Artificial intelligence has become a national-level infrastructure priority, and the capital flowing into it is reshaping an entire industrial chain — from AI chips to data center construction to the large models and applications running on top of them. Within that chain, data center construction plays a foundational role: servers and clusters, optical modules and switches, liquid cooling, and — increasingly the bottleneck — power supply and electrical protection.
This article works through where that power demand is coming from, why NVIDIA’s next-generation GPU platforms are forcing a rethink of data center electrical architecture, why DC fault protection is fundamentally harder than AC, and how solid-state circuit breakers (SSCBs) and solid-state transformers (SSTs) are becoming the answer. Along the way we’ll reference HIITIO’s own SSCB, SST, and power semiconductor product lines, which sit directly inside this shift.
The Power Demand Behind the AI Buildout
The scale of investment behind AI computing infrastructure is difficult to overstate. China’s 15th Five-Year Plan period alone is expected to bring an estimated RMB 4 trillion in new direct investment toward building a nationwide computing power network — a systemic, government-backed buildout rather than a single company’s capital expenditure cycle.
Global data center power demand tells a similar story. According to a joint report from Zhonglian Fund and Cushman & Wakefield (“Investment and Value Insights in the Data Center Industry”), global data center power demand is projected to rise from roughly 49 GW in 2023 to 96 GW in 2026. Of that 47 GW increase, newly built AI computing centers account for roughly 40 GW — about 85% of the total increase — with a compound annual growth rate exceeding 30%.
The International Energy Agency’s most recent analysis puts real numbers behind that trajectory: electricity demand from data centers grew 17% in 2025, while demand from AI-focused facilities specifically grew faster still, and overall data center electricity consumption is on pace to double by 2030. That kind of growth curve means power is no longer a background utility cost for data center operators — it is a primary design constraint, and in many markets, the primary bottleneck to how fast a facility can be brought online.
| Metric | 2023 | 2026 (Projected) | Driver |
|---|---|---|---|
| Global data center power demand | ~49 GW | ~96 GW | Overall data center buildout |
| AI computing center share of the increase | — | ~40 GW (~85% of total increase) | New AI-specific facilities |
| Compound annual growth rate | — | 30%+ | AI computing center demand |
| Data center electricity demand growth (2025, IEA) | — | +17% year-over-year | AI-focused facilities growing fastest |
That constraint shows up directly in the capital stack. Building a data center currently costs somewhere between USD 7–12 million per MW of operational IT load. Using a reference project in Northern Virginia — a 700,000-square-foot facility with 60 MW of IT load — total construction cost (excluding ICT equipment) runs between USD 420–770 million, of which the electrical power system accounts for more than 40%. Power infrastructure isn’t a line item anymore; it’s close to half the build.
| Cost Driver | Reference Range |
|---|---|
| Data center construction cost | USD 7–12 million per MW of IT load |
| Reference project (Northern Virginia, 700,000 sq ft, 60MW IT load) | USD 420–770 million total (excl. ICT equipment) |
| Electrical power system share of construction cost | 40%+ |
Two Zones, One Coordinated Architecture
Inside an AI data center (AIDC), infrastructure is typically split into a “white zone” and a “gray zone” based on function and equipment type:
| White Zone | Gray Zone | |
|---|---|---|
| Function | Core compute area | Power intake, supply, distribution, protection |
| Key equipment | Compute racks, thermal management systems, rack-level power distribution | Switchgear, transformers, UPS systems, cooling equipment, backup generators, fuel systems |
| Primary operators | IT engineers, HVAC engineers, electrical engineers | Power engineers |
| Voltage class | Skews toward low-voltage / weak-current equipment | Medium- and high-voltage equipment |
The power supply architecture connecting these zones has gone through several generations: from traditional AC UPS, to high-voltage DC (HVDC), to Panama-style power architectures, and now toward solid-state transformers (SST). Each step has targeted the same three goals — higher efficiency, higher reliability, and higher power density.
| Architecture Generation | Typical Efficiency | Key Characteristic |
|---|---|---|
| Traditional AC UPS | ~85% | Multiple AC/DC conversion stages |
| High-Voltage DC (HVDC) | 95%+ | ~10 points higher than traditional UPS |
| Panama-style architecture | Transitional | Bridges AC UPS and full DC bus designs |
| Solid-State Transformer (SST) | 97.5%+ | Single-stage MV AC to LVDC conversion |
See HIITIO’s Data Center Power Solution for how these layers fit together in a single coordinated architecture.
NVIDIA’s Next-Generation GPUs Are Forcing the Issue
Chip-level power consumption is the proximate cause of this architectural shift. NVIDIA’s per-chip power draw has already broken through 2kW, and the roadmap for the next-generation Vera Rubin platform pushes further still:
| Platform | Representative Model | Per-Chip TDP | Rack Power |
|---|---|---|---|
| Hopper | H100 | ~700W | ~10–15kW |
| Blackwell | GB300 | ~1,400W | ~120–150kW |
| Vera Rubin | VR200 | 2,300W | ~225kW |
| Vera Rubin Ultra | VR300 | 4,000W+ | 660kW–1.3MW |
Vera Rubin (VR200), expected in the second half of 2026, locks in a per-GPU TDP of 2,300W. Vera Rubin Ultra (VR300), targeted for 2027, is expected to push single-chip power beyond 4,000W. At that scale, a single rack can approach or exceed a megawatt of continuous load — territory that traditional 48V/54V rack-level DC distribution simply cannot support without prohibitive copper and conversion losses.
NVIDIA’s answer is an 800V HVDC Power Rack architecture, targeted for availability to Vera Rubin customers in Q3 2026. The physics are straightforward: at equal power, raising the bus voltage from 48V to 800V cuts current to roughly 1/16, which in turn cuts resistive line losses to roughly 1/256 — directly addressing the heat and loss problems that come with extreme power density.
This shift was made explicit in October 2025, when NVIDIA presented its whitepaper, 800 VDC Architecture for Next-Generation AI Infrastructure, at the OCP Global Summit in North America. The paper positions 800V HVDC as the mainstream power architecture for future AI data centers, marking an industry-wide shift away from traditional AC/UPS architecture toward high-voltage DC. Reported test results for NVIDIA’s Kyber rack architecture show a 26% reduction in physical volume, roughly 8% lower energy consumption, and sustained load capability above 1MW per rack. As IEEE Spectrum has covered, the appeal of 800V DC goes beyond efficiency — it also helps facilities manage the extreme, sub-second load swings that GPU clusters produce as they move between idle and full utilization, something traditional architectures struggle to absorb gracefully.
Why the GPU Revolution Changes the Power Math Inside the Building
The shift isn’t only about voltage class — it’s about how much of a data center’s footprint and power budget compute now consumes. Historically, data center design centered on the compute space, with servers occupying most of the physical hall and power/cooling systems taking a comparatively modest share. The GPU revolution has inverted that balance: compared to a typical web server, GPU rack power density is now nearly 100 times higher and growing at a near-exponential rate, effectively turning data centers into “AI factories.”
That density brings a second problem: load volatility. Without intervention, rack power can swing rapidly between roughly 30% idle and 100% full load. At small scale this is a rack-level nuisance; at cluster scale, it can ripple out to affect data-center-level and even grid-level power supply. Four mitigation strategies are commonly deployed: software-based idle-cycle optimization, on-site energy storage systems, redundant power consumption, and capping GPU performance to reduce peak draw.
Why DC Fault Protection Is a Different Problem Than AC
As data centers move toward 400V/800V DC distribution, the protection layer has to solve problems that simply don’t exist — or exist in much milder form — in AC systems. Four challenges define the difficulty:
| Challenge | What Happens | Why It’s Hard |
|---|---|---|
| 1. Fast-rising fault current | Parallel converters and filter capacitors on the DC bus discharge almost simultaneously the instant a fault occurs | Traditional switchgear can’t keep up with the current spike |
| 2. Arc without zero-crossing | DC current never naturally crosses zero the way AC does twice per cycle | The arc has no self-extinguishing point, making it inherently hard to break |
| 3. Capacitor inrush | Capacitor charging current spikes when power is first connected | Can trip protection devices never meant to see that transient, causing nuisance trips |
| 4. Selectivity difficulty | Short-circuit current must be interrupted almost instantly | Traditional mechanical switches react too slowly to isolate just the faulted branch |
A closer look at the fault-current problem specifically illustrates why the device category matters:
| Device Type | Current-Limiting Capability | Typical Response Time |
|---|---|---|
| Fuse | None — melts only once short-circuit energy reaches a set threshold | Depends on fault energy; no active limiting |
| Thermal-magnetic MCCB | Some, but limited by mechanical structure | Millisecond range |
| Hybrid solid-state breaker (SSHCB) | Limited; mechanical stage must act before the electronic stage | Slower than a pure SSCB — mechanical stage is the bottleneck |
| Solid-state circuit breaker (SSCB) | Full electronic current limiting | Microsecond range |
Solid-State Circuit Breakers: How They Solve Each Problem
| Challenge | How an SSCB Solves It |
|---|---|
| Arc without zero-crossing | No mechanical contacts to open, so there’s no arc to suppress in the first place — no oversized opening gap or arc chamber needed |
| Capacitor inrush | A current-limiting control circuit charges the capacitor gradually during conduction before the switch moves to full conduction |
| Selectivity difficulty | Action time as fast as 10 microseconds, combined with time-delay coordination, enables differential protection — the breaker closest to the fault trips first |
| Fast-rising fault current | Electronic current limiting reacts before the fault current can climb to damaging levels |
For selectivity specifically, the fastest SSCB platforms can act in as little as 10 microseconds. Combined with a time-delay coordination function, this allows differential protection across a distribution tree — the breaker closest to the fault trips first, containing the impact to the smallest possible section of the system rather than cascading upstream.
HIITIO’s HBS1 Series Solid-State Circuit Breaker — full specs on the Solid-State Circuit Breaker product page — is built around this principle. The series spans multiple current classes to match different points in a DC distribution architecture:
| Model Class | Cooling | Isolation |
|---|---|---|
| 80A | Air-cooled | With isolation |
| 250A | Air-cooled | With isolation |
| 800A | Air-cooled | With isolation |
| 2,500A | Liquid-cooled | Without isolation |
In field deployment, HBS1 units have logged multiple years of continuous operation and more than one million switching cycles at a metallurgical and chemical processing site — a data point that matters because SSCB performance is often quoted from lab testing rather than sustained field conditions. This is field-verified life data, not only a laboratory projection.
| Deployment | Application | Field Track Record |
|---|---|---|
| Metallurgical / chemical processing site | Industrial process protection | Several years continuous operation; 1,000,000+ switching cycles |
| China Southern Power Grid, Zhuhai | DC distribution system | 3+ years continuous operation |
| Inner Mongolia | PV-plus-storage installation | In service |
| Metro system | DC220V protection | In service |
| Data center | 800V DC distribution project | In service |
For a deeper technical walkthrough of how the fault-interruption sequence actually works, see HIITIO’s companion guide on how solid-state and integrated architectures are replacing the discrete component stack.
SSCB vs. Traditional Circuit Breakers
The performance gap between solid-state and traditional mechanical DC breakers isn’t incremental — it’s a step-change. Traditional mechanical DC breakers rely on physical contact separation and arc extinction to interrupt current; their action time sits in the millisecond range, meaning the DC fault current has often already climbed substantially before it’s cleared, and every interruption involves arcing and contact erosion that shortens device life. Solid-state breakers interrupt electronically, without contact separation, at speeds two to three orders of magnitude faster — with no arc-related wear to manage over the device’s lifetime.
| Traditional Mechanical Breaker | Solid-State Circuit Breaker (SSCB) | |
|---|---|---|
| Interruption method | Physical contact separation + arc extinction | Fully electronic, no moving contacts |
| Response time | Millisecond range | Microsecond range (as fast as 10µs) |
| Arcing | Present on every interruption; causes contact erosion | None — no arc is formed |
| Conduction loss | Low (closed metal contact) | Higher than a closed contact |
| Relative cost per amp | Lower | Higher |
| Maintenance | Contact wear requires periodic servicing | No mechanical wear to service |
The trade-off is that solid-state devices carry higher conduction losses than a closed mechanical contact, and the semiconductor content still makes them more expensive at scale — which is exactly why so much of the surrounding industry (from national research programs to component suppliers) is focused on closing that efficiency and cost gap.
That trade-off is also why hybrid solid-state circuit breakers (SSHCBs) remain relevant in the current market: they combine a mechanical stage (low conduction loss, higher cost efficiency) with an electronic interruption stage (fast, arc-free breaking), at the cost of slower response and more complex control than a pure SSCB. Domestic low-voltage electrical equipment leaders in China are actively developing SSCB product lines, an indication that the device category is moving from a specialty component toward a mainstream part of the DC protection catalog.
Where This Shows Up in Real Deployments
The clearest application case is an 800V DC power distribution system for a data center, where the SSCB provides fast protection for the 800V power bus feeding compute racks — the exact scenario NVIDIA’s 800V roadmap is pushing the industry toward. Beyond circuit breakers, the same architecture typically pairs SSCBs with high-speed dual-power transfer switches (HTS) for source redundancy, HVDC contactors for isolation and maintenance switching, and — increasingly — a solid-state transformer stepping medium-voltage AC directly down to 800V DC output, eliminating a chain of separate conversion stages.
| HIITIO Solid-State Transformer — Key Specs | Value |
|---|---|
| Rated power | 1.2MW–5MW |
| Input voltage | 10kV / 13.8kV / 35kV |
| Output voltage | 240–1000VDC |
| Peak efficiency | >97.5% |
This is the kind of footprint reduction and efficiency gain that a legacy transformer-plus-UPS chain can’t match. For a broader view of how SST platforms are being positioned across the industry for data center use, see HIITIO’s overview of solid-state transformer suppliers for data centers.
HIITIO’s Position Across the Power Chain
What distinguishes HIITIO’s approach is coverage across multiple layers of the same architecture rather than a single point product. That includes solid-state transformers and solid-state circuit breakers at the system level, power semiconductor modules (IGBT and SiC, also available through hiitiosemi.com) at the component level feeding those systems, and the surrounding protection layer — HVDC contactors, semiconductor fuses, and DC MCBs — that a full AI data center power architecture still needs regardless of how far the facility has moved toward solid-state and 800V DC. HIITIO supplies this stack both as a complete system for data center operators and as coordinated components for third-party SST integrators and system builders, which matters for procurement teams evaluating whether to source a full solution or fill specific gaps in an existing design. For more on how these protection layers coordinate with telecom-grade and legacy -48VDC systems as well, see HIITIO’s guide to DC power protection in telecom and data center infrastructure.
FAQ
Why is 800V DC replacing 400V and 48V architectures in AI data centers?
As GPU rack power climbs toward and past 1MW, lower-voltage DC buses require impractical amounts of copper to carry the current and lose more energy to resistive heating. Raising the bus voltage to 800V cuts current — and resistive losses — dramatically for the same delivered power, which is why NVIDIA and much of the surrounding ecosystem are converging on 800V HVDC as the target architecture for next-generation AI factories.
What makes DC fault protection harder than AC protection?
DC current has no natural zero-crossing, so an arc drawn during a DC fault doesn’t self-extinguish the way an AC arc does twice per cycle. Combined with the very fast current rise from parallel converters and capacitors discharging into a fault, and inrush current at initial energization, DC protection devices need current-limiting and interruption characteristics that traditional AC-derived switchgear wasn’t designed to provide.
How fast can a solid-state circuit breaker actually interrupt a fault?
The fastest platforms in HIITIO’s HBS1 Series clear a fault within roughly 10–20 microseconds of detection, compared to the millisecond-range response of a thermal-magnetic breaker — a difference of two to three orders of magnitude that meaningfully limits how far a fault can propagate before it’s isolated.
Is a solid-state circuit breaker a drop-in replacement for a mechanical breaker?
Not exactly. SSCBs eliminate arcing and mechanical wear and interrupt far faster, but they carry higher steady-state conduction losses than a closed metal contact and typically cost more per amp of rating. Many current DC distribution designs use SSCBs specifically at the points where fault-clearing speed matters most, while continuing to use mechanical or hybrid devices elsewhere in the architecture.
Does HIITIO supply complete SST or SSCB systems, or only components?
Both. HIITIO manufactures complete solid-state transformer and solid-state circuit breaker systems for data center operators, and also supplies the underlying power semiconductor modules, HVDC contactors, and semiconductor fuses to third-party SST integrators and system builders who need a coordinated, certified protection layer rather than a full turnkey system.
Talk to HIITIO About Your AI Data Center Power Architecture
If your team is evaluating an 800V DC migration, specifying protection for a new AI computing center, or comparing solid-state circuit breaker platforms against traditional mechanical or hybrid designs, HIITIO’s engineering team can walk through current class, cooling method, and system-level coordination against your specific rack and cluster power profile. Whether you need a complete solid-state transformer and SSCB power architecture or a coordinated protection layer to slot into an existing design, HIITIO’s HBS1 Series, SST platform, and power semiconductor modules are built to work together — reach out through the Data Center Power Solution page to start a conversation with our team about your project’s power and protection requirements.
