution, private data processing, and open model inference, as well as work together with cloud models.

But it should be noted that Apple Mac-based desktop systems being "cheaper" does not mean their “computing power/capacity is equivalent to an entire AI GPU data center.” From an economic perspective, Apple's cost-saving proposition holds only if the local AI large model system achieves the required task quality and response speed, and the cloud service fees saved from continuous use exceed the costs of device depreciation, electricity, and maintenance.

The Mac itself relies on the integrated GPU in Apple Silicon to run model calculations, with the advantage that unified memory allows the CPU and GPU to share a large memory pool, reducing data copying and accommodating model weights and key-value (KV) cache. The decoding stage of large models with low concurrency and word-by-word generation is often constrained by memory bandwidth, so "large memory capacity + high bandwidth + low power consumption", paired with quantization and MLX software optimization, can run models suitable for local deployment economically. This is especially suitable for always-on code assistants, document processing, and internal agents for individuals or small teams.

Especially as Apple’s management mentions “running a trillion-parameter model on four Macs”, it proves that the model can be loaded and inference completed, but does not directly prove throughput, latency, and concurrency are equal to cloud GPU/TPU clusters. Actual performance also depends on quantization precision, the number of actually activated parameters, context length, and inter-machine communication. Thunderbolt RDMA can reduce communication overhead, but it will not make several Macs act as one giant GPU without communication bottlenecks. Large-scale training and high-concurrency services are where Nvidia’s AI GPU-led data centers, with professional accelerators, high-bandwidth memory, and high-speed cluster interconnects, truly excel.

Apple Joins the AI Cost Reduction Race with New Macs, Aiming to Challenge Microsoft and Nvidia

The upgraded Mac mini and Mac Studio from Apple, priced up to $20,000, can process cumbersome AI inference tasks locally and will compete directly with new desktops from Nvidia and other PC manufacturers. These new machines are expected to become a focal point at Microsoft’s Windows event in San Francisco next month.

The new Macs are geared toward high-intensity AI tasks such as coding or executing complex business operations, and users don’t need to pay OpenAI, Anthropic, or other leading cloud service providers “token” fees—tokens being the basic unit of AI computation.

Apple executives hope that their experience in maximizing performance and energy efficiency on battery-powered devices such as the iPhone will help them carve a niche in Microsoft’s traditionally dominant market. But this challenge is considerable: according to the latest data from renowned market research firm IDC’s Linn Huang, Apple’s share of the enterprise desktop computer market is about 4.6%, while Windows accounts for 91.3%.

It is well known that Apple’s co-founder Steve Jobs always had mixed feelings about the enterprise large computing market, as end users can’t choose their favorite products freely. However, Apple now holds a favorable position in the AI desktop field, partly due to its long-standing pursuit of energy saving and high energy efficiency.

When Apple launched its first batch of Apple Silicon chips in 2020, it tightly integrated two previously separate types of chips in personal computers—the computation chip and the high-performance memory chip—into Apple’s exclusive unified memory architecture to improve battery life.

This close connection between computation and memory also unexpectedly made Macs adept at handling AI tasks—a method that top global AI chip companies like Nvidia only recently began to adopt. As open-source AI agent tool OpenClaw quickly gained popularity in markets such as China, Apple’s Mac mini began to frequently sell out.

Although the Mac Studio was initially aimed at content creators editing videos or producing music, over the past two years Apple has quietly added some unusual AI features, such as a specially designed chip interconnect technology called Thunderbolt-based remote direct memory access (RDMA).

At this month’s press conference, Apple’s senior management showcased the impressive feat of linking four Mac Studios together to run a massive AI system with a trillion parameters to locate and fix a graphics code bug. The number of parameters is an indicator of a model’s complexity. Such tasks usually require data centers, but this group of Macs can operate with just one wall outlet.

“Once this machine is on your desk, you've already paid for it. I firmly believe the value we provide is outstanding, both in terms of performance and cost,” said Apple’s Chief Hardware Officer Johny Srouji. “There’s no need to pay by the token. You just use the machine over and over again.”

Microsoft is also competing with Apple for the same massive incremental market; CEO Satya Nadella refers to device-side AI inference workloads as “intelligence that doesn’t charge by actual token usage.” Nadella also stated that Microsoft plans to integrate many of its AI features into a Windows “super app.”

However, Microsoft’s long-standing dominance in the enterprise computing sector means it has to support hardware from many different major manufacturers, potentially causing Windows developers who want to maximize specific chip performance to do extra work. In other words, Microsoft’s dominant market position means Windows must be compatible with hardware from numerous suppliers, so developers typically need to invest more in optimization and adaptation to unlock the full potential of any given chip.

For example, running the same AI program on different vendors’ CPUs, GPUs, or neural processing units (NPUs) may require optimizing for different architectures, drivers, compute interfaces, and memory management methods. Windows’s massive hardware ecosystem broadens compatibility but increases complexity for performance optimization. Apple, on the other hand, can unify chip, OS, and developer tool design for easier end-to-end optimization.

In response to requests for comment, Microsoft stated that it has always worked with chip partners to simplify AI workloads with Windows ML tools and is actively investing in features like RDMA. Nvidia declined to comment, but when launching new PC chips this summer, CEO Jensen Huang downplayed direct competition with Apple, saying Nvidia is focused on expanding what Windows PCs can do.

Nvidia’s stronghold remains the world’s large-scale AI data centers. Srouji said Apple is pitching enterprise buyers the concept that AI large models developed on Apple’s high-performance devices can scale up to its most expensive Mac Studio “super desktops” or down to its cheapest iPhones and iPads, as those devices’ chips are built on the same principles and designs.

“You can pour massive amounts of electricity into a data center,” Srouji said in an interview. “But we are committed to offering a variety of products, enabling customers to choose according to their needs: Which computer, which product do I need?”

As AI Agents Accelerate onto the Desktop, Mac Series Will See Explosive Computer Power Demand! Apple’s AI Growth Potential Expands

Apple’s management clearly hopes enterprises will shift some of their cloud AI computing capital expenditures—which Wall Street expects to surpass $3 trillion globally by 2030—towards budgets for purchasing Apple’s own computing devices: for workloads that are frequently invoked and can be deployed locally, companies can spread device, electricity, and maintenance costs across many tasks, improving the predictability of long-term costs. Echoing the recent stock surge of semiconductor equipment giants like ASML (ASML.US) fueled by the AI investment boom, device manufacturers benefit from increased chip-making capacity investments, while Apple is fighting for end-user computing procurement budgets as AI investment extends from chipmaking to the enterprise desktop.

As outlined above, the new market for Macs unlocked by AI agents comes from the local computing power needed for continuous task execution. Apple’s technical advantage lies in unified memory, large memory capacity, energy efficiency, and hardware-software synergy. In the low-concurrency, word-by-word generation stage of large models, memory bandwidth for reading model weights is crucial; long context and concurrent tasks further increase the memory required for key value caching. Unified memory allows CPU and GPU to share a memory pool, reducing data copying, enabling more local workloads to run efficiently.

From an investment transmission perspective, Apple’s most direct opportunities are increased Mac sales, a higher proportion of high-memory configurations, and deeper enterprise customer penetration. Lightweight tools and always-on agents can expand Mac mini’s usage scenarios, while larger local models, concurrent workflows, and mini-clusters may boost demand for Mac Studio and high-end configurations.

As desktop and other edge devices can continually handle economically valuable tasks, the criterion for users evaluating returns on device purchases will shift further towards how much work time is saved, how many tasks are completed, and how much external compute cost is avoided. At the same time, Apple has already provided cloud model support for Apple Intelligence through its private cloud infrastructure built on self-developed chips, and device-cloud collaboration also allows for optimized service costs, response times, and data processing methods.

Enterprise AI servers represent Apple’s management’s longer-term growth potential. On September 16, The Information cited sources reporting that Apple is exploring enterprise AI inference servers potentially powered by M8 Ultra chips and Nvidia’s NVLink Fusion technology, with a tentative timeline set for 2029, though this plan has not been confirmed by the company. If this direction materializes, Apple will have the opportunity to extend the synergy of its chips, memory systems, software toolchains, and whole-machine design to larger-scale enterprise inference deployments. Its commercial appeal lies in: After agents are always running, customers will focus more on the total cost, power consumption, and deployment convenience for each successful task—precisely where Apple can compete. The immediate Mac demand and longer-term server prospects thus form a connected growth story.