Comparison · Databricks

Databricks AI Gateway alternative: what to use

Databricks now documents AI Gateway as Unity Gateway and built it around Unity Catalog. That is excellent if your data and identities already live there, and expensive if they do not.

Short answer. Stay with Databricks Unity Gateway when Unity Catalog is your control plane and you need grants, service policies, inference tables and budgets on one platform. Consider Plugsky for a managed API without DBU math; Snowflake Cortex AI Gateway if your data is in Snowflake; Microsoft Foundry, AWS Bedrock or Google's platform for hyperscaler governance; OpenRouter for breadth. The decision is usually about where data and identities live.

What Databricks Unity Gateway actually is

Databricks renamed AI Gateway to Unity Gateway. It is a governance layer on Unity Catalog: models, services, agents, MCP servers and tools are securables, granted with the same privileges as tables. It serves Databricks-hosted models pay-per-token, connects external providers with BYOK, and adds rate limits, traffic splitting, fallbacks, budgets, service policies, inference tables and cost observability.

CapabilityWhat it gives you
Access controlUnity Catalog grants on models, services, MCP tools and functions
PolicyService policies for requests and responses
TrafficRate limits, traffic splitting, fallbacks
ObservabilityInference tables in Delta, usage tracking, cost attribution
ProvidersDatabricks-hosted models plus external providers with BYOK

Where Unity Catalog wins, and where it creates friction

Unity Catalog wins when data is in Delta Lake, identities are in Databricks and security understands grants. A model becomes another governed asset. Lineage, audit and cost attribution come from the same platform, and inference tables give you request history without a separate logging pipeline.

Friction appears when the app or data lives elsewhere. Requests still reach Databricks and answers travel back. DBU pricing is less intuitive than per-token, especially with provisioned throughput and inference tables. The workspace is coupled to a cloud and region, and the catalogue is narrower than an aggregator's. If your app is on AWS and data is in Snowflake, Databricks may be a third control plane.

The six alternatives

OptionGovernance centerData movementBillingBest fit
PlugskyWorkspace RBAC, SSO, audit exportAPI call; private deployment optionalFlat $8–$120/monthManaged multi-model API without DBUs
Snowflake Cortex AI GatewaySnowflake roles, USAGE/MONITOR, tracesIn-perimeter inferenceAI Credits per tokenData already in Snowflake
Microsoft FoundryEntra ID, RBAC, Azure PolicyAzure tenant and networkAzure consumptionMicrosoft-centric enterprises
AWS BedrockIAM, Guardrails, PrivateLinkAWS account and regionAWS consumptionAWS-native governance
Google Gemini Enterprise Agent PlatformGoogle Cloud IAM, Model GardenGoogle Cloud projectGoogle Cloud consumptionGoogle Cloud data and models
OpenRouterWorkspace keys and policiesProvider-dependentList price + credit feeCatalogue breadth and fallback

Microsoft Foundry is the current name for the Azure AI platform, with Entra identity, RBAC, content filters, network isolation and Azure Policy. AWS Bedrock Guardrails cover content filters, denied topics, word filters, sensitive information, contextual grounding and automated reasoning. Google's platform provides Model Garden, model-as-a-service, safety filters and evaluation. Snowflake's Cortex AI Gateway runs inference inside its perimeter. Plugsky is the flat-bill managed option with 39 models.

Cost: DBUs without the mystery

Databricks prices serving in DBUs per million tokens, multiplied by a region-specific dollar-per-DBU rate. At $0.07 per DBU, a model consuming 42.857 DBU per million input tokens costs $3.00 per million input tokens. Provisioned throughput at 142.857 DBU per hour costs $10.00 per hour. Regional processing can add a 10% uplift, and inference tables are charged in 1KB increments.

Convert DBUs to per-token and per-hour numbers before comparing. Add the warehouse or compute that calls the model, storage for inference tables and egress. A flat plan can be cheaper at steady volume; a per-token aggregator can be cheaper at low volume.

Data movement and residency

Ask where the request is processed, where responses and logs are stored. Databricks can write inference tables to Delta; that helps audit, but prompts and responses may be persisted. Set retention and access deliberately. If regional processing is on, confirm the uplift, region map and whether failover crosses a border.

For an external app, estimate egress and latency of sending every request into the lakehouse. For a Databricks app, estimate the cost of moving data out. Snowflake makes the same trade in reverse: in-perimeter inference reduces movement but ties you to its catalogue and AI Credit pricing. The right answer is usually the platform that already holds the data and identities.

Migration

Inventory every model service, external provider, grant, service policy, budget and inference table. Map Unity Catalog grants to roles in the target platform, then rebuild the policies as code. Export the inference tables you need for audit, and decide how much history to keep. Test OpenAI-compatible chat, streaming, tool calling and structured output before you move production traffic. Cut over by workload, keep the Databricks path as a fallback, and compare cost per completed task for at least one billing cycle.

For the governance model, see the enterprise LLM governance framework and the Enterprise overview. For the Snowflake comparison, see Snowflake Cortex alternative. The point is not to replace Unity Catalog with a worse catalogue; it is to avoid paying for a control plane you do not use.

FAQ

Did Databricks rename AI Gateway?

Yes. Current Databricks documentation describes AI Gateway as Unity Gateway, built on Unity Catalog. The underlying capabilities — grants, service policies, rate limits, inference tables and cost observability — are documented under the new name.

Is Unity Gateway a model provider?

It is a governance and routing layer. Databricks serves foundation models natively and also connects external providers with BYOK. The gateway controls access and traffic; the model still runs on Databricks or the external provider.

How is Databricks foundation model serving priced?

In DBUs per million tokens or per hour, multiplied by a region-specific dollar-per-DBU rate. Provisioned throughput is sold in model units, with one-month and three-month reservations. Convert the DBU factors to per-token and per-hour numbers before comparing.

What is the best alternative if our data is in Snowflake?

Snowflake Cortex AI Gateway keeps inference inside the Snowflake perimeter and bills in AI Credits per token. It reduces data movement but ties you to Snowflake's model catalogue and credit pricing.

How do we migrate off Unity Gateway?

Inventory model services, grants, service policies, budgets and inference tables. Map grants to the target, rebuild policies as code, export audit history, test OpenAI compatibility and cut over by workload with a fallback.

Sources

Primary sources checked 20 November 2026. Provider terms, limits and prices change; verify the current version before you commit.

  • Databricks Unity Gateway documentation — describes Unity Catalog governance for models, agents, MCP servers and tools, plus grants, service policies, rate limits, traffic splitting, budgets and inference tables.
  • Databricks foundation model serving pricing — lists DBU factors for pay-per-token and provisioned throughput, the DBU conversion formula and regional processing uplift.
  • Snowflake Cortex AI Gateway — describes the single governed endpoint, USAGE/MONITOR privileges, trace logging and AI Credit billing.
  • Snowflake AI pricing — lists AI Credit prices of $2.00 for global routing and $2.20 for regional routing, separate from platform credits.

Reviewed 20 November 2026. Databricks renamed AI Gateway to Unity Gateway in current documentation. DBU rates vary by cloud and region; confirm the current rate card and your order form.