Power consumption surges 100-fold! Bank of America: The era of 1.5 MW data center cabinets has arrived, disrupting the traditional power distribution system

The unlimited expansion of AI computing power is hitting a real wall—electricity.

According to Chasing Wind Trading Desk and Bank of America’s latest research report, as NVIDIA’s GPU platforms iterate, data center cabinet power consumption will surge from 10-15 kW for traditional servers to over 1.5 MW in the Feynman platform era of 2029-2030, an increase of nearly 100 times, and existing power infrastructure can no longer support this demand.

According to Bank of America's global research team estimates, power demand by AI data centers will add a total of 233 GW between 2025 and 2030, with annual additions expanding from about 17 GW in 2025 to about 60 GW in 2030. This scale far exceeds the International Energy Agency (IEA)'s projected doubling path for data center installed capacity based on current project pipelines. Electricity has become the most crucial limiting factor for AI expansion.

Breaking the power bottleneck will give rise to a vast new market for analog semiconductors. Bank of America estimates that the addressable market (TAM) for analog semiconductors in AI data centers will expand from $790 million in 2025 to about $2.7 billion in 2030, with a five-year compound annual growth rate of 28%. Analog chip manufacturers will be the most direct beneficiaries, while wide bandgap semiconductor materials such as silicon carbide (SiC) and gallium nitride (GaN) will accelerate their migration from cyclical demand in automotive and industrial sectors to long-term structural demand in AI data centers.

100-fold surge in power consumption: The cost of computing power from kilowatts to megawatts

The increase in AI computing power density is boosting cabinet power consumption geometrically.

Bank of America’s report breaks down the power evolution of various NVIDIA platforms: The Hopper H100 HGX cabinet released in 2022 has a total power consumption of about 32 kW; moving to the Blackwell GB200 NVL72 era, as the number of GPUs increases from 32 to 72 and GPU thermal design power (TDP) rises sharply, total cabinet consumption surges to 100-120 kW. The upcoming Rubin Ultra NVL576 platform is expected to consume over 646 kW per cabinet; in the Feynman era (expected 2029-2030), 576 GPUs will be integrated into a single node, with cabinet power consumption breaking 1.5 MW—enough to power about 1,000 U.S. households.

The core driving force behind the surge is the physical constraints of GPU scale networking. NVIDIA calls this the “performance density trap”: To maximize computing performance, GPUs must be tightly integrated through copper interconnects over very short distances, directly linking maximum power density with maximum performance. From Hopper to Blackwell, GPU TDP increased by 75%, but cabinet power density increased by 3.4 times, and performance increased by 50 times. Bank of America expects each expansion in scale networking domains will lead to a 2-4 times increase in total power consumption.

This trend is not unique to NVIDIA. The AMD Helios platform already consumes over 100 kW, and custom ASIC platforms like AWS Trainium 3 and Google Ironwood also continue to rise with increases in computing power and network density. Bank of America believes that all future platforms will generally converge towards higher power consumption, which is a necessary condition for competitiveness with NVIDIA.

Current architecture hits the ceiling: Threefold failure in traditional power distribution systems

The current data center power distribution architecture is simultaneously hitting physical limits in several dimensions.

The traditional architecture uses a 48V/54V DC power distribution scheme: Grid high-voltage AC is stepped down through multiple stages, converted to 54V DC at the cabinet level by the power supply unit (PSU), and then stepped down 1-2 times more to reach the sub-1V voltage rails required by GPU cores. This pathway has three fundamental flaws.

Space constraint: A GB300 NVL72 cabinet requires up to 8 power shelves. If the 54V DC distribution is used, Kyber cabinets (Rubin Ultra and subsequent platforms) will require 64U rack space for power supplies, severely compressing space for computing resources.

Copper material bottleneck: In a 1 MW cabinet, 54V DC power distribution needs up to 200 kg of copper bars to transmit power, which is completely unsustainable at gigawatt scale.

Conversion efficiency loss: Each AC/DC conversion loses about 1–2% energy; multiple-stage conversion not only reduces overall efficiency but also increases the number of failure nodes.

800V DC: Redesigning the chain from grid to chip

To address these challenges, an 800V DC (800 VDC) architecture is seen as the next-generation standard for data center power distribution. The core logic: Move the AC-to-DC conversion node as far upstream as possible, reducing intermediate conversion levels, enhancing efficiency, lowering costs, and freeing up cabinet space.

With 800 VDC, 13.8 kV AC is directly rectified to 800V DC upon entry to the site, eliminating multiple intermediate conversion steps of traditional architecture. NVIDIA data shows that compared to the 54V system, 800 VDC can increase end-to-end efficiency by up to 5%; transmit 85% more power with the same wire cross-section; reduce copper usage by about 45%; decrease maintenance costs by up to 70%; and improve total cost of ownership (TCO) by up to 30%.

800 VDC implementation will progress in phases. The current transitional solution is to move AC-to-DC conversion to the “sidecar” power rack outside the cabinet, with Kyber cabinets as an example; the mid-term solution is to install large rectifiers at the facility level to convert low-voltage AC directly to 800V DC; the long-term final solution is a hybrid microgrid architecture centered on solid-state transformers (SST), expected to roll out gradually with greenfield projects by 2028–2030.

Moreover, highly synchronized AI training loads can cause cabinet power usage to spike from 30% to 100% utilization within milliseconds, creating violent grid fluctuations. The solution is multi time-scale energy storage: supercapacitors handle millisecond-scale spikes, large battery energy storage systems (BESS) smooth minute-scale load fluctuations, isolating the volatility of AI infrastructure demand from grid stability needs.

$2.7 billion new market: Structural opportunity for analog semiconductors

The comprehensive restructuring of the power architecture will create an unprecedented incremental market for the analog semiconductor industry. Bank of America has built a bottom-up industry demand model, translating accelerator and cabinet demand into content pools for each component, divided into two dimensions: low power (<200 kW) and high power (>600 kW) cabinets.

Market size: AI analog semiconductor TAM is expected to grow from $790 million in 2025 to about $2.7 billion in 2030 (28% compound annual growth rate), including data centers growing from $760 million to $2.5 billion (about 26% CAGR), and strategic power infrastructure increasing from $24.5 million to $180 million (49% CAGR).

Single cabinet content value: As cabinet power levels rise, analog semiconductor content value climbs sharply—100–160 kW cabinets are about $36,000, cabinets above 600 kW are about $290,000, and 1 MW cabinets nearly $920,000. The value focus is shifting toward intermediate bus converters (IBC), GPU board-level power supplies, CPU added content, and optical infrastructure.

Material structure change: Analog ICs remain the largest market, expected to reach about $1.59 billion in 2030, but SiC and GaN will be the fastest-growing segments, with five-year CAGR of 63% and 69%, respectively. Both will rise from edge applications in data centers to become core materials for high-voltage conversion and protection.

Competitive landscape: Bank of America estimates TXN will have the highest share in the AI analog semiconductor market, expected to maintain about 21% in 2030; Infineon's share will rise most sharply, from about 12% in 2025 to about 17% in 2030, likely becoming the second-largest AI supplier; ADI ranks third, benefiting from its Empower acquisition and enhanced competitiveness in processor near-end power delivery; ON is rapidly expanding its share in high-power markets thanks to SiC and vertical GaN (vGaN) technologies.

Infrastructure layer: Solid-state transformers and circuit breakers open new tracks

Outside the data center server room, the power infrastructure layer will also undergo profound transformation, creating a new market for analog semiconductor manufacturers that previously barely existed.

Solid-state transformers (SST): Traditional transformers have delivery cycles as long as 2–3 years, already a bottleneck for data center construction. SST can convert medium-voltage AC (usually 13.8–35 kV) directly to 800V DC, shrinking size by about 14 times and weight by about 40 times compared to traditional transformers, and shortening construction cycle by about 50%. Bank of America expects SST’s analog semiconductor opportunity will be released intensively with the proliferation of hybrid microgrid architectures between 2028–2030, with a market size of about $500 million then. SiC is the core material for SST, with Infineon, Wolfspeed, and Navitas all actively deploying.

Solid-state circuit breakers (SSCB): Traditional mechanical circuit breakers cannot respond fast enough (millisecond level) to meet fast isolation requirements in high-voltage DC distribution environments. SSCBs can interrupt currents in nanoseconds to microseconds and integrate monitoring and remote control functions. Bank of America expects the SSCB analog semiconductor market to reach about $400 million by 2030, with Infineon and ON in strong positions via their SiC JFET to MOSFET product lines.

Energy storage systems (ESS/UPS): AI data center energy storage demand has evolved from backup power to a core part of power distribution architecture. Bank of America estimates this sub-market will grow from about $15.6 million in 2025 to nearly $80 million in 2030 (38% CAGR), with Infineon, TXN, and Renesas all having strong presences.

~~~~~~~~~~~~~~~~~~~~~~~~

The above wonderful content comes from Chasing Wind Trading Desk.

For more detailed interpretation, including real-time insights and frontline research, please join [Chasing Wind Trading Desk ▪ Annual Membership]

Risk warning and disclaimerThe market involves risk; investment should be cautious. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial condition or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are suitable for their specific circumstances. Investment based on this article is at your own risk.