Global Power Semiconductors: AI Infrastructure Creates a New Profit Pool, but the Industry Remains Bifurcated Between Mature Silicon, SiC Digestion, and Grid-to-Core Winners

1. Executive Overview

The most useful way to frame the global power semiconductor industry is as 2 overlapping markets. The narrowest common denominator is power discretes and modules, which Infineon cites at $32.8bn in 2024. The broader and more investable definition includes PMICs, gate drivers, controllers, battery-management ICs, smart power stages, protection, and wide-bandgap devices, which places the market closer to roughly $56.9bn to $57.0bn in 2025 and about $59.9bn in 2026. That broader framing is the right one for stock selection because capital and competitive advantage are spread across IDMs, fabless PMIC vendors, module vendors, SiC and GaN specialists, specialty foundries, materials suppliers, and adjacent analog names rather than only the classic discrete and module leaderboard.

The two-lens market framework also needs a hard competitive anchor. In the narrow discrete and module view, Infineon leads with 17.4% share, followed by onsemi at 8.5%, STMicroelectronics at 6.9%, Mitsubishi Electric at 4.6%, and Fuji Electric at 3.9%. That snapshot does not define the full investable universe, but it is the cleanest reality check on who controls the traditional core before the market broadens into PMICs, drivers, controllers, foundry exposure, and materials leverage.

The current industry condition is bifurcated. Broad silicon power remains a mature, manufacturing-intensive market tied to automotive, industrial, renewables, appliances, computing, and communications. Within that broad market, SiC is working through a cyclical digestion phase after an aggressive 2019 to 2024 capacity build, while AI infrastructure is creating localized tightness in low-voltage, high-current, high-frequency power delivery and in higher-voltage front-end conversion for data centers. In practice, that means the sector is simultaneously dealing with upstream overcapacity in some SiC layers and rising strategic value in select AI-relevant BCD, GaN, smart-power-stage, packaging, and electrical-infrastructure niches.

• Most important market conclusion: the core market is large, but not hypergrowth. Structural upside comes from electrification, efficiency, and wide-bandgap mix shift rather than from aggregate unit growth alone.
• Most important architecture conclusion: AI is reshaping power semiconductors less through logic demand and more through electrical-system redesign from grid interface to GPU core.
• Most important competitive conclusion: the highest-value layers remain IDM-led because power devices compete on device physics, packaging, module design, reliability, and application engineering at least as much as on process node.
• Most important near-term debate: broad SiC oversupply is real, but it should not be mistaken for broad power-semiconductor weakness. The profits are moving toward the stages of the stack where efficiency, thermal management, and conversion-stage reduction matter most.

2. Core Evidence

The practical reading is to start with the narrow share snapshot to anchor traditional merchant leadership, then broaden back out to the larger investable power stack where PMIC suppliers, specialty foundries, and materials companies capture incremental value. That prevents the analysis from being either too narrow for stock selection or too diffuse to be actionable.

The most analytically useful framing is to treat the industry as a layered stack rather than a single chip category. Layer 1 is large integrated power vendors such as Infineon, STMicroelectronics, onsemi, ROHM, Mitsubishi Electric, Fuji Electric, Renesas, Toshiba, Vishay, NXP, Microchip, Littelfuse, and Alpha and Omega. Layer 2 is analog and power-management IC suppliers such as Texas Instruments, Analog Devices, Monolithic Power Systems, Vicor, Power Integrations, Silergy, and Richtek. Layer 3 is wide-bandgap specialists such as Wolfspeed, Navitas, Efficient Power Conversion, and Innoscience. Layer 4 is specialty foundries such as TSMC, GlobalFoundries, Intel, Tower, X-FAB, UMC, VIS, and PSMC. Layer 5 is the upstream materials stack, where Coherent, SiCrystal, Ferrotec, Entegris, and Mersen matter. Layer 6 is assembly, modules, test, and system integration, where Delta, Vertiv, PSU vendors, and test specialists such as Aehr become relevant.

3. Industry Structure and End Markets

Customers should be separated into direct customers and economic end customers. Direct customers are often Tier-1 automotive suppliers, industrial OEMs, PSU manufacturers, server ODMs, contract manufacturers, distributors, and module integrators. Economic end customers are automotive OEMs, renewable-energy developers, utilities, telecom operators, appliance brands, hyperscalers, cloud providers, and enterprises building AI clusters. That distinction matters because semiconductor design wins often happen one layer upstream from where the economic demand signal is actually created.

The economic center of gravity remains automotive and industrial even as AI becomes strategically important. Infineon’s FY25 mix illustrates that clearly: 13% e-mobility, 11% power infrastructure, 8% software-defined vehicle, 6% IoT, and 3% AI/data center, with other applications making up the remainder. The implication is not that AI is unimportant. It is that diversified power leaders still derive the bulk of their economic relevance from auto, industrial, and infrastructure applications, while AI acts as the fastest-growing and most strategically valuable incremental pool.

4. TAM and Growth Framework

The market is large, but the growth rate depends almost entirely on where one sits in the stack. The broad power semiconductor market is still a mid-single-digit grower. That means the sector should not be pitched as a monolithic secular hypergrowth category. Instead, the right underwriting question is which material system wins at which voltage, current, frequency, thermal, and packaging point, and whether the company has exposure to the fastest-growing application layers rather than to the category average.

The right investment framing is therefore selective rather than monolithic. Broad power semis should not be bought or sold as a single thematic basket. The differentiation comes from exposure to EV power content, grid modernization, AI server power chains, wide-bandgap transitions, and the specialty foundry and packaging ecosystems that enable those design wins.

5. AI Data Center Power Architecture

Generative AI is changing the power industry less through the logic device itself than through the electrical architecture around the rack. Hyperscaler capex is now large enough to pull power semiconductors into a strategic role. Infineon’s late-2025 AI briefing cited more than $300bn of 2025 AI spending by Meta, Microsoft, Amazon, and Alphabet, reported that its AI server revenue exceeded €700m in FY25 and nearly tripled versus FY24, and guided to about €1.5bn in FY26. The IEA, meanwhile, projects global electricity consumption by data centers to more than double to roughly 945 TWh by 2030, with AI as the most important driver. The implication is that AI is forcing a redesign of the electrical system from grid interface to GPU core, not merely creating another semiconductor end market.

The immediate technical problem is efficiency at much higher rack power. onsemi states that hyperscale AI racks already require about 120kW and that conversion from grid power to GPU voltage is only about 88% efficient, leaving roughly 15kW of waste heat to remove. NVIDIA’s 800 VDC architecture is a direct response. NVIDIA says 800 VDC reduces conversion stages and routing losses, supports future AI servers, can improve end-to-end power efficiency by up to 5%, and can reduce copper use and cable bulk meaningfully, with its May 2025 technical blog tying full-scale production to Kyber systems in 2027 and citing up to 30% lower TCO. Each incremental point of efficiency is monetized twice, once in avoided electricity and again in avoided cooling and facility capex.

AI-specific TAM remains small relative to the full power market, but the scaling is already material enough to influence capital allocation. Separate Yole-derived reporting points to a data-center PSU market above $14bn by 2030, growing at about 15.5% CAGR, with the more than 3kW segment growing above 25%. Just as important, ST has been explicit that 50V, 12V, and 6V intermediate buses are likely to coexist depending on GPU generation, rack density, server form factor, and thermal envelope. The implication is that the AI power opportunity is real, but it is not a single-device bet. It is a multi-stage architecture transition with multiple winners across front-end conversion, intermediate buses, point-of-load regulation, protection, sensing, packaging, and thermal management.

The architecture shift is already visible in public company roadmaps. TI’s 2025 white paper describes AI servers transitioning from 400V AC toward 800V DC architectures. ST’s APEC 2025 materials map the chain as AC-to-800V, 800V-to-54V or 12V or 6V, and then 12V-to-GPU, and ST expanded its NVIDIA-aligned 800 VDC portfolio in March 2026 to include 800V-to-12V and 800V-to-6V solutions in addition to the original 800V-to-50V stage. Infineon similarly describes both 2-stage and 3-stage 800V-to-GPU paths in its public AI roadmap. The investment implication is straightforward: the best-positioned vendors are the ones with content at multiple stages of the chain rather than the vendors selling one preferred transistor type in isolation.

6. Material Systems and Device Winners

Silicon remains the incumbent by both volume and economic relevance across most of the market. It still dominates low-cost MOSFETs, IGBTs, PMICs, drivers, controllers, and many server and client power rails. Even in AI data centers, current 48V environments still offer only moderate GaN and SiC opportunity in many vendor frameworks. That matters because investor enthusiasm can sometimes overstate how quickly the entire industry will pivot away from silicon. The category is still built on mature silicon platforms, especially where cost and reliability dominate the trade-off.

Wide-bandgap becomes compelling when efficiency, switching frequency, power density, temperature tolerance, and high-voltage operation justify the cost premium. In practical terms, SiC is strongest in the roughly 650V to multi-kV range and in hard-switched, high-power applications such as traction, fast charging, UPS, energy storage, solid-state transformers, and potentially front-end 800V data-center conversion. GaN is strongest where density and frequency matter most, especially in lower-to-mid-voltage conversion, fast transient response, and compact high-frequency stages such as server PSUs, intermediate bus converters, and GPU-proximate power. This is why the strongest vendors increasingly market silicon, SiC, and GaN together rather than arguing for one universal material winner.

The cleaner analytical rule is to match each material system to a voltage, current, frequency, thermal, and packaging regime rather than to ask whether silicon, SiC, or GaN wins the market outright. Silicon remains the economic base, SiC remains strongest in high-voltage and high-power hard-switched domains, and GaN remains strongest where switching frequency, transient response, and power density translate most directly into system-level economics.

The leading public vendors are converging on a grid-to-core portfolio strategy. Infineon explicitly markets data-center solutions from grid to core using silicon, SiC, and GaN and is already quantifying AI server revenue separately. ST combines silicon, SiC, GaN, and analog and mixed-signal around 800 VDC architecture. onsemi markets EliteSiC, low- and mid-voltage MOSFETs, GaN, smart fuses, and Vcore solutions for AI data centers and has also launched vertical GaN aimed at AI data centers, energy, EVs, and industrial systems. Navitas is pursuing the cleanest pure-play wide-bandgap strategy around both 650V GaN and high-voltage SiC for NVIDIA’s 800V ecosystem. The competitive battle is therefore shifting from single-device superiority toward solution completeness, design-win intimacy, and the ability to collapse conversion stages safely and repeatably.

7. Foundry and Manufacturing Positioning, and Who Makes What

Power semiconductors are not a leading-edge-node industry in the digital-logic sense. They are a specialty-process, yield, packaging, and qualification industry. Silicon power discretes and many power ICs are built on mature 150mm, 200mm, and increasingly 300mm lines using high-voltage, BCD, trench, superjunction, and related process families optimized for Rdson, breakdown voltage, switching loss, thermal behavior, and reliability. The bottlenecks are therefore specialty line qualification, 200mm and 300mm analog and power capacity, wide-bandgap material quality, module packaging, and application-specific process transfer rather than EUV or cutting-edge logic capacity.

The capability split matters because the merchant foundry opportunity in power is not generic. TSMC is most relevant in high-voltage and BCD integration, Tower is the clearest public lever on AI smart-power and DrMOS proliferation, GlobalFoundries is becoming the most important U.S. GaN and high-voltage power-management platform, Intel matters mainly as a domestic capacity and packaging enabler, and X-FAB is the specialist route into niche SiC and GaN expansion. That is a very different competitive map from digital-logic foundry leadership.

The key manufacturing inference is that merchant foundry opportunity is real, but concentrated. It is most visible in BCD, HV analog, smart-power stages, GaN, and some SiC-related niches. It is not evidence that merchant foundries are about to displace integrated power vendors across the heavy discrete and module stack, where device, process, package, and application engineering are still tightly bundled.

8. Bottlenecks, Choke Points, and Material Risk

The most important current choke point for AI-driven data-center buildout is not the semiconductor die itself but external power infrastructure. The IEA has indicated that data-center electricity demand is set to double by 2030, and Reuters’ reporting on the IEA work said that 20% of planned projects face delays due to grid strain and shortages of essential infrastructure. Uptime’s 2025 survey also pointed to worsening power constraints as operators modernize for AI density. That is a crucial sector point because suppliers exposed to grid modernization, solid-state protection, UPS, storage, transformer replacement, and broader electrical architecture may capture more value than suppliers focused only on server-board PMICs.

For SiC specifically, the bottleneck has shifted from outright scarcity toward execution. Yole-derived reporting points to roughly 50% utilization upstream and roughly 70% at the device level in 2025, with the downturn potentially lasting until 2027 to 2028. That does not eliminate strategic risk; it relocates it into 200mm transition quality, defect density, epitaxy control, packaging reliability, and the ability to operate profitably through weaker pricing. At the same time, packaging layers such as DBC and AMB substrates, silver-sintering materials, copper clips, leadframes, mold compounds, and thermal interfaces remain underappreciated gating items for qualified shipment scale.

The most important risk translation is that location matters. The sector is no longer constrained primarily by one obvious wafer bottleneck. Instead, constraints are scattered across power availability, conversion efficiency, packaging, materials, HVDC ecosystem formation, and a handful of specialty manufacturing flows. That favors companies with broad system exposure and punishes narrow players whose thesis depends on a single scarcity narrative staying intact.

9. Investment Implications

The highest-quality franchises increasingly combine 4 attributes: control of multiple material systems, broad content from grid to load, manufacturing leverage in specialty processes and packaging, and close customer engagement at the architecture level. By that standard, Infineon appears to have the broadest public grid-to-core strategic positioning today. ST is becoming increasingly credible in 800V AI architecture while retaining strong SiC and smart-power assets. onsemi sits at the intersection of cyclical automotive and industrial digestion and secular AI server demand, with a growing GaN posture through its relationship with GF. Tower is a high-quality foundry lever on AI smart-power and PMIC proliferation. GF is emerging as a key U.S. GaN manufacturing platform. Intel is a second-order enabler rather than a first-order merchant power supplier. Wolfspeed and Coherent remain disproportionately important to the upstream SiC ecosystem, but the economics of that layer are now driven less by scarcity and more by utilization, pricing, and execution.

The central investment judgment is that global power semiconductors are no longer a purely cyclical auto-industrial analog and discrete sector. They are becoming a strategic infrastructure layer for electrification and AI. That does not make the whole market a high-growth AI market. It does mean the highest-value portions of the market are migrating toward applications where a 1% efficiency gain, one conversion stage removed, or one reduction in copper and thermal overhead translates into meaningful system-level economics. The best businesses will be the ones that can convert that system value into design wins, manufacturable packaging, and durable gross margins.

10. Risks and Disconfirming Evidence

There are several important reasons not to overstate the thesis. First, AI is strategically important but still a minority revenue stream for diversified power leaders. Infineon’s mix is the clearest reminder that automotive and industrial still dominate the economic base. Second, the broad power semiconductor market remains a mid-single-digit grower, so a clean secular narrative can hide the reality that much of the category is still cyclical and manufacturing-intensive. Third, AI power architectures are still evolving. The industry may not converge as quickly as bulls expect around a single 800V pathway, which means some product roadmaps may prove less monetizable than current investor materials imply.

• If AI rack and power-density growth slows, the premium multiples attached to AI-adjacent power names could compress faster than fundamental revenue exposure would suggest.
• If SiC utilization remains depressed through 2027 to 2028 or pricing deteriorates further, upstream ecosystem names may struggle even if long-term EV and energy-transition content remains intact.
• If foundries and packaging vendors capture a greater share of the value pool than expected, some integrated device names may see strong design activity without commensurate margin expansion.
• If grid constraints delay data-center projects more than expected, the timing of AI power-semiconductor monetization could slip even if the long-run architecture thesis remains correct.
• If low-cost silicon remains good enough for more applications than bulls assume, the adoption curve for GaN and some SiC use cases could be shallower than current narratives imply.

The most important disconfirming evidence would be a world in which AI remains too small to move the earnings power of diversified leaders, SiC oversupply lasts longer than the market expects, and the best economic value accrues mainly to upstream infrastructure and commodity electrical equipment rather than to semiconductor content. In that outcome, the sector would look much more like a traditional cyclical analog and power group with pockets of excitement rather than a structurally re-rated strategic infrastructure layer.

11. Catalysts and Watchlist

The practical watchlist is therefore straightforward. Track AI power-architecture disclosures, foundry and packaging expansion, 800V ecosystem milestones, SiC utilization and pricing, GaN manufacturing progress, and the pace at which utilities and data-center developers solve the upstream power bottleneck. Those variables will determine whether the sector earns a durable re-rating or simply oscillates between cyclical optimism and disappointment.