Google Cloud strengthens its AI infrastructure with TPU Ironwood and Cloud WAN
Launch of Google Cloud Next 2025 in Las Vegas
The Google Cloud Next 2025 event kicked off in Las Vegas from April 9 to 11, drawing nearly 32,000 attendees for this major gathering. Sundar Pichai, CEO of Alphabet, opened the conference by emphasizing the growing importance of high-performance, low-latency infrastructure to support advancements in artificial intelligence.
In this context, he announced massive investments of $75 billion in the development of new data centers, aiming to counter competing initiatives such as Stargate (led by OpenAI, Oracle, and SoftBank) and to meet the increasing demand for AI computing power.
Ironwood: The New Generation of TPUs for Inference
Google Cloud has introduced Ironwood, the seventh generation of its Tensor Processing Units (TPUs), specifically designed for inference tasks—operations that generate outputs from pre-trained AI models. This chip is the successor to the Trillium model, released in 2024.
Key features of Ironwood:
Available in pods ranging from 256 to 9,126 chips
Up to 45.2 exaflops of compute power
Each chip delivers 4.16 teraflops
Up to 192 GB of HBM memory per pod, with a bandwidth of 7.2 TB/s
Enhanced chip-to-chip communication using ICI (Inter-Chip Interconnect) at 1.2 TB/s
Ironwood also integrates SparseCore, a specialized accelerator for large-scale embeddings, along with Pathway, a TPU-distributed execution environment developed by DeepMind.
Enhanced AI Instances with Nvidia Blackwell Chips
Although Google is highlighting its own chips, Nvidia remains a key strategic partner. Google Cloud now supports:
A4 instances powered by Blackwell B200 accelerators
New A4X instances (currently in preview), built on the GB200 NVL72 architecture, which combines Grace CPUs (Arm-based) with Blackwell GPUs
The infrastructure also benefits from Cluster Director, a tool designed to manage clusters of AI accelerators as a single, optimized unit.
Additionally, Google Cloud has announced it will be the first provider to offer Nvidia’s Rubin GPUs, which were recently unveiled at GTC.
Cloud WAN: A High-Performance Global Network for AI
To support its AI offerings, Google introduces Cloud WAN, a networking solution built on Google’s global infrastructure:
Over 2 million kilometers of fiber optic cable
33 submarine cables
202 points of presence worldwide
Cloud WAN is designed for enterprises with inter-regional or multi-site connectivity needs. It offers:
Cloud Interconnect: connects enterprise data centers directly to Google Cloud
Cross-Cloud Interconnect: enables multicloud setups with direct links to other cloud providers
Cross-Site Interconnect (in preview): provides private, point-to-point connections at speeds of up to 100 Gbps, enabling fast data exchange between different locations
Conclusion: A Bold AI Strategy in the Face of Competition
With the introduction of Ironwood, the expansion of Nvidia-powered offerings, and the strengthening of its network through Cloud WAN, Google Cloud clearly demonstrates its ambition to lead the global AI infrastructure landscape, while gearing up for fierce competition in a rapidly growing sector.
source of the news: LeMondeInformatique
