06-reference/research

snowflake cortex training preview status

2026-11-05·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)

Cortex Training Entered Public Preview on 2026-10-30, Open Stage, AWS Only, No Named Regions and No Documented Edition Gate

The question

"Has Snowflake Cortex Training actually entered public preview, and in which regions, clouds and editions, per its feature doc preview banner and release notes?"

Context: this closes the first open follow-up from [[2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix]], which could not find any Snowflake release note for Cortex Training and could not confirm region, cloud or edition caveats inside its research caps. The founder needs a ledger he can cite verbatim in a client conversation without being wrong.

Headline: Yes. Cortex Training entered public preview on 2026-10-30 per its own dated Snowflake release note, at the Open preview stage, and the feature doc banner reads "Available in all commercial Amazon Web Services (AWS) regions. Requires enablement." The docs name no individual regions, state no Snowflake edition requirement, and say nothing about cross-region inference. Two primary-source inconsistencies are worth knowing before quoting any of it: the "Open" badge contradicts the same page's "contact your account team to enable" instruction, and Cortex Training does not appear at all in Snowflake's central "Features currently in preview" table.

What we already know (from the vault)

What the web says

All statuses below were observed 2026-11-05 on docs.snowflake.com. The feature doc and the get-started page were read as raw page text (Playwright), so the quoted banner wording is literal, not a fetch summary.

The verified ledger (the five questions, each answered or explicitly unresolved)

# Question Answer as of 2026-11-05 Evidence Confidence
1 Public preview, private preview, GA, or not released? Public preview. Release-note label "(Preview)"; feature-doc badge PREVIEW FEATURE — OPEN. Not GA: the doc explicitly says specifications are "subject to change before general availability". Not private preview: it has a public release note and a public feature doc. Release note 2026-10-30; Cortex Training overview page High
2 As of what date, per which release note? 2026-10-30, per the release note "Oct 30, 2026: Cortex Training (Preview)" at /en/release-notes/2026/other/2026-10-30-cortex-training-preview. Same, plus the current feature-updates index High
3 Which clouds, which named regions? AWS only. Verbatim: "Available in all commercial Amazon Web Services (AWS) regions." No Azure, no Google Cloud. No individual region is named anywhere - not on the feature doc, not on the get-started page, and not on the Cortex AI model-availability matrix, where Cortex Training is not a column. Whether "commercial" excludes AWS US Commercial Gov, FedRAMP High Plus and DoD regions is not stated on these pages. Feature doc banner; release note; model-availability page (absence) High on "AWS only"; the per-region list is a documented absence, not a finding
4 Which editions? Any Enterprise-or-above gate? The docs do not state this. No edition requirement appears on the Cortex Training overview or get-started pages. The only account-side prerequisites documented are an AWS-hosted Snowflake account, enablement by the account team, and a PAT. Not evidence that Standard Edition works - evidence that Snowflake has not published a gate. Overview + get-started pages (keyword sweep returned zero hits for edition/Enterprise/Standard) High that the docs are silent; unresolved on the actual entitlement
5 Any cross-region inference requirement? The docs do not state this. Neither page mentions cross-region inference or CORTEX_ENABLED_CROSS_REGION. Cortex Training is also absent from the cross-region scope columns on the model-availability page, which cover AI Functions, Cortex Inference, Cortex Agents / CoWork, Agent Evaluations, CoCo and Cortex Search. Plausibly moot, since training runs in-region on AWS rather than routing inference across regions, but that is inference and not documented. Overview + get-started pages; model-availability page High that the docs are silent; the reason is unverified

Convergences and contradictions

Synthesis for RDCO

The one-line version the founder can say out loud: "Cortex Training went to public preview on October 30, 2026, AWS commercial regions only, and it needs Snowflake to switch it on for your account." That sentence is fully backed by two primary sources with a date, which is the standard the parent brief set for anything said in a client conversation. Everything beyond it - which specific regions, which edition, whether cross-region inference matters - should be said as "Snowflake has not published that," not guessed. The negative findings are the valuable half of this brief, because they are the sentences a Deal Solutions Architect (DSA) would otherwise improvise under pressure. A client on Azure or Google Cloud is a hard no today, and that is the single most useful disqualifier here: it removes Cortex Training from scope for a large share of accounts with one documented line rather than a procurement cycle.

For a Statement of Work (SOW), the "Requires enablement" clause is the load-bearing risk, not the preview label. A preview dependency is a known pattern and can be papered over with a preview-dependency clause. An account-team enablement gate is different: it puts a Snowflake-internal approval on the critical path of a phData deliverable, with no published eligibility criteria, no stated turnaround, and no self-service path. Any pencilled Cortex Training scope needs enablement confirmed in writing on the client's account before the scope is signed, not after. The practical sequencing is: client asks their Snowflake account team to enable Cortex Training, enablement is confirmed, then phData scopes. Reversing that order is how a fine-tuning workstream stalls on something phData cannot escalate.

The architecture is a real differentiator worth understanding, even though it is not sellable yet. Cortex Training's split - client-side training loop on a laptop or CPU-only virtual machine, server-side GPU work on Snowflake, with weights never downloaded - is a genuinely different shape from both Cortex Fine-tuning (a SQL function, generally available, narrowing in region coverage) and from bring-your-own-GPU training in Snowpark Container Services (SPCS). It targets teams who already have a training script and want governed GPUs without moving data, which is a credible wedge for the "agents in production" positioning and a better conversation than a feature-availability debate. The per-second AI Credits metering with no minimum also makes a sandbox experiment genuinely cheap to propose. That argues for a hands-on experiment in an internal or sandbox account once enablement is available, specifically to have a first-hand answer about the edition and enablement questions the docs refuse to answer.

Correction discipline on the vault. The parent matrix row for Cortex Training now reads as stale rather than wrong: it said "unconfirmed public preview" on 2026-09-14 and that was accurate then. This brief supersedes that single row as of 2026-11-05, and should be the row cited going forward. The rest of the parent matrix is untouched and still the reference for the CoWork and Cortex Sense items.

Why this is in the vault

It closes the first open follow-up of [[2026-09-14-snowflake-cowork-cortex-ga-vs-preview-matrix]] with a primary-source, date-stamped answer, so the founder's phData DSA work can name Cortex Training's status, cloud restriction and enablement gate in a client conversation or SOW review without reaching for a third-party explainer. It also records two documented silences (edition, cross-region inference) so nobody fills them in by inference later.

Open follow-ups

Related

Sources

Vault:

Web (all observed 2026-11-05):

No paywalled sources were encountered. Method notes: WebFetch summarised the release-note body under a quote-length limit, so the release-note wording above is a close paraphrase with short exact quotes; the feature doc, get-started, preview-features and model-availability pages were read as raw page text, so quotes from those four are literal. "The docs do not state this" means the keyword sweep over those raw page texts returned no hits on 2026-11-05; it is not proof that no Snowflake document anywhere states it.