404K SEMI-AI 2026-07-17 AI Supply Chain Weekly — TSMC Raises Capex, 800V and Liquid Cooling Accelerate, Optical Interconnect and PCB Capacity Expands
目录
Executive Summary
Weekly View
AI Chips, Advanced Packaging, and Servers
Optical Communications, CPO, PCBs, and High-Speed Interconnects
Data-Center Cooling, Power Systems, and Electricity
Robotics, Edge AI, and Hardware Applications
Space, Quantum Computing, and AI Software Applications
Debates, Bear Cases, and Next Week’s Watchlist
AI demand is not cooling, but the investment thesis has shifted from “is there enough compute?” to “who can deliver power, cooling, interconnects, and production capacity on schedule?”
Executive Summary
TSMC provided the strongest demand validation this week: Q2 2026 revenue reached US$40.2 billion, up 12% quarter on quarter and 34% year on year, with Q3 guidance of US$44.6 billion–US$45.8 billion. The company raised its 2026 capex budget to US$60 billion–US$64 billion. Demand has broadened from GPUs to CPUs, custom chips, and networking chips, while advanced nodes and advanced packaging remain undersupplied. There are no typical signs of manufacturing capacity being expanded merely to build inventory.
As rack power rises from tens of kilowatts toward 600 kilowatts and ultimately 1 megawatt, power delivery, cooling, and data transmission increasingly determine system delivery more than any individual chip. 800V DC can reduce current, copper usage, and losses from multiple conversion stages, while liquid cooling addresses the heat-flux challenges of high-density racks. Neither is a distant concept; both are prerequisites for next-generation racks to launch on schedule.
“Using less water” does not mean “using less electricity.” By 2035, an estimated 56% of new global data-center capacity and 55% of new US capacity will be exposed to extreme heat, humid heat, or water constraints. Some near-zero-water heat-rejection solutions may consume roughly 25%–45% more electricity. Cooling, grid access, on-site power, and community permitting must therefore be valued as one integrated system.
High-speed interconnect competition is shifting from port counts to power consumption, density, and packaging. The Ethernet switch market is projected to expand from approximately US$50 billion in 2025 to more than US$200 billion in 2035, while the data-center interconnect market could grow from roughly US$9 billion in 2025 to approximately US$33 billion in 2030. This creates a full upgrade cycle across 800G, 1.6T, co-packaged optics, silicon photonics, fiber-array units, and high-specification optical fiber.
PCBs and copper-clad laminates are not merely following the cycle upward; they are direct beneficiaries of rising AI-system specifications. Shennan Circuits guided for Q2 net profit growth of 44%–67% year on year, supported by Wuxi capacity expansion, 1.6T optical-module boards, and server and switch boards. Shengyi Technology’s M9/M10, PTFE, and high-layer-count board roadmap, Han’s Laser’s drilling equipment, and Ibiden’s high-end ABF substrates respectively capture rising content value in materials, equipment, and package substrates.
Evidence at the application layer is clearly diverging. Electronic design automation agents already have a quantifiable labor shortage and monetization model; robotics still depends primarily on the pace of mass production; and AI PCs face the contradiction of rising penetration but declining total shipments. Not all “AI applications” should be treated as having the same earnings-realization timeline.
The principal bear case is no longer a sudden disappearance of compute demand, but a mismatch between delivery speed and returns on capital. Local project suspensions, disputes over residential electricity rates and water use, platform delays, supplier-share fragmentation, rising memory and materials prices, and excessive valuations could all turn strong demand into below-expectation profits or revenue recognized later than expected.
Weekly View
The most important development this week is another increase in the credibility of AI capex, alongside a reordering of the beneficiaries. Previously, the market focused mainly on how many GPUs cloud service providers would purchase. The more important questions now are whether advanced-node and packaging capacity can be expanded, whether racks have sufficient power, whether heat can be removed, whether switching networks and optical interconnects are properly matched, and whether complete racks can ultimately be delivered on time. If any link falls behind, chip orders become inventory awaiting installation. If any link remains structurally tight, the profit pool spreads from accelerators into power, cooling, interconnects, board materials, and manufacturing equipment.
TSMC’s quarterly results provide a manufacturing-side anchor for this view. Q2 revenue reached US$40.2 billion, at the upper end of its original guidance, while Q3 guidance implies further sequential growth of approximately 12%. More importantly, the company raised 2026 capex to US$60 billion–US$64 billion. Management explicitly broadened its demand assessment to include GPUs, CPUs, custom chips, and networking chips, while incorporating cloud deployment plans, power availability, and data-center construction progress into capacity-expansion decisions. This indicates that manufacturers are seeing a medium-term roadmap spanning platforms and customers, rather than a one-off rush order from a single customer.
However, strong demand does not mean every segment monetizes simultaneously. Advanced-node capacity takes years to build, and advanced packaging continues to constrain customer shipments. Server systems and networking equipment are also limited by the availability of memory, optical components, substrates, racks, and power systems. Supply bottlenecks produce two opposing outcomes: scarce segments gain pricing power and stronger bargaining leverage, while system integrators may delay revenue recognition for complete racks because a single component is missing. Investors should separate strong orders, strong shipments, and strong profits rather than treating backlog as current-period earnings.
Capex has also entered a return-validation phase. Surveys of enterprise chief information officers show that AI budgets remain directionally positive, but their share of spending is still modest at many companies, with deployment initially favoring the cloud. PC replacement demand is being constrained by higher memory and processor costs and the fading replacement cycle. Infrastructure investment can lead application revenue by several years. That is both the opportunity in the current cycle and its greatest valuation risk. Companies deserving a premium are not merely those associated with AI, but those capable of converting scarce supply, customer qualification, and technical specifications into cash flow.
The preferred allocation order this week should therefore follow the “hard constraints.” The first tier comprises advanced nodes, packaging, substrates, and high-end board materials. The second includes power, cooling, networking, and optical interconnects. The third consists of server and system suppliers capable of delivering complete racks on schedule. Only the fourth includes robotics, edge devices, and software applications whose commercialization pace remains unproven. The closer a company is to a hard constraint, the higher its revenue visibility; the closer it is to a long-dated application, the more its valuation depends on product timing and user willingness to pay.
AI Chips, Advanced Packaging, and Servers
TSMC remains the first place to assess the broader cycle. AI demand has expanded from GPUs to CPUs, custom chips, and networking chips, meaning wafer demand is no longer concentrated in a single product line. The company expects revenue growth to exceed 40% in 2026 and believes supply will remain tight for several years. Capacity for the 2-nanometer family could expand at a compound rate of more than 70% from 2026 to 2028, while advanced packaging such as CoWoS remains a major constraint on customer shipments. The near-term cost is gross-margin dilution from the 2-nanometer ramp and overseas fabs, but this is better understood as the cost of expanding supply than as evidence of weakening demand.
Incremental advanced-packaging demand does not accrue solely to foundries. Higher high-bandwidth-memory layer counts, CoWoS expansion, EMIB-T, and investment by outsourced semiconductor assembly and test providers are collectively expanding the thermocompression-bonding equipment market. The central debate around ASMPT is how much of the strong demand outlook is already reflected in valuation and whether the company can increase its share in memory thermocompression bonding. Target prices and ratings differ across reports, but the medium-term drivers are broadly consistent: logic-chip packaging, high-bandwidth memory, outsourced assembly and test investment, and Intel’s EMIB-T will all increase equipment demand. Investors should track actual orders, equipment utilization, and fluxless-process penetration rather than focusing only on the industry’s theoretical addressable market.
ABF substrates are also entering a phase of simultaneous supply growth and specification upgrades. Ibiden benefits not only from GPU substrates but also from standard ABF substrates used in EMIB-T and custom chips. The key questions are whether capacity will tighten again after the second half of FY2029 and whether new plants, external capacity, and recruitment can come online on schedule. Japanese component statistics also show stronger volume and pricing for ABF substrates than for most electronic components, suggesting that the cycle is not driven solely by one company’s optimistic guidance. Substrates and advanced packaging reinforce each other: the larger the chip, the more complex the package, and the denser the interconnects, the higher the required substrate specifications and content value.
Competition for custom-chip share will continue, but the addition of a second supplier should not be equated with the incumbent losing all its advantages. MediaTek’s entry into Google’s tensor processing unit supply chain confirms that cloud providers want to reduce single-supplier risk. Broadcom retains capabilities in high-bandwidth-memory integration, advanced packaging, large-scale deployment, and system execution. A more reasonable conclusion is that individual customer share may fragment while the overall custom-chip market continues to expand, with new customers partly offsetting share changes. The next indicators to watch are tape-outs, mass production, chip value per gigawatt, and new customer deployments—not a static reading of supplier lists.
The server-system opportunity comes from higher content value rather than simple unit growth. Huaqin Technology’s rack-scale projects are expected to ramp from the second half of 2026, with servers and high-speed switches becoming incremental gross-profit contributors in 2027–2028. Factory visits in Southeast Asia also show capacity expansion across PCBs, optical modules, optical engines, and complete server systems. Customers are willing to pay for geographic supply diversification and delivery certainty, including through equipment prepayments, but manufacturing in mainland China retains advantages in efficiency, materials ecosystems, and engineering talent. Whether overseas plants become profitable depends on qualification, automation, utilization, and yield—not merely on having expanded abroad.
The shared risk for system manufacturers is constrained supply. Server and networking companies broadly report strong orders, but memory, flash storage, optical components, chips, and system-level qualification will limit revenue upside. Long-term memory agreements, high-bandwidth-memory pricing, and tight enterprise SSD supply and demand can improve the durability of upstream earnings but raise system bill-of-materials costs. The ultimate winners must possess procurement security, pricing pass-through, and system-delivery capabilities simultaneously.
Optical Communications, CPO, PCBs, and High-Speed Interconnects
As compute density rises, networking is no longer an ancillary investment outside the GPU; it is central to effective throughput. The Ethernet switch market is projected to grow from approximately US$50 billion in 2025 to more than US$200 billion in 2035, while the data-center interconnect market could expand from roughly US$9 billion in 2025 to approximately US$33 billion in 2030. Growth comes through three paths: scale-out between racks, scale-up between accelerators, and inter-domain connectivity among multiple data centers. Networking revenue typically lags compute purchases because customers must first determine how much interconnection they require, but networking’s share of long-term capex should increase.
As 800G transitions to 1.6T, the limitations of copper interconnects in distance, loss, density, and heat dissipation become more pronounced. Co-packaged optics moves the optical engine closer to the switch chip. Its most immediate value is lower interconnect power consumption, with industry estimates indicating reductions of approximately 50%–80%. At the same time, the challenge shifts from “adding ports” to solving packaging, light-source, fiber-array, connector, and serviceability issues. Silicon-photonics wafers, external lasers, fiber-array units, high-specification fiber, and testing equipment should therefore benefit alongside traditional pluggable modules.
Current silicon-photonics capacity expansion is better understood as securing qualified capacity in advance than as evidence of emerging oversupply. Manufacturing initiatives by TSMC, UMC, Tower, and others will increase future supply, but photonic integration remains constrained by process stability, coupling, packaging, lasers, and testing. Largan Precision’s fiber arrays remain in sampling and qualification, with the earliest mass-production timing skewed toward mid-2027. This indicates that the long-term opportunity in optical components is real, but revenue recognition remains subject to customer qualification and pilot-line progress.
Fiber demand is also shifting from “installing more fiber” to “using higher-specification fiber.” Yangtze Optical Fibre and Cable benefits from data-center customer expansion and a better product mix, while co-packaged optics, near-packaged optics, multicore fiber, and hollow-core fiber provide further upgrade potential. Supply expansion cannot be ignored, however. Under baseline estimates, demand could grow at a compound rate of approximately 4% from 2025 to 2030, versus roughly 6% for supply, with the market potentially loosening in 2028. The fiber segment should therefore be evaluated through high-end product mix and customer qualification rather than by extrapolating current price increases indefinitely.
PCBs are among the clearest earnings-realization segments this week. Shennan Circuits guided for Q2 net profit growth of 44%–67% year on year, driven by the ramp of its Guangzhou plant, AI compute and storage demand, and a better product mix. The company plans approximately RMB6 billion of capex in 2026, up from RMB3.8 billion in 2025, and intends to use private-placement proceeds to expand AI PCB capacity in Wuxi. Servers, switches, 1.6T optical-module boards, BT substrates, and ABF operations collectively support growth, while customer and product diversification is superior to relying on a single GPU platform.
Specification upgrades in copper-clad laminates and core materials warrant even greater attention. Shengyi Technology benefits from higher prices for premium copper-clad laminates, AI customer demand, and new capacity, with M9/M10, PTFE, and high-layer-count boards forming its medium-term roadmap. Recent materials-industry data also indicate that ultra-high-layer-count boards are more likely to use M9 with Q-glass fiber, while PTFE’s drilling and lamination difficulties make it more suitable for certain lower-layer-count topologies. Q-glass fiber is expensive, hard, and brittle, while mass production of next-generation substitutes is still some way off. Strong demand and high yields are therefore not the same thing. Han’s Laser’s ultrafast drilling equipment, materials suppliers, and high-end board manufacturers will all participate in the specification upgrade, but processing bottlenecks could also delay capacity realization.
Platform delays and supplier-share fragmentation remain important counterarguments. Victory Giant Technology faces near-term earnings pressure from the ramp of high-density interconnects, customer supply-chain diversification, and shifting platform schedules, although Rubin, custom chips, and 800G/1.6T optical modules should still support medium-term growth. For companies of this type, customer share, yield, capacity utilization, and free cash flow should take precedence over order growth.
Data-Center Cooling, Power Systems, and Electricity
As rack power rises toward 600 kilowatts and ultimately 1 megawatt, the current, copper usage, and multi-stage conversion losses of traditional low-voltage AC power systems increase rapidly. The value of 800V DC lies in using higher voltage to reduce current, making cables, busbars, and conversion stages better suited to high-power racks. Data centers will not rebuild all facilities at once. Upgrades are more likely to proceed in three stages: first strengthening power equipment around the rack, then expanding the use of 800V DC across the power-delivery chain, and finally redesigning facility-level power distribution. This creates incremental demand for analog chips, silicon-carbide and gallium-nitride power devices, solid-state transformers, circuit breakers, power racks, and connectors before full adoption.
Texas Instruments benefits from data-center power sockets and an industrial-cycle recovery, while 800V DC and high-power racks increase content value per system. Sungrow is bringing energy storage, solid-state transformers, and 800V products into data centers. Sungrow’s energy-storage gross margin may face pressure in Q2, but lower lithium prices should support costs in Q3, while energy-storage shipments continue to grow rapidly. The key issue is not merely one quarter’s margin, but whether energy storage can expand from peak shaving and load shifting into a component of data-center power stability.
Cooling and power must be assessed together. By 2035, an estimated 56% of new global capacity and 55% of new US capacity will be located in regions exposed to extreme heat, humid heat, or water constraints. Near-zero-water heat-rejection systems may consume approximately 25%–45% more electricity. Simply replacing evaporative cooling with dry cooling therefore does not eliminate resource constraints; it transfers pressure from water to electricity. Liquid cooling can reduce heat-rejection demands in the white space and support higher rack density, but non-white-space mechanical equipment, the grid, and water treatment remain indispensable.
Local resistance has evolved from a reputational risk into a construction-schedule risk. Approximately US$156 billion of US projects were canceled or delayed in 2025, and roughly US$130 billion of projects had already been affected in Q1 2026. Residents’ objections center on electricity rates, water use and waste heat, as well as noise, dust, and traffic. Most current policies involve temporary moratoriums rather than permanent bans, while the federal government is more likely to provide conditional support than impose a nationwide prohibition. Nevertheless, costs, timelines, and geographic distribution will all change.
