Snapshot — the group at a glance

"Power semiconductors" are the chips that convert and regulate electricity on its way from the grid to a computer chip. The newest kind use wide-bandgap materials (semiconductor materials that can tolerate higher voltages and switch faster than ordinary silicon) — silicon carbide (SiC) and gallium nitride (GaN), both of which switch electricity faster and waste less of it as heat than ordinary silicon. In an AI data center, electricity is stepped down many times (grid → rack → board → chip), and at each step a slice is lost as heat. Better power chips mean less waste, less cooling, and more of the building's megawatts reaching the actual compute. The companies here either grow the raw SiC crystals and sell the wafers (Wolfspeed), design GaN switching chips (Navitas), design the small high-efficiency power-management chips that sit next to every GPU and CPU (Monolithic Power), or do several of these across many end markets (ON Semiconductor).

Demand for power chips rises mechanically with every megawatt of AI compute added, and the AGI build-out is adding megawatts fast — so the raw demand curve points up. The supply bottleneck here, per the scan, is milder than in compute or memory: GaN capacity is scaling quickly and SiC's main use is still electric vehicles, not servers, so a glut in one end market can swing prices. In money terms, the prices paid across the group are very uneven — the asset-light, profitable chip designers (MPWR) have historically traded at high multiples of sales, while the capital-heavy substrate maker (WOLF) has been a cash-burning, restructuring situation. There is no single group-wide multiple; the names trade very differently from one another. est.

The product & how money is made

The "product" is a chip (or the wafer a chip is built on) that handles electrical power rather than computation. There are three distinct money-making shapes inside this one group:

• Substrate / wafer makers (Wolfspeed): They grow large single crystals of silicon carbide, slice them into wafers, and sell those wafers (or finished SiC devices) to chipmakers and carmakers. The unit sold is a wafer (priced per wafer, with bigger 200mm wafers worth more than 150mm). This is a heavy-industry, factory-building business — you spend cash for years building furnaces and clean rooms (this upfront factory spending is capex, capital expenditure) before the revenue shows up.
• GaN device designers (Navitas): They design gallium-nitride power switches and the control chips around them, then have them manufactured (often by an outside foundry — a contract chip factory). The unit sold is a chip / power IC (integrated circuit), priced per part. Asset-lighter than substrate, but at small scale a designer can run at a loss while building volume.
• Power-management chip designers (Monolithic Power; the power slice of ON): They sell the small, high-efficiency voltage-conversion chips that sit on the server board right next to the GPU/CPU and step the voltage down to the exact level the processor needs. The unit is again a chip, but content-per-server (the dollar value of these chips inside one server) rises as power density rises. This is the most directly "AI server" exposed of the three shapes, and an asset-light design model — design once, sell millions, low factory cost.

So "money in" for an owner comes from selling more units (wafers or chips) and/or selling them at a higher price per unit; "money out" is R&D (chip design) plus, for the substrate maker, very large factory capex. The fundamental unit of demand for this fact sheet is dollars of power-semi content per AI server — that figure grows as racks pull more kilowatts.

Source: 500-stocks scan, /Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/02-semiconductors.html (section 15, "Power Semiconductors (SiC & GaN)"); business-model shapes from general knowledge.

Demand — how much the world will want this

Demand has two engines, and only one is AI:

• The AI / data-center engine (the AGI lens): Per the scan, AI rack power density is "exploding (30–120 kW per rack)" and a single AI facility draws "50–100 MW." Every one of those watts must be converted and regulated, and the conversion losses cost real electricity money — so the buyer (the data-center operator and the server OEM, the company that builds the server) is motivated to pay for more efficient power chips. Given AGI is arriving, total installed AI compute keeps rising for years, which means the number of racks — and therefore the number of power-conversion stages needing chips — rises with it. Higher density per rack also raises the content per server (more, and more advanced, power chips per box). GaN in particular is, per the scan, "gaining traction in server power supplies." Per the scan's own rating, the AI-demand pull for this sub-sector is "moderate/medium" — power chips are a smaller slice of a server's bill-of-materials (the parts list and their costs) than the GPU or the memory.
• The non-AI engine (still the bigger one today): SiC's largest end market is electric vehicles and grid/industrial power, not servers. So overall demand for this group is dominated by autos and industrial, and the AI data-center build-out (including the grid infrastructure needed to feed those data centers) is an incremental — and, per general knowledge, the fastest-growing — sliver on top. est.

Forecast (label: forecast, not contracted): The SiC and GaN sub-markets are widely expected to grow faster than the overall power-semi market — frequently cited at roughly 20–30%/yr for SiC and faster off a smaller base for GaN — driven by EVs plus the data-center add-on. est. These growth figures are directional forecasts, not contracted backlog (orders already signed).

Who the buyers are: server power-supply makers and ODMs (original design manufacturers — the firms that build the actual boxes, e.g., Delta), data-center operators (the hyperscalers, indirectly), and — for the bulk of volume — automakers and industrial-equipment firms.

Supply — how much can be made, and what limits it

The scan is explicit that the bottleneck here is milder than elsewhere in the chip stack: "Overall, the bottleneck here is less severe than in compute or memory." Within the group, supply splits by material:

• SiC substrate — the genuine tight point. Per the scan, "SiC substrate supply is tight (dominated by Wolfspeed and Coherent)." Growing silicon-carbide crystals is slow, hot, low-yield work; the limiting factors are physical (crystal growth time and defect yield — the share of usable material per batch), capital (furnaces and 200mm fabs cost billions), and know-how (decades of process learning). This is the part of the group that behaves like a constrained commodity.
• GaN — capacity is scaling fast. Per the scan, "GaN capacity is scaling rapidly." GaN can often be made on standard silicon wafers in existing or near-existing fabs, so the physical barrier to adding supply is lower than for SiC. Rapid capacity addition is the opposite of a shortage — it caps how short the product can get.
• Power-management chips (MPWR / ON slice) — design-limited, not capacity-limited. These ride mainstream foundry capacity; the constraint is design skill and customer sockets (the designed-in slots on a customer's board), not scarce wafers.

Market-share structure (who controls supply): SiC substrate is concentrated — Wolfspeed and Coherent are named in the scan as the dominant western substrate suppliers, with STMicro, Infineon and onsemi as large integrated SiC device players. In the broader power-discrete market the leaders are, by general knowledge, Infineon (the largest), onsemi, STMicro, and Toshiba/Renesas — a fragmented oligopoly (several large competitors, no single gatekeeper) rather than one controller. GaN is even more fragmented, with Navitas, Power Integrations, EPC, and Infineon (which bought GaN Systems) all competing. est.

Source: 500-stocks scan section 15 (substrate tightness, GaN scaling, named players); market-share ordering from general knowledge, not-live-verified.

The gap — demand vs supply

Putting the two sides together, this group is not uniformly short — per the scan and general knowledge it is tight in one narrow place (SiC substrate) and roughly balanced-to-loosening elsewhere. That is the distinction from compute (GPUs) and memory (HBM, high-bandwidth memory), where the whole product is reported sold out. Here, the demand trend is up (AGI build-out plus EVs), but supply can respond in GaN and in power-management chips, so any pricing power is local to SiC substrate rather than group-wide.

Evidence and the flip risk: Three signals to verify against current filings: (1) SiC wafer pricing — through 2025 SiC prices were reportedly softening because EV demand grew slower than the industry had built capacity for; that EV-led overbuild is a reason SiC could be oversupplied even while AI demand rises. est. (2) Wolfspeed's situation — the dominant western substrate maker entered a deep financial distress / restructuring phase reportedly in 2024–2025 after spending large capex ahead of demand that arrived slower than planned, a piece of evidence that this corner has been closer to over-supply than shortage. est. (3) Lead times — power-discrete lead times reportedly normalized off their 2021–2022 highs. est. Read factually: the AGI-driven data-center pull is genuine and growing, but per general knowledge it is currently smaller than the EV-driven swing, so the group's near-term gap is currently set more by the EV cycle than by AI; the AI add-on shifts the demand mix over time rather than creating an acute shortage today. The reader can weigh how that mix evolves.

Source: directional pricing/lead-time/Wolfspeed-distress observations from general knowledge (cutoff early 2026), not-live-verified; segment demand/supply framing from the 500-stocks scan.

The players — who captures the money

Adjacent names the scan also lists in this group: Infineon (IFNNY, largest power-semi maker, bought GaN Systems), Power Integrations (POWI), Vishay (VSH), and MACOM (MTSI). These are diversified or non-US-primary and sit outside the four key tickers above.

Source: company list and "key US-listed companies" from the 500-stocks scan section 15; revenue/size brackets from general knowledge, not-live-verified.

The price of exposure

The cleanest way to think about "what you pay" is market value per dollar of this year's sales (the price-to-sales multiple — market cap divided by annual revenue; it tells you how many dollars of company value you buy per $1 of yearly revenue) and, separately, who actually turns sales into owner cash (free cash flow — cash left after operating costs and capex). These names sit at opposite ends, and the figures below are approximate, not-live-verified:

• Monolithic Power (MPWR) — has historically carried a high multiple of sales (roughly 15–20×+ at various points) est. alongside high margins and capital-light operations: it has converted a large share of revenue into free cash flow with little factory spend. The arithmetic: at ~15–20× sales, an owner pays $15–20 of market value per $1 of current annual revenue, so the price embeds substantial future revenue growth.
• Wolfspeed (WOLF) — has traded at a low multiple of sales while reportedly burning cash, carrying heavy debt and capex, and restructuring. est. A low price-to-sales here reflects negative current owner cash and dilution risk (new shares or debt that reduce each existing owner's share); it is not the same arithmetic as a profitable company at a low multiple. The reader judges what that combination is worth.
• ON Semiconductor (ON) — a mid multiple of sales est.; profitable and cash-generative but cyclically tied to autos, so its AI exposure is a slice of a large non-AI base.
• Navitas (NVTS) — small revenue and reportedly not yet profitable, so a sales multiple is the only available handle; what an owner pays for is future GaN volume, not current cash. est.

Money-in / money-out shape of the group: it is barbelled. On one side are capital-light designers (MPWR, and the design parts of NVTS/ON) where money out is mostly R&D and money in can become high free cash flow at scale. On the other is the capital-heavy substrate (WOLF) where money out is years of multi-billion-dollar factory capex before steady money in — the profile of a business that can be starved of cash if demand arrives later than the factories. So "buying the group" blends a high-margin cash generator, a restructuring/distress situation, a diversified cyclical, and a pre-profit name — they do not move or earn alike.

Source: multiple ranges and cash-flow characterizations from general knowledge, not-live-verified; confirm against current market data and latest filings before use.

What to deep-dive next

• Most direct AI-server power exposure: Monolithic Power (MPWR) — its content per server rises directly with rack power density; a deep-dive would examine data-center revenue mix and free-cash-flow conversion to size how much of the AI demand is already reflected in the price.
• Most concentrated GaN exposure: Navitas (NVTS) — small, pre-profit, leveraged to server power-supply adoption; a deep-dive would examine design wins, gross-margin path, and cash runway, because the open question is whether it reaches profitable scale rather than current earnings.
• Most supply-advantaged but financially weakest: Wolfspeed (WOLF) — controls the one genuinely tight input (SiC substrate) but is the financial weak point; a deep-dive would examine the balance sheet, debt maturities, capex commitments, and restructuring terms first, because solvency determines whether the demand story reaches owners.
• Most diversified: ON Semiconductor (ON) — a large, profitable company where AI power is a small piece of an auto/industrial base; a deep-dive would size how much AI actually moves the numbers versus the EV cycle.

This is a factual map of where a company-level look would be most informative, not a ranking or recommendation.

Sources & confidence

Hard / grounded facts (from the provided scan and well-known filings): the company list (WOLF, ON, MPWR, NVTS plus IFNNY, VSH, POWI, MTSI); the qualitative supply/demand notes — SiC substrate tight and dominated by Wolfspeed and Coherent, GaN capacity scaling rapidly, overall bottleneck "less severe than in compute or memory"; AI rack density 30–120 kW and facility draw 50–100 MW; the scan's "Medium/moderate" AI-demand rating for this sub-sector. These come from /Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/02-semiconductors.html (section 15).

Approximate / not-live-verified (general knowledge, cutoff ~early 2026, tagged est. throughout): total power-semi market size (~$30–40B) and growth rates (SiC ~20–30%/yr, GaN faster); all per-company revenue brackets and market caps; market-share ordering (Infineon largest, etc.); SiC price softening and Wolfspeed's distress/restructuring; all valuation multiples (price-to-sales) and cash-flow characterizations. Live web retrieval was unavailable when this was written, so every market-size, growth, share, pricing, and valuation figure above must be re-confirmed against current data and the latest 10-K/10-Q before any capital decision.

No prior deep-dive report exists for this group; this fact sheet is the first pass and is built only from the scan section plus labeled general knowledge.

Sources: 500-stocks scan section 15 (Power Semiconductors, SiC & GaN); general knowledge (cutoff ~early 2026), explicitly flagged where used.