A Blueprint for 800 VDC Datacenters

The Heron Link is a 4.2 MW solid-state transformer (SST) that connects directly to 34.5 kV and delivers 800 VDC output with over 98.5% conversion efficiency. It integrates MV protection, an MV disconnect switch, and a 40-channel DC power distribution panel with per-channel overcurrent protection. Its modular, fault-tolerant architecture eliminates single points of failure, enabling higher system availability and uptime. The Link's power converter modules are field-replaceable via blind-mate connectors, allowing swaps to be completed in minutes rather than hours. A single Heron Link provides 2.5 times the power density per square foot when compared to the traditional equipment it replaces.

A rack training a frontier model may swing between 30% and 150% of its thermally designed power within milliseconds, and because all racks in a cluster step through the same training iterations in lockstep, these swings are facility-wide. Left unmanaged, the entire campus load pulses in synchrony, creating severe flicker concerns and threatening utility interconnection agreements.The SuperBBU addresses this at the source. A high C-rate battery, skid-integrated with each Heron Link, couples directly to the 800 VDC bus without an additional DC-to-DC converter. This design responds faster than any existing ripple mitigation solutions, absorbing rack-level power fluctuations and presenting a smooth, predictable load to the grid. In addition, it provides 30 seconds of full-load backup power during a grid disturbance, matched to the ~15-second start time of generators connected at the MV level. For applications requiring longer hold-up or utility-imposed ramp rate control, the SuperBBU can be substituted with long duration energy storage via external DC-DC coupling.

Reducing the complexity and footprint of electrical equipment lowers cost. With four equipment categories eliminated, total MV to rack electrical cost drops 75%, from roughly $1.77M/MW to $0.42M/MW, a saving of $1.35 million per MW installed.

The labor savings are even steeper than equipment savings: 90% versus 65%. Electrical labor falls from ~$0.73M/MW to ~$0.08M/MW. That asymmetry reflects the staggering complexity in a conventional build. UPS commissioning, switchgear installation, and PDU field wiring each require specialized trades working in strict order. Removing those steps saves not just the cost of the equipment but also collapses the commissioning dependency chain that adds months to the schedule.

By removing four equipment categories, Heron's blueprint cuts $0.65M/MW in field labor associated with MV to rack power delivery. A conventional 480V, 12 MW data hall with grey space and electrical fit-out takes approximately 14 months to build, with warm-shell labor running approximately $2.50M/MW³. The $0.65M/MW saving represents 25% of that labor cost. Assuming labor reduction maps proportionally to schedule compression, removing 25% of field labor from a 14-month schedule recovers approximately three months.

Every conversion stage in the power chain loses some energy as heat. The conventional 480 VAC chain loses 6.4% of input power before it reaches the compute tray. The Heron blueprint’s two-stage chain reduces losses by over 50%, to only 3.0%⁴. At 100 MW of grid input, this blueprint can deliver 3.4 MW more electricity directly to GPUs rather than dissipating in distribution equipment. That is 3.4 MW of additional GPU capacity from the same utility contract.

Heron Link’s power converter architecture is designed for fail- operational redundancy, ensuring continuous operation even in the presence of component failures. Each 4.2 MW Link consists of thirty 140 kW power cells, with ten cells connected in series per phase to achieve a 34.5 kV output. The system can tolerate up to two power cell failures per phase. Fast-acting bypass devices with sub-cycle response times automatically isolate and bypass faulted cells. Each cell failure reduces power handling capacity by 3.3%, but full output voltage and grid fault ride-through compliance is preserved.

This fail operational architecture allows flexibility in designing the redundancy architecture of the MV to LV power train. Depending upon the application and uptime requirements, N+1 redundancy can be reduced to N+0 where each compute rack is fed by only N+0 Heron Links. A power cell failure results in compute racks de-rating their performance proportionally. NVIDIA’s Vera Rubin GPU is already capable of limiting its power draw via software controls (Max-P and Max-Q modes). This architecture provides a compelling alternative to the traditional approach of overbuilding MV to LV conversion infrastructure.

Each Heron Link includes a solid-state circuit breaker (SSCB) per rack output that limits fault let through current to less than 2 kA and clears faults in under 5 microseconds. A fault on any output is isolated before it can propagate, bounding failures to a single power shelf, a single panel output, or a single Heron Link unit. A mechanical contactor provides galvanic isolation after clearing.On the MV side, UL-listed current-limiting fuses and an integrated disconnect rated for 40 kA at 34.5 kV coordinate with upstream utility protection without a dedicated switchgear enclosure. The 4-to-make-3 topology ensures that even a complete Heron Link failure leaves every rack running on its remaining three power shelves.

As AI campus loads grow to hundreds of megawatts on single feeders, grid operators are raising the bar for interconnection. ERCOT’s Large Electrical Load framework, which is increasingly a reference point for other markets, requires large loads to demonstrate fault ride-through capability and stable, controllable power draw. Facilities built on conventional architecture cannot meet these requirements without significant additional equipment.

The Heron Link with integrated SuperBBU addresses these requirements natively. The SuperBBU absorbs AI training load transients directly on the 800 VDC bus, preventing ripple from propagating upstream and reducing thermal stress on MV cables, switchgear, and protection equipment. During grid disturbances, the Heron Link's grid-forming capability enables seamless fault ride-through while providing sufficient hold-up time for backup generator start-up or feeder reconfiguration.

One of the most significant risks in data center construction today is building power infrastructure for current rack densities only to find it cannot serve the next generation of rack density without costly rebuilds. This blueprint eliminates that risk by design.

A data hall built to this blueprint in 2027 will serve NVIDIA's 1 MW+ Kyber racks with no changes to the MV feeders, Heron Link units, or protection system. The Link's power whips can be reallocated to higher power density racks through row-level combiners or mechanical adapters. Looking further ahead, the same first principles that drove the industry from 48 to 800 VDC are pointing toward 1500 VDC as the next preferred distribution voltage for the ultra dense compute racks. Heron Link is natively designed to support 1500 VDC applications, thereby ensuring capital deployed today remains relevant through multiple generations of compute.