NVIDIA, Google, and Microsoft launch 800 VDC power architecture for AI data centers

NVIDIA, Google, and Microsoft have jointly developed an 800-volt direct-current (VDC) power architecture for AI data centers, aiming to remove power-delivery bottlenecks that are increasingly limiting how much compute can be packed into a facility, NVIDIA said in an August 11 blog post. The specification was built through the Open Compute Project (OCP), and more than 80 equipment makers and infrastructure companies are already building products to it.

What's new

NVIDIA's announcement lays out the shift from traditional AC power distribution to high-voltage DC: "NVIDIA, Google and Microsoft have been developing the 800 VDC architecture together through the Open Compute Project (OCP)." Distributing power as 800 VDC instead of AC cuts the number of conversion stages between the grid and the GPU, reducing energy loss along the way.

Key specifics from the announcement:

• The architecture supports up to 2 megawatts per row, with initial availability expected in 2027 for row power centers.
• An MGX-compatible 800 VDC power rack is planned for the second half of 2026, ahead of the row-level rollout.
• Three deployment options are offered: a power rack (compatible with hybrid AC/DC setups), a centralized row power center, and a facility-scale DC power block.
• Vladimir Troy, NVIDIA's vice president of data center infrastructure, said: "800 VDC unlocks the compute performance and power density required for AI at scale."

Context

Power delivery, not just chip supply, has become one of the binding constraints on scaling AI data centers — racks packed with next-generation GPUs draw far more power than legacy AC infrastructure was designed to distribute efficiently, and every AC-to-DC conversion step along the way wastes energy as heat. NVIDIA has been pushing DC power architectures for its GPU racks for several release cycles, and formalizing the approach through OCP — the open-hardware standards body originally spun out of Facebook — signals an attempt to get the wider data center supply chain, not just NVIDIA's own hardware, aligned on a common standard rather than proprietary designs.

The involvement of Google and Microsoft, two of the largest hyperscale data center operators, is notable: both companies operate power infrastructure at a scale that makes standardized components economically meaningful, and their participation increases the odds that other server and equipment vendors converge on the same 800 VDC spec rather than fragmenting into competing designs.

Why it matters

Data center power constraints are increasingly cited as a limiting factor on AI capacity growth — sites are being sited and built around available grid power as much as around chip availability. An industry-standard 800 VDC architecture, backed by three of the largest players in cloud infrastructure and over 80 supporting vendors, could meaningfully raise the ceiling on how much compute a single facility can support, and reduce the energy overhead of running it. With Wood Mackenzie projecting $9 trillion in global AI and data infrastructure investment through 2040, per NVIDIA's post, standardizing the electrical backbone that all of that hardware plugs into is a substantial, if unglamorous, piece of AI infrastructure history — the kind of change that shapes build costs and timelines for years after the announcement fades from headlines.