Total cost of ownership
What does your current AI search infrastructure actually cost you?
Most teams replacing a traditional vector database save between $15,379 and $126,695 a year on a 50 GB corpus, depending on which database they are replacing, and more as the corpus grows. Here is the worked example, with every number sourced.
A worked example for three knowledge-base sizes against the vector databases buyers most commonly evaluate. Each side uses its best-practice cost-saving configuration. Pick a vector database provider below, every number on this page is derived from the assumptions and the picked vendor's recommended setup, validated against our measured July 2026 head-to-head benchmark (identical corpus and embeddings across GraphANN™, Qdrant and Neo4j; methodology and raw results available under NDA), and priced against actual AWS on-demand instance list prices (us-east-1, June 2026).
Assumptions
The basis for every number on this page
Embedding model
bge-small, 384-dim
GraphANN's bundled default model, no external embedding service
Chunk size
~512 chars
~0.5 KB raw text per chunk (default chunker; configurable)
Scale unit: raw corpus
5 GB ≈ 10M chunks
Corpus sized at the default ~0.5 KB/chunk: 5 GB ≈ 10M chunks, 50 GB ≈ 100M, 500 GB ≈ 1B. Cost is driven by the resulting vector count; a larger configured chunk size means fewer vectors, and lower cost, for the same corpus.
Pricing basis
AWS on-demand, us-east-1, June 2026
Minimum-viable instances: for every node, the cheapest EC2 / RDS class meeting its RAM and vCPU floors (families considered: m8g, c8g, r8g, x8g, x2gd; RDS db.r6g / r8g / r6i / m6g multi-AZ). List prices from the AWS pricing API, June 2026; 8,760 hr/yr; no reserved-instance discounts.
Disk provisioning
2× data per node, 50 GB min
Disks sized at 2× resident data (compaction + growth headroom), 50 GB practical floor. Replication is priced via node count, each replica's disk is billed on its node. EBS gp3 $0.08/GB-mo; RDS gp3 $0.23/GB-mo multi-AZ (AWS bills the standby copy). Backups (S3) excluded, small and symmetric.
Compute price
Actual instance pricing, not $/GB
Nodes sized so instance RAM ≥ 1.25× the serving hot set; data nodes ≥ 4 vCPU (GraphANN's covers the in-process ONNX embedder), support nodes ≥ 2 vCPU. Effective list price per GB-month of the picks: x2gd ≈ $3.80, r8g ≈ $5.40, RDS db.r6g multi-AZ ≈ $20.50.
Serving-RAM model
Resident graph, not full index
Sized from the measured resident-critical footprint — the navigation graph (~52 B/vector) plus the coarse locate codes (~92 B/vector), ~144 B/vector, plus fixed runtime. Source text and the rerank tier are memory-mapped and paged on demand, not sized into RAM. GraphANN detects the host budget and fits the instance, shedding cold pages under pressure rather than over-allocating. Competitor figures size the full in-RAM vectors/index their architectures require for the same latency.
GraphANN setup
Bundled ONNX + HA
Single node at 10M, 3-node HA cluster at 100M+
Embedding source
Self-hosted, both sides
GraphANN: bundled ONNX. Competitors: equivalent local model
Model swaps / yr
1 swap
Industry pace 2023-2026 (ada → 3-small → 3-large)
Accuracy basis
recall@10 ≥ 0.99 (measured)
July 2026 head-to-head, same corpus and embeddings, exact-KNN ground truth: GraphANN™ 0.994, Qdrant 0.999, Neo4j 0.875 (recall@10)
Benchmark validation
Measured head-to-head, July 2026
52K-chunk document set, identical bge-small (384-dim) embeddings, measured on-disk index: GraphANN™ 40 MB (efficient profile) vs Qdrant 182 MB vs Neo4j 805 MB. The GraphANN storage lines below are projected from this measured footprint (~0.76 KB/vector, client-supplied vectors) — a conservative upper bound, since text-ingested corpora carry no rerank sidecar and store less.
Compare against
Self-hosted Qdrant with vector compression and on-disk vectors enabled, the recommended cost-balanced production config.
5 GB
raw corpus
GraphANN / yr
$1.6k
Qdrant / yr
$2.2k
$577 saved per year (26% lower)
Cluster1 × m8g.xlarge (16 GB)1 × r8g.xlarge (32 GB)
RAM footprint~3 GB~20 GB
| Cost component | GraphANN | Qdrant |
|---|---|---|
| Storage (disks @ 2× data) | $48 | $134 |
| Compute (EC2) | $1.6k | $2.1k |
Single-node Qdrant with vector compression + on-disk; HA via snapshot, not cluster.
50 GB
raw corpus
GraphANN / yr
$4.9k
Qdrant / yr
$20.2k
$15.4k saved per year (76% lower)
Cluster3 × m8g.xlarge (16 GB), HA3 × x2gd.2xlarge (128 GB), HA
RAM footprint~15 GB~160 GB
| Cost component | GraphANN | Qdrant |
|---|---|---|
| Storage (disks @ 2× data) | $145 | $2.7k |
| Compute (EC2) | $4.7k | $17.6k |
3-node Qdrant cluster (RF2), vector compression keeps RAM bounded. 160 GB across 3 nodes needs ~67 GB/node with headroom, just over the 64 GB tier, forcing 128 GB instances.
500 GB
raw corpus
GraphANN / yr
$10.2k
Qdrant / yr
$120.5k
$110.3k saved per year (92% lower)
Cluster3 × x2gd.xlarge (64 GB), HA4 × x2gd.8xlarge (512 GB), HA
RAM footprint~145 GB~1.6 TB
| Cost component | GraphANN | Qdrant |
|---|---|---|
| Storage (disks @ 2× data) | $1.5k | $26.9k |
| Compute (EC2) | $8.8k | $93.6k |
Full vectors on disk (RF2), compressed layer in RAM; 4 nodes is the cheapest valid shape for 1.6 TB. Aggressive compression would shrink this further at recall cost.
Caveats
What these numbers don't include
- Embedding API cost is excluded on both sides. The comparison assumes self-hosted embedding (e.g. bge-small via GraphANN™'s bundled ONNX, or via a local Sentence-Transformers process for competitors). If your team uses OpenAI / Voyage / Cohere for embeddings, add the same per-corpus cost to both columns.
- Re-indexing cost is excluded from competitor figures. GraphANN™ eliminates re-indexing entirely. For traditional vector databases, switching to a better embedding model requires a full re-index of your knowledge base, typically 1 to 4 engineering weeks depending on scale, plus associated downtime. This cost is not included in the infrastructure figures above.
- Managed-service markup. Where pricing is given for self-hosted, switching to a managed offering (Qdrant Cloud, Zilliz Cloud, Aura, Neon) typically adds 2-4× provider markup. The GraphANN™ column assumes self-host; we don't yet offer a managed tier.
- QPS and tail latency. High-QPS workloads inflate RAM requirements for both products. The numbers assume a steady-state read pattern with cache locality. Burst-heavy workloads need more RAM headroom on both sides.
- Reserved-instance discounts and on-prem. Cloud on-demand prices are used; reserved instances and on-prem hardware can cut both columns 30-60%, but the ratio between them stays similar.
- Compliance, support, IP indemnity. Not modelled here. Talk to us about sovereign-region deployment and commercial support pricing.
- Disk-tier indexes narrow the gap. These numbers assume RAM-resident low-latency (sub-10ms median) graph-based search where appropriate. The competitor configs above already opt into disk-tier where it's the recommended best-practice (Milvus MMap, Qdrant on-disk). On configurations that go further (aggressive vector compression, disk-resident graph indexes), the competitor RAM cost drops further and the savings ratio narrows, typically at the cost of higher tail latency or lower recall.
- GraphANN™ query-side embedding compute. GraphANN™ recomputes embeddings per query. With the bundled local ONNX path this is included in the per-node compute already priced. If you use an API-based embedding for queries, expect an additional per-query API cost that scales with QPS, talk to us for a sized estimate.
- Neo4j Enterprise licensing is excluded from these figures. Enterprise licensing is commercially negotiated and not publicly listed, contact Neo4j directly for a quote. All Neo4j figures above reflect infrastructure costs only: compute and storage on equivalent AWS hardware.
Your numbers
Want this calculated for your actual knowledge base?
Send us your scale (corpus size, embedding model, deployment target) and we'll come back with a deployment plan and a fixed-price quote.