AI smart device frenzy brings fire to Mac! Apple (AAPL.US) competes for AI computing power budgets and invades the AI infrastructure sector

The Zhitong Finance App learned that when the new desktop computers of the US consumer electronics giant Apple (AAPL.US), which recently switched to AI and expensive folding screen iPhones, began shipping new desktop computers on Tuesday EST, the company's executives began to present an unusual and cost-effective marketing reason to large enterprise buyers: it is cheaper to buy these high-performance computer devices than to rent large AI data centers.

According to information, Apple's newly upgraded Mac mini and Mac Studio series Apple desktop computer devices will be able to handle some complex AI tasks/AI inference workloads locally, and the price can be close to 20,000 US dollars. They will go head-to-head with Nvidia (NVDA.US) high-performance workstation GPUs and new desktop computers being launched by US PC manufacturers such as Microsoft (MSFT.US), HP (HPQ.US), and Dell (DELL.US). It is expected that these new devices will be the focus of the Windows event to be held in San Francisco next month by the American software technology giant Microsoft.

AI agents have opened up a new market for Macs, coming from the local computing power needed to continuously perform tasks. The intelligent workflow represented by OpenClaw repeatedly performs model calls, file reading, code operation, browser operation, and result verification, transforming the computer from a terminal waiting for user operation to a task execution platform that can continue to work. Recently, Muse and Astra, which are popular around the world, have undoubtedly further expanded the scope of application of complex task automation; deducing from engineering and business logic, this is expected to increase users' demand for local computing devices that operate continuously, are highly energy efficient, and easy to deploy. In particular, Apple's Mac series products can handle local tool execution, private data processing, and open model inference, and work in collaboration with cloud models.

However, it should be noted that the “lower cost” of Apple Mac-led desktop systems does not equal “sufficient computational capacity/capacity to match an entire AI GPU data center.” From an economic perspective, Apple's cost reduction proposition was founded on the condition that the local AI large model system achieved the required task quality and response speed, and that the cloud service costs saved by continuous use exceeded equipment depreciation, electricity, and operation and maintenance costs.

The Mac itself also relies on the integrated GPU in Apple Silicon to perform model calculations. Its advantage is that unified memory allows the CPU and GPU to share a larger memory pool, reduce data replication, and accommodate model weights and key value caches (KV Cache). The decoding stage of large models with low concurrency and word-by-word generation is often limited by memory bandwidth. Therefore, “large memory+higher bandwidth+low power consumption”, along with quantization and MLX software optimization, allows models suitable for local deployment to operate economically. This is particularly suitable for code assistants, document processors, and internal agents that individuals or small teams use continuously.

In particular, the “four-Mac operating trillion parameter model” mentioned by Apple management proves that the model can be loaded and completed inference, and cannot directly prove that its throughput, latency, and concurrency capabilities are equivalent to cloud GPU/TPU clusters. Actual performance also depends on quantitative accuracy, actual activation parameters, context length, and cross-machine communication; Thunderbolt RDMA can reduce communication expenses, but it will not turn multiple Macs into a huge GPU without communication bottlenecks. Large-scale training and high concurrency services can better utilize the advantages of professional accelerators, high-bandwidth video memory, and high-speed cluster interconnection in data centers dominated by Nvidia AI GPUs.

Apple joins AI cost reduction competition with new Macs to challenge Microsoft and Nvidia

Apple's upgraded Mac Mini and Mac Studio can handle tedious AI inference tasks locally, and can cost close to $20,000. They will go head-to-head with the new desktop computers that Nvidia and PC manufacturers are launching. These new devices are expected to be the focus of Microsoft's Windows event in San Francisco next month.

The new Mac is aimed at more intense AI tasks such as writing code or performing complex business tasks, and users do not need to pay “tokens” (so-called “tokens”) to leading cloud service companies such as OpenAI or Anthropic for this — tokens (tokens) are the basic unit of AI calculation.

Apple executives all hope that the experience they have accumulated in fully exploiting the performance and outstanding energy efficiency ratio of battery-powered devices such as iPhones can help them get a share of Microsoft's traditional dominant market. But the challenge is very difficult: According to the latest statistics provided by Linn Huang, a researcher at IDC, a well-known market research agency, Apple's share of the enterprise desktop computer market is about 4.6%, compared to 91.3% for Windows.

As we all know, Apple co-founder Steve Jobs has always had a conflicting attitude about the large-scale enterprise computing market because end users are unable to choose their favorite products on their own. However, Apple now has an advantageous position in the field of AI desktop computers, partly due to the company's long-standing obsessive pursuit of energy efficiency and energy efficiency.

When Apple launched the first batch of Apple Silicon chips in 2020, the two types of chips originally separated from each other in personal computers — computing chips and high-performance memory chips — were closely integrated into Apple's unique unified memory architecture to improve battery life.

This close connection between computing and memory also unexpectedly made Macs good at handling AI tasks; the world's top AI chip companies, such as Nvidia, have only recently begun to switch to this method. As the open source AI smart tool OpenClaw quickly became popular in markets such as China, Apple's Mac mini began to be sold out frequently.

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

At this month's press conference, Apple's senior management stunningly demonstrated connecting four Mac studios to run a super AI model system with 1 trillion parameters to find and fix a graphics code bug. The number of parameters is an indicator of the complexity of a model. Such tasks usually require a data center, but this set of Macs only needs a wall outlet to run on power.

“Once this machine is on your desk, you've already paid for it. I believe the value we provide is absolutely outstanding, both in terms of performance and cost,” said Johnny Slugge, Apple's chief hardware officer. “No need to pay per word. You just need to use this machine over and over again.”

Microsoft is also competing with Apple for a large-scale incremental market at the same blue ocean level. Its CEO Satya Nadella called device-side AI inference workloads “intelligence that is not billed based on actual token usage.” Nadella also said that Microsoft plans to integrate many of its AI functions into a “super application” for Windows.

However, Microsoft's long-standing leadership in enterprise computing means that it needs to support hardware from many different major vendors, which may push Windows developers who want to take full advantage of the performance of specific chips to take on more work. In other words, Microsoft has long occupied a leading position in the enterprise computing market. Windows needs to be compatible with hardware provided by many manufacturers, so if developers want to make full use of the performance of a certain chip, they usually need to do more adaptation and optimization work.

For example, the same AI program runs on CPUs, GPUs, or neural network processors (NPUs) from different vendors, and may need to be optimized for their respective architectures, drivers, computing interfaces, and memory management methods. The vast hardware ecosystem of Windows has expanded the scope of compatibility and increased the complexity of performance optimization; Apple can unify the design of chips, operating systems, and development tools, making it easier to collaborate and optimize the entire system.

In response to a request for comment, Microsoft said the company has been working with chip partners to streamline AI workloads through Windows ML tools and is actively investing in features such as RDMA. Nvidia declined to comment. However, when the new PC chip was released this summer, Nvidia CEO Hwang In-hoon downplayed the intention to directly compete with Apple, saying that Nvidia is focusing on expanding the functions that Windows PCs can achieve.

Nvidia's dominant position remains the world's hyperscale AI data centers. Slugi said that Apple is promoting this idea to corporate buyers: the large end-side AI model developed on the Apple series of high-performance devices can be expanded up to its most expensive Mac Studio super desktop, or downwards to its cheapest iPhone and iPad, because the chips of these devices use a common principle and design.

“You can put a lot of electricity into the data center,” Sluji said in an interview. “But we insist on providing a variety of products, so that customers can choose according to their needs: which computer and which product do I need?”

As AI smart devices accelerate into desks, the Mac series welcomed a huge explosion in computing power demand! Apple's AI growth space extends outward

Apple management clearly hopes that global companies will use part of the cloud computing AI capital expenditure — Wall Street unanimously expects global AI computing power infrastructure spending to exceed 3 trillion US dollars by 2030, and turn it into a budget to buy Apple's own computing equipment: for workloads with frequent calls and models that can be deployed locally, companies can allocate machine purchase, electricity, and operation and maintenance costs to a large number of tasks to improve the predictability of long-term costs. In line with the recent rise in AI investment in semiconductor equipment stocks such as lithography giant ASML.US (ASML.US), equipment manufacturers are benefiting from the expansion of chip manufacturing capacity, while Apple is seeking a complete machine procurement budget for additional downstream computing requirements, forming another path for AI investment extending from chip manufacturing to corporate desktops.

As shown above, the new market opened up by AI agents for Macs comes from the local computing power needed to continuously perform tasks. Apple's technical advantages focus on unified memory, large memory capacity, energy efficiency, and software and hardware collaboration. In the process of generating large models with low concurrency word by word, the memory bandwidth for reading model weights is often critical; long contexts and concurrent tasks further increase the memory capacity required for the key-value cache. Unified memory allows the CPU and GPU to share a memory pool, reduce data replication, and enable more local workloads to run efficiently.

Judging from investment transmission, Apple's most direct opportunities are an increase in Mac sales, an increase in the share of high memory configurations, and an increase in the penetration rate of enterprise customers. Lightweight tool execution and resident agents can expand the usage scenarios of Mac mini, and larger local models, concurrent workflows, and small clusters are expected to increase the demand for Mac Studio and high-configuration products.

When end-side systems such as desktop computer equipment can continue to undertake economically valuable tasks, the basis for users to measure the return on machine purchases will further shift to how much work time they save, how many tasks to complete, and how much external computing power expenses to reduce. At the same time, Apple has provided cloud model support for Apple Intelligence through a private cloud computing system using self-developed chips, and collaboration between the end side and the cloud also provides space for it to optimize service costs, response speed, and data processing methods.

Enterprise AI servers represent the potential growth direction of Apple's management in the longer term. On September 16, The Information quoted people familiar with the matter as reporting that Apple is exploring an enterprise AI inference server using the M8 Ultra chip and possibly incorporating Nvidia's NVLink Fusion technology. The tentative time point is 2029, and the relevant plans have not yet been confirmed by the company. If implemented in this direction, Apple will have the opportunity to extend the synergetic advantages of chips, memory systems, software tool chains, and machine design to larger enterprise reasoning deployments. Its commercial appeal is that after the smart device continues to operate, customers will pay more attention to the total cost, power consumption, and ease of deployment for each successful task, and these can become important dimensions of Apple's competition; current Mac demand and long-term server exploration also form an interconnected growth logic.