TSMC’s top customer has changed.

01 The AI Wave Is Reshaping the Economic Landscape of the Wafer Foundry Industry

Nvidia’s rise to the top of TSMC’s customer rankings reflects an unprecedented economic structure driven by the AI revolution. Nvidia’s H100 and H200 data center GPUs, manufactured with TSMC’s advanced 4nm and 5nm processes, command an average unit price of over $30,000, with some configurations fetching more than $40,000. Such profit margins far outpace those of smartphone processors – even Apple’s high-end A-series and M-series chips pale in comparison – a shift that has fundamentally reshaped the incentive structure within the semiconductor value chain.

Industry insiders reveal that Nvidia’s wafer orders have grown exponentially over the past 18 months, with the company locking in a substantial volume of TSMC’s advanced process capacity for nodes including N3 and N4P. Demand has been so robust that TSMC has been forced to expand its advanced packaging capacity exclusively for Nvidia to meet its CoWoS (Chip-on-Wafer-on-Substrate) requirements. This packaging technology is critical to high-bandwidth memory (HBM) configurations and serves as a core enabler for modern AI accelerators; it has also emerged as a key bottleneck in the AI supply chain, compelling TSMC to invest billions of US dollars in ramping up related production capacity.

The significance of this shift becomes even more apparent when weighing price against scale. Apple ships hundreds of millions of chips each year for products such as the iPhone, iPad and Mac, yet these consumer-oriented processors carry a relatively low unit price. By contrast, Nvidia, despite its lower shipment volumes, outperforms with an exceptionally high per-unit chip value.

A single AI training cluster deployed by a major tech firm may require thousands of Nvidia GPUs, corresponding to a semiconductor value of hundreds of millions of US dollars – equivalent to the total chip value used in millions of smartphones.

02 Apple’s Evolving Semiconductor Strategy

Apple’s being overtaken by Nvidia does not signal a decline in the company’s strength; rather, it largely reflects that its product portfolio has entered a mature phase and that it has adopted a more restrained approach to the pace of hardware upgrades.

In 2022, Apple completed the full transition of its Mac product line to in-house designed chips, which created a one-off sharp surge in chip demand that has since gradually returned to normal. Additionally, Apple’s strategy to extend product lifecycles and bolster services revenue has naturally slowed the growth rate of its semiconductor demand.

Apple remains a key customer of TSMC and continues to secure priority access to the most advanced manufacturing processes for its flagship products. Its upcoming A19 and M5 chips are expected to be built on TSMC’s N3E process, and may even adopt the N2 process – ensuring Apple remains at the technological forefront of the industry, even as its relative share of TSMC’s revenue declines.

This highlights the critical distinction between absolute scale and relative position: Apple’s total chip procurement volume is still growing; it is simply that Nvidia’s demand is growing at a far faster pace.

Meanwhile, Apple’s semiconductor strategy is becoming increasingly diversified. The company is continuously developing custom chips for specific applications, including the Neural Engine for on-device AI, custom image signal processors (ISPs), and dedicated chips for accessories such as AirPods and Apple Watch.

While this diversification is technically impressive, it also scatters Apple’s manufacturing demand across more process nodes and even multiple foundries, thereby reducing its concentration on the most advanced process nodes – the very area where Nvidia’s demand is most highly concentrated.

03 TSMC’s Dilemma in Capacity Allocation

It must not only meet Nvidia’s pressing demand for AI chip capacity and align with the launch cadence of Apple’s consumer electronics products, but also serve AMD, Qualcomm, MediaTek and the fast-growing cohort of AI chip startups – all against the backdrop of the company’s massive capital expenditures. TSMC’s capital expenditure exceeded $40 billion in 2024 and is expected to remain elevated through 2026, earmarked for expanding production capacity in Taiwan China, Arizona (US), Japan, and potentially Europe.

The shift in its customer mix toward AI is also reshaping TSMC’s technological R&D priorities. Smartphone processors prioritize energy efficiency and incremental performance gains, while AI accelerators pursue extreme performance with a higher tolerance for power consumption – a dynamic that demands a fundamental overhaul of its process optimization strategies. TSMC’s technology roadmap now features dedicated process variants for high-performance computing (HPC), engineered to boost power delivery and thermal performance, making them far better suited for data centers than mobile devices.

Industry insiders, however, point out that the risk of customer concentration is actually on the rise. Unlike Apple’s stable, predictable demand cadence, AI chip demand is far more volatile, being heavily swayed by corporate IT investment cycles and the uncertainties surrounding the commercialization of AI applications. A slowdown in AI infrastructure investment could leave TSMC with overcapacity at its most advanced nodes – a scenario that would bring significant financial pressure amid the company’s hefty capital outlays.

04 Competitive Landscape and Market Impact

The strengthening of the relationship between Nvidia and TSMC is taking place against the backdrop that other wafer foundries still struggle to catch up with advanced process technologies. While Intel’s foundry ambitions are backed by government subsidies and internal resources, it will take years for the company to close the gap on leading process nodes; Samsung, despite pouring heavy investment into advanced processes, has been hindered by yield issues and a lack of customer confidence, making it unable to secure high-value product orders for its most advanced manufacturing technologies. This has endowed TSMC with formidable bargaining power in the realms of advanced processes and advanced packaging – particularly in the technological segments that Nvidia relies on heavily.

Geopolitical uncertainties have also added complexity to the long-term capacity planning of both Nvidia and TSMC. As TSMC prioritizes securing capacity for high-margin AI chips, other customers are facing extended lead times and potential pricing hikes for advanced process access. This trend is driving major tech companies to accelerate the development of their in-house chips and lock in production capacity through long-term agreements or strategic investments.

Companies including Amazon, Google, Microsoft and Meta have all ramped up their efforts in developing custom chips – a move aimed not only at reducing their reliance on Nvidia, but also at securing manufacturing resources in an environment of tightening global semiconductor capacity.

05 The Future Trajectory of Semiconductor Manufacturing

Looking ahead, the relationship between Nvidia, Apple and TSMC will continue to evolve. TSMC’s 2nm process, slated to enter volume production in 2025, will once again test its ability to allocate scarce advanced manufacturing capacity. Preliminary signs indicate that both Nvidia and Apple have locked in a portion of N2 capacity, though the final allocation will hinge on negotiations and market conditions.

There remain divergent views on the long-term trajectory of AI chip demand. If investment in AI infrastructure cools, the relative positions of Nvidia and Apple at TSMC could shift once more; if Apple launches groundbreaking new products that drive a new wave of consumer hardware upgrades, its chip demand could stage a notable rebound. Long-rumored AR/VR products, automotive projects and new device form factors all represent potential variables in this equation.