Taiwan Semiconductor Manufacturing Company reported revenue of $33.7 billion last quarter, up 35.9% year over year. The high-performance computing segment, which includes every AI chip that matters, grew 48% and now represents 58% of the company's total revenue. TSMC has committed $45 billion in capital expenditure for 2026. The company commenced mass production of 2-nanometer chips in late 2025 and projects overall revenue growth of roughly 30% this year, with AI-specific revenue growing at a mid-to-high 50% annual rate through 2029.
None of this is about TSMC's stock price, which sits at $342 and a $1.7 trillion market cap. For the cooling industry, TSMC's numbers are the closest thing to a forward-looking demand model that exists. Every chip TSMC produces for AI applications ends up in a server that needs thermal management. The company's production volume is, in effect, a leading indicator for cooling infrastructure demand 12 to 18 months into the future.
TSMC controls over 70% of the global foundry market. Nvidia, AMD, Apple, Qualcomm, and Broadcom all rely on TSMC to manufacture their most advanced processors. For AI data centers specifically, TSMC fabricates the Nvidia H200, B200, and the upcoming Rubin GPUs. It produces AMD's MI300X and MI400 accelerators. It builds the custom AI chips that Google, Amazon, and Microsoft design in-house.
The company's node breakdown tells the thermal story. Three-nanometer chips accounted for 24% of wafer revenue last quarter. Five-nanometer and 7-nanometer combined for another 50%. Advanced nodes at 7nm and below now represent 74% of total wafer revenue. Each node shrink allows more transistors per die, which means more compute per chip, which means more heat per chip. The 2nm generation entering production this year will push that curve further.
TSMC's advanced packaging technology, CoWoS, is equally relevant. CoWoS packages multiple dies, including GPU, memory, and interconnect chiplets, into a single module. The thermal density of a CoWoS package exceeds what any individual die generates because you are stacking heat sources in close physical proximity. Nvidia's Blackwell architecture uses CoWoS extensively. So does every next-generation AI accelerator in the pipeline.
TSMC's capital expenditure goes primarily toward building and equipping new fabrication facilities. A 2nm fab is among the most thermally demanding industrial facilities that exist. EUV lithography tools, each consuming roughly a megawatt, operate in clean rooms maintained at precisely controlled temperatures. Process chemicals must be heated and cooled through multiple stages. The wafers themselves cycle through thermal steps that require cooling loops with stability measured in hundredths of a degree.
TSMC is building new fabs in Arizona, Japan, and Germany in addition to its ongoing expansion in Taiwan. Each of these facilities represents a multi-billion-dollar cooling infrastructure project before a single production wafer is processed. The Arizona fabs alone have required the construction of dedicated water treatment and cooling facilities in a desert climate where evaporative cooling capacity competes with municipal water needs.
The cooling vendors serving TSMC's fab construction projects are a different tier from those serving data centers. Industrial process cooling for semiconductor manufacturing involves ultra-pure water systems, chemical-resistant heat exchangers, and temperature stability requirements that make data center cooling look approximate by comparison. But the demand driver is the same: more compute capacity being built at more advanced nodes generates more heat at every stage of the production and deployment chain.
Here is the connection that matters for cooling professionals. TSMC's revenue growth directly translates to chips shipping into AI data centers. A 48% growth rate in the HPC segment means 48% more high-wattage processors leaving TSMC's fabs and arriving at hyperscale and enterprise data centers within two to three quarters. Those chips need cold plates, CDUs, chilled water loops, and facility-level heat rejection infrastructure.
TSMC projects AI revenue growing at 50% or more annually through 2029. Compounded over four years, that roughly triples the volume of AI chips entering the data center market. The cooling infrastructure to support those chips needs to be designed, procured, manufactured, and installed on a timeline that matches or leads the chip delivery schedule. It does not. Lead times for CDUs have stretched past six months. Cold plate manufacturing is running near capacity at the major suppliers. The supply gap that exists today widens with every TSMC earnings beat.
The $45 billion capex number also represents TSMC's confidence that the demand is sustained. Semiconductor fabrication equipment has a payback period measured in years. When TSMC commits that level of capital, it has multi-year purchase commitments from its largest customers backing the investment. Those commitments mean the chip volumes are coming. The cooling demand behind those volumes is as close to guaranteed as anything in this industry gets.
Cooling vendors have been expanding manufacturing capacity throughout 2025 and into 2026. Motivair opened a fourth production facility after being acquired by Schneider Electric. CoolIT, pending the Ecolab acquisition, has been scaling production to meet hyperscale orders. Vertiv, Boyd (now Eaton), and a growing number of Chinese suppliers including Envicool are adding lines.
The question is whether the expansion is fast enough. TSMC's production ramp follows a schedule set by its equipment installation and yield improvement curves, both of which the company executes with a consistency that borders on mechanical. The cooling industry's expansion depends on hiring skilled labor, qualifying new manufacturing lines, and securing raw material supply chains for copper, aluminum, and specialty alloys. Those constraints are softer, slower, and less predictable than TSMC's.
TSMC's numbers tell the cooling industry what is coming. The industry's job is to be ready when it arrives. Based on current lead times and capacity expansion schedules, it will not be.