The AI electric wall: Mitsubishi $9 bn, SEA under pressure, and solid-state transformers as saviors

In plain terms

The energy demand of AI has shifted from a theoretical concern to a measurable industrial constraint. This week, four announcements converge: Mitsubishi Estate is investing $9 billion in data centers in Japan, IEEE Spectrum documents that AI load is testing the limits of the US grid, Southeast Asia is heading towards a power wall, and Pandaily documents the shift from conventional transformers to solid-state transformers (SST) in hyperscale data centers.

The story (the 4 signals)

1. Mitsubishi Estate - $9 bn to build DCs in Japan. (Nikkei) The real estate developer is shifting a significant portion of its 2026-2029 capex to data center campuses in the Kanto and Kansai regions. This is the first Japanese real estate developer to make a move of this magnitude. Signal: In Japan, the constraint is no longer fiber, but the combined availability of land and electricity.

2. IEEE Spectrum - the US grid bends but does not break. AI consumption peaks (training runs, desynchronized from the diurnal cycle) create transients of several hundred MW in a few minutes that the grid was not designed to absorb. PJM and ERCOT are beginning to impose "demand response" contracts on hyperscalers - obligation to shed load at peak.

3. Southeast Asia - race for PPAs. The DC boom in Singapore (already saturated, moratorium lifted at the end of 2025), Johor (Malaysia), and Batam (Indonesia) hits generation. The solar+battery PPAs signed in 2024 are being renegotiated today at significantly higher levels (order of magnitude: +30 to +60% according to market channels, not attributable to a single public report). The power wall is not theoretical - it is already in the contracts.

4. Solid-state transformers become the hardware solution. (Pandaily) SSTs replace conventional 60 Hz transformers with electronic power conversion (SiC/GaN). Benefit: power density ×3, losses halved, millisecond reactivity (relevant for AI transients). The Chinese are leading the way - CRRC, Sungrow, TBEA equip Sugon and Alibaba.

Under the hood

• Order of magnitude: an H100 GPU draws ~700 W, a B200 ~1000 W. A rack of 72 B200 = ~120 kW load - 3× a conventional rack.
• Long-term PPA: the recent trend in new hyperscaler contracts points to a marked increase, without a single consensus figure being publicly established to date.
• STS cost vs conventional transformer: ×3 at purchase, but significantly more favorable on total cost over 15 years (TCO) thanks to reduced losses.
• Thread (#1001, #994): this issue consolidates several signals already covered under the same hardware prism.

So what

Three consequences. One: the real estate lever (Mitsubishi, GLP, Blackstone, Digital Realty) becomes a "power-adjacent" asset class more than real estate - the real question is the connection. Two: hyperscalers will have to sign demand-response contracts, making AI workloads more complex (geographic displacement of trainings). Three: SiC/GaN + SST suppliers (CRRC, Sungrow, Wolfspeed, onsemi) are the picks-and-shovels of the wave - to watch. Signal to monitor: the first outage imposed on a hyperscaler by a US grid operator.