According to Google DeepMind on X, Decoupled DiLoCo combines Pathways—an AI system that orchestrates heterogeneous chips at independent speeds—with DiLoCo, a bandwidth-minimizing distributed training approach, to enable scalable multi-datacenter training of large models (source: Google DeepMind, April 23, 2026). As reported by Google DeepMind, Pathways allows asynchronous coordination across diverse accelerators, while DiLoCo reduces cross-site communication, together improving efficiency and reliability for frontier model training at global scale. According to Google DeepMind, this architecture targets practical bottlenecks in interconnect bandwidth and straggler effects, creating business opportunities in cost-optimized LLM and multimodal model training, geographically resilient ML ops, and elastic capacity pooling across cloud regions.

According to Google DeepMind on X, its latest research details AI training that scales across geographies, capacities, and heterogeneous chips, removing locality and hardware lock‑in constraints. As reported by Google DeepMind’s research post linked in the tweet, the system coordinates distributed training over multiple data centers and mixed accelerators, using techniques such as elastic scheduling, topology‑aware communication, and fault‑tolerant aggregation to keep utilization high and costs predictable. According to Google DeepMind, this approach targets vendor‑agnostic training on GPUs and specialized accelerators, enabling enterprises to pool idle capacity, shorten time‑to‑train, and reduce queuing risk for large jobs. As noted by Google DeepMind, the business impact includes higher effective throughput, improved resilience to regional outages, and better price performance by matching jobs to the most cost‑efficient chips and regions.

According to Google DeepMind on X, the team successfully trained a 12B Google Gemma model across four US regions over low-bandwidth networks and demonstrated heterogeneous training across TPU6e and TPUv5p without performance regressions. As reported by Google DeepMind, this cross-region, low-bandwidth orchestration suggests large language model training can be decoupled from single datacenters, enabling cost-efficient multi-region capacity pooling, improved resiliency, and better utilization of stranded compute. According to Google DeepMind, the ability to mix TPU generations without slowdown opens procurement flexibility and reduces upgrade friction for enterprises planning phased hardware refreshes.

According to @krea_ai, the upcoming Infra Talks event in San Francisco will feature CTOs from Chroma (@HammadTime) and Krea (@asciidiego) discussing advanced AI GPU infrastructure topics, including distributed training, strategies to maximize GPU utilization, optimizing inference paths, and designing highly-concurrent systems for reinforcement learning rollouts. This event is targeted at professionals interested in AI infrastructure, systems engineering, and backend development, providing valuable insights into scaling AI workloads and building performant, low-latency AI platforms. Attendees can expect to learn practical solutions for managing GPU clusters, accelerating model inference, and supporting large-scale AI deployments (Source: @krea_ai, Twitter, Nov 14, 2025).