MISSION

We build the coordination layer between AI rack power and grid behavior — so new AI capacity can connect without making the grid less stable.

THE PROBLEM

Two power crises inside every AI rack.

Compute stranded by power conservatism

AI workloads create fast, high-amplitude power swings. To keep facility protection from tripping, operators hold back electrical headroom and cap power conservatively — capacity that is paid for but not used.

The ORv3 / ORW interconnection deadlock

The OCP ORv3 / ORW rack architecture distributes battery backup (BBU) across individual racks for energy efficiency, but provides no orchestration layer. On a grid sag, PSUs disconnect and BBUs operate independently; on grid restoration, PSU startup follows a randomized 0–5.5 s delay (ORv3 PSU specification) plus a cold-boot of several seconds. That is incompatible with the recovery windows North American grid codes now set — ERCOT requires at least 90% of load back within 2 seconds of voltage recovery (PFAPR).

What happened on 10 July 2024

On 10 July 2024 a faulted 230 kV line in Northern Virginia reclosed six times in 82 seconds, each time pulling local voltage down to 0.25–0.40 p.u. Nearby data centers moved about 1,500 MW of load to backup power on their own — customer-side controls, not utility equipment, took it off the grid — and most of it stayed off for hours. On 22 July 2026 a similar fault moved about 3,800 MW, the largest load transfer PJM has recorded.

NERC's January 2025 Incident Review — Considering Simultaneous Voltage-Sensitive Load Reductions made the point: grids have long planned for losing a large generator, not for large blocks of load leaving at once. Event background and sources →

Which is why ERCOT's NOGRR 282 is now in force, and other operators are following:

  • Ride-through and recovery — mandatory for large computational loads in ERCOT since 1 August 2026.
  • Failure to ride through — triggers a root-cause investigation and a remediation plan; if ERCOT determines that continued operation poses an imminent risk to local or system reliability, it can require the load to disconnect until compliance is shown.

WHAT WE BUILD

XMight POM — active power orchestration at the rack.

XMight builds two products on POM, the open rack-level coordination interface it contributes to OCP:

XMight POM Manager

Facility-level layer. Sets the power envelope for each rack from grid conditions and site policy, and records the per-rack evidence grid operators increasingly ask for.

XMight POM Module

Rack-mounted module in the ORv3 power shelf's PMI-class slot, with its own control plane. Coordinates BBU discharge and PSU output in real time, and keeps operating autonomously if the facility layer is unreachable.

XMight's POM implementation adds the coordination layer ORv3 / ORW racks are missing — orchestrating BBU discharge to absorb workload transients, and sequencing PSU restart so the rack can recover inside the windows NOGRR 282 and the rules following it define.

WHY THIS APPROACH

Three technical paths. Where each one fits.

What grid codes now ask Facility UPS GPU throttling XMight POM
Stays on the grid through a faultWhat failed in July 2024 and July 2026 Transfers the load to battery; whether it returns to the grid depends on each UPS's settings. Acts on GPU power draw; how PSUs and UPS react to a voltage dip is unchanged. By designThe Module is designed to hold the rack busbar from local storage and keep the rack connected.
Recovers inside the windowERCOT: back to 90% within 2 s Set by each UPS's re-transfer logic, facility by facility. PSU restart timing is unchanged: a random 0–5.5 s delay on top of a cold boot. By designPSU restart is sequenced so rack power can return inside the window.
Repeated faults in a rowSix in 82 seconds in 2024 Depends on protection settings; some schemes count dips and lock the load off. Outside its scope. By designEach disturbance is handled on its own — no strike counting.
Evidence for the grid operatorGrid operators now ask for it Facility-level logs. Job-level logs. By designA per-rack record of every event, as input to the site operator's filing.
Impact on AI jobs Compute is untouched. Works by reducing GPU power, which can slow synchronous training. By designActs on the power layer and does not command GPU throttling; throughput through a disturbance depends on its envelope and the stored energy available.
Fit with OCP racks Upstream of the rack; rack BBUs act on their own. Integrated per server vendor, through BIOS / BMC. By designPOM is an open interface in review with OCP Rack & Power — ORv3 today, designed for Mt Diablo.
Retrofit on a running site Facility works: switchgear, distribution and battery rooms. Software change; no hardware. By designOne module in each power shelf; no facility shutdown.

Where each fits: facility UPS and BESS still carry bulk energy and long outages, and throttling can trim peaks in software. XMight POM is the layer that coordinates each rack's ride-through and recovery and records it — alongside them, not instead of them.

The XMight POM column describes the design of the XMight POM Module and POM Manager. UPS behaviour: NERC, Incident Review — Considering Simultaneous Voltage-Sensitive Load Reductions, Jan. 2025. PSU restart timing: OCP ORv3 PSU specification.

INDUSTRY CONTRIBUTION

Where we sit in OCP.

  • OCP Startup Program 2026 member

    Admitted to the OCP Startup Program 2026 cohort.

  • POM in review with OCP Rack & Power

    POM (Power Orchestration Module) — an open rack-level coordination interface for grid ride-through and recovery — is in review with the OCP Rack & Power project. The XMight POM Module is XMight's implementation of POM.

  • Contributor, OCP DCF LVDC Power Distribution white paper v1.1

    Named contributor to the OCP Data Center Facility LVDC Power Distribution white paper v1.1.0 (September 2026), section 8.1 Utility Grid — ride-through and recovery requirements as they reach the rack power shelf.

  • Presenter, OCP Global Summit 2025 Future Technologies Symposium

    "The Active BBU: Dynamic Power Orchestration for Stable and Efficient ORv3 AI Racks" — the technical foundation of POM.

  • Exhibitor, OCP Global Summit 2026 Innovation Village

    Demonstrated "No Central UPS, Still Grid-Code-Ready" in the Innovation Village (FTS – Power), San Jose, 12–15 October 2026.

WHAT WE BELIEVE

Open standards win

We contribute to OCP because rack power specs should be open. Power orchestration belongs in a community-maintained spec, not a vendor moat.

Coordination closes the gap

The gap between GPU power transients and grid protection is a coordination problem. XMight's POM implementation addresses it with rack-level coordination and a record of how each rack responded.

Compliance is continuous

ERCOT NOGRR 282 makes ride-through an operating requirement, not a one-time test. Every disturbance needs the right response — and the record to show it.

Legal entity
XMight Smart Energy Corp.
Sector
AI data center power infrastructure
Founded
2022
Focus
Rack-level power orchestration (POM) — ORv3 to Mt Diablo

Founded by Young Liang, 2022 · LinkedIn ↗

Reach the team at service@xmight.com or visit the contact page to discuss grid-code ride-through or rack integration.