From Torque to Tokens: Power Semiconductors - The Other AI Trade | Long-term Investing : Thematic

There is a peculiar irony at the heart of AI infrastructure. The chips generating the most excitement in financial markets — the GPUs at the centre of every earnings call and capital allocation decision — cannot function until an entirely different set of semiconductors solves a problem most investors have never considered.

The Re-Rating Thesis

Every watt of power consumed by an AI accelerator must be transformed, regulated, protected, and delivered with extraordinary precision before a single calculation occurs. The companies making this possible are, in aggregate, less discussed than the GPU designers they serve. They are also, in several important cases, more defensibly positioned — because their revenues accrue regardless of which GPU vendor wins the AI compute race. Every rack pays the same toll to the same supply chain.

The physics is straightforward. Two equations explain why the entire power semiconductor supply chain must be re-engineered. First: P = V × I — power equals voltage multiplied by current. As AI racks push from 120 kilowatts toward one megawatt, the industry faces a choice: deliver that power by increasing current, or by increasing voltage. The previous generation chose current — running at 54 volts and pushing thousands of amps through copper. But the second equation determines why that approach has hit a wall: P_loss = I² × R — waste heat in copper scales with the square of the current. Double the current, quadruple the loss. At megawatt scale, the copper alone would weigh 200 kilograms per rack. The only viable path forward is to raise the voltage — from 54V to 800V — which cuts the required current by roughly 15×, slashing resistive losses by over 200×. This voltage transition is not a design preference. It is a law of physics, and it structurally requires silicon carbide and gallium nitride semiconductors that silicon cannot replace.

The sector was mispriced as an electric vehicle cyclical. When EV demand slowed in 2023, the market treated the entire power semiconductor industry as broken. It was not broken — it was incomplete. The EV recovery remains on track, with next-generation 800V vehicle platforms driving meaningfully higher semiconductor content per car. But two additional demand vectors — AI infrastructure and grid decarbonisation — have since emerged that were absent from every 2022 bull case. The market is currently pricing AI infrastructure power semiconductors as one recovering cyclical. The report argues that mispricing is the investment case.

A note on valuation: Several of the names highlighted in this report have already re-rated meaningfully from their 2023–2024 troughs. We acknowledge the initial move. Our thesis, however, is that the earnings inflection ahead — driven by the structural 54V-to-800V transition and the emergence of AI and grid demand — will dwarf the share-price recovery seen so far. More importantly, if this demand proves secular rather than cyclical, these companies deserve a fundamentally different multiple. The absence of the boom-bust pattern that has historically compressed power semiconductor valuations is itself a source of future capital appreciation that the market has not yet priced in.

Market at a Glance

Total Power Semi TAM~$55–57B2025A global market (Mordor Intelligence, Global Market Insights, SNS Insider)→ $77–95B by 2031–2035E (~5–6% CAGR)SiC Power Devices~$3B2024A; correction persisting through 2025–26 (Yole: ~50% upstream utilisation)→ ~$10.3B by 2030E (20% CAGR, Yole 2025)GaN Power Devices~$1.2B2024A est. (broader market incl. consumer; Yole power-device segment: ~$355M)→ ~$3B by 2030E (~42% CAGR, Yole 2025)Hyperscaler CapEx FY2025A~$384BAmazon $132B · Alphabet $91B · Microsoft $89B · Meta $72B (FY2025 actuals)Four largest hyperscalers; all five spent ~$450B+

Sources: Mordor Intelligence (2026), Global Market Insights (Mar 2026), SNS Insider (Apr 2026) for TAM; Yole Group Power SiC 2025 and Power GaN 2025 for sub-segment forecasts; company reported financials for hyperscaler capex. GaN figure reflects broader power market including consumer; Yole power-device segment only ~$355M in 2024. All forecasts subject to revision.

Why the Transition Is Happening Now — Four Forcing Functions

01The GPU power curve went parabolicThe electricity demand of a single NVIDIA AI server rack has grown roughly 60-fold in five years — from about 10 kilowatts (the A100 generation, 2020) to a projected 600 kilowatts for the next generation arriving in 2027. That is enough power to run several hundred homes from a single equipment rack. At these extremes, routing electricity sideways across a circuit board generates so much heat that the board itself becomes the bottleneck. The solution — moving the final power conversion step directly underneath the chip — is no longer optional engineering. It is a physical necessity.02Silicon ran out of roadStandard silicon semiconductors work well up to around 600 volts. Above that, they waste too much energy as heat — a fundamental property of the material that no amount of clever engineering can fix. Silicon carbide (SiC) handles three times the voltage at a fraction of the heat loss. Gallium nitride (GaN) switches on and off 50 to 100 times faster than silicon, enabling far more compact power supplies. Both are essential for the voltage levels that modern AI data centres and electric vehicles now demand.03The EV correction mispriced the whole sectorThe 2023–2025 SiC inventory cycle compressed multiples across the entire sector — including companies whose AI data centre revenues were growing rapidly. The market treated them as one story. They are not. Critically, AI data centre power demand is non-cyclical in the way EV is — hyperscaler capex commitments are multi-year and largely insensitive to consumer sentiment. The entry opportunity exists because investors remain anchored to the wrong demand driver.04Grid decarbonisation adds a third, uncorrelated pillarEuropean and North American policy-driven build-out of wind, solar, and grid-scale storage all require SiC and GaN at the inverter stage. This demand is uncorrelated with EV cycles and AI capex — it structurally widens the earnings floor and justifies a higher mid-cycle multiple.

The Full Value Chain — Where Value Concentrates and Where It Doesn't

Where Value Actually Concentrates

Highest-quality moats: Equipment incumbency (ASMI, AIXTRON) — 12–18 month switching costs protect revenue across cycle turns, entry at trough. Protection infrastructure (ADI, TXN) — grows with GPU power density, generation-agnostic, no socket reset risk. System integration (Delta) — less volatile than pure device names, direct AI CapEx linkage.Hidden chokepoint: Maruwa's ceramic substrates — a near-oligopoly upstream of the chip manufacturers that standard analysis ignores. Worth tracking before the next SiC expansion wave makes the constraint visible.Avoid: Commodity Si MOSFET and standard packaging — Chinese vendors closing the quality gap with state subsidy. Margin compression is structural, not cyclical.Highest near-term leverage: Chip-level GPU power positions (Infineon, MPWR, Vicor) carry the highest earnings upside from platform wins — and the highest risk of disruption when NVIDIA changes generations. These are event-driven positions, not long-term compounders.

Silicon Hits a Wall — and Wide-Bandgap Materials Step In

Most electronics run on standard silicon chips. Silicon works well for most applications. But it has a hard ceiling: above a certain voltage and temperature, it wastes too much energy as heat and simply stops working efficiently. AI data centres and electric vehicles have now pushed past that ceiling. Two newer materials — silicon carbide (SiC) and gallium nitride (GaN) — handle much higher power levels with far less heat loss. They are not upgrades to silicon; they are replacements built for a different class of problem.

Silicon (Si)Industry standard — approaching its limitsPower ceilingLowHeat handlingBaselineMax voltage<600VSwitch speed20–100k/secIndustry standard — hitting limitsSilicon Carbide (SiC)Handles the heavy lifting at high voltagePower ceiling3× siliconHeat handling3× siliconMax voltage650V–3,300VSwitch speed20–300k/secIn production · rapidly scalingGallium Nitride (GaN)Switches faster — smaller, more efficientPower ceiling3× siliconHeat handling†1.3× siliconMax voltage650–900V+Switch speed1–10M/secIn production · AI demand accelerating

† GaN voltage range reflects current commercial devices (650V lateral GaN widely available; 900V devices shipping). Vertical GaN architectures — including onsemi's GaN-on-GaN platform announced October 2025 — are pushing beyond 1.2kV, directly targeting 800V data centre and EV applications.

NVIDIA's GPU Roadmap — Power Demand Doubles With Every Generation

GenerationRack PowerPeak CurrentPower ArchitectureSocket / StatusBlackwell (GB200/GB300)~120 kW~1,200A54V DC bus, horizontal deliveryInfineon primary · Renesas secondaryVera Rubin (R100/R200)~240–300 kW~1,500A+800V DC transition begins · VPDRequalification open — MPWR binary Q2–Q3 2026Rubin Ultra / Kyber NVL576~600 kW~2,000A+VPD mandated — new rack architectureVicor FPA best-positioned architecturallyFeynman NVL1152~1.2 MW>3,000AFull VPD + CPO optical scale-upArchitecture TBD · 2028+ horizon

Why the Next Generation of AI Servers Needs a Fundamentally Different Power Design

A GPU chip runs at less than one volt — roughly the voltage of a AA battery — but a next-generation AI rack consumes up to 600,000 watts. Delivering that much power at such a low voltage means pushing an enormous amount of electrical current through the system. The problem: copper wiring resists current flow, generating heat. Even a short section of circuit board trace wastes 800 watts at these current levels — roughly the heat output of a small electric heater, concentrated in a few centimetres of wire.

The underlying constraint is the relationship between power, voltage, and current: P = V × I. At today's 54-volt rack architecture, delivering 120 kilowatts requires over 2,200 amps. The heat wasted in copper conductors scales with the square of the current (I²R), so every increase in rack power drives a disproportionate increase in waste heat and copper mass. At megawatt scale, 54V distribution would require up to 200 kilograms of copper busbar per rack — the weight of the rack structure itself. The industry's response is to raise the distribution voltage from 54V to 800V, reducing the required current by approximately 15× and cutting I²R losses by over 200×. This is the same principle that drives long-distance electricity transmission at hundreds of thousands of volts. It is also why this transition structurally requires wide-bandgap semiconductors — silicon carbide and gallium nitride — which can operate efficiently at these higher voltages where standard silicon physically cannot.

The solution — now being adopted across the industry — is to move the final voltage conversion step from the edge of the circuit board to directly beneath the chip itself, shortening the current path from several centimetres to a few millimetres and eliminating most of that heat loss. This is called Vertical Power Delivery (VPD). It is not a design choice; it is what the physics demands at the power levels the next generation of AI hardware requires.

Six Conversion Stages, Six Toll Booths — GPU-Vendor Agnostic Revenue

Electricity from the grid cannot go straight into an AI chip. It must pass through six separate conversion and regulation stages — each one stepping down the voltage and managing the power more precisely — before reaching the chip at the correct level. Think of it as a series of six toll booths on the road from the electricity grid to the GPU. Each toll booth is a separate commercial opportunity, served by different companies, generating revenue regardless of which chip manufacturer wins the AI race.

The Layer 4 Socket Race — The Single Most Proximate Catalyst

Q2–Q3 2026 · MPWR Vera Rubin Requalification WindowMonolithic Power held the primary supplier position at NVIDIA's H100 generation, lost significant share on Blackwell after performance concerns, but retained some confirmed orders. By early 2026, the analyst consensus had shifted meaningfully positive on Vera Rubin — with some estimates suggesting MPWR could hold a dominant position on that platform. The key question for investors is no longer binary win/loss but whether current valuations already reflect this recovery.

Infineon — the incumbent (Vera Rubin primary). Direct architecture collaboration with NVIDIA on next-generation 800V DC power delivery makes within-generation displacement very difficult. CoolMOS and CoolSiC platforms benchmark best-in-class for high-current GPU VRM synchronous rectification. Kyber path depends on VPD compatibility of next-generation platform.Renesas — the qualified second. Co-qualified across Blackwell and Vera Rubin. Automotive VRM heritage translates directly to AI demands. Path to primary requires Infineon execution failure or a VPD transition that resets the competitive field.Vicor — the long-duration VPD option (Kyber H2 2026). Factorised Power Architecture places the Voltage Transformation Module directly beneath the GPU package substrate — the exact topology Vertical Power Delivery demands. Architecture-matched. Execution unproven at hyperscaler volume. Kyber specification finalisation in H2 2026 is the defining binary event.Tiber (private) — the wildcard. Substrate-embedded voltage conversion could change the physical form of the Layer 4 socket entirely if adopted at Vera Rubin or Kyber scale. A monitoring item for all VRM positions with a 2027–2028 risk horizon.

Six Risks That Could Break the Thesis

AI CapEx PauseHyperscaler monetisation shortfall reprices the entire supply chain simultaneously. Leverage is higher than "picks and shovels" framing implies — device names have significant earnings sensitivity to platform timing. Amazon, Google, and Microsoft have committed capital but have redirected CapEx before when return profiles disappointed.Chinese Vendor QualificationChinese semiconductor manufacturers — supported by substantial government funding — are rapidly closing the performance gap at the lower end of the power semiconductor market. Reaching the quality level required for premium AI applications is a harder target, but the 2028–2030 timeframe is credible, not speculative.System Integrator Margin RiskDelta has explicitly warned of rising costs from oil and AI-demand-driven materials shortages. Gross margin expansion is not linear — dual-sided dynamic warrants explicit position-sizing discipline rather than a momentum approach.SiC Cycle ExtensionOvercapacity extending beyond 2027 delays ASMI/AIXTRON recovery. The Si₃N₄ constraint becomes more visible during the trough as remaining high-utilisation fabs compete for limited Maruwa substrate supply.Tiber / Substrate EmbeddingA private company called Tiber is developing technology that would embed the voltage conversion function directly inside the GPU's own packaging — potentially eliminating the need for a separate external component at the highest-value position in the power chain entirely. A 2027–2028 risk horizon for Infineon, Renesas, MPWR, and Vicor positions.Geopolitical / Export ControlsChina's 2023 gallium/germanium controls established the precedent for materials-level disruption. Escalation to SiC substrates would hit the entire wide-bandgap (SiC and GaN) supply chain simultaneously. Taiwan Strait risk reprices the sector overnight — there is no Plan B at scale.

AuthorsDrishtant Chakraberty, CFAVice President – Equity ResearchLighthouse CantonSaniya SalunkeAssociateLighthouse Canton

Please refer to our full disclaimer for important disclosures and regulatory information.