Power Semiconductors & Power Electronics

Overview / thesis

Power semiconductors and power electronics convert, regulate, and deliver electrical energy — and the sector is splitting into two halves with very different economics. The commodity half (low-voltage silicon MOSFETs, generic power-management ICs) trades at index multiples. The differentiated half — high-voltage silicon-carbide (SiC) and gallium-nitride (GaN) wide-bandgap devices, integrated rack power delivery, and the equipment that feeds AI datacenters and 800V vehicle architectures — has structural reasons to earn a premium. The investment "so what" is that the market still prices much of the category as undifferentiated even where physics, supply constraints, and footprint-limited sockets create real moats.

The core thesis

High-voltage SiC and GaN (650V+ continuous, 800V+ transient) are legitimately differentiated in a power-semi market that is otherwise commoditized. The framing that this is "the most miscovered sector in semis" is correct at the sector level — power semi as a category trades at index multiples while the high-V SiC/GaN slice has structural reasons to earn a premium. The framing is less clean at the stock-pick level, because the publicly traded vehicles carry very different risk profiles (vertical SiC leaders, GaN pure-plays, broad diversified players, picks-and-shovels upstream). Three demand vectors drive a 2026–2030 inflection.

Demand drivers

1. EV traction inverters and onboard chargers. Migration from silicon IGBT to SiC at 800V battery architectures (Porsche Taycan, Hyundai E-GMP, Tesla Plaid, BMW Neue Klasse). SiC content per vehicle runs $150–350 versus $30–80 for IGBT — a 4–5x content step-up per car as 800V architectures spread.

2. 800V DC datacenter rack power. Nvidia's GB200 NVL72 reference design and rumored Stargate-scale AI buildouts force a 48V→800V DC bus transition because copper bus-bar I²R losses scale catastrophically at MW-class rack power. This is the contested overlap zone where GaN (650V class, high frequency) and SiC (1200V+ class, high power) compete, and where footprint-constrained sockets create the clearest moats. AI accelerator racks now require >200 kW at the rack versus 15–20 kW for conventional CPU servers — a 10x increase in power delivery requirement, with NVTS and ON cited as device-side beneficiaries and DELTA / Monolithic Power System winning the conversion sockets.

3. DC fast charging and industrial drives. 50–350 kW DC fast-charger inverters use SiC heavily; industrial drives are migrating to wide-bandgap to meet efficiency mandates. Grid-scale equipment (transformers, switchgear) sits behind the same datacenter and electrification demand wave.

Opportunity sizing and the entry problem

Content economics anchor the bull case: SiC per-vehicle content of $150–350 (vs IGBT $30–80), datacenter rack power as a significant and growing BoM cost line as power per rack rises 10x, and recurring consumption of upstream materials, substrates, and test on every device shipped. The cleanest expression of "volume growth regardless of who wins the device share war" is the picks-and-shovels upstream — MOCVD reactors, engineered substrates, and wafer-level burn-in test — because that volume flows through them no matter which device-maker takes share.

The honest caveat, as of the 2026-05-15 thesis snapshot: the thesis is right but the entry was wrong on all four primary device-side vehicles (NVTS, ON, TXN, IFX) — every name was at or near 52-week highs after meaningful runs. Pink's existing AIXA (Aixtron) position was flagged as the cleanest current exposure precisely because MOCVD reactors are the upstream bottleneck regardless of which device-maker wins. See players for the comp-set verdicts and monitor for the catalysts that change them.

How it works

Power electronics is fundamentally about moving energy from the grid to a chip operating at roughly 1V DC, while losing as little of it as possible to heat. Every design decision in the sector — from wide-bandgap device chemistry to where you place the voltage regulator — traces back to one equation: power dissipated in a resistive element scales with the square of current.

Electrical fundamentals

Four parameters govern the whole chain. Power (P) is energy used per unit time, in Watts. Current (I) is the flow rate of electrons, in Amps. Voltage (V) is electric potential difference, in Volts. Resistance (R) is difficulty of electron flow, in Ohms (Ω). Three equations tie them together:

• P = I × V — power equals current times voltage
• P = R × I² — power dissipation in resistive elements
• Ohm's Law: V = I × R

The load-bearing insight is P = R × I². Resistive (I²R) loss rises with the square of current. So if you can deliver the same power at higher voltage and lower current, losses fall dramatically. This single fact drives both the rack-level 48V→800V transition and the device-level case for wide-bandgap.

The grid-to-chip power path

Energy crosses four conversion stages on its way to silicon:

1. Grid — hundreds of thousands of volts AC
2. Transformers — convert to lower voltages
3. Power Supply Units (PSUs) — convert AC to DC
4. Voltage Regulator Modules (VRMs) — final step-down to chip voltage

Silicon operates at around 1V DC or less, and the design trend is toward even lower voltages and lower clock speeds for efficiency, with silicon running most efficiently in a specific performance/power zone. The high-current/low-voltage problem at the end of the chain is the crux: transporting power at low voltage and high current creates large I²R losses. The solution is to transport at higher voltage and lower current, then step down as close as physically possible to the active silicon. Minimizing the distance (and resistance) of the final high-current leg is the central engineering battle of rack power delivery.

Why power consumption is escalating

The driver is the GPU. An Nvidia H100 draws 700W TDP; the H100→H200 progression pushes even higher; GB200 is expected higher still. By contrast, the most common datacenter CPUs — Intel Skylake/Cascade Lake — draw under 200W. At the rack level this compounds: conventional CPU servers need 15–20 kW per rack, while AI accelerator racks require more than 200 kW — roughly a 10x increase in power delivery requirement. At MW-class rack power, copper bus-bar I²R losses scale catastrophically, which is precisely why the GB200 NVL72 reference design forces a 48V→800V DC bus transition. Higher bus voltage means lower current for the same power, which means tolerable copper losses.

Voltage Regulator Module (VRM) design

The VRM takes input voltage from the PSU and converts it to the correct voltage for the System-on-Chip (SoC). It has three main parts:

1. Capacitors — store electrical energy, release it at a constant rate, smooth power delivery to the processor, and handle transient current demands.
2. Inductors — resist changes in current, prevent massive current spikes, and protect the processor from power surges.
3. Power Stages — the active conversion components: MOSFETs and drivers that convert input voltage to output voltage.

VRM placement matters because it determines the length and resistance of that final high-current path. Typical placement is on the PCB housing the chip; alternatives are on the package itself or integrated on silicon. Placement directly affects efficiency and heat dissipation — moving the regulator closer to the die shortens the lossy leg.

On the capacitor side, the AI cycle is pulling in passive-component categories beyond conventional MLCCs. Silicon capacitors serve IC-package decoupling in advanced-packaging contexts (EMIB, CoWoS-S, I-Cube) — a tightly held supply base (a four-supplier oligopoly of Murata, Samsung Electro-Mechanics, AP Memory, and TSMC). Supercapacitors and lithium-ion capacitors (LIC) address rack-level energy buffering, smoothing the large transient current swings AI workloads impose on the rack; that competitive set is fragmented (Skeleton private, Musashi consolidated). These are passive-component adjacencies to power delivery rather than power-semiconductor devices, but they sit in the same rack socket and the same transient-management problem.

SiC vs GaN — the device physics

Wide-bandgap materials switch faster and tolerate higher voltages and temperatures than silicon, which is what lets a power stage shrink and run cooler. The two materials specialize:

Property	GaN (650V class)	SiC (1200V+ class)
Switching frequency	1 MHz commercial, 50 MHz theoretical	100 kHz–1 MHz
Power handling per die	~120W (typical 650V part)	~240W (typical 650V part)
Gate parasitic loss	17.4× lower than SiC at same V	Higher
Optimal use case	Compact high-density power (laptop chargers, AI rack PSUs, RF)	EV traction inverter, industrial drive, DC fast charge, grid 10kV
Boule/wafer supply	GaN-on-Si grown on standard 200mm silicon	SiC boule = physical vapor transport, ~1 wk/boule, defect-prone

The market settles roughly on GaN for ≤650V high-frequency applications and SiC for ≥1200V high-power applications. The 800V DC datacenter falls in the overlap zone — which is exactly why it is the contested, moat-creating market. The supply-chain asymmetry matters too: GaN-on-Si rides standard 200mm silicon wafers, while SiC boules are grown by physical vapor transport at roughly one week per boule and are defect-prone — a structural cost and yield disadvantage for SiC, and the reason engineered substrates (e.g., SmartSiC, claimed ~30% wafer-cost reduction) and MOCVD epitaxy capacity are genuine bottlenecks. See value_chain for where that capacity sits and who captures the margin.

Subsectors

The sector breaks into distinct sub-areas with different physics, different players, and different ways to capture value. Below, each is described by what it is, who plays, and the angle.

SiC / GaN wide-bandgap devices

What it is: high-voltage power switches built from silicon carbide (SiC, 1200V+ high-power) and gallium nitride (GaN, ≤650V high-frequency), displacing silicon IGBTs and MOSFETs where efficiency, switching speed, or power density matter. The 650V–1200V range is the contested overlap zone, surfacing most sharply in 800V DC datacenter rack power.

Who plays: on the device side, the high-V GaN field is six viable players — Navitas/Infineon (cross-licensed mid-2025), Renesas, ROHM, Nexperia (Wingtech-owned, China), Innoscience (HK-listed Chinese pure-play, a volume disruptor running a ~30% pricing discount), and Texas Instruments (LMG3422 family). On SiC, ON (Onsemi) is the vertical leader, Wolfspeed has restructured, STMicro is the Tesla supplier, and IFX (Infineon) plays broad.

The angle: the publicly traded vehicles have very different risk profiles, so "right thesis, wrong vehicle" is the recurring trap. By disclosed product quality, NVTS (Navitas) and TXN (Texas Instruments) lead on integrated driver + GaNSafe protection + power dissipation. The picks-and-shovels alternative — MOCVD epitaxy, engineered substrates, wafer-level burn-in test — captures volume regardless of which device-maker wins the share war. See players and value_chain.

Power-management ICs (PMIC) and low-voltage power

What it is: the broad, largely commoditized category of voltage regulation, DC-DC conversion, and power-management silicon that sits below the high-voltage frontier. It trades at index multiples because it is undifferentiated — the inverse of the high-V SiC/GaN slice.

Who plays: the analog/PMIC majors — TXN, ADI, Renesas, Infineon, Monolithic Power System — overlap heavily with the rack-power supplier set below. TXN in particular is flagged as a "wrong vehicle" for the wide-bandgap thesis: own it for the analog cycle, not for SiC/GaN, unless GaN revenue is disclosed above ~$1B annual / >5% of revenue.

The angle: this is the part of "power semi" that justifiably trades at index multiples; the alpha is in distinguishing it from the differentiated high-V slice that the market lumps in with it.

Rack power delivery / VRM

What it is: the conversion chain from PSU output down to the ~1V the GPU needs — VRMs (capacitors, inductors, power stages of MOSFETs and drivers) and the 48V/800V DC-DC converters that feed them. AI accelerator racks need >200 kW (vs 15–20 kW for CPU racks), forcing both higher bus voltages and denser, closer-to-die regulation.

Who plays: Monolithic Power System (MPS) replaced Vicor as the primary supplier for H100 GPU power; the winners list also includes DELTA, Renesas, Infineon, and ADI, all making share gains against Vicor (the 2017 market leader when Google drove first 48V DC adoption). DELTA's High Power Density Bi-Directional DC-DC Converter U50SU4P162 (a Vicor NBM equivalent) is winning massive hyperscaler deals; its subsidiary Cyntec is involved in power modules, with Foxconn and Quanta as server-OEM supply-chain partners. 2308 is Delta Electronics' Taiwan listing; FPS is referenced in the rack-power supplier set.

The angle: footprint-constrained sockets and design wins are where margin is realized — power delivery is a significant BoM cost, and once designed in, it is sticky. Vicor's incumbency eroding is the live competitive story.

Datacenter PSU

What it is: the AC-DC conversion stage (and increasingly DC-DC at the shelf level) that takes facility power down to the rack DC bus. As racks move from 48V to 800V DC, the PSU and its downstream converters carry more of the conversion burden and more of the loss budget.

Who plays: power-supply and conversion specialists — DELTA / 2308 (Delta Electronics, with Cyntec) and FPS among them — supplying hyperscale datacenter rack power.

The angle: the 48V→800V transition is a once-a-decade architecture reset; whoever owns the high-density bi-directional converter socket at the new voltage rides the AI rack power ramp.

Grid power equipment / switchgear / transformers

What it is: the upstream electrical infrastructure — transformers, switchgear, grid-scale power equipment — that feeds datacenters and the broader electrification wave. Distinct from chip-level power but pulled by the same demand: hundreds of thousands of volts AC at the grid stepped down through transformers before it ever reaches a PSU.

Who plays: grid-equipment majors. GEV (GE Vernova) and HPS.A are the linked vehicles for this sub-area.

The angle: thin dedicated source material here — this sub-area is named in the sector scope and linked to its tickers, but the consolidated sources are device- and rack-centric. The angle is the same datacenter/electrification demand wave seen one stage upstream of the PSU. (Thin — see players.)

Fuel cells (datacenter power)

What it is: on-site fuel-cell generation as primary or backup power for datacenters, an alternative to grid interconnection where grid capacity or timelines constrain AI buildouts.

Who plays: BE (Bloom Energy) is the linked vehicle.

The angle: thin dedicated source material — named in the sector scope and linked to its ticker, but not developed in the consolidated sources. The connective tissue is datacenter power demand outrunning grid availability. (Thin — see players.)

Passive-component adjacencies (capacitors)

What it is: not power semiconductors per se, but the passive components co-located in the same rack power and IC-package sockets. Silicon capacitors handle IC-package decoupling in advanced packaging (EMIB, CoWoS-S, I-Cube); supercapacitors and lithium-ion capacitors (LIC) handle rack-level energy buffering of AI transient swings.

Who plays: silicon-cap supply is a four-supplier oligopoly (Murata, Samsung Electro-Mechanics, AP Memory, TSMC); the supercap/LIC set is fragmented (Skeleton private, Musashi consolidated). None are in this sector's linked-ticker set.

The angle: included for completeness because they sit in the same transient-management problem as the VRM; covered in depth in the dedicated non-MLCC capacitor briefing (see sources). ALGM (Allegro MicroSystems) is in this sector's ticker set on the power/sensing side rather than capacitors.

Value chain

The power-semiconductor value chain runs from raw wide-bandgap material through device fabrication, packaging and test, into the rack power-delivery stack, and finally into the system OEMs and hyperscalers that buy it. Margin pools concentrate where supply is physically constrained or where a design win locks in a footprint-limited socket.

Upstream: materials, substrates, epitaxy

This is where the SiC supply bottleneck lives. SiC boules are grown by physical vapor transport at roughly one week per boule and are defect-prone — a structural cost and yield disadvantage versus GaN, which is grown as GaN-on-Si on standard 200mm silicon wafers. That asymmetry creates two distinct upstream margin pools:

• MOCVD epitaxy reactors — the tool that grows the GaN-on-Si epitaxial layers. AIXA (Aixtron) holds an effective MOCVD reactor monopoly for GaN-on-Si epitaxy and benefits regardless of which device-maker wins. This is the cleanest "picks-and-shovels" exposure: it is the upstream bottleneck, and volume flows through it no matter who takes device share.
• Engineered substrates — SOITEC (Soitec) makes SmartSiC engineered substrate, claimed to reduce SiC wafer cost by ~30%; positioned as an under-the-radar play that attacks SiC's core cost disadvantage.

The economics: because SiC wafer supply is the binding constraint, anyone who relaxes it (substrate engineering) or who sells the irreplaceable growth tool (MOCVD) captures durable margin independent of the device share war downstream.

Midstream: device fabrication

Device-makers convert wafers into power switches. SiC margin is tied to vertical integration and wafer cost; ON (Onsemi) is the vertical leader on SiC, Wolfspeed restructured, STMicro is the Tesla SiC supplier, and IFX (Infineon) plays broad across SiC and GaN. On GaN, the six viable high-V players are Navitas/IFX (cross-licensed mid-2025), Renesas, ROHM, Nexperia, Innoscience, and TXN. Margin here is contested: Innoscience runs a ~30% pricing discount as a volume disruptor, and a China SiC ASP collapse (Innoscience, BYD pricing data) is the live margin threat to Western device-makers' gross margins (e.g., ON's 35%+ GM target). Product differentiation — integrated driver, GaNSafe protection, power dissipation — is what defends device margin against commoditization; NVTS and TXN lead on that axis.

Test and burn-in

A recurring-revenue choke point. AEHR (Aehr Test Systems) does SiC/GaN wafer-level burn-in test and earns recurring revenue on every SiC device shipped — volume growth flows through it regardless of who wins the device share war. The margin logic mirrors the upstream tool model: consumption scales with device units, not with any single vendor's share.

Rack power delivery stack

Where datacenter BoM dollars land. The VRM and DC-DC converter layer takes PSU output down to ~1V at the die. Power delivery is a significant BoM cost line, and the dollar value rises with rack power (>200 kW AI racks vs 15–20 kW CPU racks). The bottleneck and margin driver is the footprint-constrained socket: design wins are crucial for margin realization, and once a converter is designed into a hyperscaler platform it is sticky. Monolithic Power System (MPS) captured the H100 GPU power socket from incumbent Vicor; DELTA / 2308 (with subsidiary Cyntec on power modules), Renesas, IFX, and ADI are all taking share from Vicor (the 2017 leader when Google drove first 48V DC adoption). DELTA's U50SU4P162 bi-directional DC-DC converter (a Vicor NBM equivalent) is winning hyperscaler deals. FPS sits in this rack-power supplier set.

Passive-component layer (adjacent)

Co-located in the same sockets but a separate margin pool. Silicon capacitors for IC-package decoupling are a tightly held four-supplier oligopoly (Murata, Samsung Electro-Mechanics, AP Memory, TSMC) — oligopoly structure supports pricing. Supercapacitor/LIC rack energy buffering is fragmented (Skeleton private, Musashi consolidated), so margin is weaker. Detail in the non-MLCC capacitor briefing (see sources).

Downstream: OEMs, hyperscalers, grid

System assembly and demand pull. Server OEMs Foxconn and Quanta sit in DELTA's supply chain; hyperscalers (Google as the original 48V driver; the GB200 NVL72 reference design as the 800V forcing function) are the buyers whose architecture choices reset the whole chain. Upstream of the datacenter, grid power equipment — transformers, switchgear (GEV, HPS.A) — and on-site fuel-cell generation (BE) feed the same demand, though these stages are thinly developed in the consolidated sources.

Where the margin actually pools

The pattern across stages: the most defensible margin sits at the two physically constrained ends — upstream tools/substrates/test that scale with total volume (AIXA, SOITEC, AEHR), and the footprint-limited rack sockets where a design win is sticky (DELTA/2308, MPS). The middle (device fabrication) is where ASP wars compress returns, especially under Chinese GaN/SiC pricing pressure.

Players

Companies are grouped by where they sit in the chain. Industry-level positioning only — see each ticker's own page for valuation and the comp tables below for the device-side screen.

SiC / GaN device-makers

• NVTS (Navitas) — high-V GaN pure-play; one of two disclosed product-quality leaders (integrated driver + GaNSafe protection + power dissipation). Cross-licensed with IFX mid-2025. The most leveraged single-name bet on the 800V datacenter GaN socket; also the most volatile.
• ON (Onsemi) — the vertical SiC leader. Broad EV traction and datacenter exposure; gross-margin defense (35%+ target) is the key variable against Chinese SiC ASP pressure.
• TXN (Texas Instruments) — has a GaN line (LMG3422 family) but flagged as the wrong vehicle for the wide-bandgap thesis: own it for the analog cycle, not for SiC/GaN, unless GaN revenue is disclosed above ~$1B / >5% of revenue.
• IFX (Infineon) — broad, diversified power-semi proxy spanning SiC and GaN; cross-licensed with Navitas mid-2025. The defensive/diversified expression; skepticism flagged on its integrated GaN parts versus best-in-class.
• ALGM (Allegro MicroSystems) — power and magnetic-sensing semis serving EV and industrial electrification; the sensing/power-IC angle within the wide-bandgap demand wave.
• Other GaN field (not in this sector's linked set): Renesas (flagged "super high threshold voltage" issue), ROHM, Nexperia (Wingtech-owned, China), Innoscience (HK-listed Chinese pure-play, ~30% pricing discount, volume disruptor). On SiC: Wolfspeed (restructured; excluded from the device screen — recently emerged from bankruptcy and lacks GaN), STMicro (STM, Tesla SiC supplier, not in the current comp set).

Picks-and-shovels upstream

• AIXA (Aixtron) — MOCVD reactor monopoly for GaN-on-Si epitaxy; benefits regardless of which device-maker wins. Flagged as the cleanest current exposure to the sector.
• SOITEC (Soitec) — SmartSiC engineered substrate; reduces SiC wafer cost ~30%; under-the-radar play.
• AEHR (Aehr Test Systems) — SiC/GaN wafer-level burn-in test; recurring revenue on every SiC device shipped.

These three are the picks-and-shovels alternative to direct device-maker bets — volume growth in SiC/GaN flows through them regardless of who wins the device share war.

Rack power delivery / datacenter PSU

• DELTA / 2308 (Delta Electronics) — winning massive hyperscaler datacenter deals with the High Power Density Bi-Directional DC-DC Converter U50SU4P162 (Vicor NBM equivalent); subsidiary Cyntec on power modules; Foxconn and Quanta server-OEM supply chain.
• FPS — referenced in the rack-power / datacenter PSU supplier set.
• Monolithic Power System (MPS) (not in linked set) — replaced Vicor as primary supplier for H100 GPU power; the share-gain story against Vicor.
• Also gaining against Vicor: Renesas, IFX, ADI. Vicor was the 2017 market leader when Google drove first 48V DC adoption; competitors now aggressively entering and taking share.

Grid equipment and fuel cells

• GEV (GE Vernova) — grid power equipment / transformers / switchgear; datacenter and electrification demand pull. (Thin source coverage.)
• HPS.A — grid power equipment vehicle. (Thin source coverage.)
• BE (Bloom Energy) — fuel cells for datacenter power. (Thin source coverage.)

Device-side comp screen (Pink's verdict, 2026-05-15)

Ticker	Verdict	Price	At 52w high?	Better entry
NVTS	PASS	$22.32	YES (+1087% off low)	$14 starter, $11 add, $7 conviction
ON	WATCH	$118.37	YES (+191% off low)	$100-105 starter, $90-95 core, $75-85 conviction
TXN	WRONG VEHICLE	$308.17	YES (+102% off low)	Own for analog cycle, not this thesis
IFX	DIVERSIFIED PROXY	$78.33 (ADR)	YES (+118% off low)	Fully valued — buy if you want broad power-semi exposure

The single highest-conviction line: every name is at or near 52-week highs after meaningful runs. The thesis is right; the entry is wrong on all four. AIXA is the cleanest current exposure — MOCVD reactors are the upstream bottleneck regardless of which device-maker wins.

Rack-power supplier landscape (competitive pressure on Vicor)

Role	Players
Incumbent losing share	Vicor (2017 leader, first Google 48V adopter)
Winners taking share	MPS, DELTA, Renesas, IFX, ADI
H100 GPU power socket	MPS replaced Vicor
Hyperscaler DC-DC wins	DELTA U50SU4P162 (Cyntec modules; Foxconn/Quanta OEMs)

Monitor

Dated developments, catalysts, and watch-items that move the sector thesis.

Dated developments

• 2026-05-12 — Irrational Analysis (@insane_analyst) tweet: "High voltage SiC and GaN is the most miscovered sector in semis right now," listing TI, Navitas, ON, IFX, Wolfspeed. Trigger for the sector thesis.
• 2026-05-15 — Pink's comp-set verdict: NVTS PASS, ON WATCH, TXN WRONG VEHICLE, IFX DIVERSIFIED PROXY. All four at/near 52-week highs; thesis right, entry wrong on all four. AIXA flagged cleanest current exposure.
• 2026-05-28 — non-MLCC capacitor briefing: silicon caps (4-supplier oligopoly Murata/SEMCO/AP Memory/TSMC) and supercaps/LIC (rack energy buffering, fragmented) mapped as AI-cycle passive adjacencies.
• Mid-2025 — Navitas / Infineon GaN cross-license signed.
• 2017 — Vicor was rack-power market leader as Google drove first 48V DC adoption (baseline for the current share-loss story).

Catalysts — what changes the verdict

For NVTS:

• Bull fires: Aug 4, 2026 Q2 print delivers $11M+ revenue (vs $9.8M consensus) AND a named 800V hyperscaler customer → $22 entry rationalized.
• Bear: Aug 4 print in-line or below ($9–10M) AND no named customer → round-trip to the $10–14 zone.

For ON:

• Entry window closes: Q2 2026 print delivers +10%+ YoY revenue AND 35%+ gross margin → entry closes at $105.
• Re-rate bull: named 800V DC datacenter customer disclosure → $135+ becomes base case.
• Bear: China SiC ASP collapse (Innoscience, BYD pricing data) → $75–85 zone becomes likely.

For TXN:

• Thesis re-activates: GaN revenue disclosed >$1B annual / >5% of revenue → re-activates the SiC/GaN vehicle thesis.
• Separate trade: $250 or below → re-rates on analog cycle alone.

For IFX:

• Re-rates as conviction vehicle: Infineon delivers performance-competitive integrated GaN (closes the flagged gap) → high-conviction, not just proxy.
• Defensive pick: SiC ASP wars compress NVTS/ON → IFX's diversified base becomes the defensive choice.

Structural watch-items

• China GaN/SiC ASP pressure — Innoscience's ~30% pricing discount and BYD pricing data are the live margin threat to Western device-makers' gross margins.
• 800V DC datacenter socket — which device-maker and which rack-power vendor lock the GB200-class 800V sockets; footprint-constrained design wins are the moat.
• Vicor share erosion — track MPS, DELTA, Renesas, IFX, ADI gains in hyperscaler GPU power.
• SiC wafer cost curve — SOITEC SmartSiC adoption and AIXA MOCVD capacity as the upstream bottleneck signals.

Open follow-ups

• [ ] Sivers Semiconductor — IA-adjacent compound-semi pure-play; needs /profile.
• [ ] Innoscience (HK) — Chinese GaN volume disruptor; not in vault yet.
• [ ] STMicro (STM) — Tesla SiC supplier; not in current comp set.
• [ ] Full sector primer on SiC/GaN from first principles (physics, supply chain, value capture) — extends AIXA and SOITEC context.

Sources

Authors and publications

• irrational-analysis (Irrational Analysis, @insane_analyst) — Substack + Twitter. Origin of the "most miscovered sector in semis" framing; technical SiC-vs-GaN comparison and the 6-player GaN market-structure / winner picks. Disclosed positions: long NVTS + ON; explicit skeptic on IFX integrated GaN parts.
  • 2026-04-25-power-semis-part-2-sic-vs-gan-wolfspeed — IA Substack, technical SiC-vs-GaN comparison.
  • 2026-04-21-power-semis-800v-dc-and-gan-mini-note — IA Substack, 6-player market structure + winner picks.
  • IA tweet 2026-05-12 (@insane_analyst) — "miscovered sector" pitch listing TI, Navitas, ON, IFX, Wolfspeed.

Consolidated source files

• sic-gan-high-voltage-thesis — KB/wiki/themes/sic-gan-high-voltage-thesis.md — Register D theme page (2026-05-15). Core SiC/GaN high-voltage thesis, device physics table, 6-player GaN structure, comp-set verdict, catalysts, picks-and-shovels alternatives.
• rack-power-delivery-primer-gb200 — KB/wiki/rack-power-delivery-primer-gb200.md — rack power-delivery primer (dated 2023-08-01). Electrical fundamentals, grid-to-chip path, VRM design, GPU power escalation, Vicor competitive landscape, Delta/MPS share story.
• 2026-05-28-non-mlcc-capacitor-primer — KB/wiki/briefings/2026-05-28-non-mlcc-capacitor-primer.md — non-MLCC capacitor primer briefing (2026-05-28). Silicon-cap 4-supplier oligopoly and supercap/LIC rack-energy-buffering adjacency; referenced for passive-component context only (company specifics — SEMCO, Musashi 7220.T, Samsung Electronics — not re-hosted here; tickers fall outside this sector's set). Portal: https://pink.sakdiarpa.com/reports/handoff/non-mlcc-capacitor-primer.html

---

Consolidation queue (merged 2026-05-30 — section-scoped rebuild)

Industry-wide content folded in from these source files. They stay live pending Pink's archive confirm.

• [ ] themes/sic-gan-high-voltage-thesis.md
• [ ] rack-power-delivery-primer-gb200.md
• [ ] briefings/2026-05-28-non-mlcc-capacitor-primer.md