The 100 MW Blueprint Rewiring AI Data Center Energy

The architecture is not a single product launch, it is a vendor-defined blueprint that may become the de facto standard operators are evaluated against

Decision Focus

In early 2026, Siemens and nVent Electric announced a joint reference architecture for large-scale AI data centers built on NVIDIA infrastructure. According to Data Centre News, the blueprint targets facilities up to 100 MW, supports Tier III-capable and modular deployment, and integrates Siemens electrical and automation systems with nVent’s rack-level liquid cooling technology. The operational signal for Global Heads of Data Center Energy: a significant portion of your supply chain, efficiency benchmarking, and facility design conversations are now being pre-shaped by a vendor-defined architecture — before you reach the procurement table.

90-Second Brief

In recent days, siemens and nVent have published a joint reference architecture targeting 100 MW hyperscale AI data centers using NVIDIA compute. The design integrates rack-level liquid cooling with industrial power and automation systems and centers on tokens-per-watt as its core efficiency metric. NVent reportedly recorded data center sales of approximately $1 billion in 2025 and carried a $2.3 billion backlog into 2026, while expanding manufacturing capacity in Minnesota specifically for liquid cooling production. The architecture is not a single product launch, it is a vendor-defined blueprint that may become the de facto standard operators are evaluated against.

What Is Really Happening?

The Siemens-nVent-NVIDIA partnership reflects a structural shift that has been building for several years. As AI compute density surpasses the thermal limits of air cooling — GPU clusters drawing 60 kW or more per rack cannot be efficiently managed with traditional CRAC units — the infrastructure stack is being redefined from the rack outward, and that constraint is now driving architectural standardization, not just vendor selection.

What is significant about this blueprint is that it collapses the typical design decision tree. Rather than operators independently selecting cooling vendors, electrical integrators, and automation platforms, a pre-integrated reference architecture is being positioned as the validated path for 100 MW AI deployments. This is the pattern that has emerged in other infrastructure markets where complexity and speed-to-deployment pressure eventually forced standardization — telecom switching, semiconductor fab tooling — and it tends to happen faster than operators expect.

The commercial signals support this read. nVent’s reported revenue growth and the near-threefold increase in backlog suggest hyperscalers are committing to specific vendors at significant scale and lead time. That level of forward commitment does not indicate customers browsing options; it indicates supply chain lock-in already underway. The new Minnesota manufacturing facility, reportedly coming online in early 2026, points to the same underlying pressure: liquid cooling component availability is tightening as demand accelerates ahead of available production capacity.

Why It Matters for Global Heads of Data Center Energy

The tokens-per-watt metric embedded in this architecture deserves close attention. Until now, energy efficiency in AI data centers has largely been reported through PUE — a facility-level overhead metric. Tokens-per-watt is a workload-specific measure: AI output extracted per unit of electricity consumed by the compute stack. When a vendor architecture is designed around that metric from the start, accountability shifts from the facility energy team toward the integrated infrastructure stack, and it changes what you must justify to your CFO and board when selecting infrastructure vendors.

For energy portfolios, cooling load is not a passive planning input. Estimates cited in the source material suggest cooling accounts for roughly 38–40% of total data center electricity consumption. A material shift toward rack-level liquid cooling across a significant share of a hyperscale fleet has a direct effect on total facility load, load shape, and the demand baseline underlying PPA sizing assumptions. If liquid cooling delivers meaningful efficiency improvement over air systems — as the source claims — then forward energy procurement models built on air-cooled baselines may be over-hedged on cooling load while simultaneously being under-hedged on raw compute power draw. Both errors compound budget and procurement risk in opposite directions.

The supply chain dimension carries the same urgency. If the Siemens-nVent blueprint becomes the preferred reference for hyperscale NVIDIA deployments, liquid cooling component lead times will tighten regardless of your own build schedule. This follows the same dynamic that appeared in large power transformer procurement: by the time scarcity is visible in the market, the queue is already years deep. Monitoring nVent’s reported backlog trajectory and the manufacturing ramp in Minnesota provides a leading indicator on this constraint.

Forward View

If architectural standardization continues along this path, three fronts merit active watching. First, whether competing infrastructure vendors — Vertiv, Eaton, Schneider Electric — respond with rival integrated reference architectures or instead seek certification within the NVIDIA partner ecosystem, which would signal the ecosystem approach is winning. Second, how tokens-per-watt propagates into colocation service agreements and energy SLAs; if hyperscale customers begin specifying minimum tokens-per-watt thresholds in lease negotiations, colo operators face new infrastructure obligations that flow back to energy procurement assumptions. Third, whether increasing liquid cooling efficiency materially changes near-term load growth projections for ERCOT, PJM, and Northern Virginia markets, where utilities and ISOs are currently planning capacity against air-cooled demand baselines.

What Is Still Uncertain

The primary source for financial figures is analyst commentary, not audited operator disclosure, so key variables remain unconfirmed. The performance delta between this reference architecture and air-cooled alternatives at 100 MW scale has not been independently validated at full deployment. The degree to which hyperscalers running AMD-based or custom-silicon compute will adopt this blueprint is unestablished — the architecture is explicitly built around NVIDIA DGX infrastructure. Projections for liquid cooling penetration beyond its currently reported sub-30% share carry the inherent uncertainty of vendor-aligned market forecasts. Whether nVent’s Minnesota production ramp is sufficient to meet the implied demand from its own reported backlog also remains an open question that could affect lead times across the supply chain.

One Question for Your Team

Does your forward energy procurement model — PPA sizing, load forecasting, utility capacity reservations — still assume an air-cooled demand profile, and if so, when did you last stress-test those assumptions against a liquid-cooled scenario at the scale you are planning to build?