Google Cloud: Storage is the hidden brake on enterprise AI

Storage delays are burning GPU budgets

Storage is no longer a back-office chore—it now dictates whether expensive GPU clusters sit idle or churn out models. Ronen Schwartz, who runs product for Google Cloud Storage, told SiliconANGLE that enterprises only notice the bottleneck after GPUs underperform and training timelines slip.

Schwartz put it bluntly: every second an accelerator waits for data is cash burned. That reality is pushing storage from a procurement checkbox to a frontline performance lever.

Training, inference and agents need different storage speeds

AI workloads split into three distinct patterns—training, inference, and agent tasks—and each demands its own storage playbook. Training clusters need rapid checkpoint snapshots every 15 minutes to avoid freezing million-dollar GPU farms. Inference tasks, meanwhile, live or die on single-digit millisecond latency. Agents, on the other hand, must sift through petabytes of images, PDFs and videos that often lack any metadata at all.

Traditional storage stacks weren’t built for any of these. That mismatch leaves accelerators twiddling their thumbs while data crawls across the bus.

Rapid storage cuts checkpoint times 3–5×

Google’s Rapid object storage tackles the training pain point first: checkpoint writes finish three times faster, and restores clock in at five times the speed of legacy object storage. On managed clusters, Google delivers 10 TB/s; for Rapid object storage, the number jumps to 15 TB/s.

The secret sauce is twofold. First, zonal colocation places data physically next to GPU and TPU clusters, slashing travel time for every byte. Second, Rapid Bucket technology handles 20 million queries per second at sub-millisecond latency—without forcing enterprises to rebuild their entire pipeline.

Nine of the top ten AI labs rely on Google storage

Google’s internal AI workloads have already proven the model. That experience gives the company’s enterprise pitch credibility: nine of the ten top AI frontier model companies now run their storage on Google Cloud.

For startups and enterprises alike, the math is straightforward—faster storage translates directly into shorter training cycles and lower effective costs per run.

AI-powered metadata turns messy archives into agent fuel

Agents can’t work without clean metadata. Google’s Smart Storage Initiative tackles that problem by using AI models to tag images, videos and PDFs automatically. The result is a lakehouse that agents can query instantly, even if the underlying files live outside Google Cloud.

Torin, an autonomous-vehicle startup, is already using the tooling to feed unstructured sensor logs into agent workflows without manual scrubbing.

Storage push mirrors Agentic Data Cloud updates

Last month’s Agentic Data Cloud update set the stage for this storage push. Both moves target the same problem: enterprises stalling because their data layers weren’t designed for AI.

Taken together, the two releases show Google weaving storage, data accessibility and compute into a single AI-optimized stack.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Startups that treat storage as an afterthought risk tying their most expensive compute assets to slow data pipelines—locking in higher costs and slower iteration loops for years.

Key takeaways

• GPU idle time from slow storage adds thousands of dollars to every training run.
• Training demands fast snapshots, inference needs low latency, and agents need instant metadata.
• Zonal colocation and Rapid Buckets now set the ceiling for AI scale.
• Automated metadata tagging lets startups turn messy archives into ready-to-use agent datasets without manual prep.

Related on Lazyfounder

• Google LLC
• Google Cloud
• More ai news
• Google Cloud updates Agentic Data Cloud with borderless Lakehouse and autonomous AI tools