The economics of Tigris object storage for AI training
Comparing Tigris, Cloudflare R2, and AWS S3 reveals significant cost differences for AI workloads, such as a 10 TB storage and 50 TB egress task costing only $154 on R2 versus $4,630 on S3. Tigris offers superior performance for small objects with sub-8 ms p90 read latency.
Egress savings drive the decision
Tigris and Cloudflare R2 eliminate egress fees. AWS S3 charges $0.09 per GB to move data out. For an AI training workload that requires 10 TB of storage and 50 TB of egress, AWS S3 costs $4,630 while Cloudflare R2 costs only $154 per month. Tigris storage costs $0.02 per GB. R2 storage costs $0.015 per GB. R2 provides 10 GB of free storage, 1 million Class A operations, and 10 million Class B operations monthly. If you handle massive AI datasets, look at these specific rates. For a user with 100,000 files of 100 KB each and 10,000,000 reads per day, R2 costs $104.40. For a user with 1,000 objects of 1 GB each and 1,000 reads per month, R2 costs $14.85 in Standard storage. If the user uses Infrequent Access for those same 1,000 objects, the cost rises to $29.90 because R2 charges $0.01 per GB for data retrieval. R2 Infrequent Access also costs $0.90 per million Class B operations. For a 1 TB stored and 10 TB egress workload, R2 costs $15 whereas S3 costs $936.
R2 lacks versioning.
| Component | Tigris | Cloudflare R2 | AWS S3 |
|---|---|---|---|
| Storage (per GB) | $0.02 | $0.015 | $0.023 |
| Egress (per GB) | $0.00 | $0.00 | $0.09 |
| Class A Ops (per 1M) | $5.00 | $4.50 | $5.00 |
| Class B Ops (per 1M) | $0.50 | $0.36 | $0.40 |
Performance advantage for small objects
Tigris provides superior performance for small object workloads. In benchmarks using 10 million 1 KB objects, Tigris finishes ingestion in 6711 seconds. This is 31 percent faster than S3 and an order of magnitude faster than R2, which takes 72063 seconds. Tigris sustains 3.3k operations per second, which is 20x the 170 operations per second from R2 and nearly 4x the 892 operations per second from S3. Tigris keeps p90 read latency below 8 ms. R2 p90 read latency stretches beyond 199 ms. Tigris delivers 828 write operations per second, while R2 delivers only 43 and S3 delivers 224. These gaps exist because Tigris inlines small objects in metadata and uses an LSM-backed cache. The write-side tail latency for Tigris is less than 17 ms, while R2 exceeds 680 ms and S3 sits at 41 ms.
Tigris delivers speed.
| Metric | Tigris | Cloudflare R2 | AWS S3 |
|---|---|---|---|
| Max Object Size | 5 TB | 5 TB | 5 TB |
| Read Latency (p90) | < 8 ms | 199 ms | 42 ms |
| Write Latency (p90) | < 17 ms | 340 ms | 38 ms |
Managing AI datasets and replication
Tigris provides instant, zero-copy bucket forks for parallel experiments and dataset versioning. R2 does not have this feature. You can move data from R2 to Tigris without downtime using shadow buckets. Tigris fetches objects from R2 on demand and caches them. However, users report replication lag in Tigris that reaches four hours. This lag causes files to remain missing from certain geographic locations. Some users also encounter 500 internal errors or invalid argument errors during uploads. Replication issues in US regions have forced teams to migrate whole production infrastructures to different locations. Tigris guarantees strong consistency within each region and eventual consistency across all regions with sub-second replication lag. R2 lacks object lock and versioning. R2 users often use Sippy or Super Slurper to manage migrations.
Replication lag occurs.
You should evaluate if your workflow requires native Cloudflare Workers bindings for low latency or if you prefer a multi-cloud approach that is not tied to the Cloudflare ecosystem.
Does replication lag affect production stability?
Technewsdaily
Senior tech writer covering AI, gadgets and cybersecurity. Breaking down the news that matters, every day.
