NVTS is getting all the love. But 650V GaN cannot survive an 800V rack — and POWI is the only chip on Earth that can.

Two GaN companies. One physics problem. Only POWI's chips survive the voltage.

An 800VDC server rack is not a more powerful version of a 48V rack. It is a different physical regime. A standalone 650V GaN switch — the voltage class Navitas, Innoscience, EPC, and ST all top out at — does not survive being placed directly on an 800V bus. To participate in the next generation of Nvidia's AI infrastructure, those suppliers have to combine multiple lower-voltage devices in stacked-half-bridge topologies: clunky workarounds, wasted board space, severe single-point-of-failure risk, and added BOM cost. POWI's InnoMux-2 is the only chip on Earth shipping a 1700V switch on a single piece of silicon. Native 800V operation with a 900V safety buffer built in. This isn't an analyst rating debate. It's a materials science fact.

Nvidia is going from 120kW to 600kW per rack in three years. The physics force the architecture. The architecture forces POWI.

This is the most important power-electronics transition since the data center moved off 12V a decade ago. The numbers are not subtle. GB200 (today): 120-130kW per rack. Vera Rubin NVL144 (H2 2026): 130-190kW. Rubin NVL144 CPX: 370kW. Rubin Ultra Kyber NVL576 (H2 2027): 600kW. Global average data center rack power in 2024 was 8kW. Nvidia is asking the industry to deliver 75x that within 36 months. The 48V/54V rack distribution architecture cannot scale to those numbers — copper losses become prohibitive long before you get there. Nvidia's published reference design converts medium-voltage AC at the facility level directly to 800VDC at the rack. There is no middle path. Either the entire power tree moves to 800V, or Kyber doesn't ship.

Why NVTS Cannot Solve This Problem Alone

Navitas (NVTS) is the AI rack power darling right now and the love is fair — they're an official Nvidia 800VDC partner, they have GaNFast and GaNSafe, and they're scaling 100V GaN for board-level GPU power. But NVTS flagship GaN tops out at 650V. It was designed for high-speed switching in consumer fast-chargers, not megawatt rack architectures. To participate in 800VDC, NVTS must combine multiple 650V devices in stacked half-bridge topologies — four devices in series, isolated gate drivers per device, voltage-balancing capacitors, dual fault domains, and dramatically more board space. Every component added is a new failure mode. A power spike on an 800V bus instantly destroys a lone 650V chip. The workarounds work — but they are workarounds.

POWI's InnoMux-2 with 1700V PowiGaN does the same job with a single die. No stacking. No isolated drivers. No balancing capacitors. A 900V buffer between the operating voltage and the breakdown voltage that absorbs grid transients natively. If Nvidia decided tomorrow to skip 800V and go straight to 1200V DC for the next generation, POWI's 1700V chips still handle it — with a safety buffer. NVTS has to redesign from scratch. That is what a moat looks like when the physics is on your side.

The Real Origin Story Of POWI's 1700V Chip

POWI did not build the 1700V chip for AI data centers. They built it for unstable power grids in developing markets and for heavy electric vehicle architectures — applications where voltage transients and reliability are existential. That is the entire bull case in one sentence. When Nvidia moved Vera Rubin to an 800V baseline, their engineers realized they needed a battle-tested single-chip solution to safely drive the rack's background cooling infrastructure (fans, liquid pumps, logic controllers, isolated aux supplies). POWI was the only company in the world that had spent decades perfecting single-chip high-voltage integration at this node. POWI is co-designing power blueprints alongside Nvidia today — public OCP Summit collaboration, published 1250V/1700V reference architectures, two new InnoMux2-EP auxiliary-power design wins in Q1 alone.

Per-rack power infrastructure spend estimates compiled from sell-side commentary on Nvidia roadmap. Cooling alone for a single NVL576 rack is estimated at $60K-$195K (Teradata Labs, Apr 2026).

The SAM Anchor

CEO Jen Lloyd quantified the data center opportunity for the first time on the Q1 call: combined rack and grid serviceable market exceeding $1 billion by 2030. That is the framing the multiple is being asked to discount, and it's a number stamped by the CFO seat — not analyst extrapolation. Against current run-rate revenue of ~$440M, capturing meaningful share of a $1B data center SAM functionally redefines the company. Combine with a doubling auto segment by 2029 ($100M target) and the continued industrial pivot, and the run-rate math gets to $650-700M+ in two years — at meaningfully better margins than the current cyclical trough.

The smart money signal isn't a senator. It's Nvidia engineers picking POWI for the cooling stack.

POWI doesn't have a Trump-bought-it photo op or a Jensen-on-stage moment. It has something more durable for an industrial semi: Nvidia engineers chose POWI's single-chip 1700V solution for the cooling and auxiliary power infrastructure of an architecture that has to ship in 2027. Engineering selections in power semiconductors are sticky — design cycles are 18-24 months, and once a chip is on the BOM it ships for the life of the platform. The market signal stack is sell-side catching up, institutional accumulation, and management putting real numbers behind the data center opportunity for the first time.

What Actually Moved

What The Signal Stack Actually Means

This is not a meme-stock catalyst chain — Trump didn't buy POWI, Jensen isn't going to do an interview about a power-management IC vendor, and the CEO isn't going to do an open-market $3M buy because that's not how industrial semi management teams operate. The signal that matters is design wins, and design wins are confirmed by management language plus reference architecture publications. Two Q1 InnoMux2-EP auxiliary-power wins. Public Nvidia co-design papers. A workforce restructuring whose press release explicitly cited reallocation toward data center. A CEO putting a hard $1B SAM number on the slide deck for the first time. This is what the inflection looks like before it shows in the P&L.

A 38-year-old high-voltage analog company. Hiding a 40%+ growth GaN engine inside a flat legacy wrapper.

Power Integrations, Inc. (NASDAQ: POWI). Founded 1988 by Klas Eklund, Art Fury, and Steven J. Sharp. Headquartered in San Jose, CA. ~877 employees post the strategic 7% workforce reduction announced with Q4 2025 earnings — explicitly framed as reallocation toward higher-growth segments including AI data center and automotive. Fabless model with three foundry partners. ~98% of revenue from customers outside the U.S. Strong EcoSmart® IP heritage in standby power efficiency that quietly underpins billions of devices globally.

The Four Strategic Layers

The Mix-Shift Story The Market Hasn't Indexed

POWI's headline corporate revenue looks flat — and that's the bull thesis hiding in plain sight. The company reports by end market (industrial/consumer/communications/computer), not by technology platform. Internally, PowiGaN and high-power industrial products are growing well north of the corporate average — community-tracked estimates and management commentary point to ~40%+ growth for the GaN portfolio. That growth is being entirely masked at the consolidated level by deflation in legacy consumer chargers, white goods, and TV adapters — the dying business that used to be the entire company. Q1 2026 made the inversion visible: industrial revenue at 41% of mix (+23% YoY), consumer down to 38%. Consumer is shrinking; GaN and industrial are accelerating. Once the legacy drag stops netting against the high-voltage growth (a 2026-2027 phenomenon), corporate-level numbers re-accelerate sharply.

Customer Base & Automotive Pipeline

POWI is in production or design engagements with 17 of the top 20 EV automakers globally. Auto revenue tracking to roughly double in 2026, $100M annual target by 2029. Active sockets include inverter emergency power supplies, gate drivers, battery management, and CMU applications. Long automotive design cycles (3-5 years) mean once a chip wins the platform, it ships for the life of the model — revenue is structurally locked.

Capital Position — Pristine

$257.2M in cash and short-term investments. Zero debt. $391.8M in working capital. Current ratio 6.88x. New $100M revolving credit facility with PNC Bank (undrawn, matures 2031). Quarterly dividend of $0.215/share, ~1.2% yield. Q1 free cash flow $18M. The balance sheet is a fortress and removes existential risk during any back-half air pocket — POWI can fund the GaN scaling and dividend without touching debt markets while the data center revenue ramps.

Q1 beat. Q2 guide above consensus. The trailing P/E hides the trade — you're paying for a cyclical trough and getting the AI ramp free.

Q1 2026 Results (Reported May 7)

Why The Headline Multiple Is Lying To You

POWI screens at a trailing P/E of ~236x and forward non-GAAP P/E of ~40x — and on those numbers alone, half the screeners in the world will flag it as overvalued. Here is what the screeners don't see: 2022 revenue peaked at $651M with operating income of $180M and net income of $171M. The company is currently running ~$446M TTM revenue with operating income of ~$11M — meaning it is approximately at the cyclical bottom of its historical earnings power. Forward analyst estimates project 2026 revenue ~$475M with non-GAAP EPS ~$1.35, and 2027 revenue ~$544M with EPS ~$1.83. Those numbers do not yet meaningfully model data center revenue. If the company merely returns to its 2022 earnings power on a higher revenue base, EPS goes to ~$3.00-3.50 without any AI contribution. Add even modest data center share — say $100M of incremental high-margin GaN revenue by 2028 — and EPS comfortably clears $4.

At 30x a $4 EPS, the stock is $120. At 35x, $140. You are not paying 40x today's earnings — you are paying 17-20x what the company actually earns at mid-cycle plus the GaN ramp. That is the inversion the trailing multiple hides.

Analyst Coverage

Consensus is roughly flat to current price. That's not a sell signal — that's sell-side waiting for the data center revenue to show up in the print before raising numbers. Susquehanna is already there ahead of the room at $85. Deutsche Bank just raised $20 in one move on Q1. As 2H 2026 data center wins announce and 2027 Kyber ramp gets closer, consensus catches up — and the upgrades come in chunks of $15-$20 each, not gradual.

NVTS is solving 100V for the GPU board. POWI is solving 800V for the rack. They are not the same trade.

The GaN landscape gets flattened into "GaN suppliers compete for AI rack power" — and that flattening is the entire mispricing. The voltage tiers in an 800VDC rack are distinct trades with distinct winners:

NVTS is fairly priced for being the participant they are: an excellent 650V GaN player, an Nvidia partner, getting a meaningful share of high-frequency switching sockets. None of that is in dispute. What is mispriced is POWI being valued as if it competes with NVTS at the same node, when in reality POWI is the only viable supplier at a different node entirely. NVTS's 650V technology cannot physically operate at 800V standalone — they need stacked workarounds. POWI's 1700V technology has 900V of headroom above 800V operation. Different problem, different chip, different trade.

The "Lone Chip vs Stacked Workaround" Math

To hit equivalent 800V capability, a 650V supplier must place at least two devices in series — often four, given the safety margin engineers demand for transient suppression. That means: 4x the die area, 4x the gate driver cost, isolated drivers per device, voltage-balancing capacitors, dual fault detection circuits, and significantly more PCB real estate. A single PowiGaN die replaces all of it. On a 600kW rack with hundreds of conversion stages, the cost and reliability delta of "lone chip vs stacked four" compounds into a real BOM advantage — exactly the kind of advantage data center customers underwrite at scale.

The Other Comp Pictures

POWI vs MPWR. MPWR is the established power IC compounder — 30x earnings, 20% growth, 26% margins, deeply embedded in GPU power trees. MPWR is the high-quality, low-beta way to express AI rack power. POWI is the higher-beta single-name with the unique voltage node. Owning both is reasonable. Choosing POWI over MPWR if you want asymmetric upside; MPWR if you want the safer compound. POWI vs the diversified analog peers (ON, NXPI, Infineon). These trade 17-22x earnings with diversified end markets. They are the safe-and-boring power semi exposure. POWI's premium to this group is real and justified by the GaN technology lead — but the gap closes if POWI revenue accelerates as expected.

The only chip on Earth at this voltage. The risks are timing, not technology.

The chart is coiled under the HTF downtrend. Moving averages catching up. Breakout is when, not if.

The Technical Setup

POWI is trading just under its high-timeframe downtrend line from the 2021 ATH of $104. The 50-day and 200-day moving averages are climbing into the price, compressing the trading range. This is the classic coiled-spring setup — falling trendline above, rising MAs below, declining volatility into the wedge. Resolutions out of this pattern are typically violent. Given the fundamental setup (Q2 guide above consensus, Nvidia co-design confirmation building, sell-side starting to raise targets), the direction of the inevitable resolution heavily favors the upside. The question is the trigger — Q2 print in late July is the obvious candidate.

How To Structure The Position

The trade is being long now and adding on confirmation. Starter position at $70 makes sense given the asymmetric setup. Add aggressively on either (a) breakout through $82-$85 with volume, confirming the HTF downtrend break, or (b) any retest of $60-$62 area that holds the rising 50-day MA. Stop loss on a clean close below $55 — that breaks the MA structure and pauses the thesis. For options-inclined: Jan 2027 $80 and $90 calls capture the breakout window through Q3/Q4 earnings cycle. IV is elevated post-Q1 run but the catalyst density (Q2 print, OCP Summit, Q3 print, FY27 guide) supports paying it.

Catalyst Calendar

NVTS is a participant. POWI is in a category of one. The market is comparing them at the wrong voltage.

The Three Comps That Matter

POWI vs MPWR. MPWR is the high-quality power semi compound — embedded everywhere, predictable, 30x. POWI is the asymmetric voltage-niche play with the unique 1700V product. Different risk/reward profiles, both can win. POWI vs NVTS. NVTS is correctly priced for being an Nvidia partner at 650V and 100V GPU board level — high multiple, no earnings, real participation. POWI is incorrectly priced as if it competes at the same node when it doesn't. POWI captures the 800V bus and 1700V auxiliary sockets that NVTS physically cannot address without complex workarounds. POWI vs the diversified analog peers (ON, NXPI, IFX). These trade 17-22x and are the safe-and-boring power semi exposure. POWI's premium to this group is fully justified by the GaN moat and the rack-power TAM multiplication — and the gap will close further as the data center revenue ramps.

Long POWI. Planting the flag in the only chip that survives 800V.

I added to POWI today, planting my flag in what I believe is a massive capex wave transforming the entire power semi space. NVTS is getting all the love right now — and that's fine, NVTS is a real participant. But the 800V architecture forces a physics-driven separation between voltage tiers that the market hasn't priced yet. NVTS's flagship GaN tops out at 650V and was originally built for consumer electronics. POWI's InnoMux-2 is the only chip on Earth shipping a 1700V switch on a single piece of silicon, with a 900V safety buffer over 800V operation. If Nvidia decides tomorrow to skip 800V and go straight to 1200V, POWI handles it natively. NVTS has to redesign. That asymmetry is the trade. The chart is coiled under the high-timeframe downtrend with moving averages catching up — this is a "when not if" breakout setup, and I want to own it before Q2 earnings in late July.