Buy Databricks Mosaic AI if the model catalogue decides your roadmap. Buy Snowflake Cortex AI if getting your data back out decides it. Everything else in this comparison is closer than either sales team will admit — and the one number that genuinely swings your bill by 40x is not the platform at all. It is which model tier you point at.
Here is the whole decision on one screen, then the arithmetic behind it.
| Snowflake Cortex AI | Databricks Mosaic AI | Microsoft Fabric AI | |
|---|---|---|---|
| Models you can call | ~30 listed, mostly Claude, Llama, Mistral, plus OpenAI up to GPT-5.2 (docs) | 47 listed, including the GPT-5.6 family, Gemini 3.x, Claude Opus 5, Qwen, GLM (docs) | 2 chat models — gpt-5.1, gpt-5-mini — still in preview (docs) |
| Bring your own weights | Model Registry + container services | Provisioned throughput serves fine-tuned and custom pre-trained models in supported families | Bring-your-own-key to Azure only; nothing hosted on Fabric capacity |
| Billing unit | AI Credits at $2.00 (global) / $2.20 (regional), plus warehouse compute (docs) | DBUs at $0.070 for Model Serving (Flexera, 2026) | Capacity Units off the same meter as your BI, at $0.18/CU-hour list |
| Where the prompt travels | ANY_REGION by default for new accounts created after 9 March 2026 (docs) |
Your workspace region unless you route out through the gateway | Cross-geo processing is a tenant setting; data residency treated as primary constraint |
| Outside engine can write your tables | No — third-party clients cannot append, delete or upsert against Iceberg tables that use Snowflake as the catalog (docs) | Managed Iceberg only — foreign Iceberg and all Delta tables stay read-only (docs) | OneLake shortcuts; Fabric is the write path |
| Writes into a catalog you own | Yes — AWS Glue, Snowflake Open Catalog, Unity Catalog, OneLake, BigLake (docs) | No — foreign catalogs are read and govern only | No |
| AI assets as governed objects | Model Registry under the same RBAC as tables (Horizon) | Models, model services and MCP services are securables (Unity Catalog) | Governed by workspace and capacity, not a data catalogue |
| Verdict | Best exit, narrowest catalogue | Widest catalogue, exit depends on your table format | Loses on both unless you are already all-in |
The Same Job Costs $2,100 or $51,000, and the Logo Barely Matters
Normalise everything to one workload and the platform stops being the variable. Take a job every enterprise actually runs: classify and summarise 10 million support tickets — 800 input tokens and 200 output tokens each, so 8 billion input tokens and 2 billion output tokens, once.
Databricks publishes DBU rates per million tokens on its Foundation Model Serving pricing page, and independent FinOps analysis puts Model Serving at $0.070 per DBU on AWS Premium. Microsoft publishes CU-second rates per 1,000 tokens on Microsoft Learn, which convert at the $0.18/CU-hour US pay-as-you-go list rate (the Fabric pricing page quotes per region on sign-in). Snowflake publishes its AI Credit price in the docs but keeps the per-model rates in a downloadable rate card off the pricing options page; FinOps vendor Finout puts the spread at roughly $0.12 to $5.10 per million tokens.
| Model tier | Platform | List cost for the 10M-ticket job |
|---|---|---|
| Small open-source | Databricks — Llama 3.1 8B | $2,100 |
| Small open-source | Snowflake — cheapest tier | ~$1,200 plus warehouse time |
| Mid open-source | Databricks — Qwen3.5 122B | $6,160 |
| Mid open-source | Databricks — Llama 3.3 70B | $7,000 |
| Small proprietary | Fabric — gpt-5-mini |
$10,080 |
| Large open-weight | Databricks — GLM-5.2 | $20,000 |
| Frontier | Fabric — gpt-5.1 |
~$50,400 |
| Frontier | Snowflake — top tier | ~$51,000 plus warehouse time |
The spread within Databricks is 10x. The spread within Snowflake is roughly 40x. The spread between platforms, holding model class constant, is under 2x. Anyone modelling a platform decision on inference list price is optimising the third decimal place.
One structural caveat that matters more than the rates: Snowflake charges warehouse compute on top of tokens, and its own documentation tells you to use a warehouse no larger than MEDIUM because bigger ones do not go faster — they just cost more. That is an unusually honest line in a vendor doc, and it is the single cheapest optimisation on this page.
Fabric's AI Bill Comes Out of Your Power BI Budget
This is where the third option stops being a serious contender for anything at volume, and it has nothing to do with the per-token rate.
Fabric bills AI functions in Capacity Units drawn from the same capacity that serves your reports. Microsoft's published rates are 42.02 CU-seconds per 1,000 input tokens and 336.13 CU-seconds per 1,000 output tokens for gpt-5.1. Run the 10-million-ticket job through that and you consume roughly 280,000 CU-hours. An F64 capacity provides 64 × 24 = 1,536 CU-hours a day. That single batch job is 182 days of your entire tenant's capacity — and Microsoft's own documentation notes that once capacity is exhausted, all operations shut down.
Even the cheap model does not rescue it: gpt-5-mini on the same job is about 56,000 CU-hours, or 36 days of an F64 running nothing else. On Snowflake or Databricks you would provision more serverless capacity and eat the cost. On Fabric the contention is avoidable — Microsoft's own field guidance is to give AI its own capacity, and workspaces can be assigned to separate F-SKUs — but the meter is not. You have removed the risk of throttling the CFO's dashboard by paying for a second capacity to run the same job at the same CU cost.
Fabric data agents — the natural-language query product — bill at 100 CU-seconds per 1,000 input tokens and 400 per 1,000 output tokens, which works out to $5.00 per million input and $20.00 per million output at list. That is frontier-model pricing regardless of what runs underneath, and the queries the agent generates are billed separately to whichever engine executes them.
None of this makes Fabric a bad product. It makes Fabric a conversational AI product with a capacity meter, not a bulk inference platform. Buy it for a few thousand analyst questions a month. Do not buy it to process a corpus.
Databricks Wins the Catalogue, and the Gap Sits at the Frontier
Count what each vendor's own documentation lists and the gap is real but narrower than the decks suggest — and it concentrates almost entirely at the top of the range.
Databricks Mosaic AI's supported models page enumerates 47 endpoints: the full OpenAI GPT-5.6 family (Sol, Terra, Luna), GPT-5.5 Pro, Gemini 3.6 Flash and Gemini 3.1 Pro, Claude Opus 5 and Sonnet 5, Llama 4 Maverick, Alibaba's Qwen3.5 122B, Zhipu's GLM 5.2, and open-weight GPT OSS 120B. Provisioned throughput extends the same serving path to fine-tuned and custom pre-trained models in supported architecture families — so a model you trained is a first-class citizen, not an integration project.
Snowflake Cortex AI's regional availability page lists roughly 30 for AI_COMPLETE — the exact number moves depending on whether you count preview and legacy entries: nine Claude variants including Opus 5 in preview, Gemini 3.1 Pro in preview, Llama and Mistral open models — and on the OpenAI side, gpt-5.2, gpt-5.1, gpt-5, gpt-5.4-mini, gpt-5.4-nano, gpt-5-mini, gpt-5-nano and gpt-4.1. Full-size OpenAI models are there, up to gpt-5.2. What is missing is the newest generation — GPT-5.5 Pro and the entire GPT-5.6 family. If your evaluation concluded that GPT-5.6 Terra is the only model that clears your accuracy bar, Cortex cannot run it and no amount of procurement pressure changes that this quarter.
Fabric's prebuilt path hosts two chat models and text-embedding-ada-002, a 2022-vintage embedding model, and the whole surface is marked preview. Bring-your-own-key to Azure AI Foundry is the escape hatch, and it is a real one — but then you are buying Azure AI Foundry and using Fabric as a data source, which is a different purchase than the one the Fabric sales deck describes.
The steel-man for Snowflake here is genuine: about 30 models spanning five labs covers the overwhelming majority of enterprise text work, and a shorter list is a shorter security review. Most teams that insist on model breadth end up running three models in production and arguing about the other thirty.
Governance Is Nearly a Tie, So Stop Scoring It
Both vendors have converged, and the RFP row that asks "does it support column-level lineage" no longer separates anyone.
Unity Catalog governs tables, views, volumes, functions, models and services in one namespace, automatically tracks how assets "flow and connect, from source data through to models, services, and dashboards", and treats models and model services as securable objects you grant permissions on. Horizon Catalog does column-level lineage across Snowflake, external databases, BI tools and OpenLineage feeds, versions and governs ML models under the same RBAC as tables with training-data lineage, and enforces masking and row-access policies across any Iceberg REST-compatible engine.
The one live difference is maturity of the AI-traffic layer. Databricks' Unity AI Gateway — rate limits, payload logging to Delta tables, failover across backends, token and latency tracking in system tables — is documented as Beta, with no charge during the beta. Snowflake's equivalent guardrails, including PII and PHI redaction from agent outputs, carry no preview label in Horizon. If you need a governed, logged, rate-limited chokepoint in front of model traffic this quarter, Snowflake is further along. If you need it in 2027, this row will have evened out.
A definition worth pinning down, because both vendors use the word loosely: an AI gateway is the policy chokepoint every model call passes through — it authenticates the caller, applies content and rate policy, logs the request and response, and routes to a backend. If your platform's gateway is in beta, your audit trail is in beta.
The Exit Cost Is Real, but Not the One the Decks Argue About
Everything above is a preference. This is a fact, and both vendors document it themselves — including the part that cuts against Snowflake.
Snowflake supports writes to externally managed Iceberg tables backed by a remote Iceberg REST catalog, including AWS Glue and Snowflake Open Catalog, and catalog-linked databases keep automatically in sync with the remote catalog's namespaces and tables — per Snowflake's own documentation, with real limitations attached (autocommit transactions only, and no CREATE TABLE AS SELECT against AWS Glue or Unity Catalog).
Read the direction carefully, because this is where most comparisons go wrong. That is Snowflake writing outward, into a catalog somebody else owns. Pointed the other way, Snowflake is at least as closed as Databricks: the same documentation states plainly that "third-party clients can't append to, delete from, or upsert data to Iceberg tables that use Snowflake as the catalog." Snowflake's exit advantage is not that its own tables are open to outside writers. They are not. It is that you never have to put the catalog inside Snowflake to begin with.
Databricks' Iceberg REST Catalog API documentation publishes the matrix plainly. External engines can read and write managed Iceberg tables. Foreign Iceberg tables: read, no write. Managed Delta with Iceberg reads enabled: read, no write. External Delta with Iceberg reads enabled: read, no write. The page even warns that "attempting to write from Snowflake to read-only Databricks tables can result in errors."
Read that as a buyer, not an architect. If your estate is Delta tables in Unity Catalog — which is the default outcome of buying Databricks — then every other engine in your company is a read replica. Trino can query it. Snowflake can query it. Neither can write back. The moment a second team needs to write, they need Databricks compute, and the platform decision you thought you made for one workload has quietly become the platform decision for all of them.
The escape hatch is real, and it is a table-format decision rather than a vendor one. Create the table as managed Iceberg instead of Delta and outside engines can write it — Databricks has shipped that path, and Spark, Flink and Trino can use it. Most estates will not be on it, because Delta is the default, the conversion is not free, and nobody chooses a table format in year one on the strength of an exit argument. But "Databricks is read-only to everyone else" is too strong, and any vendor rep will correct you on it in the room.
That is not a scandal. It is a product choice, disclosed in the documentation, and it buys real things: a single write path is easier to make consistent, easier to govern, easier to optimise. But it is the exact mechanism by which a two-year pilot becomes a nine-figure migration, and it belongs in the contract negotiation rather than in a footnote discovered in year three.
Do not source this from either vendor's comparison microsite. Both publish one. Read the two documentation pages linked above instead — they are the vendors describing their own products under their own names, which is the only version that survives a procurement challenge.
Who Should Not Buy Each One
Do not buy Snowflake Cortex AI if your accuracy bar depends on the newest frontier OpenAI generation, if you need to serve your own fine-tuned weights as a first-class endpoint, or if you are a regulated EU or UK institution that has not audited the cross-region default. Snowflake's documentation is explicit that for new accounts created in new organizations in commercial regions after 9 March 2026, ANY_REGION is on by default — inference payloads transit to whichever region serves the model, though no customer data is persisted there. Nothing is stored outside your home region, but your prompts leave it, and "the prompt is transient" is a sentence a data protection officer will want to test before you find out in an audit.
Do not buy Databricks Mosaic AI if a second engine needs write access to Delta tables and you are not willing to standardise on managed Iceberg, if your team's centre of gravity is SQL analysts rather than engineers who are comfortable with clusters and notebooks, or if you need a GA'd, audited AI gateway in front of model traffic before your next compliance review. Also skip it if your inference is genuinely small — provisioned throughput reserves capacity, and GPU serving starts at 10.48 DBUs an hour for a T4-class unit and reaches 628 DBUs an hour for eight A100 80GBs, which is a fixed bill against a variable workload.
Do not buy Microsoft Fabric as your AI platform for anything beyond conversational analytics on data already in OneLake. Two preview chat models, a 2022 embedding model, and an AI meter that competes with your BI reports for the same capacity is not an inference platform. It is a very good analytics platform with AI features attached, and the pricing model tells you which one Microsoft built.
The Four Criteria That Actually Predict Regret
Score these. Ignore the rest of the RFP.
- Does a second engine need to write — and to whose catalog? If your tables are Delta in Unity Catalog, every other engine is a read replica, and standardising on managed Iceberg instead is the fix Databricks already ships. Snowflake's own catalog never opens to outside writers at all; its answer is to leave the catalog outside Snowflake. Decide which of those two shapes you want before you decide the vendor.
- Is your model shortlist inside Snowflake's catalogue? Run your own eval first, then check the regional availability list. If the winner is a GPT-5.6, the decision is made for you.
- Where must the prompt physically be processed? This is a config value on Snowflake, a tenant setting on Fabric, and a routing decision on Databricks. All three are answerable in an afternoon and all three are usually answered wrong by default.
- Who pays when AI usage spikes? On Databricks it is a separate DBU line. On Snowflake it is a separate AI Credit line. On Fabric it comes out of the same capacity as your reports, which means the first person to notice your AI success is the person whose dashboard stopped loading.
What changes the answer: Databricks shipping bidirectional Iceberg writes for Delta and foreign tables would erase the strongest reason to prefer Snowflake. Snowflake adding the GPT-5.6 family would erase the strongest reason to prefer Databricks. Both are plausible within 12 months. Neither has happened, and you are buying today.
What to Do in the Next 30 Days
This week:
- Run
SHOW PARAMETERS LIKE 'CORTEX_ENABLED_CROSS_REGION' IN ACCOUNTon every Snowflake account created since March. If it returnsANY_REGIONand you are under GDPR, UK GDPR or a sector residency rule, decide deliberately whether to keep it — do not inherit it. - Pull your last 90 days of AI spend and split it into input tokens, output tokens and compute. Most teams discover output tokens are 70% of the bill, which makes prompt design a bigger lever than platform choice.
This month:
- Take your top three production prompts and price them on all three platforms using the rates above. Do it in a spreadsheet, not a POC. If the gap is under 2x — it usually is — remove cost from the decision entirely and say so out loud in the steering committee.
- Ask your Databricks account team, in writing, for the ship date on Iceberg REST write support for Delta and foreign tables. The answer, or the absence of one, is worth more than any benchmark in the deck.
Before your next renewal:
- Write the exit clause before you need it. Name the format, the catalogue API, and the maximum acceptable time to stand up a second engine with write access. A vendor that cannot meet that clause has told you the exit cost without you having to model it.
The Bottom Line
Every platform cycle sells the same trade and prices it differently. In 2012 the question was which database; the answer determined your schema for a decade. In 2019 it was which cloud; the answer determined your egress bill. In 2026 it is which lakehouse runs your inference — and the industry has done a remarkable job of making that argument about model catalogues and per-token rates, because those are the two things that will be commoditised fastest.
The catalogue gap closes as fast as vendors can sign model providers. The price gap is already inside the noise. What does not close on its own is who holds the write path to your tables, because that is not a feature gap — it is the business model.
Pick the platform whose documentation tells you how to leave.
Continue Reading
- Pinecone vs Weaviate vs pgvector: Stay on Postgres
- LangGraph vs CrewAI vs AutoGen: One of Them Is Retired
- AI Agents Have Admin Access. 57% of Firms Aren't Ready.
- Databricks $188B: Stop Token Waste, Start AI Governance
- Snowflake Bids for the Agentic Enterprise Control Plane
- Your AI Router Is Trading a 10x Discount for a 2.5x One
- Enterprise AI Costs Drop 67%: Multi-Model Is the New Default
- Snowflake's $6B AWS Bet: Why Agentic AI Stays on Data