Software advice
Best AI Time Series Databases in 2026
An AI time series database is a database that stores timestamped measurements such as sensor readings, metrics and market ticks, and adds features for AI work, such as forecasting, anomaly detection, vector search or an MCP interface for assistants. It is not a general data warehouse, a dashboard tool or a monitoring service, though several products here overlap with those.
Quick answer
TDengine ranks first because it prints a free tier and paid prices and names the most concrete AI features. Tiger Data fits PostgreSQL shops that want vector search next to metrics. InfluxDB fits teams that want a mature ecosystem and a free core.
A buyer choosing a time series database faces two questions at once: which engine copes with the write volume, and whether any of the AI claims on the homepage are real. Many vendors in this group name no AI feature at all, and the older open source projects name none. This ranking weighs published prices, stated limits, named AI features, data handling statements and how plainly a vendor says who it is. We read vendor pages in November 2026 and ran none of the databases, so every speed figure here is the vendor's own claim. Thirteen products are listed, from free open source projects to custom-quote enterprise systems.
The 13 best AI time series databases in 2026, ranked
There is no score. Vendors can pay for more visibility on this website, for example a higher place in a ranking or more room on a page. Commercial relationships can also affect which products we feature and how prominently. How we rank.
| # | Tool | Best for | Free plan | Entry price |
|---|---|---|---|---|
| 1 | TDengine | Industrial teams that ingest OPC, MQTT or Kafka data and want AI forecasting built into the database | Yes | Free (self-hosted, up to 5,000 tags) |
| 2 | Tiger Data | Teams that already run PostgreSQL and want time-series storage plus vector search without leaving SQL | Trial | Starting at $30/month |
| 3 | InfluxDB | DevOps and IoT teams that want a purpose-built time-series engine with a large plugin ecosystem | Yes | Free Forever (InfluxDB 3 Core) |
| 4 | QuestDB | Finance and trading teams that need fast ingestion and time-series SQL such as ASOF JOIN | Trial | Quoted |
| 5 | ClickHouse | Teams that treat time series as one workload inside a large analytical database | Trial | Quoted |
| 6 | kdb+ | Capital markets and quant teams with in-house q skills and heavy tick data | No | Quoted |
| 7 | VictoriaMetrics | Teams that outgrow a single Prometheus server and want a compatible, scalable replacement | Yes | Quoted |
| 8 | CrateDB | Teams that mix time series, JSON documents and full-text search and prefer a SQL interface | Trial | From $0.073 per hour |
| 9 | GreptimeDB | Observability teams that want metrics, logs and traces in one database speaking SQL and PromQL | Yes | From $290 per month |
| 10 | Prometheus | Engineers who monitor services and containers and want PromQL alerting they run themselves | Yes | Quoted |
| 11 | Apache IoTDB | Industrial IoT operators with very large device fleets and an existing Hadoop, Spark or Flink stack | Yes | Quoted |
| 12 | Apache Druid | Analytics teams that query streaming Kafka or Kinesis data across billions of rows | Yes | Quoted |
| 13 | OpenTSDB | Teams that already operate Hadoop and HBase and want raw metrics kept without aggregation | Yes | Quoted |
- 1
TDengine ranks first because it is the only tool here that prints both a free tier and priced paid tiers while also naming the most concrete AI features. Forecasting, anomaly detection, an MCP interface and built-in assistants are all described. The weak spot is the vendor identity, which the pages we read do not state.
Strong on
- A free self-hosted tier up to 5,000 tags, with paid tiers priced by tag count on the pricing page
- SQL-based AI and ML forecasting and anomaly detection are named, along with more than 50 capabilities exposed over MCP
- Ingests OPC, MQTT and Kafka with built-in ETL, which fits factory and energy data
Weak on
- Company name and headquarters are not stated on the pages we read, so due diligence takes extra work
- High availability costs extra, 30% on self-hosted plans and 50% on managed plans
Skip it if: Teams that want a general-purpose SQL warehouse or a pure Prometheus replacement, since the product is shaped around industrial assets. Paid self-hosted starts at $12,500 per year for 7,500 tags; managed starts at $970 per month for 5,000 tags.
- 2
Tiger Data takes second place because its three plans are laid out with CPU, memory, disk and IOPS ceilings, and the Scale plan names a SOC 2 report. Vector search sits inside a hybrid retrieval feature, so AI workloads and metrics share one engine. It trails TDengine only because its AI description is narrower. Choose it when your team already knows SQL.
Strong on
- Plan limits are explicit: up to 32 CPU and 128 GB memory per service on Scale, with 64 TB disk
- Row-columnar storage and tiered storage to object storage keep older data cheap to hold
- Hybrid retrieval combines keywords, vectors, filters and ranking in one feature
Weak on
- Headquarters is not stated on the pages we read, and the vendor lists no region for hosted data
- Extras such as HA replicas, tiered storage and production support are billed on top of the starting price
Skip it if: Buyers who need a self-managed license with a flat fee, because the pages we read describe a hosted service billed by usage. Compute is billed hourly and storage by average GB per hour; HA replicas, IO boost and tiered storage cost extra.
- 3
InfluxDB comes third because its free core, usage-priced serverless cloud and 400-plus Telegraf plugins cover most monitoring and sensor cases. The vendor positions the engine for feeding AI and ML pipelines but names no assistant or agent. Its free core also lacks high availability. Teams that need replicas or enterprise security must move to the paid Dedicated or Enterprise products, which are quoted individually.
Strong on
- Per-unit cloud prices are printed, from $0.0025 per MB written to $0.09 per GB of data out
- Telegraf offers 400-plus input plugins, which shortens collection work
- Cold data can move to Parquet files that data lakes and warehouses can read
Weak on
- The free core has no high availability and offers community support only
- Cloud Serverless has service quotas and no high availability either
Skip it if: Teams that want a named AI assistant or agent inside the database, because none is described. Cloud Serverless is pay-as-you-go with a $250 credit; Dedicated and Enterprise are custom quotes.
- 4
QuestDB sits fourth because its SQL extensions for market data, ASOF JOIN and SAMPLE BY, are documented with examples, and the site shows a coding-agent demo that builds a materialized view. Price transparency is thin, since Enterprise is quote-only, and the vendor's identity is not stated. For a trading team the SQL features matter more than the missing price list, but a finance buyer should expect a sales call.
Strong on
- Apache 2.0 open source edition that anyone can self-host and inspect
- ASOF JOIN, SAMPLE BY and materialized views for candle data suit market-data work
- PostgreSQL wire protocol access lets existing clients and tools connect
Weak on
- Enterprise pricing requires contacting sales and the trial terms are not given
- The vendor describes open source use mainly as evaluation, prototyping and pilots
Skip it if: Buyers who need published paid prices or a named hosted service with stated compliance. The open source edition is free under Apache 2.0; Enterprise is a custom quote with an unspecified trial.
- 5
ClickHouse ranks fifth because it is a general analytical database more than a time-series specialist, even though many telemetry stacks run on it. Its cloud has three clear tiers and credits for trial use. AI appears through Langfuse tracing and agent workloads, not through a database-side model feature. Teams already on it for logs or product analytics can add time series without a second system, which is its real advantage here.
Strong on
- Three cloud tiers with stated limits, from 1 TB storage on Basic to unlimited on Scale
- Compute is metered per minute, so idle capacity is not billed as a whole month
- Enterprise adds SAML SSO, private regions, CMEK, HIPAA and PCI compliance
Weak on
- No price string is published on the pricing document, so cost needs the calculator
- Basic is capped at 1 TB of storage and a single availability zone
Skip it if: Teams that want a point-and-click time-series product with built-in downsampling rules, since ClickHouse is a general SQL engine. Usage-based with no monthly minimum; the calculator shows prices, and new accounts get $300 in credits for 30 days.
- 6
kdb+ lands sixth because its newer KDB-X engine pairs vector search with an MCP server and agent orchestration, a real AI story for trading data. The cost is opacity: no price, no free tier and a language, q, that few engineers know. The speed claims on the site are the vendor's own. Expect a sales process and a hiring problem before you expect a quick pilot.
Strong on
- KDB-X names native vector search, an MCP server and AI agent orchestration
- Queries run in Python, SQL or q, and Parquet files are supported
- The vendor claims microsecond queries on billions of rows, a claim we did not test
Weak on
- No price is published and there is no free license on the pages we read
- Customer data hosting region and certifications are not stated
Skip it if: Small teams that want to start with a free tier and a gentle learning curve. No kdb+ price or free license is shown on the pages read; the deal is a custom quote.
- 7
VictoriaMetrics sits seventh. It accepts Prometheus scraping and remote write, adds MetricsQL, and its site names MCP and agent skills for querying telemetry from an AI assistant plus ML-based anomaly detection. Four support tiers are described, but none carries a price, and the maker is not named on the pages we read. It is a sensible step up for teams already running Prometheus.
Strong on
- Prometheus-compatible scraping and remote write ease migration from an existing setup
- Anomaly Detection with AI-assisted setup and an MCP route for assistants are both named
- Enterprise tiers are differentiated by support response time, from best effort to 2 hours for P1 and P2
Weak on
- Enterprise and Cloud prices are quote-only, with no amounts published
- The company behind the product is not named on the pages we read
Skip it if: Buyers who want a fixed public price list or a hosted free tier with a stated allowance. The open source versions list no price; four Enterprise tiers and the Cloud product are quote requests.
- 8
CrateDB ranks eighth because its cloud pricing is printed per hour and the tiers have stated ceilings, yet the AI offer is still mostly links: a vector data model page, an AI operations use case and MCP documentation, none tied to a plan. It is one of few vendors here with a European-registered maker. That makes it a sound choice for teams that value stated limits and a hosted trial over breadth of AI claims.
Strong on
- Hourly prices are listed, from $0.073 for the shared card to $0.246 for Dedicated
- Dedicated clusters reach up to 9 nodes, 144 vCPUs and 72 TB of storage
- AWS and Azure Private Link are offered on the paid tiers
Weak on
- The privacy policy says personal data is processed outside the EEA, including the United States
- No SOC 2 or ISO 27001 is stated on the pages we read
Skip it if: Buyers who need only a narrow metrics store, because CrateDB's multi-model scope adds surface they will not use. CrateDB Cloud bills hourly per node; the shared free card starts at $0.073 per hour and Dedicated at $0.246.
- 9
GreptimeDB comes ninth. A single database that speaks SQL, PromQL and OpenTelemetry is attractive for observability, and the managed plan prints a starting price with unlimited retention. The AI claim is thin, limited to investigation agents exploring data in parallel, with no model or MCP feature named. It suits teams comfortable buying a managed plan for observability data.
Strong on
- SQL, PromQL and OpenTelemetry are supported in one database
- The Apache-2.0 core can run in your own environment at no license cost
- Automatic region load balancing and splitting is named as Intelligent Autopilot
Weak on
- The $290 starting price is for the managed Enterprise plan only; no smaller paid tier is listed
- No SOC 2 or ISO certification appears in the page text
Skip it if: Teams that need a published free managed tier or a named AI assistant. The managed Enterprise plan starts at $290 per month; bring your own cloud is custom; the Apache-2.0 core is free.
- 10
Prometheus holds tenth place because it is the reference monitoring toolkit, free and widely supported, yet it is a metrics server rather than a long-term time-series store. Each server node is standalone, and the project pages name no AI feature. Many other tools on this list accept its data format. Pair it with a long-term store if retention matters, and treat it as the baseline the others are measured against.
Strong on
- PromQL queries a multi-dimensional data model of metric names and key-value labels
- Servers run standalone with no dependency on distributed storage
- Service discovery or static configuration finds the targets to scrape
Weak on
- Data is limited to each node's local storage, so long retention needs another system
- Pull-based scraping depends on reachable HTTP targets, with a push gateway only for short jobs
Skip it if: Buyers who want built-in AI analysis, clustering or a managed service. Released under the Apache 2 license; the project page lists no price.
- 11
Apache IoTDB ranks eleventh. It is built for large device fleets, with time alignment across sensors and a cloud-to-on-premise sync tool, and it connects to Hadoop, Spark, Flink and Grafana. The page names no AI, ML or AINode feature, and it prints no commercial plan. It is a fit for teams with engineers who can run Apache software on their own servers.
Strong on
- Time alignment lines up series from different devices and sensors
- A cloud-to-on-premise sync tool and one-click cloud installation are described
- Hadoop, Spark, Flink and Grafana integrations are listed
Weak on
- The page states a storage cost claim for hard disk only
- It targets IoT fleets rather than general analytics
Skip it if: Buyers who need vendor support, a price list or an AI feature. Apache open source project with no price on the site; the page names no license.
- 12
Apache Druid takes twelfth place. It is a real-time analytics database more than a point-by-point time-series store, strong at fast aggregate queries over streams but built for different workloads than sensor writes. The site names no AI feature and no commercial plan. Choose it for dashboards over event streams, and look elsewhere for per-device history or AI features.
Strong on
- Native ingestion from Apache Kafka and Amazon Kinesis
- Millisecond OLAP queries on billions to trillions of rows are claimed on the site
- Automatic schema discovery and a SQL API simplify loading
Weak on
- It is built for analytics rather than writes of individual points
- Joins run fastest when tables are pre-joined at ingestion
Skip it if: Teams that need per-device sensor storage, vendor support or a named AI feature. Apache open source project with no price on the site; the page does not name the license.
- 13
OpenTSDB closes the ranking because its page lists 2.4.1, dated September 2021, as the latest release, and it needs Hadoop and HBase underneath. It does keep raw data at millisecond resolution indefinitely. No AI feature is named, and no support or price is published. Newer projects on this list are easier to adopt for most teams starting today.
Strong on
- Raw, unaggregated data is kept indefinitely at millisecond resolution
- Writes scale by adding nodes, with millions of writes per second claimed
- A GUI for graphing and an HTTP API come with the project
Weak on
- It requires Hadoop and HBase as the storage layer
- The latest release listed on the project page dates from 2021, with no newer version
Skip it if: Buyers who do not already run Hadoop, or who need AI features and active vendor support. Open source project; the page states no license or price, and the latest listed release is 2.4.1 from 2021.
How these AI time series databases were judged
Which time series databases name a real AI feature?
Only a minority do. TDengine names SQL-based AI and ML forecasting, anomaly detection, thirteen built-in assistants and more than 50 capabilities exposed over MCP. kdb+ describes KDB-X with native vector search, an MCP server and agent orchestration. Tiger Data lists vector search inside its hybrid retrieval feature, and VictoriaMetrics names MCP access and ML-based anomaly detection. QuestDB shows a coding-agent demo, CrateDB links vector and MCP pages, and InfluxDB positions itself for ML pipelines without an assistant. Prometheus, Apache IoTDB, Apache Druid and OpenTSDB name nothing. Ask each vendor for a worked example on data shaped like yours, and treat all speed figures as claims we did not test.
What does a time series database cost to run?
Cost splits into three shapes. Printed prices exist for TDengine (free to 5,000 tags, then $12,500 per year self-hosted or $970 per month managed), Tiger Data (from $30 a month), GreptimeDB (from $290 a month) and CrateDB (from $0.073 per hour). InfluxDB prints per-unit cloud rates. ClickHouse is usage-based, with prices in a calculator. kdb+, QuestDB Enterprise and VictoriaMetrics Enterprise are quote-only. The open source projects cost nothing to license but need engineers. Compare on your own volume, because tag counts, storage hours and queries are metered differently across vendors.
What do these vendors say about data handling?
Very little, which is the main finding. No vendor in this ranking states whether customer data trains models. Tiger Data names a SOC 2 report on Scale and HIPAA on Enterprise. ClickHouse names HIPAA and PCI on Enterprise and a Data Privacy Framework certification. InfluxData and Crate.io also state the EU-U.S. Data Privacy Framework, and CrateDB's policy says personal data is processed outside the EEA. Open source projects leave hosting to you, which gives full control and full responsibility. For regulated data, ask for hosting region, retention and certifications in writing before a pilot.
How hard is it to leave a time series database?
Lock-in depends on the query language and storage format. Prometheus-compatible tools such as VictoriaMetrics and GreptimeDB share PromQL, which eases moves between them. InfluxDB and QuestDB use Parquet for older data, and Tiger Data builds on PostgreSQL, so SQL skills transfer. kdb+ uses q, which few engineers know and which makes migration costly. OpenTSDB depends on Hadoop and HBase, a heavy stack to leave. Before choosing, export a sample of your data and replay your five most common queries on a second engine, so you know the real switching effort.
Which database fits which team?
Match the engine to the data source. Factories and energy sites ingesting OPC or MQTT fit TDengine or Apache IoTDB. Trading desks with tick data fit QuestDB or kdb+. PostgreSQL shops fit Tiger Data, and teams that monitor services fit Prometheus, VictoriaMetrics or GreptimeDB. Analytics groups reading streams from Kafka fit Apache Druid or ClickHouse. A small team with no operations staff should prefer a hosted plan with printed prices over a free self-hosted project, because the hidden cost of an open source database is the person who keeps it running.
How to choose an AI time series database in three steps
- Measure your write rate and retention first. Estimate points per second, number of series and how long raw data must stay queryable. TDengine prices by tag count, InfluxDB by data written and queried, and Tiger Data by compute and storage hours, so the same workload produces very different bills. Open source projects shift the cost to hardware and staff. Write the numbers down before you open a pricing page, then run each vendor's calculator or request a quote with those exact figures. Drop any tool whose published limits fall below your peak volume, since a ceiling discovered after migration is expensive.
- Test one AI claim on your own data. Pick the single AI feature you would actually use, such as forecasting, anomaly detection or vector search, and run it on a week of your own data during a trial. Several vendors describe features on the homepage that the pricing page does not tie to any plan, so confirm which plan includes it. Read the vendor's data page too, and check whether queries leave your environment. If a feature cannot be tried within the trial, treat it as a roadmap item, not a capability, and rank the tool on its database strengths alone.
- Plan the exit before you load data. Write down how you would export a year of data and how long a replay would take. Check the query language, the storage format and the client libraries in use. Prometheus-compatible and PostgreSQL-compatible tools keep the exit cheap, while q-based or Hadoop-based stacks do not. Ask whether the vendor offers a self-hosted edition if its cloud service changes terms. A short pilot that includes an export and an import on a second engine reveals lock-in cheaply, and the results give you leverage in contract talks.
Frequently asked questions
What is the best AI time series database in 2026?
TDengine ranks first here because it prints a free self-hosted tier and priced paid tiers and names the most concrete AI features, including forecasting, anomaly detection and an MCP interface. Tiger Data is the pick for PostgreSQL teams that want vector search, and InfluxDB suits teams that want a mature ecosystem and a free core. Rankings rest on vendor pages we read in November 2026.
Which time series databases have a free plan?
InfluxDB 3 Core is free to self-host, TDengine is free up to 5,000 tags, and CrateDB shows a shared free card from $0.073 per hour with a no-credit-card tier described. QuestDB, VictoriaMetrics, GreptimeDB, Prometheus, Apache IoTDB, Apache Druid and OpenTSDB offer open source software. Tiger Data and ClickHouse offer trials or credits. kdb+ shows no free license.
Do any of these databases train AI models on my data?
No vendor in this ranking states whether customer data is used for model training, so the answer is unknown for all of them. Open source projects that you host yourself keep data in your environment. For hosted services, ask the vendor for a written statement on training, retention and hosting region before you load production data.
Is Prometheus a time series database?
Prometheus stores time series data and queries it with PromQL, but it is built as a monitoring and alerting toolkit. Each server is standalone and limited to local storage, so long retention usually needs another system. VictoriaMetrics and GreptimeDB accept Prometheus-style data and add scalable storage, which is why they appear in this ranking alongside it.
Which tool suits industrial IoT sensor data?
TDengine ingests OPC, MQTT and Kafka and models asset hierarchies, which fits factories and energy sites. Apache IoTDB targets very large device fleets, with time alignment across sensors and a cloud-to-on-premise sync tool. InfluxDB with Telegraf is the broader alternative. Check tag or series limits against your device count, since TDengine prices by tags.
How much does a managed time series database cost?
Printed starting prices vary: Tiger Data from $30 a month, GreptimeDB from $290 a month, CrateDB from $0.073 per hour, TDengine managed from $970 a month for 5,000 tags. InfluxDB Cloud Serverless meters writes, queries and storage separately. ClickHouse uses a usage calculator. Real cost depends on volume and retention, so request a quote or run a calculator with your numbers.
Which vendors name an MCP server or AI assistant access?
TDengine names more than 50 capabilities exposed to AI agents through MCP. KDB-X from kdb+ describes a built-in MCP server. VictoriaMetrics names MCP and agent skills for querying telemetry, and CrateDB links MCP server documentation. QuestDB shows a coding-agent demo and describes its SQL as native for AI agents. Confirm which plans include these features.
Is there a European option among these databases?
CrateDB is made by Crate.io AG, which the vendor pages place in Liechtenstein, though its privacy policy says personal data is processed outside the EEA, including the United States. kdb+ is made by a United Kingdom based company. Most other vendors are listed as United States based or do not state their location. Verify hosting regions with each vendor before relying on location.