Pinecone vs Qdrant Cloud vs Weaviate Cloud: Real Vector Database Cost per 100 Million Vectors in 2026
The short answer: these three do not price the same thing, and the gap between them widens dramatically as vector count grows. Pinecone meters storage, reads and writes separately and stays roughly linear with usage; Qdrant Cloud and Weaviate Cloud (at small to mid scale) charge for a fixed cluster and let you run unlimited queries inside it. At 100 million vectors (1536 dimensions, roughly 800 GB of indexed storage after overhead) with 10 million queries and 15 million writes a month, Pinecone Standard costs about $484/month, and a comparable Qdrant Cloud cluster was reported at about $480/month. Weaviate is the case that requires care: its published per-dimension Serverless rate, extrapolated naively to this scale, produces an absurd $14,592/month, while a third-party estimate of Weaviate's actual Enterprise Cloud tier at this scale (a different, dedicated-instance pricing model) puts it closer to $450/month. The per-dimension headline rate is not the number a real 100-million-vector deployment pays.
What each vendor bills
Pinecone (OFFICIAL, pinecone.io/pricing). Starter: free, 2 GB storage, 2M write units and 1M read units a month, 5 indexes. Standard: $50/month minimum, then usage-based — storage $0.33/GB-month, write units $4 per million, read units $16 per million, plus SAML SSO, RBAC, backup/restore, and dedicated read nodes. Enterprise: $500/month minimum, storage at the same $0.33/GB-month rate but read units at $24 per million and write units at $6 per million, adding a 99.95% uptime SLA, private networking, customer-managed encryption and HIPAA coverage. Rates vary somewhat by cloud and region; this article uses the published AWS/us baseline. The monthly minimum is charged even when usage falls below it, and real-world reviews report bills landing 2.5–4x over naive budget estimates once read-unit volume at production query rates is accounted for.
Qdrant Cloud (OFFICIAL for the mechanism, REPORTED for dollar figures). No per-operation charge of any kind: Free tier is a single-node 1 GB RAM / 0.5 vCPU / 4 GB disk cluster, forever. Paid tiers (Standard, Premium) bill for provisioned cluster resources (RAM, CPU, disk), not reads, writes, or storage volume directly, and Qdrant does not publish an exact per-resource rate card — sizing requires its pricing calculator or a sales conversation. Third-party benchmarking against a standardized workload reported Starter around $9/month (1 GB RAM, ~100K vectors), Performance around $65–220/month (10M vectors), and a Scale-tier cluster sized for 100M vectors around $480/month. Because you pay for the cluster, not the query, your bill does not spike during a high-traffic day or a large batch re-embedding job the way Pinecone's read/write-unit meter can.
Weaviate Cloud (OFFICIAL for the Serverless rate; REPORTED for Enterprise Cloud dollar estimates). Free: one cluster per user, 100,000 objects, 1 GB memory, 10 GB disk, 2,000 embedding requests a day, forever (a genuine permanent free tier as of 2026, replacing an older 14-day expiring sandbox). Serverless Cloud: starting at $25/month, billed at $0.095 per 1 million vector dimensions stored per month, rising to $0.145 per million on the Professional support tier. Enterprise Cloud: dedicated instances, priced from $2.64 per AI Unit (AIU), fully custom-quoted for large-scale production. One current third-party source describes a renamed, cheaper self-serve tier ("Flex," $45/month minimum, $0.00465 per million dimensions) that would be roughly 20 times cheaper per dimension than the official Serverless rate above; this article could not confirm that rate against Weaviate's own current pricing page and flags it as an unresolved conflict rather than adopting it.
The workload and its storage footprint
All ILLUSTRATIVE. This article normalizes to 1536-dimension vectors (the OpenAI text-embedding-3-small size, a common default) and a workload pattern reported for a "Scale" bucket in third-party benchmarking: 100 million vectors, 10 million queries a month, and 500,000 writes a day (15 million a month). Raw float32 storage for 100 million 1536-dimension vectors is 614.4 GB (count × dimensions × 4 bytes); with a 1.3x index-overhead multiplier (HNSW graph structures and metadata, a commonly cited rule of thumb) that becomes about 799 GB of billed storage on a byte-metered platform like Pinecone.
Cost at three scales
Formula: Pinecone = storage GB × $0.33 + (write units ÷ 1,000,000) × rate + (read units ÷ 1,000,000) × rate, floored at the plan minimum; Qdrant and Weaviate Enterprise figures are REPORTED cluster-sizing estimates for the equivalent vector count and query/write volume, not computed from a public per-unit rate.
| Vectors | Queries/mo | Writes/mo | Pinecone Standard | Qdrant Cloud (REPORTED) | Weaviate (REPORTED Enterprise Cloud sizing) |
|---|---|---|---|---|---|
| 1,000,000 | 100,000 | 300,000 | $50 (plan minimum) | ~$9 | ~$25 (Serverless minimum covers this scale) |
| 10,000,000 | 1,000,000 | 3,000,000 | $54 | ~$65–220 | ~$150 |
| 100,000,000 | 10,000,000 | 15,000,000 | $484 | ~$480 | ~$450 |
At 1 million vectors, Qdrant's free-tier-adjacent Starter cluster and Weaviate's Serverless minimum both undercut Pinecone's $50 floor; at 100 million vectors, all three land within about 8% of each other on this specific workload shape, which is a closer race than the vendors' marketing pages suggest.
The Weaviate per-dimension trap
Formula: naive Serverless cost = total vector dimensions ÷ 1,000,000 × $0.095. At 100 million 1536-dimension vectors, that is 153.6 billion dimensions, producing a naive monthly figure of $14,592 — roughly 30 times the REPORTED Enterprise Cloud estimate for the same vector count. This is not a pricing error; it reflects that Weaviate's per-dimension Serverless rate is designed for a much smaller scale, and a deployment approaching 100 million vectors is expected to move to Enterprise Cloud's dedicated-instance model, which is priced entirely differently (per AI Unit, custom-quoted) and does not carry the per-dimension meter at all. Any calculator or spreadsheet that extrapolates the public Serverless per-dimension rate past roughly 10–20 million vectors will overstate Weaviate's real cost by an order of magnitude.
Read-heavy versus write-heavy workloads
Pinecone is the only one of the three whose bill moves directly with query and write volume, because it is the only one metering operations rather than provisioned capacity. Formula: read-unit cost share = read units × $16/1M ÷ total cost. In the 100-million-vector scenario above, reads are $160 of Pinecone's $484 total (33%) and writes are $60 (12%); doubling query volume to 20 million a month would add another $160, pushing Pinecone past Qdrant's and Weaviate's REPORTED cluster costs for the same vector count, while a genuinely read-heavy chatbot workload at high queries-per-second is exactly the scenario independent reviews flag as where Pinecone's bill "surprises" teams that budgeted off the storage line alone.
Quantization as the biggest lever
Across every vendor, the single largest cost lever is not which platform you choose but whether you use vector quantization (binary or scalar compression of the stored vectors), reported to cut storage and RAM requirements by 4 to 32 times depending on the compression level and acceptable recall loss. Applying even a modest 4x reduction to the 799 GB storage figure above would cut Pinecone's storage line from $264 to $66 a month, and would proportionally shrink the RAM requirement driving Qdrant's and Weaviate's cluster-sizing costs.
Sensitivity
- Vector dimension. A workload using 768-dimension embeddings instead of 1536 halves every storage-based cost in this article; 3072-dimension embeddings (some newer, higher-fidelity models) double them.
- Query volume. Only Pinecone's bill moves directly with this; Qdrant and Weaviate absorb query growth into the same cluster cost until capacity limits require a resize.
- Index overhead multiplier. The 1.3x used here is a common estimate; some HNSW configurations and payload-heavy metadata can push this considerably higher.
- Which Weaviate rate is current. The gap between the official $0.095/1M-dimension Serverless rate and a third-party report of a cheaper "Flex" tier at $0.00465/1M is unresolved and should be checked directly before budgeting.
Budgeting traps
- Extrapolating a per-unit rate past the scale it was designed for. Weaviate's Serverless per-dimension rate is the clearest example, but any consumption-based meter can do this.
- Ignoring read-unit costs at production query volume. Pinecone's storage line looks cheap in isolation; the read-unit line is where high-QPS applications get surprised.
- Comparing a metered price (Pinecone) to a capacity price (Qdrant, Weaviate) without normalizing the workload. A cluster sized for low query volume looks cheap next to Pinecone's usage-based bill until traffic grows.
- Sizing storage off raw vector bytes only. Index overhead reliably adds 20–40% or more on top of the raw float32 size.
What to ask before you buy
Ask Qdrant and Weaviate for a sized quote against your actual vector count, dimension, and query/write volume, since neither publishes a full rate card for its paid cluster tiers. Ask Pinecone for a projected bill at your expected read-unit volume specifically, not just storage, since that is the line most likely to exceed a naive budget.