` through `

` appears twice. I need to deduplicate, keeping the better version (the second copy appears to be the refined one in some places, but they're nearly identical). Let me also check for the 9 required sections. Looking at the IDs: s-snapshot, s-product, s-demand, s-supply, s-gap, s-players, s-price, s-next, s-sources. That's 9. Now let me identify all the issues: 1. **Opinion/verdict to remove:** - "the most extreme price-to-current-cash pairing" — evaluative - "substantially cheaper per dollar of revenue" — comparative judgment that implies preference - "most of the re-rating to 'AI power play' has already occurred" — verdict on future returns - "The bet is that..." — framing as a bet is opinion - "notoriously cyclical" — "notoriously" is editorializing - Various phrasings that imply what a buyer should think 2. **Boilerplate to remove:** - The preamble "Now I have comprehensive data. Let me write the fact sheet..." — this is agent internal monologue, not part of the fact sheet - "The task says 'AAON' is in the group..." — same 3. **Unverified figures needing est.:** Check all figures — most company financials are from yfinance (verified). Market-size figures already tagged. Need to check rack power density figures, growth rate projections. 4. **Dead-weight sentences:** Sentences that restate what was just said, or add no new information. Let me also look for the "showing lumpiness" type of editorial commentary. Now let me produce the corrected HTML.

Snapshot — the group at a glance

This group makes the equipment that converts, regulates, and manages electrical power on its way from the grid to the computing chip, plus the cooling systems that remove the heat those conversions generate. The products range from voltage-regulator chips on a server motherboard (Monolithic Power Systems) to modular DC-DC converter bricks inside GPU racks (Vicor), to room-scale precision power supplies for semiconductor fabs and data centers (Advanced Energy), to building-scale liquid-cooling and HVAC systems (AAON, via its BASX division). The unit of demand is dollars of power conversion and thermal equipment per megawatt of installed compute. Every watt an AI chip consumes must first be converted from grid AC to the chip's precise DC voltage, losing some energy as heat at each step — and that heat must then be removed.

Every dollar spent on an AI GPU mechanically pulls dollars of power-conversion and cooling equipment behind it. Rack power density has jumped roughly 10x from legacy servers to current AI configurations, which means more conversion stages, higher-efficiency semiconductor switches, and far more cooling capacity per rack. The two chip designers (MPWR, VICR) trade at roughly 28–32x trailing revenue; the two equipment/systems makers (AEIS, AAON) trade at roughly 7–8x trailing revenue. The chip designers carry essentially no debt; AAON carries $451M of net debt against effectively zero cash. All four are profitable and growing revenue 20%+ year-over-year.

Source: yfinance live data Jun 2 2026; 500-stocks scan section 09 ("Power Electronics & Power Conversion").

The product & how money is made

The four companies sell different products along the power-and-thermal chain:

• Monolithic Power Systems (MPWR) designs DC-DC converter chips — integrated circuits that step voltage down from one level to another — used on server motherboards, in AI accelerator power delivery, storage, networking, automotive, and consumer electronics. Fabless model: designs chips, contracts fabrication to foundries. Money in = chip sales; money out = R&D plus foundry fees. TTM revenue $2,957M, gross margin 55.2%, operating margin 27.1%. Cash $1,367M, debt $20M, 4,501 employees. Data-center/enterprise is the fastest-growing segment; also sells into automotive, industrial, communications, and consumer.
• Vicor (VICR) designs and manufactures high-density modular power components — DC-DC converters, power-on-package modules, and factorized power architecture (FPA) systems — that convert and regulate power at high efficiency in a small footprint. Used in AI accelerators, high-performance computing, and defense/aerospace. Unlike MPWR, Vicor fabricates most products internally in Andover, Massachusetts. TTM revenue $472M, gross margin 54.5%, operating margin 14.9%. Cash $404M, debt $7M, 1,092 employees. Revenue swings with GPU-rack build timing — Q2-2025 was $141M, Q1-2026 was $113M.
• Advanced Energy Industries (AEIS) makes precision power-conversion equipment — power supplies, plasma generators, and measurement instruments — used in semiconductor manufacturing equipment, data centers, industrial processes, and medical devices. Serves the fabs that build GPUs as well as the data centers that run them. TTM revenue $1,905M, gross margin 39.0%, operating margin 11.5%. Cash $700M, debt $683M, 13,000 employees. Acquisition-related goodwill ($300M) and intangibles ($112M).
• AAON (AAON) engineers and manufactures commercial HVAC equipment — rooftop units, air handlers — and data-center liquid-cooling systems through its BASX division (acquired 2024). BASX builds custom liquid-cooling solutions for hyperscale data centers, addressing the thermal load from high-density AI racks that air cooling alone cannot handle. TTM revenue $1,617M (Q1-2026 up 54.3% YoY, driven by BASX ramp), gross margin 26.2%, operating margin 10.6%. Cash ~$0, debt $451M (largely from the BASX acquisition). 5,897 employees.

Source: yfinance company profiles and financial data; 500-stocks scan section 09; 10-K/10-Q filings (segment descriptions).

Demand — how much the world will want this

Demand already contracted or being delivered

None of these four companies disclose RPO (remaining performance obligations) in the same way a neocloud or utility does. Demand evidence comes from order backlog and revenue trajectory: contracted

• MPWR Q1-2026 revenue was $804M, up 26.2% YoY ($638M in Q1-2025). Data-center/enterprise segment has been the growth driver, reflecting design wins in AI server power delivery.
• VICR Q2-2025 revenue spiked to $141M (vs. $94M in Q1-2025 and $96M in Q2-2024), consistent with a large delivery tied to a GPU-rack deployment cycle. Q3 and Q4 pulled back to $110M and $107M, then Q1-2026 was $113M. FY2025 total was $453M, up 26% from FY2024's $359M.
• AEIS Q1-2026 revenue was $511M, up 26.3% YoY ($405M in Q1-2025). Semiconductor equipment and data-center power are the two growth engines. FY2025 was $1,799M, up 21% from FY2024's $1,482M.
• AAON Q1-2026 revenue was $497M, up 54.3% YoY ($322M in Q1-2025). BASX data-center cooling is the source of acceleration. FY2025 was $1,442M, up 20% from FY2024's $1,201M. Q1-2026 annualized run-rate ($1,988M) is above FY2025, showing continued acceleration.

The broader demand picture

The demand drivers for this group are mechanical links to the AI compute build-out: est.

• Power density per rack is the multiplier. AI GPU racks consume 40–120 kW each est., versus 5–10 kW for traditional server racks. That 8–12x increase means proportionally more power-conversion hardware (voltage regulators, DC-DC converters, power modules) and cooling capacity per rack.
• The shift from 12V to 48V rack distribution reduces copper losses and cable weight but requires new converter topologies, creating fresh design-win opportunities for MPWR and VICR.
• Liquid cooling adoption. Air cooling hits practical limits above roughly 30–40 kW per rack est.. Current AI racks at 40–120+ kW require direct liquid cooling (cold plates, rear-door heat exchangers, or immersion). This is AAON/BASX's addressable demand — a product category that barely existed at scale before the AI build-out.
• Semiconductor fab equipment spending. AI chip demand drives investment in new fabs and fab tools. AEIS sells precision power supplies to those tools. Gartner projects worldwide data-center systems spending at $788B in 2026, up 55.8% YoY. NVIDIA's data-center revenue hit $75.2B in Q1 FY2027 (quarter ended Apr 2026), up 92% YoY.

Source: 500-stocks scan section 09 (rack power density, 12V-to-48V shift); Gartner via TechEdge AI (May 20 2026); NVIDIA Q1 FY27 press release (May 28 2026); company quarterly filings.

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

Capacity by product type

• Power-management chips (MPWR): fabless model using third-party foundries (TSMC, others). Supply rides mainstream wafer availability, not proprietary factory capacity. The bottleneck is design wins and customer sockets, not wafer supply. MPWR can scale revenue without building factories.
• Power modules (VICR): internally manufactured in Andover, MA. Capacity is limited by Vicor's own factory footprint. Revenue lumpiness ($141M Q2-2025 vs. $94M Q1-2025) suggests production can ramp to meet bursts but may face throughput limits on sustained high volume.
• Precision power supplies (AEIS): manufactured across multiple global sites with 13,000 employees. Running at 39% gross margins — neither severe input constraints (which would compress margins) nor excess capacity (which would force price cuts).
• Data-center cooling systems (AAON/BASX): manufactured in Redmond, Oregon and Longview, Texas. FY2025 free cash flow was -$215M even while revenue grew 54%, indicating capital is being deployed on factory buildout faster than cash is generated. Bottlenecks: factory construction speed, skilled labor for custom-engineered systems, and supply of specialized components (heat exchangers, pumps, coolant distribution manifolds).

The real bottlenecks

• For chip designers (MPWR, VICR): design wins, not factory capacity. Once designed into a GPU platform, revenue is locked for that product generation (typically 2–4 years). NVIDIA's power-delivery architecture choices (voltage level, module form factor) directly determine which suppliers win.
• For equipment makers (AEIS, AAON): factory throughput and capital investment. AAON's negative free cash flow is the cost of building capacity. AEIS's roughly neutral cash/debt position ($700M cash vs. $683M debt) suggests more measured capacity growth.
• Shared input constraints: power semiconductors (SiC MOSFETs, GaN switches) are inputs to power supplies and UPS systems across the industry. SiC wafer capacity is dominated by Wolfspeed and Coherent — supply is tight but the main demand driver is EVs, not servers. GaN capacity is scaling rapidly est..

Source: 500-stocks scan section 09 (SiC/GaN bottlenecks); yfinance financial data (capex, employees, margins); power-semiconductors sector report.

The gap — demand vs supply, and the pricing signal

Pricing direction. MPWR and VICR have maintained or expanded gross margins through 2025 and into Q1-2026 — MPWR at 55.2%, VICR at 54.5% — consistent with pricing power rather than commoditization. AEIS at 39.0% gross margin is stable but lower, reflecting more diversified end markets. AAON's gross margin at 26.2% is the lowest in the group, typical of manufacturing-heavy businesses; BASX's contribution is still ramping and may carry different unit margins at scale.

What could flip the gap. For MPWR and VICR, a design loss — displacement from the next NVIDIA GPU platform — would compress revenue regardless of market demand. For AAON/BASX, hyperscalers bringing cooling in-house or competitors (Vertiv, Schneider) scaling liquid-cooling offerings faster would slow backlog-driven growth. For AEIS, a semiconductor-equipment spending downturn (cyclical every 3–5 years) would hit fab-tool power-supply revenue.

Source: yfinance gross margins (TTM); 500-stocks scan demand/supply framing.

The players — who captures the money

Balance-sheet contrast. MPWR and VICR are asset-light, debt-free, and sitting on large cash piles relative to their size. AEIS is roughly net-neutral ($700M cash vs. $683M debt). AAON is the only name with material net debt ($451M) and effectively no cash — it is financing BASX factory expansion with debt and operating cash flow. AAON's negative free cash flow (-$215M TTM) reflects the cost of building manufacturing capacity.

Source: yfinance live data Jun 2 2026; quarterly filings (Q1-2026 10-Q).

The price of exposure — neutral arithmetic

MPWR: $83B market cap on $492M TTM free cash flow = 0.6% cash yield. At 26% annual revenue growth (current rate), revenue triples in about 4.5 years if the rate holds and margins stay flat.

VICR: $15.1B market cap on $9M TTM free cash flow = 0.06% cash yield. 52-week range spans $42 to $362, a roughly 8x move from low to current price. Revenue would need to grow several-fold and margins expand for current price to be supported by cash flow.

AEIS: $12.9B market cap on $1.9B revenue (6.4x EV/revenue). Forward P/E of 27x. Semiconductor equipment spending is cyclical (3–5 year swings); FY2024 revenue dipped to $1,482M from FY2023's $1,656M during the prior WFE downturn.

AAON: $12.1B market cap on $1.6B revenue (7.8x EV/revenue). Negative free cash flow (-$215M) and $451M debt reflect ongoing factory buildout. Q1-2026 annualized revenue run-rate ($1,988M) is 38% above FY2025 ($1,442M).

Source: yfinance live data Jun 2 2026; forward EPS consensus from yfinance.

What to deep-dive next

• MPWR — design-win pipeline and GPU platform dependency. Which NVIDIA, AMD, and custom ASIC platforms is MPWR designed into for next-generation GPUs? Is content per server rising or plateauing? Data-center/enterprise segment revenue breakout in the 10-K, customer concentration, and competitive pressure from Texas Instruments, Renesas, and Infineon on the same sockets.
• VICR — revenue lumpiness and factory capacity. Is the spike-and-retreat pattern (Q2-2025 $141M, Q1-2026 $113M) tied to specific customer deployment schedules (NVIDIA DGX/HGX rack builds), or is Vicor capacity-limited? Long-term supply agreements vs. purchase-order basis. Factory utilization rate.
• AEIS — semiconductor cycle exposure vs. data-center diversification. Revenue split between WFE (wafer fab equipment) cycle-tied business and data-center power. The FY2024 dip mirrored the 2023 WFE downturn — exposure to the next cycle.
• AAON — BASX ramp and path to positive FCF. Revenue split between BASX and legacy HVAC. BASX backlog if disclosed. When capex peaks and free cash flow turns positive. Hyperscaler customer identity and contract structure (take-or-pay vs. purchase-order).

Sources & confidence

• Primary grounding: 500-stocks sector scan, section 09 "Power Electronics & Power Conversion" — rack power density (40–120 kW), 12V-to-48V transition, SiC/GaN bottleneck description, ticker assignments.
• Existing sector reports: sectors/power-semiconductors.html (design-win dynamics, SiC substrate tightness, GaN capacity) and sectors/ups-power-distribution.html (adjacent power-delivery chain).
• Company financials: yfinance live data Jun 2 2026 — prices, market caps, TTM revenue/margins/earnings, balance sheets, forward consensus EPS. Reflects Q1-2026 10-Q filings for all four companies.
• Verified external: NVIDIA Q1 FY27 data-center revenue $75.2B (press release May 28 2026); Gartner data-center systems spending $788B forecast (via TechEdge AI May 20 2026).

Hard vs. approximate: Company financial figures (revenue, margins, earnings, balance sheet, prices) come from yfinance pulling filed SEC data. Market-size estimates ($30–40B total power-semi market), rack-power figures (40–120 kW), and SiC/GaN growth rates are from the sector scan — directional, not live-verified, and marked est. where they appear. The "short vs. long" gap assessments are analytical inferences from the financial and industry data, not contracted facts.