AI Server Power Modules Market Size, Share & Forecast to 2035

The AI Server Power Modules Market is valued at USD 3.57 billion in 2026 and is projected to reach USD 8.84 billion by 2035, up from USD 1.80 billion in 2025, a CAGR of 17.24% from 2025 to 2035 (DC Market Insights estimate). Current, not unit count, sets the money here: a 1,000 watt (W) accelerator running its core near 0.8 volts (V) draws about 1,000 amperes (A), so every jump in board power adds power stages, phases and converter bricks to the tray. Monolithic Power Systems (MPS) shows how fast this moved: its Enterprise Data revenue reached USD 380.6 million in the second quarter of 2026, up 164.3% on a year earlier, on sales of power management for AI and server applications (MPS second-quarter 2026 results, July 2026). DC Market Insights builds the 2025 figure from 11.48 million AI accelerators shipped, about USD 212 of board-level conversion content per graphics processing unit (GPU) and USD 89 per custom accelerator, giving USD 1.80 billion. Vertical power delivery modules grow fastest, at 51.13% a year, and the Middle East and Africa is the fastest region at 21.15% a year as Saudi Arabia and the United Arab Emirates (UAE) fill new AI campuses.

Key Market Statistics: How Big Is the AI Server Power Modules Market in 2026?

DC Market Insights values the AI Server Power Modules Market at USD 3.57 billion in 2026, 98.23% above 2025, because accelerator shipments rise from 11.48 million to 16.20 million while average GPU board power climbs from about 850 W to 1,150 W.

• Market size 2025: USD 1.80 billion; 2026: USD 3.57 billion; 2030: USD 6.93 billion; 2035: USD 8.84 billion.
• CAGR 2025-2035: 17.24% in the base case; the 2035 range runs from USD 6.36 billion (low, 13.45%) to USD 11.13 billion (high, 19.98%).
• Content per accelerator: about USD 157 on average in 2025, rising to USD 248 by 2035 (DC Market Insights estimate).
• Largest product type: multiphase voltage regulator (VR) power stages and controllers, USD 1.04 billion or 58.0% of 2025 value.
• Fastest product type: vertical power delivery (VPD) modules, from USD 54.03 million in 2025 to USD 3.36 billion by 2035.
• Fastest input class: 48 V to 54 V input, 20.50% a year; 800 V direct input reaches USD 1.94 billion by 2035 from zero in 2025.
• Custom application-specific integrated circuit (ASIC) servers grow 20.44% a year, against 15.93% for GPU servers.
• Leading region: North America, USD 1.07 billion (59.4%) in 2025; fastest region: Middle East and Africa, 21.15% a year.
• Largest country: United States, USD 986.99 million in 2025; fastest large country: India, 24.94% a year.
• Market leaders: MPS and Infineon Technologies, each with about a quarter of 2025 value; the top five suppliers hold about 74% (DC Market Insights estimate).

Executive Summary: How Fast Is Spending on AI Server Power Modules Growing?

The Global AI Server Power Modules Market size was valued at USD 94.20 million in 2020, reached USD 1.80 billion in 2025, and is anticipated to reach USD 8.84 billion by 2035, at a CAGR of 17.24% during the forecast period.

Attribute	Value
Historical period	2020-2024
Base year	2025
Forecast period	2026-2035
Market size 2025	USD 1.80 billion
Market size 2026	USD 3.57 billion
CAGR 2025-2035	17.24%
Market size 2035	USD 8.84 billion

Spending grew 80.42% a year from 2020 to 2025 because AI accelerator shipments rose from about 0.9 million to 11.48 million and board power per GPU more than doubled, from about 400 W for the A100 generation to 1,000 W for a B200 in a DGX B200 system. Growth slows to 17.24% a year to 2035: unit shipments flatten after 2028, and the price per watt of conversion falls from USD 0.250 to USD 0.150 as power stages integrate into modules. The 2026 step of 98.23% is the steepest of the decade; Infineon Technologies said on 5 August 2026 that it expects more than EUR 1.6 billion of dedicated AI power revenue in its fiscal year 2026 and that it is in allocation (Infineon third-quarter FY2026 analyst call, August 2026).

What Counts as an AI Server Power Module?

AI server power modules were a USD 1.80 billion market in 2025, counting only board-level conversion inside AI servers.

The market covers direct current to direct current (DC-DC) conversion bought for AI accelerator trays, AI server host boards and AI switch trays, valued at the module or component maker’s selling price.

Included: multiphase VR controllers and smart power stages (driver plus MOSFET in one package) for GPU, ASIC, high bandwidth memory (HBM) and input-output rails; vertical and dual-phase power modules mounted under or beside the processor; 48 V to 12 V and 48 V to 6 V intermediate bus converters (IBCs); 800 V to 12 V or 6 V on-board converters from 2027; point-of-load (PoL) modules for central processing unit (CPU), network interface and switch rails.

Excluded: alternating current to direct current power supply units and rack power shelves, which DC Market Insights covers in the Data Center Power Supply Unit Market; 800 V rack power and sidecars, covered in the Data Center High-Voltage Direct Current (HVDC) Market; battery backup units, uninterruptible power supplies, power distribution units, busbars, and power devices sold for general servers without accelerators.

How the segments fit together: every dollar sits once in each dimension: product type (USD 1.04 billion VR, USD 432.26 million IBC, USD 270.16 million PoL and USD 54.03 million VPD in 2025), input voltage, accelerator platform and device technology; regions and countries follow where the AI servers are installed.

Why Do AI Server Power Module Market Estimates Differ?

AI server power module estimates differ by boundary; DC Market Insights puts the market at USD 3.57 billion in 2026.

Broader definitions add the AC-DC power supplies and 54 V power shelves at the top of the rack, discrete power semiconductors sold for any server, telecom and industrial DC-DC bricks, or all data center power electronics, and those totals can be several times larger. Narrower definitions count only vertical power modules or only GPU core regulators and land well under USD 1 billion for 2025. DC Market Insights counts the conversion chain from the 48 V or 54 V tray input, and from 2027 the 800 V input, down to the processor rail, including the controller and the power stages even when they are sold as separate chips rather than as one module. Enterprise and smaller AI buyers, such as neoclouds and university clusters, are in scope through the servers they buy.

Element	In this report	Why
VR controllers and smart power stages	Yes, USD 1.04 billion in 2025	They are the core-rail power stage on every accelerator board
Vertical power delivery modules	Yes, USD 54.03 million in 2025	Same function, mounted under the processor
48 V and 800 V intermediate bus converters on the tray	Yes, USD 432.26 million in 2025	First board-level conversion step
AC-DC power supplies and power shelves	No	Rack-level product, covered in a separate report
Battery backup units and UPS	No	Energy storage, not conversion to the processor
General server and telecom DC-DC	No	No AI accelerator on the board

Market Drivers: Why Do AI Chips Need More Power Conversion Every Year?

Three forces add USD 7.04 billion to the market between 2025 and 2035: rising amps per accelerator, the move of conversion closer to the die, and the 800 V rack.

Board Power Rising from 700 W (H100) to 1,400 W (MI355X)

NVIDIA lists the H100 SXM at up to 700 W; NVIDIA’s DGX B200 user guide gives 1,000 W per GPU, and AMD’s datasheet gives the Instinct MI355X a maximum board power of 1,400 W. At about 0.8 V, a 1,400 W part needs about 1,400 A on its core rail, or about 20 smart power stages rated at 90 A running at 80% load, against about 10 for an H100. DC Market Insights estimates average GPU board power for units shipped at 850 W in 2025 and 1,900 W by 2030, which lifts conversion content per GPU from USD 212 to USD 361.

Allocation at Infineon and a 164% Jump at MPS in 2026

Supply is the binding constraint in 2026. Infineon said in August 2026 that it has signed capacity reservation agreements worth a high single-digit EUR billion amount over several years with prepayments, and that it will raise its EUR 2.5 billion fiscal 2027 AI data center target in November 2026. MPS guided third-quarter 2026 revenue to USD 1.14 billion to USD 1.16 billion after Enterprise Data revenue of USD 262.8 million in the first quarter and USD 380.6 million in the second. Those two suppliers alone show why DC Market Insights puts 2026 growth at 98.23%.

Vertical Power Delivery Moving Under the Processor

Lateral power stages beside the package lose power in the board copper as currents pass 1,000 A, so module makers are moving the stage under the processor. Vicor reported third-quarter 2026 revenue of USD 197.1 million and a USD 406 million backlog, and said it granted its first VPD licence to an AI original equipment manufacturer (OEM) at Stage-2 unit royalty rates (Vicor third-quarter 2026 results, October 2026). DC Market Insights expects Vertical Power Delivery Modules to take 38.0% of the market by 2035, against 3.0% in 2025.

Market Trends: Which Design Shifts Reshape AI Server Power Modules?

Three trends move about USD 5.66 billion of 2035 AI server power module value into product classes that barely existed in 2025.

800 V Racks from 2027 Remove the 54 V Shelf

NVIDIA wrote on 20 May 2025 that 54 V in-rack power hits its limits above 200 kilowatt (kW) racks and that 800 V direct current (VDC) power for 1 megawatt racks starts in 2027, with DC-DC conversion in the compute rack down to 12 V for the GPUs (NVIDIA technical blog, May 2025). NVIDIA also said the change cuts copper by up to 45% and lifts end-to-end efficiency by up to 5%. For this market it means a new high-ratio converter on each tray: DC Market Insights models 800 V Direct Input at 1% of value in 2027, 10% in 2030 and 22.0% in 2035.

Gallium Nitride in the First Conversion Stage

Infineon told analysts in August 2026 that it pairs gallium nitride (GaN) for dense, high-frequency intermediate stages with silicon at the processor. Renesas Electronics began sampling a 650 V GaN field-effect transistor in a dual-side-cooled 8 x 8 millimetre package for 800 V intermediate stages on 30 September 2026. DC Market Insights puts Gallium Nitride Devices at USD 72.04 million in 2025 and USD 1.77 billion by 2035, 37.71% a year.

Integrated Voltage Regulators Change Who Captures the Value

Analog Devices said on 19 May 2026 that it agreed to buy Empower Semiconductor, a maker of integrated voltage regulators for AI processors, for USD 1.5 billion in cash, and onsemi completed the purchase of Vcore power technology rights from Aura Semiconductor on 27 October 2025. DC Market Insights puts Integrated Voltage Regulators and Other at 22.0% of 2035 value, against 10.0% in 2025.

Market Challenges and Restraints: What Slows AI Server Power Module Spending?

Price erosion and patent risk could take about USD 2.47 billion off the 2035 base case, the gap between the base and low scenarios.

Price per Watt Falling 40% in Ten Years

DC Market Insights models conversion content falling from USD 0.250 per W in 2025 to USD 0.150 in 2035 as dual-phase and quad-phase modules replace discrete stages and Chinese suppliers qualify. A 90 A smart power stage such as Infineon’s TDA21490 lists at about EUR 2.34 on a 5,000-unit reel at a European distributor, so a 20-stage GPU rail costs under EUR 50 in stages before the controller and inductors.

Vertical Power Patent Dispute at the US International Trade Commission

According to the US International Trade Commission, it instituted investigation 337-TA-1526 into vertical power delivery systems on 9 October 2026 on a complaint by Vicor, which named module makers and server assemblers including Delta Electronics, Infineon, MPS, Flex and Luxshare. An exclusion order would hit USD 512.91 million of 2027 VPD spending modelled by DC Market Insights, and the low scenario assumes a two-year delay in VPD adoption.

Market Opportunities: Which AI Server Power Module Classes Add the Most Value?

The two largest AI server power module openings are worth USD 5.30 billion together by 2035.

VPD Modules: USD 3.36 billion by 2035

Vertical Power Delivery Modules grow from USD 54.03 million in 2025 to USD 3.36 billion by 2035 in the DC Market Insights base case. Infineon’s TDM24745T quad-phase module puts four power stages, a trans-inductor voltage regulator (TLVR) inductor and capacitors in 9 x 10 millimetres; suppliers that pair such modules with a licence position under the Vicor patents win this segment.

800 V Tray Converters: USD 1.94 billion by 2035

800 V Direct Input reaches USD 1.94 billion by 2035. Navitas Semiconductor showed an 800 V to 6 V power delivery board at NVIDIA GTC in March 2026 that removes the 48 V intermediate stage, and Texas Instruments’ LMM104RM0 1.6 kW 48 V to 12 V regulated converter shows the density level the 800 V parts must beat.

Pricing Analysis: How Much Does Power Conversion Cost per AI Accelerator?

AI server power modules cost about USD 212 per GPU in 2025 (DC Market Insights estimate).

That compares with USD 89 per custom ASIC and an average USD 0.250 per W of board power.

Year	GPU board power (fleet average)	Content per W	Content per GPU	Content per custom ASIC
2020	400 W	USD 0.320	USD 128	USD 68
2025	850 W	USD 0.250	USD 212	USD 89
2026	1,150 W	USD 0.270	USD 310	USD 118
2030	1,900 W	USD 0.190	USD 361	USD 145
2035	2,400 W	USD 0.150	USD 360	USD 153

Public price anchors are thin because hyperscalers and server makers buy under contract. The best public anchor is distributor pricing for a 90 A stage: about EUR 4.30 for a single Infineon TDA21490 and EUR 2.34 on a 5,000-unit reel. Content per W rose in 2026, to USD 0.270, because Infineon and MPS were supply-limited and because 2026 boards mix early VPD modules with lateral stages; DC Market Insights expects the decline to resume from 2027.

Which Specifications, Efficiency Laws and Tariffs Govern AI Server Power Modules?

Six named rules shape AI server power modules, a USD 3.57 billion market in 2026.

They cover rack voltage, the power management bus, server efficiency law and US and Chinese trade measures.

The Open Compute Project (OCP) Open Rack V3 specification defines the 48 V busbar that feeds most AI trays today, and the GB200 NVL72 busbar NVIDIA contributed to OCP carries 1,400 A. GPU VR controllers such as the MPS MP2891 report telemetry over the Power Management Bus (PMBus) 1.3. European Union Ecodesign Regulation (EU) 2019/424 sets server power supply efficiency of 96% at 50% load from January 2023, which pushes losses down the chain to the board converters. ENERGY STAR Computer Servers Version 4.0 applies to servers made from 12 January 2024. In trade, the US Department of Commerce opened a Section 232 investigation into semiconductors on 1 April 2025, and a proclamation of 14 January 2026 set a 25% duty on certain advanced computing chips with exemptions that include chips for US data centers (White House proclamation, January 2026). China’s Ministry of Commerce (MOFCOM) Announcement No. 46 of December 2024 restricted gallium exports to the United States; the ban is suspended until 27 November 2026, a risk for GaN stages.

Market Segmentation: Which AI Server Power Module Types Take the Most Spend?

Multiphase VR Power Stages and Controllers lead AI server power modules with USD 1.04 billion in 2025.

By Product Type

VR stages and controllers hold 58.0% in 2025 but grow only 6.88% a year as VPD modules, at 51.13%, take the core rail. IBCs grow 18.63% a year as every tray gains converters.

Product type	2025 (USD million)	Share 2025	2035 (USD million)	CAGR 2025-2035
Multiphase VR Power Stages and Controllers	USD 1,044.62 million	58.0%	USD 2,032.33 million	6.88%
Vertical Power Delivery Modules	USD 54.03 million	3.0%	USD 3,357.76 million	51.13%
Intermediate Bus Converters	USD 432.26 million	24.0%	USD 2,385.77 million	18.63%
Point-of-Load and Auxiliary Rail Modules	USD 270.16 million	15.0%	USD 1,060.34 million	14.65%

By Input Voltage

12 V input parts still carry 62.0% of 2025 value; 48 V to 54 V input grows 20.50% a year and 800 V direct input appears in 2027.

Input voltage	2025 (USD million)	Share 2025	2035 (USD million)	CAGR 2025-2035
12 V Input	USD 1,116.67 million	62.0%	USD 2,474.14 million	8.28%
48 V to 54 V Input	USD 684.40 million	38.0%	USD 4,418.10 million	20.50%
800 V Direct Input	USD 0.00 million	0.0%	USD 1,943.96 million	new from 2027

By Accelerator Platform

GPU servers account for 74.3% of 2025 value; custom ASIC servers grow faster, 20.44% a year, as Google, Amazon Web Services and Meta add their own accelerators.

Accelerator platform	2025 (USD million)	Share 2025	2035 (USD million)	CAGR 2025-2035
GPU-Based AI Servers	USD 1,338.76 million	74.3%	USD 5,868.00 million	15.93%
Custom ASIC AI Servers	USD 462.31 million	25.7%	USD 2,968.20 million	20.44%

By Device Technology

Silicon MOSFET stages hold 86.0% of 2025 value; GaN devices grow 37.71% a year, mostly in 48 V and 800 V converters.

Device technology	2025 (USD million)	Share 2025	2035 (USD million)	CAGR 2025-2035
Silicon MOSFET Power Stages	USD 1,548.92 million	86.0%	USD 5,125.00 million	12.71%
Gallium Nitride Devices	USD 72.04 million	4.0%	USD 1,767.24 million	37.71%
Integrated Voltage Regulators and Other	USD 180.11 million	10.0%	USD 1,943.96 million	26.86%

DC Exclusive: Amps-per-Accelerator Teardown, About 20 Power Stages for a 1,400 W Chip

AI server power modules per GPU rise from about USD 128 (A100 class) to USD 360 (2035 fleet).

The Amps-per-Accelerator Teardown is a DC Market Insights model built on 7 inputs: vendor board power, core voltage, stage rating, load factor and the content per W series.

Accelerator class	Board power	Core-rail current (approx.)	90 A stages at 80% load	Content per W	Conversion content per accelerator
NVIDIA A100 SXM class (2020)	400 W	400 A	6	USD 0.320	USD 128
NVIDIA H100 SXM (2023)	700 W	700 A	10	USD 0.305	USD 214
AMD Instinct MI300X (2024)	750 W	750 A	10	USD 0.300	USD 225
NVIDIA B200 in DGX B200 (2025)	1,000 W	1,000 A	14	USD 0.250	USD 250
AMD Instinct MI355X (2025)	1,400 W	1,400 A	19	USD 0.250	USD 350
Fleet average, units shipped 2030	1,900 W	1,900 A	26	USD 0.190	USD 361
Fleet average, units shipped 2035	2,400 W	2,400 A	33	USD 0.150	USD 360

The teardown assumes about 80% of board power sits on the core rail at about 0.8 V and that each 90 A stage runs at 80% of its rating; HBM, input-output and host rails take the rest. Two readings follow. First, stage count roughly tracks board power, so the 2,400 W fleet average of 2035 needs about 33 stage-equivalents, three times an H100. Second, content per accelerator rises far less than current, from about USD 212 to about USD 360 between 2025 and 2035, because integration lowers the price per W by 40%.

Regional Insights: Where Are AI Server Power Modules Installed?

North America leads with USD 1.07 billion, 59.4% of 2025 value by place of installation, and the Middle East and Africa grows fastest at 21.15% a year.

North America

North America grows from USD 1.07 billion in 2025 to USD 4.86 billion by 2035, 16.34% a year, as the large US cloud operators install most GB200 and GB300 racks; the United States alone is USD 986.99 million in 2025.

Asia Pacific

Asia Pacific grows from USD 443.06 million to USD 2.26 billion, 17.71% a year. Most modules are assembled into trays in Taiwan and mainland China, but DC Market Insights counts value where servers are installed: China at USD 199.92 million and India, the fastest large country, at 24.94% a year.

Europe

Europe grows from USD 165.70 million to USD 883.62 million, 18.22% a year, led by the United Kingdom at USD 41.42 million and Germany at USD 32.42 million in 2025.

Middle East and Africa

The Middle East and Africa rises from USD 88.25 million to USD 600.86 million, 21.15% a year, the fastest of the five regions, with Saudi Arabia at 21.63% and the UAE at 21.57% a year.

Latin America

Latin America is the smallest region, USD 34.22 million in 2025, and grows 20.98% a year to USD 229.74 million by 2035 as Brazil and Mexico add AI halls.

Which Countries Spend Most on AI Server Power Modules?

The United States leads AI server power module spending with USD 986.99 million in 2025, 54.8% of the market, and India is the fastest large country at 24.94% a year.

Rank	Country	Region	2025 (USD million)	2035 (USD million)	CAGR 2025-2035
1	United States	North America	USD 986.99 million	USD 4,453.45 million	16.26%
2	China	Asia Pacific	USD 199.92 million	USD 795.26 million	14.81%
3	Canada	North America	USD 82.85 million	USD 406.46 million	17.24%
4	Japan	Asia Pacific	USD 63.04 million	USD 300.43 million	16.90%
5	South Korea	Asia Pacific	USD 41.42 million	USD 220.91 million	18.22%
6	United Kingdom	Europe	USD 41.42 million	USD 203.23 million	17.24%
7	India	Asia Pacific	USD 32.42 million	USD 300.43 million	24.94%
8	Saudi Arabia	Middle East and Africa	USD 32.42 million	USD 229.74 million	21.63%
9	Germany	Europe	USD 32.42 million	USD 167.89 million	17.87%
10	United Arab Emirates	Middle East and Africa	USD 28.82 million	USD 203.23 million	21.57%
11	Australia	Asia Pacific	USD 27.02 million	USD 141.38 million	18.00%
12	France	Europe	USD 23.41 million	USD 132.54 million	18.93%

China’s 14.81% a year is below the global rate because US export rules limit the highest-power accelerators it can buy; its domestic accelerators run at lower board power, which means less conversion content per chip.

Forecast Scenarios: How Large Could AI Server Power Modules Be by 2035?

DC Market Insights puts AI server power modules in 2035 between USD 6.36 billion and USD 11.13 billion, with USD 8.84 billion as the base case.

Scenario	2035 (USD million)	CAGR 2025-2035	Main assumption
Low	USD 6,362.06 million	13.45%	Accelerator units flat after 2028 and VPD adoption delayed two years
Base	USD 8,836.20 million	17.24%	35.7 million accelerators in 2035 at 2,400 W per GPU
High	USD 11,133.61 million	19.98%	Board power reaches 3,000 W per GPU and 800 V trays take 35% of value

Who Are the Leading AI Server Power Module Suppliers?

MPS and Infineon each hold about 25% of 2025 value, and the top five suppliers about 74% (DC Market Insights estimate based on company disclosures).

• Monolithic Power Systems
• Infineon Technologies
• Vicor
• Renesas Electronics
• Delta Electronics
• Texas Instruments
• Analog Devices
• onsemi
• Alpha and Omega Semiconductor
• Flex
• Murata Manufacturing
• Navitas Semiconductor
• Luxshare Precision
• Empower Semiconductor

MPS reported 2025 revenue of USD 2.8 billion, up 26.4%, with Enterprise Data at USD 701.8 million; DC Market Insights attributes about USD 450 million of that, 25% of this market, to AI accelerator boards. Infineon gives the largest disclosed AI power figure, more than EUR 1.6 billion in fiscal 2026, but that includes power supplies and silicon carbide, so DC Market Insights counts only its board-level share, about 24% of 2025 value. Vicor’s 2025 revenue was USD 452.7 million, including a USD 45 million patent settlement, and its royalties grow with VPD. Renesas and Delta Electronics complete the top five at about 7% and 6%. Texas Instruments, Analog Devices and onsemi are building positions through new modules and acquisitions, and Murata showed its MYC0409 48 V divide-by-four module for AI hardware at PCIM in May 2026. The AI Server Market report covers the server makers that buy these parts, and the Data Center Accelerator Market report gives the unit base used here.

What Has Happened Recently in AI Server Power Modules?

Six dated AI server power module events since October 2025 point to tighter supply and a fight over vertical power.

• In October 2026, Vicor reported third-quarter revenue of USD 197.1 million, a USD 406 million backlog and its first VPD licence to an AI OEM, and bought two more sites for ChiP fabs (Source: Vicor).
• In October 2026, the US International Trade Commission instituted investigation 337-TA-1526 into vertical power delivery systems (Source: US International Trade Commission).
• In September 2026, Renesas Electronics began sampling a 650 V GaN transistor for 800 V intermediate bus converters (Source: Renesas Electronics).
• In August 2026, Infineon raised its fiscal 2026 AI power revenue outlook to more than EUR 1.6 billion and reported multi-year capacity reservation deals (Source: Infineon Technologies).
• In May 2026, Analog Devices agreed to buy Empower Semiconductor for USD 1.5 billion in cash (Source: Analog Devices).
• In October 2025, onsemi completed its purchase of Vcore power technology from Aura Semiconductor (Source: onsemi).

Key Takeaways

Four AI server power module findings matter most, starting from a USD 3.57 billion market in 2026.

• 2026 market size: USD 3.57 billion, up 98.23%, because accelerator units rise 41% and GPU board power rises 35% in the same year.
• Vertical Power Delivery Modules reach USD 3.36 billion by 2035, because lateral stages lose too much power in board copper above 1,000 A.
• 800 V Direct Input reaches USD 1.94 billion by 2035, because NVIDIA’s 800 VDC racks from 2027 need a new converter on every tray.
• Content per W falls 40% to USD 0.150 by 2035, because power stages integrate into dual-phase and quad-phase modules.

Research Methodology: How Did DC Market Insights Size the AI Server Power Modules Market?

DC Market Insights sized AI server power modules bottom-up at USD 1.80 billion for 2025.

The receipt: 11.48 million accelerators, 6.30 million GPUs at 850 W and 5.18 million custom ASICs at 420 W, times USD 0.250 of conversion content per W (15% lower for ASIC boards). Unit shipments come from the DC Market Insights accelerator model. GPUs: 6.30 million x 850 W x USD 0.250 = USD 1,338.76 million; custom ASICs: 5.18 million x 420 W x USD 0.2125 = USD 462.31 million; total USD 1,801.07 million.

Three bound checks hold. First ceiling: USD 1.80 billion is 0.93% of NVIDIA’s USD 193.7 billion data center revenue for fiscal 2026, against a bill-of-materials share of about 1% for board power in published teardowns of GPU servers. Second ceiling: the sum of MPS Enterprise Data (USD 701.8 million), Vicor (USD 452.7 million) and Infineon’s board-level AI share stays within twice the 2025 estimate, even though part of each figure is non-AI. Floor: MPS’s AI share and Vicor’s product revenue alone give about USD 0.80 billion in 2025, so the market cannot be under USD 0.8 billion. Full method notes are on the research methodology page.