Amazon Nova Forge Turns Your AWS Spend Into a Proprietary AI Asset Without Sending Your Data Anywhere
The Problem Nova Addresses
Most enterprise AI deployments are not successful, as general-purpose models are trained on everything, which means they are not optimized for that organization. A frontier large language model knows a great deal about the world but relatively little about your customer base, your deal cycle, or your churn patterns and needs high customization and data science resources to make that work for the organization. This results in a pattern where organizations deploy general AI tools against business problems, get unreliable outputs, and lose confidence in the investment. The issue is rarely the model. It is that the model was never built for the specificity the problem requires.
Nova's custom training approach inverts this. You bring your data. AWS supplies the infrastructure, compute, and training pipeline. The resulting model is scoped narrowly to the problem you defined. A mid-sized distributor trying to predict which leads convert, or a services firm modeling churn, does not need a frontier model. It needs one that knows its own customers.
How the Training Workflow Works
The workflow, designed for organizations without a dedicated ML research team, has a narrow use case. AWS has made deliberate choices to reduce the operational burden at each stage, abstracting away infrastructure management while preserving the control enterprises need.
Data preparation begins in S3, where most AWS enterprise customers already store operational data. Nova ingests training datasets directly from S3 buckets, which eliminates the export-and-reformat step that has historically been a friction point in managed fine-tuning workflows. Customer data stays within the customer's S3 environment and VPC throughout the process and is not used beyond their own training run. Organizations with structured CRM exports, transaction logs, or historical customer records in S3 can move from data to training with minimal preprocessing.
The training pipeline runs in four discrete phases: continued pre-training, mid-training, supervised fine-tuning, and reinforcement learning. Each phase produces a checkpoint, and AWS advises evaluating the model after each one before proceeding. Because each phase runs as a separate SageMaker job with observable compute usage, teams have cost visibility at every stage rather than committing to a single open-ended run. For IT buyers managing AI budgets, this structure makes the investment easier to govern and easier to present to finance.
Once training is completed, the resulting model, which Amazon calls a Novella, is deployed as a private endpoint on Amazon Bedrock, visible only to the customer's AWS account and deployed as a private Bedrock endpoint accessible only to the customer’s account. For regulated industries, this closed architecture is a meaningful compliance advantage.
Accuracy and Data Scope
Because the model is trained exclusively on customer-supplied data and does not retrieve from external sources at query time, its outputs are tightly scoped to what you taught it. For use cases where reliability and auditability matter, such as financial forecasting, compliance-adjacent analytics, and customer scoring, that is a meaningful advantage. Prediction errors trace back to training data rather than to opaque retrieval behavior, which makes the model easier to validate internally and easier to present to compliance teams.
The quality of results is directly tied to the quality of the training data. Organizations that invest in data labeling and validation before initiating training get better outcomes. The staged checkpoint design has been built in part to support this: the phased structure gives teams natural stopping points to review data quality before committing to the next phase of compute spend.
IT teams should plan for data preparation as a distinct workstream, not a preliminary step. The time spent there directly determines how well the trained model performs against the target use case.
Key Terms
Nova Forge is Amazon's branded custom model training service within the Nova family, introduced at AWS re:Invent 2025.
A Novella is the term Amazon uses for a custom model produced through Nova Forge. It is a private model artifact trained on the customer's data and hosted exclusively within their AWS account on Amazon Bedrock.
Fine-tuning adjusts the weights of an existing pre-trained model. Nova Forge goes further, allowing data to be introduced at earlier stages of the training process, which produces deeper domain integration than post-training fine-tuning alone.
S3 integration means training datasets are ingested directly from Amazon S3 within the customer's own environment. Data does not leave the customer's VPC during training.
Dedicated endpoint refers to the deployment model: the Novella runs as a private Bedrock endpoint accessible only to the customer's AWS account. AWS documentation does not describe a weight export mechanism.
Who This Is Built For
Customers pay an annual subscription fee ... Learn more about the subscription pricing in the Nova Forge console. AWS positions their pricing as substantially cheaper than building an equivalent model from scratch, which would require dedicated GPU infrastructure and research staff at far greater expense.
The pricing contrast with Nova's standard services is significant. Using Nova models via Bedrock is consumption-based, billed per token with no upfront commitment. Nova Act agents, which automate tasks using Nova models, are priced at $4.75 per agent-hour. For organizations with modest or experimental AI needs, those paths are far more cost-effective. Nova Forge's $100,000 entry is justified when the use case demands deep domain expertise baked into the model itself, not a general model applied to a narrow task.
The fee effectively limits the buyer pool to three profiles. The first is the data-rich enterprise on AWS: a company that has been accumulating customer transactions or operational data for years, runs its infrastructure on AWS, and has a specific prediction problem it cannot solve adequately with general tools. For this buyer, Nova Forge's activation cost is lower than introducing a new vendor because the data and procurement relationship already exists.
The second is the regulated industry buyer – financial services, healthcare, and legal technology companies that need AI output they can audit and attribute. A model with traceable training provenance is easier to defend to regulators and internal compliance teams than a general model whose retrieval behavior they cannot inspect.
The third is the enterprise with existing AWS committed spend. Organizations with large AWS agreements often have flexibility to apply that spend to new services. For those buyers, Nova Forge may fit within an existing financial commitment rather than requiring a new budget line.
Smaller organizations with modest datasets and broader AI needs are better served by Nova's standard consumption-based agents. The custom training tier is not a better version of the standard product. It is a different product for a different problem.
Competitive Landscape
AWS is not alone in this space, and the competitive framing matters for organizations evaluating alternatives.
Databricks MosaicML is the closest direct comparison for organizations that want custom model training with their own data. MosaicML integrates well for teams already running data pipelines on the lakehouse architecture, with tight Delta Lake and MLflow integration. It supports fine-tuning and continued pre-training of open-source models and, unlike Nova Forge, model weights can be exported and run outside the platform. The key difference is that MosaicML cannot offer what Nova Forge calls mid-training injection – the ability to introduce proprietary data partway through a foundation model's training cycle. AWS describes this as unique to Nova Forge among managed providers. For AWS-native organizations, MosaicML also adds a separate vendor relationship and procurement step that Nova Forge does not.
Abacus.ai targets a broader buyer profile with a low-code, fully managed approach that requires less technical overhead to get started. Its entry pricing is lower than Nova Forge's subscription fee, which makes it a reasonable option for organizations that want to evaluate AI capabilities before committing to a larger investment. At enterprise scale, however, Abacus uses a credits model that some customers find difficult to forecast, and it does not offer the kind of foundation model training that Nova Forge or MosaicML provide. Abacus.ai platform includes model fine-tuning and custom model capabilities.
Azure ML offer comparable capabilities within their respective clouds. Organizations running hybrid or multi-cloud environments may find the vendor-neutral framing of those platforms more flexible, though neither has simplified the workflow for non-engineering teams to the same degree AWS has with Nova.
The AWS advantage is reach and integration depth. If your data already lives in S3, your security model is built around IAM, and procurement runs through AWS – the activation cost of Nova Forge is lower than introducing a new vendor relationship.
Our Take
Nova Forge is a practical on-ramp for enterprises that have been sitting on proprietary data without a clear path to acting on it. The combination of S3-native ingestion, phased training with cost visibility at each stage, and a private Bedrock endpoint removes three barriers that have historically made custom model training inaccessible outside organizations with dedicated ML research teams.
The clearest fit is a defined prediction problem with historical data to support it. A Novella trained on your domain will outperform a general model for that specific task, and the output is easier to audit and explain to leadership than a broad AI deployment. For regulated industries particularly, the closed architecture, no external data at inference, and full training lineage within your AWS account is a compliance argument that general AI tools cannot match.
The phased pipeline design is worth highlighting internally. Because each training stage is a discrete SageMaker job, teams can evaluate model quality and observe costs before committing to the next phase. For organizations building the case for AI investment, this structure makes the governance conversation straightforward.
Organizations new to custom model training should assess data readiness and problem definition before engaging. The consumption-based Nova agents are the right starting point for teams that want to build familiarity with the platform before committing.
Want to Know More?
Amazon Bets Big on Agentic AI Development With $50 Billion OpenAI Investment
Amazon's Bedrock: Unlocking the Gateway to Powerful Open LLMs
Amazon Throws Its “Cat” Into the Quantum Computing Ring With Ocelot Announcement