Per-GB Storage Pricing: Examples & Companies

What is it

Per-GB storage pricing is a billing unit where customers are charged per gigabyte of data stored on the platform per month.

Storage is the persistent-state dimension of cloud pricing. Where compute meters time and APIs meter requests, the storage meter captures data at rest — the bytes that survive between sessions and keep occupying disk whether or not anything is actively running against them. The unit is almost always expressed as a rate per gigabyte over a billing window, most commonly a GB-month, though marketplace vendors like Vast.ai bill it per GB-hour for finer-grained, continuous accounting.

The model shows up in two distinct places in the corpus. The first is cloud and GPU infrastructure, where storage is a secondary meter layered on top of per-second or per-hour compute: RunPod, Modal, E2B, CoreWeave, Nebius, and Lightning AI all bill persistent disk separately from the machines that use it. The second is data and retrieval platforms — vector databases and serverless search — where the stored index is the product. Pinecone, Weaviate, Qdrant, Chroma, Upstash, and turbopuffer all meter stored data per GB because the embedding corpus a customer holds is the asset they pay to keep online.

The defining characteristic of per-GB storage pricing is that the rate is low but compounds. A few cents per gigabyte feels negligible at signup, but storage never sleeps: unlike a paused container or an idle endpoint, retained data keeps accruing charges for as long as it exists. That asymmetry — cheap per unit, relentless over time — is what makes storage the line item buyers most often underestimate. And the range is wider than any other billing unit in the corpus: object-storage tiers like Nebius ($0.0147/GiB-month) sit more than three orders of magnitude below enterprise SaaS add-ons like Kustomer ($50/GB-month), so the same nominal unit can mean radically different economics depending on the product it wraps.

How it works

The core formula is simple: a vendor measures how much data a customer is keeping, multiplies it by a per-GB rate, and prorates it over the billing window. The variation lives in three dimensions — the time basis, the storage class, and the free allowance.

The headline rates cluster into recognizable bands. Object-storage-backed tiers are the cheapest: Nebius Object Storage Standard is $0.0147/GiB-month and Zilliz Cloud cut managed storage to $0.04/GB-month in late 2025 (an 87% drop from $0.30). GPU block and network disk runs a notch higher — RunPod prices Network Storage at $0.05–$0.14/GB-month, Modal persistent volumes at $0.09/GiB-month, and CoreWeave object storage from $0.06/GB-month Hot. Database and vector storage sits higher still: Upstash charges a flat $0.25/GB, Pinecone is ~$0.33/GB-month on Standard, and Chroma is $0.33/GiB-month. At the top, enterprise SaaS folds storage into premium add-ons — Kustomer bills additional data storage at $50/GB/month.

Unit math: Monthly storage bill = (stored_GB − free_allowance_GB) × rate_per_GB_month. On Upstash Redis, keeping 50 GB costs (50 − 1) × $0.25 = $12.25/mo on top of per-command charges. The same 50 GB on Nebius object storage is roughly 50 × $0.0147 = $0.74/mo — the storage class matters far more than the volume.

The subtlety is when the meter runs. GB-month vendors snapshot or average your volume across the month. Vast.ai’s per-second model is stricter: storage accrues for every second an instance exists, including while it is stopped, so the only way to stop paying is to delete the instance. CoreWeave has the same property — object and file storage “accrues whether or not compute runs.” Upstash softens this with a daily-average basis, so a brief spike in stored data doesn’t set a high-water mark for the whole period. For buyers, the time basis matters as much as the headline rate — a “cheaper” GB-hour rate that bills paused resources can cost more than a higher GB-month rate that doesn’t.

Companies using this

Twenty-nine companies in the corpus meter storage per GB, spanning GPU and sandbox infrastructure (RunPod, Modal, Vast.ai, E2B, CoreWeave, Nebius, Lightning AI), vector and retrieval databases (Pinecone, Weaviate, Qdrant, Chroma, Milvus/Zilliz, LanceDB, Upstash, turbopuffer, Voyage AI), LLM observability and eval platforms (Helicone, Braintrust, Comet, Weights & Biases, LangChain, Vellum), and application platforms that fold storage into credits, compute units, or SaaS add-ons (Apify, Dify, Genspark, Vectara, GitLab, Krisp, Kustomer). The table below lists all of them with their pricing model and other billing units.

Patterns observed

Across the twenty-nine companies, per-GB storage almost never stands alone — it is the second, third, or fourth meter in a stack, and how it’s positioned reveals the vendor’s product shape.

• Storage is a secondary meter at compute vendors. For RunPod, Modal, E2B, CoreWeave, Nebius, Lightning AI, and Vast.ai, the headline product is per-second or per-hour GPU/CPU time; storage is the line item that captures the state customers want to keep between runs. E2B’s pricing history is telling — its per-vCPU compute rates and $150/mo Pro fee stayed fixed for the entire tracked range, but RAM and storage were added as separate meters ($0.0000045/GiB/s) once customers needed persistent sandboxes. Modal went the opposite direction on generosity, launching persistent volumes with 1 TiB/month free precisely to absorb model-weight storage without pushing customers back to S3.

• Storage is a primary, co-equal meter at retrieval platforms. At vector databases the stored index is the product, so storage sits alongside reads/writes/queries as a first-class usage axis. Pinecone bills read units, write units, and storage as three independently-scaling layers; Chroma meters four dimensions ($2.50/GiB written, $0.33/GiB-month stored, $0.0075/TiB queried, $0.09/GiB egress); and Weaviate charges per 1M vector dimensions stored plus $0.10–$0.12/GiB storage. turbopuffer goes furthest by making cheap object storage its entire architectural bet against in-memory rivals.

• Vector-storage rates are deflating fast. The corpus captures a clear downward trend in the price of keeping an embedding index online. Milvus/Zilliz cut managed storage 87% from $0.30 to $0.04/GB-month in October 2025, and Weaviate’s per-1M-dimension rate fell roughly 95% from $0.095 (Jan 2024) to $0.00465 (2026) as object-storage architectures replaced RAM-resident indexes. Storage-per-GB is becoming a commodity line at the retrieval layer even as it stays sticky elsewhere.

• Free allowances are small and deliberate at infra, generous at research. Upstash (1 GB Redis free), E2B (1 GiB Hobby, 5 GiB Pro), and Pinecone (2 GB Starter) include just enough to remove signup friction. The exception is research land-grabs: Weights & Biases hands academic users 200 GB free, and Qdrant offers a permanent free 4 GB-disk cluster with no credit card.

• Storage tiers proliferate by durability. CoreWeave’s four-tier object storage (Hot/Warm/Cold/Archive) and RunPod’s three-way split (Container Disk, Volume Disk, Network Storage) show vendors charging different rates for ephemeral vs. persistent vs. archival durability — the same way hyperscalers separate instance store from block volumes and cold archive.

Counterexamples & variants

The clearest variant is the gated rate. turbopuffer meters storage per GB-month and writes per GB written, but the base per-GB rate appears only inside its interactive cost calculator — the static pricing page does not expose it. The corpus records this as “unknown” rather than guessing, which is itself instructive: a usage-based vendor can lean entirely on storage economics for its competitive story (object storage vs. RAM) while keeping the actual rate behind a calculator, so buyers must model their own corpus to learn the number. LanceDB is a softer version of the same pattern — storage is a real billed dimension, but the public pricing page is a contact form and the rate is free during public beta, so the number lives in sales quotes and a single AWS Marketplace listing.

The second variant is storage absorbed into a synthetic unit. Apify bills compute units at $0.13–$0.20 per CU with storage as one of four platform-usage components rolled into the CU; Dify, Genspark, and GitLab fold storage into message credits, credit tiers, or GitLab Credits respectively; and Vectara meters storage inside a credit unit and never publishes a per-GB overage rate. In these models storage consumption is real but is converted into the platform’s own currency rather than billed as raw GB. The customer still pays for data at rest, but the per-GB rate is invisible. This trades transparency for simplicity and is common where storage is a small fraction of total cost.

The sharpest counterexample to “storage is cheap and passive” runs in two directions. On the low end, Vast.ai and CoreWeave bill storage continuously — including on stopped instances or while no compute runs — so a customer who spins down a GPU to save on compute keeps paying for the attached disk indefinitely; deleting the resource is the only way to stop the meter. This inverts the usual mental model where pausing a resource pauses its cost, and it’s the most common surprise on the bill. On the high end, Kustomer’s $50/GB/month data-storage add-on shows that “storage is cheap” is a claim about infrastructure, not about SaaS: wrapped inside an enterprise support platform, the same gigabyte costs orders of magnitude more than raw object storage. The enterprise RAG case (Vectara) is the opposite failure again — storage is metered but its rate is never published, so buyers cannot self-estimate where a deployment lands and must negotiate the size band per deal.

What this means for buyers vs vendors

For buyers

Treat the headline per-GB rate as the least important storage variable. Ask three questions instead: What is the time basis (GB-month, GB-hour, or daily-average)? Does the meter keep running on stopped or idle resources — as it does on Vast.ai and CoreWeave? And what storage class am I actually buying — object storage at $0.015–$0.04/GB, GPU block disk at $0.05–$0.20, database storage at $0.25–$0.33, or a SaaS add-on at $50? The 3,000x spread in this corpus means the class you land in dominates the bill far more than your volume does. Storage compounds silently, so model it over the full retention period, not just month one, and use the guide to choosing the right usage metric to weigh whether storage or requests will dominate. Where rates are gated (turbopuffer, LanceDB) or bundled into credits (Vectara, Apify), insist on a worked estimate for your own data volume before committing.

For vendors

Per-GB storage works best as a secondary meter that captures persistent state without distorting your primary value metric — the pattern RunPod, Modal, CoreWeave, and E2B follow by keeping compute as the headline. Decide your time basis deliberately: daily-average (like Upstash) reduces bill-shock from spikes, while per-second continuous metering (like Vast.ai) maximizes capture but generates support tickets when customers don’t realize stopped instances still cost money. If you sell retrieval, expect the storage rate itself to keep deflating — Milvus/Zilliz’s 87% cut and Weaviate’s ~95% drop are the direction of travel, so plan to differentiate on architecture (object vs. RAM) rather than on the rate card alone. A small free allowance removes onboarding friction cheaply. For the mechanics of metering and prorating data at rest across a billing cycle, see the guides on usage-based pricing models and usage invoicing and billing cycles.

Explore this theme in the knowledge graph →