Cloudflare R2 jitters starve H100s

The sharp take: zero egress is not the same as training-ready storage. For GPU-heavy workloads, predictable latency and locality matter more than headline storage pricing.

• –The post highlights a common infra trap: object storage that is cheap on paper can still waste expensive GPU time if read latency is noisy.
• –Alternatives like Wasabi and Backblaze B2 can remove egress fees too, but they come with their own pricing or policy tradeoffs, so “free egress” is not a complete answer.
• –At 40TB scale, the likely winning pattern is regional replication plus a local cache tier, not direct-from-object-store streaming as the only data path.
• –This is less a Cloudflare-specific complaint than a reminder that AI training pipelines need storage engineered for throughput consistency, not just S3 compatibility.