BACKBLAZEFIELD GUIDE · BY NIX UNOFFICIAL
SecurityMigrationPartnersFAQGlossaryQuizAboutUnofficial. Not affiliated with Backblaze, Inc.
LEARN / FILE 07

07. Overdrive, B2 Neo, B2 Reserve, MCP, and AI tooling

GPUs are the expensive part. Storage’s job is to never be the reason they wait.

B2 Overdrive

What it is (backblaze.com/cloud-storage/b2-overdrive, checked 2026-10-14):

  • Built on B2, with dedicated storage vaults and networking walled off from shared infrastructure.
  • Throughput tiers from 100 Gbps up to 1 Tbps.
  • Starts at $15/TB/month. Multi-petabyte commitment required.
  • Unlimited free egress to any destination, including other clouds. Zero transaction fees.
  • Private networking integration, white-glove onboarding, “PETA” support tier (highest offered).
  • Enabled by Sales, including for testing.

When to lead with it:

  • Read multiple well above 3x stored per month (repeated epochs from a remote GPU cloud).
  • Aggregate throughput needs in the tens of GB/s.
  • Training spread across multiple GPU providers.

Break-even vs standard B2 (list prices): Standard B2 monthly = 6.95·S + 10·max(E − 3S, 0) where S = stored TB, E = egress TB, $10/TB overage. Overdrive monthly = 15·S. Set equal: 6.95S + 10(E − 3S) = 15S gives E = 3.805·S. Above about 3.8x monthly read multiple to non-partner destinations, Overdrive is cheaper on list price alone, before counting the dedicated throughput. Below that, standard B2 is cheaper unless throughput is the constraint.

Proof points:

  • Mirage (formerly listed as Captions on the case-study URL): “0 GPUs waiting for data”, “95% estimated cost savings”, “$0 egress”, H100 clusters, after a competitive bakeoff.
  • Decart: 0 to 16 PB in 90 days; “10x more efficient than competition”; “$0 egress spend”; “the only technical thing we used in training these models that didn’t crash the first time.”

B2 Neo (B2 for Neoclouds)

What it is (backblaze.com/cloud-storage/b2-neo, checked 2026-10-14):

  • White-label object storage for GPU-as-a-Service providers. Launched Feb 2026.
  • Partner API to provision accounts, usage, reporting, billing from the neocloud’s own portal.
  • Custom domains so storage looks native.
  • SSO integration with the operator’s identity provider.
  • Up to 1 Tbps throughput claim (meta description), SOC 2.
  • “No egress fees” and “less operational overhead than building it yourself.”
  • Backblaze white paper: “Why Neoclouds Should Not Build Their Own Object Storage” covering the split-stack problem (customer data in a hyperscaler, compute on the neocloud, egress friction) and the cost of running Ceph or custom hardware.

The pitch to a neocloud: “Every customer who keeps data in a hyperscaler pays an egress tax to use your GPUs. That tax is your churn risk. B2 Neo puts S3 storage under your brand in weeks, and you never hire a storage ops team.”

CoreWeave is the reference pattern, but be precise: the SEC 8-K does not label CoreWeave as a B2 Neo customer. Say “CoreWeave uses Backblaze to power HDD-based tiers in CoreWeave AI Object Storage”, not “CoreWeave runs on B2 Neo.”

Discovery for neoclouds:

  1. What percent of your customers’ training data lives outside your platform?
  2. Do you offer object storage today? Built on what?
  3. How many storage engineers would you need to run it at 10 PB? At 100 PB?
  4. What do customers ask for: throughput, price, or just “S3 next to my GPUs”?
  5. Billing: do you need per-customer usage feeds?

B2 Reserve and capacity pricing

B2 Reserve was Backblaze’s prepaid annual capacity offering sold through channel partners (TB bundles, included egress and premium support). As of 2026-10-14 the old /cloud-storage/b2-reserve URL resolves to the main B2 page, and the pricing page describes “Pay as you go monthly or for capacity annually” with 1 to 5 year terms. The API still has b2-reserve-trial-create-account in docs. Treat “B2 Reserve” as a legacy or channel SKU name and confirm the current name and terms with Backblaze deal desk before quoting. [unverified current status]

B2 MCP Server (Sep 8 2026)

What it is: open source Model Context Protocol server that lets AI agents (Claude, Gemini CLI, Cursor-style clients, etc.) operate B2 through a defined tool set.

Facts from docs (checked 2026-10-14):

  • 40 tools in three groups: 17 B2 Native (buckets, notification rules, app keys, retention and legal holds, Partner API group/account ops), 19 S3-compatible (objects, metadata, multipart, presigned URLs, lifecycle), 4 analytics (storage growth, egress leaders, largest files, unfinished multipart uploads).
  • Profiles: full (40 tools) and read-only (20 tools).
  • Local mode runs as a subprocess over STDIO; the app key stays in the server environment and is not passed to the model.
  • Tools exposed are gated by the app key’s capabilities. A read-only, single-bucket key yields a smaller tool set. The server does not invent a new permission model.
  • Read-only resources: b2://server-config, b2://capabilities, b2://bucket/{bucketName} (webhook secrets redacted).
  • Optional prompts via B2_ENABLE_MCP_PROMPTS=true.
  • Self-hosted HTTP deployments have configurable credential modes.

SE angle: this is the safest way to let agents touch storage, because the blast radius equals the key. Demo it with a read-only key first, then a bucket-scoped write key, and show the tool list shrink and grow.

Other AI integrations (backblaze.com/cloud-storage/b2-ai-integrations)

  • Artifact store for model weights and data files (MLflow-style guide).
  • Pixeltable: declarative multimodal dataset management.
  • CVAT: source raw images/video and retain labeled datasets.
  • Vibe Coding Starter Kit, a Claude skill for B2 (Agent Skills spec), OpenClaw state backups.
  • Startup program Flamethrower: up to $100,000 in storage credits (“Unlock 1PB free storage for startups”).
  • Caveat from the page: integrations are not covered by the SLA.

Positioning in the AI stack (draw this)

[ Data sources ] → [ B2: capacity tier, system of record ] ⇄ [ Hot tier: VAST / WEKA / Lustre / local NVMe ] → [ GPUs ]
                                ↘ partner egress $0 → CoreWeave, Vultr, Equinix Metal
                                ↘ Overdrive: dedicated 100G to 1T, $0 egress anywhere

Flash for the working set, Backblaze for everything else. Hume AI uses WEKA with B2 (case study integration tag). Never pitch B2 as a replacement for a parallel file system feeding GPUs at line rate; pitch it as the place the full corpus lives and the source the hot tier fills from.