The 800V DC Era Is Here — and It's Rewriting Who Builds AI Power

Global big tech companies are making pilgrimages to Korean power equipment firms, showing up personally with power grid blueprints and placing orders by any means necessary. The reason is simple and urgent: the pace of building AI data centers is so far outstripping the expansion of power grids that power supply has become a matter of survival. And at the center of this scramble is a single architectural decision that NVIDIA is now pushing as an industry standard — the shift to 800V direct current.

This is not a routine infrastructure upgrade. The shift from today's 54V DC rack architecture to 800V DC is the largest change to data center electrical infrastructure in decades. It touches every layer of the power stack, from the utility grid to the GPU die. It requires entirely new semiconductor materials — silicon carbide (SiC) and gallium nitride (GaN) replacing conventional silicon — and it is creating a scramble for supply that explains why NVIDIA engineers are sitting in conference rooms in Seoul, Suwon, and Daejeon.

This essay explains why 800V DC is necessary, how the architecture works from grid to GPU, what specific power semiconductor devices are required at each stage, who is supplying them, and why Korean firms occupy a strategically critical position in a supply chain that every major AI infrastructure operator in the world now depends on.

1. Why 54V Is Dying and 800V Is the Only Solution

To understand the urgency, start with the physics. Today's data center racks run on a 54V DC distribution architecture. Every compute tray receives 54V from rack-level power supply units (PSUs), which convert the incoming AC power (typically 415VAC from the building) to 54V DC. The GPUs then receive further conversion down to the 1V-class voltages their cores require.

This worked reasonably well when racks consumed 10–30 kW. It breaks catastrophically at the power densities that modern AI compute demands.

Power = Voltage × Current   →   Current = Power / Voltage

At 54V, a 1 MW rack requires:
  Current = 1,000,000W / 54V = 18,519 Amperes

Copper conductor constraint:
  Each mm² of copper safely carries ~5-10 A in power distribution
  18,519 A requires ~1,850 - 3,700 mm² of copper cross-section
  That's a busbar the size of your forearm — per rack

At 800V, the same 1 MW rack requires:
  Current = 1,000,000W / 800V = 1,250 Amperes
  Copper required: ~125 - 250 mm² — a standard industrial cable

Conclusion: 800V doesn't just help — it's the only way to deliver
            megawatt rack power without the building becoming copper.

The numbers are stark. Physics dictates that at higher voltage, the same power can be delivered with far less current, which means far less copper, far less resistive heating, and far less I²R loss in the distribution system. NVIDIA's own analysis shows that switching to 800V busways enables 85% more power through the same conductor size compared to 415VAC distribution — and reduces copper requirements by 45%.

But the physics argument is only part of the story. The second problem with 54V is conversion efficiency. Today's data center power chain involves multiple conversion stages, each adding loss:

TODAY — MULTI-STAGE CONVERSION CHAIN (end-to-end efficiency <90%)
─────────────────────────────────────────────────────────────────
Grid (13.8 kVAC)
    │ Step-down transformer       [loss ~1.5%]
    ▼
415 VAC distribution
    │ UPS / AC conditioner        [loss ~3-5%]
    ▼
Low-voltage AC busway
    │ Rack PDU                    [loss ~1%]
    ▼
54 VDC (rack PSU, AC→DC)         [loss ~5-8%]
    │ IBC (Intermediate Bus Conv) [loss ~3-5%]
    ▼
12 VDC
    │ VRM (Voltage Regulator)     [loss ~3%]
    ▼
~1V GPU core voltage
─────────────────────────────────────────────────────────────────
6-7 conversion stages · Cumulative loss: 12-20% of input power

FUTURE — 800VDC ARCHITECTURE (end-to-end efficiency ~95%)
─────────────────────────────────────────────────────────
Grid (13.8 kVAC)
    │ SST / Industrial Rectifier  [loss ~2-3%]  ← SiC 1200V
    ▼
800 VDC distribution (campus / floor level)
    │ 800V → 6V or 12V LLC Conv  [loss ~2-3%]  ← GaN 650V/SiC
    ▼
6-12 VDC (adjacent to GPU)
    │ Multiphase buck VRM         [loss ~2%]   ← GaN 100V
    ▼
<1V GPU core voltage
─────────────────────────────────────────────────────────
3 conversion stages · Cumulative loss: ~5-7%

The architecture diagram reveals why 800V is transformative: it collapses a 6-7 stage conversion chain into 3 stages. Each eliminated stage is not just a percentage-point efficiency gain — it is a failure point removed, a set of components no longer needed, and floor space reclaimed for compute.

"The exponential growth of AI computing demands a fundamental rethinking of how we deliver power in data centers. Traditional power distribution architectures are reaching their limits."

— Kannan Soundarapandian, VP & GM of High-Voltage Power, Texas Instruments (GTC 2026)

2. The Semiconductor Problem: Why Silicon Can't Do This

Here is where the power semiconductor story becomes critical — and where Korea enters the picture. The 800V DC architecture does not just require new power conversion topology. It requires power semiconductor devices that conventional silicon simply cannot provide.

Conventional silicon power transistors and diodes have a hard practical voltage limit of around 600-900V due to the material's bandgap (1.1 eV). Above that, the device would require silicon so thick that its on-resistance becomes prohibitive and switching becomes too slow. For 800V bus-side power conversion, you need devices rated at 1200V or higher to provide adequate safety margin. That means silicon carbide (SiC).

For the mid-rail conversion stages (800V to 6-12V), the requirement shifts to very high switching frequency — up to 1 MHz — to reduce the size of magnetic components (inductors and transformers) that would otherwise consume too much rack space. Silicon cannot switch at 1 MHz without unacceptable switching losses. That means gallium nitride (GaN).

3. Why NVIDIA Is Knocking on Doors in Seoul

The Korean angle in the news report from April 22 is not accidental. Korean industrial conglomerates and specialized power equipment manufacturers occupy a unique position in the 800V supply chain — they sit at the intersection of high-voltage electrical infrastructure (transformers, switchgear, busways, industrial rectifiers) and advanced power electronics, with manufacturing scale that Western and Japanese competitors struggle to match.

The firms NVIDIA is courting are not consumer electronics companies. They are the builders of the electrical infrastructure that power plants, industrial facilities, and now hyperscale data centers depend on. Companies like LS Electric, HD Hyundai Electric, and Hyosung Heavy Industries make the medium-voltage switchgear, transformers, and rectifier systems that sit between the utility grid and the data center — exactly the layer where 800V DC infrastructure begins.

The urgency is structural. Global demand for AI data center build-out is running so far ahead of equipment supply that lead times for medium-voltage transformers and switchgear have extended to 18-24 months in many markets. NVIDIA is not just looking for suppliers — it is trying to secure capacity before competitors lock it up, bringing its own power grid blueprints to Korean firms and structuring supply agreements that bypass normal procurement channels.

4. The Full Semiconductor Ecosystem: Who Supplies What

The 800V DC architecture has catalyzed one of the most active partnership formation cycles in the semiconductor industry. NVIDIA has publicly named its ecosystem partners, and the list spans the globe — but the specific devices each company supplies map clearly onto the architecture stages above.

4.1 The SiC Tier: Grid-to-800V Conversion

Silicon carbide devices handle the high-voltage, high-power stages of the 800V architecture. The dominant players are companies that built SiC expertise in the automotive industry (EV inverters use 1200V SiC devices for very similar reasons) and are now pivoting that expertise to data centers.

4.2 The GaN Tier: In-Rack and Point-of-Load Conversion

Gallium nitride dominates the mid-rail and point-of-load stages where high switching frequency and extreme power density are the binding constraints.

4.3 System-Level Players: The Infrastructure Tier

Above the semiconductor layer sit the system integrators and infrastructure companies that turn discrete power devices into rack-level products:

5. The NVIDIA Kyber Architecture: How 800V Actually Gets to the GPU

NVIDIA has published the forward architecture for its data center power delivery in the Kyber rack design — the successor to the NVL72 / MGX architecture that supports GB300 and Rubin Ultra platforms. Understanding Kyber's power path explains exactly which semiconductor devices matter at each stage and why the conversion ratios are what they are.

NVIDIA KYBER RACK — 800VDC POWER DELIVERY PATH
═══════════════════════════════════════════════════════════

FACILITY PERIMETER
  13.8 kVAC utility  ──→  Industrial SiC Rectifier  ──→  800 VDC bus
                          (Vertiv / onsemi SiC 1200V)
                          ~97-98% efficiency

RACK ENTRY
  800 VDC busway  ──→  800V Hot-Swap Protection
                       (ST SiC 1200V + BCD controller)
                       Prevents arc flash, enables live insertion

COMPUTE TRAY  
  800V  ──→  LLC Converter (64:1 ratio)  ──→  12V (adjacent to GPU)
             (GaN 650V, 500kHz-1MHz)
             Peak efficiency: 97.5%
             Power density: 2,500 W/in³
             26% less area than prior multi-stage approach

  OR (next-gen, TI / Navitas approach):
  800V  ──→  Single-stage LLC  ──→  6V  (eliminates 12V IBC)
             (GaN 650V)
             97.6% peak efficiency, >2,000 W/in³

GPU DIE POWER
  6V or 12V  ──→  Multiphase Buck VRM  ──→  <1V GPU core
                  (GaN 100V, multi-phase)
                  High-current delivery directly adjacent to die

═══════════════════════════════════════════════════════════
Total: 3 stages vs. 6-7 in legacy · Efficiency gain: ~5%
1MW+ rack density supported with standard conductor sizing

The 64:1 conversion ratio in the LLC converter is notable. Generating 12V from 800V in a single stage — with 97.5% efficiency and fitting in a smartphone-sized form factor — is a genuine materials science achievement made possible entirely by GaN's high-frequency switching capability. At 500kHz-1MHz, the magnetic components (transformers, inductors) shrink proportionally. Silicon switches at 50-100kHz in comparable configurations; GaN at 1MHz is a 10× frequency increase that translates directly to 10× smaller magnetics.

"By eliminating an entire conversion stage, we lower system cost and power losses while freeing up valuable board space, enabling customers to dedicate more real-estate to compute, memory and GPUs."

— Chris Allexandre, CEO, Navitas Semiconductor (GTC 2026)

6. The Korean Supply Chain Specifically

The news article's focus on Korean power firms is pointing at a specific and real supply constraint. Korea's industrial conglomerates — built over decades serving power plants, heavy industry, and submarine cable infrastructure — have capabilities that very few global manufacturers match.

The critical components Korean firms supply (or are being solicited to supply) fall into three categories:

6.1 Medium-Voltage Switchgear and Transformers

Every 800V DC data center begins with a medium-voltage grid connection (typically 13.8kV or higher in the US, similar voltages in Europe and Korea). The switchgear and step-down transformers at this interface must be industrial grade, built to IEEE and IEC standards, with the scale to handle tens of megawatts per facility. Korean companies like LS Electric, HD Hyundai Electric, and Hyosung Heavy Industries are among the few global manufacturers with the capacity, certification infrastructure, and lead-time flexibility to supply at the pace AI data center build-out requires.

6.2 Industrial Rectifier Systems

The conversion from medium-voltage AC to 800V DC at the facility perimeter requires industrial-grade rectifier systems rated for continuous megawatt-scale operation. This is distinct from the rack-level SiC devices discussed earlier — these are room-sized power conversion systems requiring custom engineering for each installation. Korean firms with backgrounds in HVDC transmission (Korea has invested heavily in HVDC undersea cable links) have directly transferable technology.

6.3 SiC Wafer and Device Manufacturing Capacity

Korea's semiconductor manufacturing ecosystem — anchored by Samsung and SK Hynix but extending to specialized fabs — is being courted for SiC device manufacturing capacity. SiC wafer supply remains a genuine global constraint: 6-inch SiC substrates are still scarce, and 8-inch production is only beginning to ramp. Korean fabs that can be qualified for SiC device production represent strategic supply chain diversification away from the current US/European/Japanese SiC supplier concentration.

7. The Transition Timeline and What It Means for Investment

NVIDIA has been explicit that the 800V DC transition will occur in phases. The current generation (Blackwell GB300 NVL72) still runs on traditional power architecture. Kyber — the rack design purpose-built for 800V DC — is targeted at the Rubin Ultra generation, with deployment expected in 2027. This timeline creates a specific investment and supply chain ramp pattern.

The investment implication is directional: the demand for SiC and GaN power semiconductors from the AI data center vertical is about to grow from a small fraction of total market to a significant share. Currently, automotive (EV inverters, onboard chargers) drives the majority of SiC revenue. Industrial applications are second. Data center has been negligible. By 2027-2028, that mix shifts substantially — and unlike automotive (where EV adoption has been choppy), data center SiC demand is backed by capital commitments that are being made today for AI infrastructure that must be built regardless of macroeconomic conditions.

8. The Solid-State Transformer: The End-State Vision

The current 800V architecture still uses traditional transformers and rectifiers at the grid interface — large, expensive, heavy equipment with 18-24 month lead times. The ultimate vision, being pursued by Navitas in partnership with EPFL (École Polytechnique Fédérale de Lausanne), is the solid-state transformer (SST).

An SST replaces the traditional grid-frequency transformer with a high-frequency switching converter that uses SiC and GaN devices to step down medium-voltage AC directly to 800V DC in a single compact unit. The advantages are significant: an SST can be 10-50× smaller than an equivalent traditional transformer, can include active power factor correction and harmonic filtering, can respond to grid events in microseconds rather than cycles, and can directly integrate battery storage into the power path.

The SST is not yet production-ready at data center scale. The SiC devices required for 6.6kV-to-800V conversion in an SST context must operate at extremely high dV/dt rates and face reliability challenges that are still being engineered. But the trajectory is clear: within 3-5 years, the entire facility-level power conversion stage may be replaceable by an SST-based system that is smaller, faster, more efficient, and easier to deploy than today's substation-scale rectifier infrastructure.

When that happens, the lead times driving NVIDIA to knock on Korean factory doors disappear. So does a significant part of the advantage those Korean firms currently possess. The scramble we are seeing today is partly a race to secure supply before the solid-state alternative matures — and partly an investment in relationships that will matter when the next architectural transition happens.

9. Bottom Line

The Korean news story is a symptom of a structural shift that is now well underway. AI data centers must reach megawatt rack densities. Megawatt racks require 800V DC distribution. 800V DC requires SiC and GaN power semiconductors at every conversion stage. Those semiconductors are in constrained supply, and the industrial-scale infrastructure to deploy them at data center scale is in even more constrained supply.

NVIDIA's courtship of Korean power firms is the supply chain consequence of a physics requirement. It is not primarily about cost or convenience — it is about the fact that the equipment needed to build 800V AI data centers at the pace NVIDIA's customers are demanding can only be manufactured by a small number of qualified global suppliers, and Korean industrial firms are among the most capable and least constrained of them.

For the power semiconductor industry, this transition represents the opening of a major new vertical market. SiC manufacturers who built their process maturity on automotive applications are finding that data center is now their fastest-growing customer segment. GaN manufacturers who proved their technology in USB-C chargers and 5G base stations are receiving qualification orders for rack-level power delivery boards that represent hundreds of times the revenue per unit of their prior products.

The 800V DC era is not coming. Based on the investments, partnerships, and factory visits already underway, it is already here — running slightly ahead of the silicon that powers it.