Nvidia's AI Factories Are Rewriting the Power Stack and the Trade With It
The 800-volt power architecture mandate creates a backbone supplier the Street has entirely wrong.
9 Ventures · Equity Research June 2026 | AI Infrastructure & Power · Thematic Series
The electrical infrastructure stack for AI data centers is being rebuilt from the ground up. Not the cooling. Not the networking. The wiring itself. For decades, large-scale computing facilities have run on alternating current distribution, typically at 415 VAC in Europe and 480 VAC in North America, with power cascading through multiple conversion stages from the utility meter to the server rack. Each stage is a loss event. Each transformer bleeds efficiency, generates heat, and adds latency to power delivery. That architecture worked because it had to, and because rack power demands were manageable enough that the inefficiency was tolerable. That tolerable threshold has now been crossed. The Vera Rubin generation of compute racks runs at up to 370 kilowatts per rack, against roughly 120 kilowatts two years ago on Hopper. At that power density, high-voltage direct current stops being an engineering preference and becomes a hard physics requirement. Running 800 VDC distribution means that at equivalent power delivery, current drops proportionally, conductors get smaller, resistive losses in the distribution network collapse, and PSUs operating on DC input run materially more efficiently than the same hardware on AC input. The industry is eliminating conversion stages it tolerated for a century because it can no longer afford them.
Nvidia formalized this transition at Computex 2026. In a joint architecture roadmap developed with Siemens, Nvidia published the blueprint for what they are calling next-generation AI factories: 800 VDC distribution fed through a medium-voltage rectifier or solid-state transformer, replacing the old chain of step-down transformer, main switchboard, AC UPS, and power distribution unit that has been standard for fifty years. The document is explicit. The 2025-era data center diagram shows battery storage with a dotted line, labeled optional. The 2027-era AI factory diagram shows battery storage as a solid block, labeled in the same weight as the medium-voltage rectifier and the 800V DC bus. No dotted line. No optional. Battery Energy Storage is now a standard, mandatory component of the reference architecture that Nvidia and Siemens are putting in front of every hyperscaler, neocloud, and sovereign AI program building to this specification. When Nvidia draws a mandatory box in an official blueprint, the procurement cycle for that box begins immediately.
The reason BESS is mandatory rather than optional in this architecture comes down to three engineering problems that all surface simultaneously as rack power crosses the 100+ kilowatt threshold. First is grid interconnection. The US interconnection queue across major ISOs and RTOs now stretches five to ten years in most territories, a direct product of the volume of large load connection requests filed by data centers, EV charging networks, and industrial electrification projects. A developer who needs two gigawatts of capacity to run a full AI campus can contract for 500 megawatts of grid connection today and deploy battery storage to buffer peak compute demand within those limits, effectively compressing a decade-long infrastructure bottleneck into an engineering workaround. Speed to power is the single most commercially urgent problem in the hyperscaler buildout right now, and BESS is the only economically viable solution at scale. Second is power quality. AI workloads generate demand fluctuations at millisecond timescales that passive grid infrastructure cannot absorb. When a GPU cluster cycles through a training run or inference burst, the instantaneous power delta destabilizes voltage on the 800V DC bus, tripping server PSUs across the cluster. BESS systems with sub-10 millisecond response capability inject or absorb power fast enough to maintain bus stability. Traditional AC UPS systems using lead-acid or VRLA chemistry operate at response times an order of magnitude too slow for this application. Third is backup power. Lithium-based BESS has reached total cost of ownership parity with diesel generator backup at multi-hour durations, while eliminating refueling logistics, emissions liability, and the maintenance overhead of large rotating machinery. For facilities drawing 8 megawatts per compute cluster, the operational and ESG case for replacing diesel backup with battery backup has closed completely.
The market that emerges from this transition is large and accelerating. Global BESS market projections from Wood Mackenzie, BloombergNEF, and S&P Capital IQ converge on $105-150 billion by 2030, built primarily on utility-scale and front-of-meter storage demand. Those models were constructed before the Hyperscaler data center demand vector was visible in the order intake data of the leading BESS manufacturers. US installed battery storage capacity is projected to exceed 170 gigawatts by 2030, from 2.4 gigawatts in 2020. The Siemens/Nvidia 800V reference architecture is now the certified specification for sovereign AI programs in the Middle East, Southeast Asia, and Europe, markets where national AI infrastructure programs are deploying capital at speed and have neither the time nor the institutional knowledge to engineer power architecture from scratch. The 800V future data centers’ procurement cycles are active today.