Plans to return to the server market after 15 years: What are Apple's (AAPL.US) AI computing power ambitions? Big bank argues about “far water” and “near thirst”
The Zhitong Finance App learned that according to reports from people familiar with the matter in the tech media The Information, Apple (AAPL.US) is considering entering the enterprise-grade server market and plans to use self-developed M-series Ultra chips with Nvidia (NVDA.US) network devices.
According to the report, Apple plans to sell servers to AI developers, enterprises, and government customers. The core driving point is to support customers to run local AI inference on privatized infrastructure. According to the disclosure, the proposed server will have two versions: the low version integrates two M8 Ultra chips that are still under development, and the high version is equipped with four M8 Ultra to form a computing cluster. To solve the problem of high-speed interconnection between multiple chips, Apple has negotiated with Nvidia to adopt its NVLink Fusion interconnection solution (covering switches, cores and supporting software).
This move, if carried out, would mark Apple's return to the server market. Between 2002 and 2011, Apple launched xServe servers equipped with Intel and IBM chips, but due to insufficient attention to the corporate market and weak customer support, the product never opened up the situation, and eventually stopped production in early 2011.
Mac sales surged 29%, fueling server ambitions
However, today's market environment is not what it used to be. In Apple's financial report for the third quarter of fiscal year 2026, sales of the Mac business surged 29% year over year to US$10.3 billion, leading the growth rate of all product lines, including iPhone and service businesses. According to information, large-scale AI laboratories such as OpenAI have purchased tens of thousands of Mac mini and Mac studios to train AI agents through intensive learning, and Anthropic is also renting Mac mini from Amazon Web Technology — strong demand even caused these models to be seriously out of stock for a while.
However, the report also pointed out that even if Apple advances the project, new products will not be launched until 2029, and there is still a possibility that the project will be cancelled. Although the details are uncertain, Apple's new CEO John Ternus (John Ternus) already expressed support when the project started more than a year ago (he was still in charge of the hardware engineering department at the time) — the plan opened the door for Apple's self-developed chips to enter the AI computing power market.
Affected by this news, Apple's stock price rose slightly by 0.7% in early and late trading on Wednesday, and Nvidia's stock price rose more than 2% during the intraday session on the same day.
Notably, in an earlier review of Apple's 3QFY26 earnings report, CICC pointed out that the quarter's revenue of US$109.417 billion (+16% YoY) and net profit of US$29.789 billion (+27% YoY) all exceeded expectations, and Mac revenue of US$10.352 billion (+29%) was the biggest contributor; it kept FY2026/27 profit forecast unchanged, and raised the target price by 10% to $340 on the grounds of “accelerated AI progress and increased valuation center”. If the server plan is implemented, this logic of “revaluing the value of self-developed chips” will undoubtedly be further strengthened.
Regarding the server message itself, several agency analysts believe that the relationship between “far water” and “near thirst” should be viewed calmly: the 2029 release window means that the project has made little substantial contribution to Apple's FY 2026-27 profit forecast, does not act as a catalyst for performance in the short term, and is better viewed as a “strategic bullish option.” According to some agency estimates, the real key underwriting issue is whether Apple can sell differentiated privatized AI computing power at a premium (based on unified memory architecture, energy efficiency ratio, and privacy and security advantages), rather than becoming another buyer that purchases expensive computing power. In other words, until definitive customer promises, capital expenditure, or pricing models are disclosed, the news is insufficient to support a major revaluation of valuations.
The trillion racetrack competition pattern: surrounded by giants, where is Apple based?
Goldman Sachs expects the global AI server market to reach 1.24 trillion US dollars by 2030, driven by the AI wave. The track is currently overcrowded, and the players' performance is still high:
Dell: Revenue for the latest fiscal quarter was 46.97 billion US dollars (+57.8% year over year), AI server orders reached 60.9 billion US dollars in a single quarter, and the backlog of orders at the end of the period reached 95 billion US dollars. The cumulative stock price increased more than three times during the year;
HPE: Revenue for the latest fiscal quarter was US$12.21 billion (+32.7%). Driven by AI and network services, it once surged 137% during the year, and only experienced a deep correction in mid-September due to a rating downgrade;
Ultra Micro Computer (SMCI): Revenue for the latest fiscal quarter was $11.12 billion (+93.2%), and the backlog of orders also set a record.
Apple's differentiated approach is not to compete head-on with Nvidia's GPU clusters for the big model training market, but to cut into AI reasoning, particularly the “private deployment” segment — targeting enterprise and government customers who want to run models on their own servers and are highly sensitive to data privacy. This card slot is a natural fit with the energy efficiency advantages of Apple hardware and the unified memory architecture. It is also a link that Nvidia is currently relatively weak but is being quickly fixed.
However, Apple's shortcomings are just as clear. Servers are essentially a low-margin, capital-heavy, and service-heavy B2B business, which is very different from Apple's usual high-margin consumer electronics approach; enterprise buyers value software ecology, remote management, compliance certification, and full life cycle support, and this is exactly the core lesson of the Xserve era. In addition, Apple currently uses self-developed interconnect technology for internal Private Cloud Compute, and there are speed and cost bottlenecks in hyperscale deployments, which is also a direct reason for turning to Nvidia's NVLink Fusion.
Industrial chain and pattern deduction: Nvidia may be an “invisible winner”
If the partnership finally comes to fruition, the market generally sees Nvidia as a more definite beneficiary. Network equipment currently accounts for 10% to 15% of the total hardware cost of AI data centers, and the network business has contributed about one-fifth of Nvidia's data center revenue. Since its launch in May 2025, NVLink Fusion has attracted partners such as MediaTek, Marvell, Fujitsu, and Qualcomm. Last month, Amazon Web Services also announced that the next generation of AI servers will expand adoption. If Apple joins, it will be the platform's most powerful customer to date — this not only means a new revenue stream, but also validates Nvidia's “even if customers use self-developed chips, they must use my interconnection” ecological strategy, thus further strengthening the moat in the face of the open standards camp.
What is intriguing is Apple's connectivity route choice: Apple is also a board member of the UALink Alliance (an open interconnect standards organization formed by AMD, Intel, Google, Microsoft, etc., with more than 65 members). If it were to eventually switch to Nvidia's proprietary solution, it would be tantamount to choosing the latter between “openness” and “performance,” and the industry connectivity standards dispute may usher in a key turning point. Also, it is important to pay attention: some Apple engineers involved in the project believe that Nvidia has the best connectivity technology at the moment, but the project may also proceed without Nvidia technology.
On the risk side, the shortage of memory chips across the industry continues to drive up hardware costs. If Apple wants to deliver servers on a large scale, it needs to lock in component supply several years in advance; at the same time, the gap between its AI software ecosystem (such as the MLX framework) and CUDA, as well as the loss of core talent during the year, are all lessons that cannot be avoided.
Conclusions
For Tenus, who has just completed the CEO handover, the server project is not only a natural extension of his hardware engineering background, but also a strategic gamble that defined the legacy of his tenure. In the short term, the news provided a new room for imagination in Apple's AI narrative, compounding that the Mac business grew beyond expectations, and the mood was warm; however, looking at the extended cycle, whether 2029 products can break the gap in the crowded reasoning market depends on whether Apple can make up for the biggest shortcoming in enterprise-level capabilities. For investors, instead of chasing rumours, it is better to keep an eye on three signals: whether data center capital expenditure has been recorded, whether customers have landed, and the final choice of the connected technology route.