Understanding MotherDuck Pricing: Ducklings, Cooldown, and AI Units
MotherDuck publishes its prices to the cent and even walks through worked billing examples in its docs. That is ahead of most data vendors on transparency. But published rates and predictable bills are different things. The bill depends on mechanics the pricing page only hints at: how ducklings meter time, what cooldown costs on your workload shape, and what an AI Unit actually buys.
This is a breakdown of how MotherDuck pricing works as of July 2026, with the math checked against their current pricing page and billing docs. At the end: when the model is a good fit, and how it compares to a flat platform price.
The short version
- Three meters: compute (per duckling instance, mostly billed per second), storage ($0.04/GB/month in US regions), and AI functions (AI Units at $1.00 each).
- Two plans that matter: Lite is free with real limits (Pulse compute only, 3 users, 10 GB). Business is $250/month per org plus usage and gates larger instances, read scaling, query history, and the 99.9% SLA.
- Compute runs $0.60 to $36 per hour across five duckling sizes in US regions. Each duckling is one machine running one DuckDB process. You scale up, not out.
- Cooldown is billable. After your last query, the instance stays warm for 1 to 10 minutes by default and you pay for that time. Fine for steady workloads, expensive for spiky ones.
- AI Units run about 3.3x OpenAI's list prices for the models they meter (as of July 2026), capped at 10 units per day by default.
- Raw hardware markup is about 2.8x to 3.7x AWS on-demand rates, normal for a managed database. You are paying for the service, not the metal.
The plans
MotherDuck runs in four AWS regions (us-east-1, us-west-2, Frankfurt, Dublin); European compute rates run about 22% higher. There is a 7-day free trial of Business with no credit card, after which you land on Lite.
| Lite | Business | Enterprise | |
|---|---|---|---|
| Platform fee | $0/mo | $250/mo per org + usage | Custom |
| Compute | 10 CU hours/mo included, Pulse only | Pay-as-you-go, all sizes | Custom |
| Storage | 10 GB included | $0.04/GB/mo | Custom |
| Users | 3 active + 2 service accounts | 10 active + unlimited service accounts | Unlimited |
| Read scaling | No | Yes | Yes |
| Query history | No | Yes | Yes |
| Snapshot retention | Up to 1 day | Up to 90 days | Up to 90 days |
| SLA | None | 99.9% | Custom |
Lite is a genuinely useful free tier, but it caps you at the smallest instance type and 3 users. The moment you want a dedicated instance, more seats, or query history, you are a Business customer, and $250/month is your floor before any usage.
Compute: the duckling model
Ducklings are MotherDuck's compute instances. Five sizes (current rates):
| Duckling | US price | EU price | Billing | Default cooldown | Startup |
|---|---|---|---|---|---|
| Pulse | $0.60/hr per CU | $0.73/hr | Per query, CU-seconds | 1 second | ~100ms |
| Standard | $2.40/hr | $2.93/hr | Per second, 1-min minimum | 1 minute | ~100ms |
| Jumbo | $4.80/hr | $5.86/hr | Per second, 1-min minimum | 1 minute | ~100ms |
| Mega | $12.00/hr | $14.65/hr | Per second, 1-min minimum | 5 minutes | A few minutes |
| Giga | $36.00/hr | $43.95/hr | Per second, 1-min minimum | 10 minutes | A few minutes |
Two things separate this model from a Snowflake-style warehouse (see our Snowflake pricing breakdown for that model).
First, each duckling is a single machine. DuckDB is a single-node engine, so MotherDuck scales by handing you a bigger box, not a cluster. At their Coalesce 2025 talk they described Giga as "192 cores and a terabyte and a half of RAM" and Mega as "64 cores and 256 gigs of RAM." Those specs map almost exactly onto AWS instances you can rent yourself:
| Duckling | MotherDuck price | Closest AWS match (inferred) | AWS on-demand | Effective markup |
|---|---|---|---|---|
| Mega (64 cores, 256 GB) | $12.00/hr | m7i.16xlarge (64 vCPU, 256 GiB) | $3.23/hr | ~3.7x |
| Giga (192 cores, 1.5 TB) | $36.00/hr | r7i.48xlarge (192 vCPU, 1,536 GiB) | $12.70/hr | ~2.8x |
The instance mapping is our inference from the published specs, not something MotherDuck states. A 3x-ish hardware markup is standard for managed databases (Snowflake's is comparable or higher). MotherDuck is not gouging; you should just know that the convenience layer costs about 2x the metal.
Second, Pulse bills differently from everything else. Pulse is elastic, multi-tenant compute metered per query in Compute Unit seconds, with a minimum of 1 CU-second per query. Two 10-second queries at 1 CU each cost 20 CU-seconds, about a third of a cent, so small reads are very cheap. The catch: a CU measures resources consumed, not wall-clock time, so a heavy query can burn multiple CUs at once, and MotherDuck's own docs note that complex queries (spatial joins, regex-heavy work) may do better on the larger fixed sizes. Lite users only get Pulse, so compute-hungry workloads meter accordingly on the free tier.
Cooldown: the part of the bill that depends on workload shape
This mechanic decides whether MotherDuck is cheap or expensive for you.
When a fixed-size duckling finishes your last query, it does not shut down. It stays warm for the cooldown period so the next query skips the cold start, and you pay full price for that warm time. Defaults are 1 minute on Standard and Jumbo, 5 minutes on Mega, 10 minutes on Giga, all configurable from 1 minute up to 24 hours.
MotherDuck's docs give the friendly case: 5 queries of 30 seconds each, run back to back on Standard, bill as 150 seconds of work plus one 60-second cooldown, 210 seconds total. Overhead: 29%, and the warm instance earns its keep by keeping caches hot.
The unfriendly case is spiky traffic. Say a dashboard fires one 10-second query every 5 minutes, 8 hours a day, on a Standard duckling. Every query wakes the instance, runs 10 seconds, then pays 60 seconds of cooldown. That is 96 wakes and about 112 minutes billed per day for 16 minutes of query time: roughly 7x your compute, $4.48 a day instead of $0.64, purely because of arrival pattern.
On Mega and Giga the defaults magnify this. A Mega instance that wakes once an hour for a 2-minute job pays 5 minutes of cooldown each time: 7 minutes billed per 2 minutes of work.
Four levers, all legitimate:
- Batch your queries. Back-to-back work shares one cooldown.
- Tune the cooldown down. The floor is 1 minute on every size. If your nightly batch runs as one block, the default 5 or 10 minutes buys you nothing.
- Route small queries to Pulse. Its 1-second cooldown suits frequent light reads.
- Watch the 1-minute minimum. Every wake bills at least 60 seconds, so many tiny isolated queries on a Standard instance is the worst shape.
Storage: cheap, with one thing to watch
Storage is $0.04/GB/month in US regions ($0.043 in Europe), billed on your average daily storage. That is about 1.7x the S3 Standard list price of $0.023/GB, a modest margin for a managed layer with versioning and point-in-time restore. Storage is rarely the line item that surprises anyone.
Two details:
- Retention counts. Business plan databases keep 7 days of history by default (configurable from 0 to 90 days), and that history bills as storage. Heavy-update tables with long retention can quietly multiply their footprint.
- Your own buckets are free. Data queried in your own S3, GCS, or Azure buckets does not count toward MotherDuck storage, and shares are zero-copy. Cold data belongs in your own bucket.
AI Units: passthrough with a consistent margin
MotherDuck prices its AI features in AI Units at $1.00 each. The FixIt SQL-repair function is free, and text-to-SQL runs 60 calls per unit, which is hard to complain about. The generative and embedding functions are where the math matters. One AI Unit buys:
| Model | Tokens per $1 AI Unit | Effective price /M tokens | OpenAI list /M | Markup |
|---|---|---|---|---|
| GPT-5 input | 240,000 | $4.17 | $1.25 | ~3.3x |
| GPT-5 output | 30,000 | $33.33 | $10.00 | ~3.3x |
| GPT-5-mini input | 1,200,000 | $0.83 | $0.25 | ~3.3x |
| GPT-5-mini output | 150,000 | $6.67 | $2.00 | ~3.3x |
| text-embedding-3-small | 15,000,000 | $0.067 | $0.02 | ~3.3x |
MotherDuck rates from their pricing docs; OpenAI list prices from OpenAI's model pages for GPT-5, GPT-5 mini, and text-embedding-3-small, current as of July 2026. The ratio is almost exactly 3.33x across the board, which suggests a deliberate margin policy. One wrinkle: AI Units are denominated in GPT-5-family tokens, and OpenAI now labels GPT-5 a previous-generation model (its current flagship is GPT-5.5), so check current rates before pointing a token-heavy workload at AI functions. A default soft cap of 10 AI Units per day stops runaway spend at $10 until you ask support to raise it. The cap is also a signal: these functions are priced for in-database convenience, not as your primary LLM gateway. If you are embedding millions of rows, calling the model APIs directly costs a third as much at list price.
What a bill actually looks like
Three illustrative shapes, computed from the list prices above (your workload will differ):
| Shape | Plan | Compute | Storage | AI | Est. total/mo |
|---|---|---|---|---|---|
| Side project, 3 users, 50 GB | Lite | ~5 Pulse CU-hrs beyond free ($3) | 40 GB billable ($1.60) | FixIt only ($0) | ~$5 |
| Mid-market team, 500 GB | Business | ~40 hrs Standard ($96) | $20 | ~5 units ($5) | ~$371 |
| Nightly 4-hr batch on Mega, 10 TB | Business | 120 hrs Mega ($1,440) + ~2.5 hrs cooldown ($30) | $400 | ~$20 | ~$2,140 |
The pattern: the platform fee dominates small Business bills, compute dominates large ones, and the gap between a good and bad bill is mostly cooldown discipline and instance sizing.
When MotherDuck is a good fit
Honest answer: for a common shape of problem, it is excellent.
- You want a fast SQL engine in the cloud and nothing else. You already have ingestion and BI sorted, or do not need them. Managed DuckDB is excellent on the small-to-medium data most companies actually have.
- You live in the DuckDB ecosystem. Local DuckDB for development, MotherDuck for sharing and cloud execution. The hybrid execution story is genuinely differentiated.
- Your workload is steady or batchable. Per-second billing with tuned cooldowns is efficient when queries arrive in blocks.
- You value pricing transparency. Prices published to the cent, worked billing examples in the docs, no capacity contracts, no edition multipliers, and a free tier. Refreshingly legible next to a Snowflake capacity negotiation.
If your friction is with the engine itself (scale, regions, ecosystem), see our guide to MotherDuck alternatives by category.
How Definite's model differs
The structural question with MotherDuck is not the rates. It is the unit of purchase. MotherDuck sells you a database. Ingestion from your SaaS tools, dashboards, a semantic layer, and an AI analyst all have to come from somewhere else, each with its own bill and its own integration to maintain. That is the stack-versus-platform decision, and it matters more than any per-hour rate.
Definite prices the whole platform flat: $250/month for the Standard plan with unlimited users, 500+ connectors, BI, a semantic layer, and Fi (our AI analyst) included. Compute and AI usage bill as credits at $1 each, with 100 included. Same base number as MotherDuck Business, different denominator: one buys a database for 10 users, the other buys the full analytics system for unlimited users.
Definite runs the same engine family MotherDuck does: a lakehouse built on DuckLake and DuckDB, open Parquet underneath, sub-second dashboards on typical company data. And for teams that want the data to stay home, Definite's private deployment runs the entire platform, lakehouse included, in your own cloud account from a single Helm chart, so the object store holding your data is yours.
If a managed database is the piece you are missing, MotherDuck is a well-built, honestly priced one. If the database is just one of four things you still have to buy, the platform math is worth running first.
FAQ
How much does MotherDuck cost per month? The Lite plan is free: 10 Compute Unit hours of Pulse compute, 10 GB of storage, 3 active users. Business is $250 per org per month plus usage: compute from $0.60 to $36 per hour by instance size, storage at $0.04 per GB per month in US regions. As a worked example, a team running about 40 Standard compute hours a month with 500 GB stored lands around $370; heavy batch work on large instances runs into the thousands.
What is a duckling in MotherDuck pricing? A duckling is MotherDuck's compute instance. Five sizes: Pulse ($0.60/hr), Standard ($2.40/hr), Jumbo ($4.80/hr), Mega ($12/hr), and Giga ($36/hr) in US regions. Pulse is elastic and metered per query in Compute Unit seconds; the rest bill wall-clock time per second with a 1-minute minimum. Each duckling is a single machine running one DuckDB process, so you scale up to a bigger duckling, not out to a cluster.
What is cooldown and do I pay for it? After your last query finishes, the duckling stays warm so the next query skips the cold start, and you pay full price for that warm time. Defaults: 1 minute on Standard and Jumbo, 5 minutes on Mega, 10 minutes on Giga, configurable from 1 minute to 24 hours. On steady workloads cooldown is negligible. On spiky ones it can be most of the bill.
Is MotherDuck cheaper than Snowflake? Often, for the workloads it targets. MotherDuck has no edition multipliers, no capacity contracts, and a free tier, and single-node DuckDB is efficient on the small-to-medium data most teams have. But the models differ more than the rates: Snowflake bills credits that vary by edition and region, MotherDuck bills flat hourly rates per instance size. Compare your workload shape, not the sticker prices.
What do MotherDuck AI Units cost compared to calling OpenAI directly? Roughly 3.3x OpenAI's list prices as of July 2026. One AI Unit costs $1.00 and buys 240,000 GPT-5 input tokens ($4.17 per million effective, versus $1.25 list) or 30,000 output tokens ($33.33 versus $10). The same ratio holds for GPT-5-mini and text-embedding-3-small. A normal managed-service margin, but GPT-5 is now a previous-generation OpenAI model, so check current rates before pointing a token-heavy workload at AI functions.
When does MotherDuck stop making sense? When the database is no longer the thing you are missing. MotherDuck is a query engine with storage; ingestion, dashboards, a semantic layer, and AI analysis all come from other tools you buy and wire up yourself. Once you are assembling three or four products around it, the coordination cost usually exceeds the MotherDuck bill, and an all-in-one platform is the better unit of purchase.