Maker

The infrastructure Amazon operates in AI

Amazon occupies two separate layers in this reference. It builds Nova itself, seven models spanning text, image, video and speech, and through Bedrock it also deploys almost every other vendor on its own infrastructure. Someone who says "we use Amazon AI" usually means the second layer, and the two share almost no architecture.

  • Seven models in the Nova line
  • A distribution layer for other vendors
  • Access by deployment region

Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log

Model families

The Nova line, from the text-only Micro unit to the speech unit Sonic.

Versions

Versions Deployed Status Model families
Amazon Nova Canvas current Amazon Nova
Amazon Nova Lite current Amazon Nova
Amazon Nova Micro current Amazon Nova
Amazon Nova Premier current Amazon Nova
Amazon Nova Pro current Amazon Nova
Amazon Nova Reel current Amazon Nova
Amazon Nova Sonic current Amazon Nova

Amazon is two layers here, and users treat them as one

When someone says Amazon AI, they could mean two entirely different layers. One is Nova: the seven models Amazon built itself. The other is Bedrock: the layer that deploys other vendors models on Amazon infrastructure.

And this layer is large. The Bedrock pricing page carries a separate pricing section for AI21 Labs, Amazon Nova, Amazon Titan, Anthropic, Cohere, DeepSeek, Luma AI, Meta, Mistral AI, OpenAI, Qwen, Stability AI, TwelveLabs and Writer. That is the largest number of vendors consolidated under a single infrastructure anywhere in this reference.

The practical consequence: if you read that a company was built on Amazon AI infrastructure, there is a good chance the model unit that actually runs the workload is not an Amazon product at all. Claude from Anthropic, Llama from Meta and Command from Cohere are all deployed on that same Bedrock.

200 languages in the specification cell, fifteen in the footnote

The Nova specification table has a row labeled Supported Languages and registers 200+ for all four text models. That cell carries a footnote marker, and the footnote text reads: optimized for these 15 languages: English, German, Spanish, French, Italian, Japanese, Korean, Arabic, Simplified Chinese, Russian, Hindi, Portuguese, Dutch, Turkish, and Hebrew.

Arabic is there. Turkish is there. Persian is not.

This is exactly the pattern we documented in the audio category too: the large figure sits in the cell the consumer reads, the real list sits in a footnote they do not. The 200+ means the model operates with those languages. The 15 means Amazon commits to their quality. For Persian text you are in the first group, not the second, and that is a distinction to know before you estimate cost, not after.

The speech unit in this line, Nova Sonic, is documented more plainly: five languages, English US and UK, French, Italian, German and Spanish. No guessing is required here.

Access in this architecture is a list of data centres, not a list of countries

Every other vendor in this reference publishes a supported-countries list. Amazon publishes a deployment-region list instead, and that is a real architectural difference: the model runs in a specific data centre and your network path has to reach that data centre.

Nova Premier is deployed in only two data centres: US East (N. Virginia) and AWS GovCloud (US-West). Pro, Lite and Micro are deployed in N. Virginia, Tokyo, Sydney, London and GovCloud, and through cross-region inference also reach Ohio, Oregon, Stockholm, Ireland, Frankfurt, Paris, Tokyo, Singapore, Sydney, Seoul and Mumbai. Canvas and Reel are deployed in N. Virginia, Ireland and Tokyo; Sonic in N. Virginia, Stockholm and Tokyo.

Not one of those lists registers a Middle East region. Amazon has Middle East regions in its infrastructure, but Nova is not served from them. The nearest data centre where a Nova model actually runs is Mumbai, or Frankfurt.

On Iran specifically: Amazon publishes no citable supported-countries list, and the AWS customer agreement assigns sanctions compliance to the customer. So we log the status as neither open nor blocked. What we can say is that we found no citable evidence and did not test the network.

The strongest unit in the line is the most restricted one

Nova Premier has a one-million-token context window and a ten-thousand-token output ceiling. That ratio, a hundred to one, is the widest in this catalogue, and the figure itself signals the workload the model was engineered for: very high input volume, short output.

But the same specification registers three other constraints few people check. Premier has no fine tuning. It has no provisioned throughput. And it is deployed in only two regions. So the strongest unit in the line is also the least reachable and least flexible one. What Amazon has defined it for is written into the same specification: Premier serves as the teacher, training smaller units through distillation.

If you want a unit to actually deploy into production, Pro and Lite are the practical choices in this line, not Premier.

What is not documentable in this architecture

The price of any Nova unit. The Bedrock pricing page has the tables but injects the numbers client-side, and the HTML our server receives carries a placeholder instead of the actual figure. There is no number to extract, so the price column for Nova models in our list stays empty.

We log this because it is a real problem for anyone trying to estimate cost before signing up, and because this is the second time we have hit this exact pattern in this reference. Google plan pricing page exhibits the same behavior.

Documented strengths

  • The largest number of vendors under a single infrastructure: the Bedrock pricing page has a separate section for fourteen model families
  • A one-million-token context window on Nova Premier, with language count, deployment regions and model ids all documented in one public table
  • Arabic and Turkish are both registered in the list of fifteen committed languages
  • The documentation registers both the API id and the available regions for every model, which is required to estimate latency

Known weaknesses and limits

  • Persian is not registered in the list of fifteen optimized languages. The 200+ figure in the cell is not a commitment; the footnote is what defines the commitment level.
  • No Nova unit is served from a Middle East region, even through cross-region inference.
  • Prices on the Bedrock page are built client-side and cannot be extracted from the page HTML, so the Nova price column stays empty.
  • Nova Premier has neither fine tuning nor provisioned throughput, so it is not suited to scalable production.
  • Nova Sonic supports only five languages and Persian is not among them.
  • We found no citable supported-countries list for Bedrock, so on access from Iran we neither confirm nor deny.

Technical verdict

If you are deployed on AWS infrastructure and want several vendors models under one bill, Bedrock is the most sensible option in this reference. But do not treat Amazon AI as equivalent to Amazon own models: Nova registers no commitment on Persian text, and for Persian you should look at a vendor that has documented its language list.

Frequently raised questions, with documented answers

Does Nova support Persian

Amazon registers 200+ languages in the table cell and names fifteen optimized languages in that cell footnote. Persian is not one of the fifteen. So the model operates with Persian, but Amazon has registered no commitment on its quality.

What is the difference between Nova and Bedrock

Nova is the seven models Amazon built itself. Bedrock is the layer that deploys those seven units alongside models from Anthropic, Meta, Mistral, Cohere and others on one infrastructure. One name, two architectures.

Is Nova reachable from Iran

We do not know and we register no claim. Amazon publishes no citable country list. What is documented is that no Nova unit is served from a Middle East region, and the nearest data centres are Mumbai and Frankfurt.

What does Nova cost

We cannot document it. The Bedrock pricing page fills in numbers with JavaScript and no figure exists in the page HTML. An empty price column means there was no data to extract.

Sources

  1. Amazon, Nova user guidevendor sourcedocs.aws.amazon.comread on 28 August 2026
  2. Amazon, Bedrock overviewvendor sourceaws.amazon.comread on 28 August 2026
  3. Amazon, Bedrock pricingvendor sourceaws.amazon.comread on 28 August 2026
  4. Amazon, AWS customer agreementvendor sourceaws.amazon.comread on 28 August 2026
  5. Amazon, AWS service termsvendor sourceaws.amazon.comread on 28 August 2026