Artificial Intelligence

AWS Marked Nova Legacy in March, Not This Week

Reports say Amazon is winding down its Nova models. AWS's own lifecycle table has carried the Legacy dates since March, with end-of-life in September.

Artificial IntelligenceBy Servola Tech Desk2026-07-284 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

AWS Marked Nova Legacy in March, Not This Week

The date on the table was March

Business Insider reported on 28 July that Amazon is winding down most of its flagship Nova line, moving engineers and compute toward a Frontier Model Research group under Pieter Abbeel, with a new flagship model expected at re:Invent later this year. The reporting is accurate. It is also four months behind the only document that carries dates.

AWS publishes a model lifecycle table for Bedrock. It has listed Nova Premier and Nova Sonic as Legacy since 13 March 2026, and Nova Canvas and Nova Reel since 30 March, each with a hard end-of-life date attached: 14 September for the first pair, 30 September for the second. Nova Reel appears twice, at v1:0 and v1:1, both on the same schedule.

None of those rows were hidden. A model on Bedrock sits in one of three states, Active, Legacy or EOL, and the table publishes only the ones already sentenced. An engineer who opened that page in March saw the same dates a reporter described this week as a strategy overhaul.

What Legacy already took away

Legacy is not a warning about the future; it is a set of removals that have already happened. Once a model enters that state you cannot start new fine-tuning jobs against it, and you cannot create new Provisioned Throughput endpoints for it.

Existing commitments survive on narrow terms. Fine-tunes and Provisioned Throughput created before the model went Legacy keep working, and you can still create a new on-demand deployment of a custom model you had already trained. New customers are blocked outright.

The quietest clause deserves reading twice: existing customers may lose access to a Legacy model after 15 days of inactivity. A model you use seasonally, or one sitting behind a service that has been idle since spring, can become unavailable well before its end-of-life date.

So if you fine-tuned Nova Premier for a production workload, the freeze started on 13 March. Every retraining cycle you have planned since then was already impossible. The constraint had simply not been tested yet.

Two September dates, and no published reprieve

End-of-life falls on 14 September 2026 for Nova Premier and Nova Sonic, and 30 September for Nova Canvas and Nova Reel. AWS states that on or soon after the EOL date a model stops being available across all regions and requests to it fail. Migration does not happen automatically; you update your application code or it breaks.

There is normally a cushion. For models with EOL dates after 1 February 2026, a Legacy model enters a public extended access period after at least three months, during which active users can keep it until EOL at pricing the provider sets, which AWS warns will be higher. AWS says it notifies customers of that date when a model first enters Legacy.

The Nova rows carry no published extended access date. The rows beside them do: Claude Sonnet 4 shows July against an October end-of-life, and Claude Opus 4.1 shows October against January 2027. Same table, same column, and Amazon's own models are the ones without a stated fallback window.

Where the roadmap actually lives

Put the lifecycle page on a monthly calendar item, the way you already treat certificate expiry. On Bedrock the state is machine-readable as well: GetFoundationModel and ListFoundationModels return a modelLifecycle field, so the check can be a scheduled job rather than something a person has to remember to do.

European teams are inside this rather than watching it. Nova Canvas and Nova Reel are served from eu-west-1 in Ireland and Nova Sonic from eu-north-1 in Stockholm, so a workload pinned to an EU region for data residency carries the same two September dates with fewer places to fall back to. Nova Premier ran only in US regions, which is its own answer to whether it was ever a candidate for a residency-bound deployment.

The general rule outlasts Nova. A vendor's marketing tells you what it wants to sell you next; its deprecation table tells you what it has already decided to stop supporting, with dates attached. When the two disagree, the table is the one with a contractual consequence behind it.

Frequently asked questions

Which Nova models are affected and when?

On the Bedrock lifecycle table, Nova Premier and Nova Sonic went Legacy on 13 March 2026 with end-of-life on 14 September 2026. Nova Canvas and Nova Reel, at both v1:0 and v1:1, went Legacy on 30 March with end-of-life on 30 September.

Can we keep using a Legacy model until its EOL date?

Generally yes, but on reduced terms. You cannot start new fine-tuning jobs or create new Provisioned Throughput, new customers cannot adopt the model at all, and existing customers may lose access after 15 days of inactivity. Existing deployments created before the Legacy date continue to work.

Does this mean Amazon is leaving frontier models?

No. The reporting describes a consolidation, with resources moving to a Frontier Model Research group and a new flagship expected at re:Invent later this year, possibly still under the Nova brand. What is fixed are the September dates already published for the current models.

Vendors announce what they are launching and publish what they are killing, and only the second one comes with dates you can plan against. The Nova schedule was readable in March by anyone who looked at the table instead of the keynote. Make that table something your team reads on a rhythm, because the next deprecation will arrive the same way.

AmazonAWSAmazon BedrockAmazon Novamodel deprecationvendor lock-inPieter Abbeel

Amazon Borrowed $25 Billion for AI Capacity

Amazon raised at least $25 billion in an eight-tranche bond sale to help fund a roughly $200 billion capex year for AI. The order book was pared and Amazon paid up to fill it. For anyone whose stack runs on AWS, the build-out is now financed with debt, and that is both a commitment and a warning.

3 min read

Artificial Intelligence

Cognizant Has Trained 30,000 Of 350,000 On Claude

Cognizant set out to put Claude in front of 350,000 people. Nine months later 30,000 are trained. What the absorption rate means for anyone buying delivery.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.