- Both ASML (ASML: NA/US) and TSMC (TSMC: 2330:TT) have raised their annual revenue guidance this week, affirming that leading American cloud service providers are disregarding short-term investment return pressures and continuing to expand their capital spending on advanced AI chips.
- Industry data indicates that tech giants such as Microsoft (MSFT:US), Meta (META:US), and Amazon (AMZN:US) are expected to spend over $600 billion on data centers this year, providing significant assurance for the performance of fabless chip designers.
- Computing power demand is undergoing a structural shift. TSMC management pointed out that as large language models (LLM) mature, the consumption of computing power is increasingly shifting from the model training end to the inference end, requiring the supply chain to deliver higher specification advanced processors and advanced packaging services.
The Battle Between Capital Expenditure Cycles and Return on Investment
The unexpected forward guidance released by TSMC and ASML this week provides strong fundamental support for the valuation center of the global semiconductor market. Previously, among Wall Street analysts, there was a growing concern about the bubble in AI assets. The core of the market's skepticism lies in the colossal investments in computing infrastructure by hyperscalers that have yet to translate into a corresponding cash flow return on the software end. However, TSMC CEO C.C. Wei's comments during an analyst call have temporarily allayed such concerns. He clearly stated that customers and their downstream major cloud service providers are continuously expressing extremely strong demand signals and optimistic outlooks.
This supply-demand dynamic indicates that in the arms race for artificial general intelligence (AGI), tech giants have regarded capital expenditure as a sunk cost to maintain their technological moat, rather than a financial tool pursuing short-term price-to-earnings ratios. The estimated $600 billion in data center capital expenditure this year not only covers core GPU procurement but also includes high-bandwidth memory (HBM), optical modules, liquid cooling systems, and supporting power infrastructure construction. Such cost-intensive investments ensure backlogs of companies like NVIDIA (NVDA:US), AMD (AMD:US), and Broadcom (AVGO:US) remain at historically high levels, with performance visibility extending to 2027.
Analysis of Advanced Manufacturing Capacity and Equipment Orders
As an absolute cornerstone asset of the global semiconductor manufacturing supply chain, ASML's extreme ultraviolet lithography (EUV) orders and TSMC's utilization rate of advanced process capacity are key proxy indicators for global technology capital liquidity. ASML's move on Wednesday to raise its annual revenue forecast directly reflects semiconductor foundries accelerating the expansion of 3nm and below advanced process capacities. The perspective of Reflexivity's co-founder Giuseppe Sette corroborates this, suggesting that ASML's positive data paints an exceptionally resilient and favorable picture for the semiconductor industry against the backdrop of a global macroeconomic slowdown.
TSMC not only raised its revenue guidance but also announced further increases in this year's capital expenditures. The core focus of this additional capital spending is likely to tilt towards expanding CoWoS (chip-on-wafer-on-substrate) advanced packaging capacity, apart from traditional wafer manufacturing nodes. In the current AI chip architecture, as lithography sizes approach their physical limits, advanced packaging technologies to heterogeneously integrate compute cores with high-bandwidth memory have become the only feasible path to boost computing density. TSMC's capacity bottleneck in advanced packaging has long been a key obstacle restricting NVIDIA's high-end chip delivery cycles. This upward revision of capital expenditure signals that this supply chain bottleneck is likely to be substantively alleviated in the coming quarters.
Evolution of Computing Power Structure: From Training to Inference
Another key trend revealed this earnings season is the structural pivot in AI computing power demand from training to inference. Over the past two years, the market's insatiable demand for computing power primarily stemmed from tech giants training large language models with billions of parameters from scratch. However, as foundational model frameworks gradually converge and commercial applications are realized, the inference tasks for processing everyday user queries, generating text, and videos are growing exponentially.
This shift in demand imposes fundamentally different requirements on underlying silicon design. Compared to the training phase's extremely high demands on the absolute computing power of a single card and cluster interconnect bandwidth, the inference phase prioritizes chip energy efficiency, latency performance, and throughput for specific tasks. This has driven computing power demand increasingly towards advanced processors and customized chips (ASIC) optimized for specific algorithms. Companies like Broadcom (AVGO:US), which have a deep technological foothold in custom network and computing chips, are witnessing new structural growth opportunities. Simultaneously, to support massive parallel inference computing, the order structure of TSMC's advanced process also becomes more diversified, further reducing excessive reliance on any single core customer.
