How Colos and Hyperscalers Co-Evolve

Author: Varun Sakulkar, Overall Tech Lead for Google Data Centers; Scott Sharp, Mechanical Engineer, Google Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Google presents a call to action for deeper technical co-design between hyperscalers and colocation providers, moving beyond the current "commercial co-design" of wholesale data centers to address the "Cambrian explosion" of hardware heterogeneity in AI infrastructure. The presentation identifies two core challenges: demand-supply mismatch (dynamic AI demand vs. long-lead hardware) and combinatorial explosion in design complexity from rapidly iterating TPU and GPU generations with constantly changing power, cooling, and network requirements deployed globally.

The solution proposed centers on designing wholesale data centers with agility and fungibility as first principles across three domains. For server halls, this means creating families of layouts with documented variances rather than single-purpose designs. For power, it requires fungible power delivery systems compatible with Mount Diablo (+/-400V DC distribution per OCP), with a call to eventually integrate the Mount Diablo rack into base electrical infrastructure to free floor rack positions. Central energy storage at medium voltage is proposed to buffer power fluctuations between AI clusters and utilities, with bidirectional signaling to utility providers. For cooling, standardization of technical cooling loop specifications (PG-25 propylene glycol vs. deionized water with biocide), flow rates, temperatures, and rack cooling interfaces is urgently needed. The real-world air-to-liquid cooling ratio across the aggregate fabric will differ significantly from the 80/20 chip-level ratio due to air-cooled networking gear and ancillary equipment. Google issues a formal open letter to the industry for collaboration, recognizing that the historical approach of planning for individual racks no longer works at gigawatt campus scale.