Physical AI · Grid, Power & Thermal Infrastructure

AI Data Center 800V DC Power Architecture — System-Level Supply Chain

800 VDC is a higher voltage distribution architecture for moving power toward dense computing racks. It changes the path between facility electricity and final processor rails; it does not mean a GPU operates directly at 800 volts.

Tiers
4
Companies
3
Updated
2026-10-20
Source state
5 source links

Supply-chain depth map

800 VDC is a higher-voltage distribution architecture for moving power toward dense computing racks. It changes the path between facility electricity and final processor rails; it does not mean a GPU operates directly at 800 volts. The supply chain should therefore be mapped by electrical function, with protection, conversion and final regulation treated separately.

NVIDIA’s August 2026 architecture update describes a hybrid power rack planned for the second half of 2026 alongside a transition toward native 800 VDC facilities. That is a manufacturer schedule and architecture statement, not independent confirmation of completed deployment at every site. It is more specific than a blanket claim that the transition begins only in 2027.

A GB300 collaboration does not by itself establish that every GB300 installation uses 800 VDC, that one infrastructure supplier is mandatory, or that all intermediate rails disappear. Legacy, hybrid and native installations can require different equipment. The supplier map below identifies functions and evidence boundaries rather than asserting an exclusive bill of materials.

For equal power, higher distribution voltage reduces current; the resistive-loss benefit depends on the actual conductors and operating point. A current-squared illustration holding resistance constant is not a measured whole-facility efficiency improvement. Conversion losses, protection, installation and operating costs must also be included before estimating customer savings.

Facility supply and protection → row/rack distribution → conversion stages → point-of-load rails → computing load
Power-semiconductor devices enable selected conversion and protection stages.
Supplier selection and intermediate rails depend on the qualified system architecture.
Tier 0

Facility supply, protection and project integration

The upstream layer connects the site to an available electrical supply and provides suitable protection and distribution. Utility access, transformers, switchgear and construction can constrain a project independently of the chosen rack voltage. The required equipment depends on the site and cannot be inferred from a generic campus size.

A delayed interconnection can postpone downstream hardware demand even when the hardware is ready. Conversely, an existing powered facility may use a hybrid retrofit rather than rebuild its entire electrical system. This is why upstream electrical spending and native 800 VDC conversion should not be counted as the same market.

Supplier exposure at this layer requires a specific product or project award. This map does not repeat universal lead-time estimates or assume that a listed industrial company owns the grid connection. Project timing, permitting and available capacity must be checked for the actual installation.

Tier 1

Row and rack distribution and integration

This layer moves power into and across the computing installation. A higher-voltage design needs appropriate conductors, connectors, protection and mechanical integration. Thermal infrastructure remains important, but electrical distribution and coolant routing are separate functions; every busbar is not necessarily a liquid-cooled assembly.

NVIDIA’s described hybrid approach provides a transition path for existing facilities. That does not establish exclusive purchasing from a single reference-design participant. Buyers may qualify multiple suppliers, use different integration partners or adopt the architecture on different schedules. Reference participation, product availability and installed production systems are distinct evidence categories.

The economic opportunity should be tied to equipment actually supplied. Integrators, enclosure manufacturers and power-equipment providers should not each receive credit for the entire rack budget. Useful evidence includes qualified configurations, funded orders, accepted deliveries and subsequent service requirements rather than unsupported current market-share rankings.

Vertiv describes its NVIDIA collaboration as work on platform designs. Eaton separately announced an architecture supporting NVIDIA’s 800 VDC direction. Together these establish multiple participating suppliers, rather than a mandatory single-vendor choice. Neither announcement alone establishes current installed share or a completed customer purchase.

Vertiv VRTAnnounced 800 VDC architecture participant; no exclusive position or installed share claimed.Vertiv: 800 VDC platform collaboration: https://investors.vertiv.com/news/news-details/2025/From-Vision-to-Readiness-Vertiv-Collaborates-with-NVIDIA-to-Advance-800-VDC-Platform-Designs-to-Power-the-Next-Generation-of-AI-Factories/default.aspxconfirmed
Eaton ETNAnnounced 800 VDC architecture participant; no exclusive position or installed share claimed.Eaton: announced 800 VDC architecture: https://www.eaton.com/gb/en-gb/company/news-insights/news-releases/2025/eaton-unveils-next-generation-architecture.htmlconfirmed
Tier 2

Power conversion and final point-of-load regulation

Power must ultimately reach the low-voltage rails required by processors and supporting electronics. Higher upstream voltage changes the conversion problem but does not remove it. An architecture can retain intermediate stages, and the most economical implementation depends on efficiency, isolation, protection, packaging and control requirements.

This map does not assume an unverified future GPU wattage, core voltage or phase count. Nor does it claim that one supplier owns every GPU regulation socket. A valid bill-of-materials model needs a defined platform and qualified component list. Sampling a device is weaker evidence than accepted production shipments.

Controllers, power stages, passive components and complete modules represent different supplier roles. Integration can raise value per component while reducing the number of discrete parts. A growth model should account for that change rather than assume that rising rack power multiplies every supplier’s revenue in the same proportion.

Tier 3

Power semiconductors and enabling device technology

Power-conversion designs can use different semiconductor technologies at different stages. The selection depends on voltage stress, switching requirements, thermal behavior, reliability and cost. A material label such as GaN or SiC is not a complete architecture and does not establish a universal winner across the power path.

Infineon’s architecture collaboration provides manufacturer evidence of work on 800 V power delivery with NVIDIA. That directly contradicts a claim that only one other vendor spans the opportunity. It does not prove exclusive selection, installed market share or the current revenue generated by the collaboration.

A supplier map should distinguish device fabrication, module integration and complete power-system delivery. Corporate gross margins cannot be assigned to an undisclosed 800 VDC product segment. Current product qualification and realized revenue remain stronger commercial evidence than broad claims of process scarcity or an assumed mandatory attachment rate.

Infineon Technologies IFNNY · not in coveragePower-semiconductor supplier with a publicly announced NVIDIA 800 V architecture collaboration; no exclusive supply or current revenue claim.Infineon manufacturer announcement, May 20, 2025: https://www.infineon.com/press-release/2025/INFXX202505-107confirmed
Risks

Invalidation risks

Deployment timing differs from the vendor roadmap; planned products or facilities may be delayed, and a hybrid installation is not the same as a fully native architecture.

Higher distribution voltage reduces current but may add protection, conversion or integration costs; the complete customer return needs a site-specific model.

Additional qualified suppliers, customer integration or open specifications can reduce incumbent content and pricing. Reference-design participation does not guarantee exclusive revenue.

Facility power availability and construction can limit installation independently of semiconductor or rack-equipment readiness. Equipment orders need to be linked to funded project stages.

Reliability or safety problems can delay acceptance. Demonstrated component efficiency is insufficient if the integrated power path fails the customer’s operating requirements.

Sources

5 cited sources preserved from the vault map.

  1. nvidia.comNVIDIA: August 2026 hybrid and native architecture updateOpen source ↗
  2. nvidia.comNVIDIA: technical 800 VDC architectureOpen source ↗
  3. infineon.comInfineon: announced NVIDIA architecture collaborationOpen source ↗
  4. vertiv.comVertiv: 800 VDC platform collaborationOpen source ↗
  5. eaton.comEaton: announced 800 VDC architectureOpen source ↗
The weekly brief

Get a weekly digest of what’s changing.

Research updates, thesis changes and company developments. One email a week.