1 · Where does your training data live today?
Which storage tier? (auto-fills Hot rate; pick Custom to enter a contracted rate)
2 · Your workload
5 PB · 64 GPUs · hourly ckpts · $23/TB·mo hot
MLPerf anchors (default = 70B): ···
⚙ Advanced assumptions — click to expose every modeling knob (concurrency, utilization, I/O wait inputs, displacement %, per-GPU bandwidth). Defaults are conservative.
modeled — measure actual training-time fraction
mode default — measure yours via
nvidia-smi dmon / NCCLmodeled outcome — pilot to validate
These defaults are conservative. If your actuals differ — measure first, dial in here — the math updates live and the picker re-snaps to the new best-Year-1 tier.
B2 throughput tier
—
Standard fits most sub-3 PB workloads. Larger active-training datasets (5+ PB) benefit from a higher tier to keep GPUs fed — Section 4 shows the per-tier math after the run.
3 · Watch the math build, validated by real B2 operations
Click to start. Hero tiles fill in as each stage finishes.
skip the demo · just populate the math
Projected savings — fill in as the analysis runs
GPU $ reclaimed / mo
—
pending Stages 1+2
—
Hot-storage $ avoided / mo
—
pending Stages 3+4+5
—
Migration egress covered (one-time) ★
—
Backblaze-covered switching incentive
—
Year-1 customer value · incl. migration coverage
—
recurring × 12 + Backblaze-covered migration egress
—
Live narration · math + real B2 operations
Click "Calculate my savings" above. The math builds line by line. Each tile fills in as the relevant stage finishes.
Pipeline stages
B2 console ↗
1
Data Lake
Pack 50k samples → 1 shard
queued
✓ What B2 does here
- Stores raw, processed, and versioned datasets in durable object storage.
- Lowers the cost of keeping large training datasets and historical versions.
- Gives training jobs a reliable place to read large shards, manifests, and dataset files.
- Keeps data portable across GPU clouds, regions, and compute environments.
✗ What B2 doesn't do here
- Replace fast local storage for active training — training reads come from a pre-staged NVMe copy; B2 is the durable source, not the working set.
2
Training Prep
Parallel range reads → NVMe
queued
✓ What B2 does here
- Provides the storage source for reading or pre-staging training data to local NVMe/cache.
- Supports parallel reads when shard size, network path, and client concurrency are tuned.
- Keeps durable data separate from short-lived GPU nodes and training clusters.
✗ What B2 doesn't do here
- Speed up CPU-bound work like tokenization, decompression, or data augmentation.
- Guarantee low I/O wait without testing your real workload.
3
Checkpoints
Train + mid-run ckpt → B2
queued
✓ What B2 does here
- Stores checkpoints off-cluster so recovery state survives node, disk, or cluster failure.
- Gives training runs a durable recovery target outside the GPU environment.
- Makes it affordable to keep more recovery points and milestone checkpoints.
- Supports checkpoint retention policies instead of forcing teams to keep only the latest copy.
✗ What B2 doesn't do here
- Make training itself faster or remove checkpoint serialization/GPU sync time.
4
Model Registry
Versioned bundle → B2
queued
✓ What B2 does here
- Stores versioned model artifacts: weights, tokenizer, config, eval files, and manifests.
- Provides durable artifact paths for deployment, rollback, and distribution.
- Makes it practical to keep previous model versions available for rollback.
✗ What B2 doesn't do here
- Replace MLflow, W&B, Comet, approval workflows, or model evaluation.
5
Inference
Cold-start: B2 → generate
queued
✓ What B2 does here
- Acts as the source of truth for model files used by inference fleets.
- Lets new or replacement nodes download and cache model artifacts before serving traffic.
- Helps keep multiple active, previous, and rollback model versions available on demand.
- Supports KV cache workflows by storing reusable prompt/context assets or cache warm-up artifacts that serving nodes can load locally, when the inference stack supports it.
✗ What B2 doesn't do here
- Serve live inference requests or reduce per-token latency after the model is loaded.
- Act as the live KV cache — per-request KV state belongs in GPU/CPU memory or the serving runtime's local cache.
Click any stage above to see what B2 helps with — and what it doesn't. Real B2 operations execute against your bucket; click "B2 console ↗" at top to verify objects landed.
Real B2 cost (bytes moved this session) $0.00
Stored 0 B
Egressed 0 B
PUTs 0
GETs 0
Estimate only — not a quote. Figures shown here are projections derived from the inputs you've entered, including estimated storage allocation percentages across workloads, and Backblaze list pricing as of June 1, 2026. Actual results may vary based on your specific workload, contracted rates, migration scope, storage distribution, retention policies, access patterns, and implementation. This dashboard is for discussion purposes only and does not constitute a sales quote, pricing commitment, service-level agreement, or warranty. Year-1 customer value is shown gross of customer-paid migration project costs, including engineering time, validation, dual-running, security review, and retraining. Final pricing and terms require a signed agreement.