OCP Global Summit 2025 | Google | Agile AI Architectures: The Fungible Data Center for the AI Era
Introduction
At OCP Global Summit 2025, Google gave a talk titled "Agile AI Architectures: The Fungible Data Center for the AI Era". As AI enters the Intelligence Revolution, Google stressed the need to build future data centers with a design mindset of agility and fungibility. This is not only about compute and network performance, but a comprehensive overhaul spanning power, cooling, modularity and sustainability.
Content
AI's broad impact on society and science
Google noted that AI is not only driving consumer applications (such as Pixel AI features, YouTube Shorts and Gemini search summaries) but has also profoundly changed how businesses operate and how science is done. From finance and retail to pharmaceuticals, and on to major scientific breakthroughs such as AlphaFold and AlphaGenome, AI's reach spans every field.
System and TPU co-design
At the core of Google's AI development is the TPU (Tensor Processing Unit), now in its seventh generation. These chips, combined with liquid cooling, optically reconfigurable network topologies and a new power-delivery architecture, form what Google calls the AI Hypercomputer. The design philosophy is full-stack co-design from chip → system → platform → ecosystem, delivering 10–100× gains in performance and energy efficiency.
Explosive growth and design challenges
Over the past 12–18 months, AI accelerator usage has grown 15×, ML storage has grown 37×, and token traffic has reached the quadrillion scale per month. This staggering growth, together with rapid TPU/GPU iteration and the diversity of data-center types (hyperscale, neo-cloud, colo), means that power, cooling, networking and system design all face unprecedented challenges.
The fungible data center concept
Google and the OCP community proposed the fungible data center concept, whose core elements include:
Power: a standardized 400V architecture, a shift from monolithic to disaggregated power supplies (the Mount Diablo project), and the introduction of solid-state transformers and microgrids, so that data centers are not just power consumers but can also feed back to the grid.
Cooling: driving liquid-cooling standardization (Project Deschutes) and exploring technologies such as rear-door heat exchangers to ensure high resilience under high AI power loads.
Facility design: unifying aisle height, weight and fiber layout, and promoting telemetry and security standards for third-party data centers.
Security and sustainability: adopting the open-source Caliptra Root of Trust, introducing post-quantum cryptography, and promoting rigorous methods for measuring AI's carbon, water and energy footprint (for example, each Gemini query uses less than five drops of water).
AI for AI: using AI to design systems
Google went further, proposing a vision of AI for AI: using AI to design AI systems themselves. For example, AlphaChip applies AI to IC layout, shortening design cycles and improving PPA (power, performance, area). In the future, AI could be applied broadly across system design flows for hardware, software, firmware, manufacturing and test.
Conclusion
In this OCP talk, Google clearly laid out the challenges and direction for future AI data centers. The fungible data center is not just about standardization; it is a comprehensive overhaul across power, cooling, networking, sustainability and security. At the same time, AI itself will become a new tool for system design, driving the next wave of massive performance gains.
Further perspectives
Technical impact
Google's fungible data center and 400V power architecture will drive further maturity in power distribution, liquid cooling and silicon photonics interconnect. As AI workloads keep multiplying, optical interconnect and high-voltage power delivery will become central to design.
Supply-chain observations
Liquid-cooling standardization (Project Deschutes) is a major opportunity for cooling-system vendors such as Vertiv and Schneider Electric.
Demand for disaggregated power and solid-state transformers may also drive market growth for power management ICs (PMICs) and power devices (SiC/GaN).
Market trends
The AI-for-AI design philosophy could evolve into a fusion of EDA and cloud AI, challenging the dominance of incumbents Synopsys and Cadence.
For tier-2 clouds and enterprise data centers beyond the hyperscalers, standardization of the fungible data center will lower the barrier to build-out and accelerate the spread of AI.
Comments