AI News Deep Dive

Amazon Web Services: AWS Unveils Nova Forge for Custom AI Training

At AWS re:Invent 2025, Amazon launched Nova Forge, a platform enabling enterprises to train AI models from scratch using their proprietary data with simplified workflows. This addresses previous complexities in custom model development, integrating seamlessly with AWS infrastructure for scalable deployment. The tool supports multimodal data and aims to accelerate AI adoption across industries.

👤 Ian Sherk 📅 December 04, 2025 ⏱️ 10 min read

AdTools Monster Mascot presenting AI news: Amazon Web Services: AWS Unveils Nova Forge for Custom AI Tr

As a developer or technical decision-maker grappling with the complexities of custom AI model training, imagine slashing the time and expertise needed to infuse your proprietary data into frontier-level models—without sacrificing scalability or control. AWS's Nova Forge promises exactly that, empowering you to build tailored AI solutions that drive competitive edges in your industry, from healthcare diagnostics to financial forecasting.

What Happened

At AWS re:Invent 2025, held December 1-5 in Las Vegas, Amazon Web Services unveiled Nova Forge, a groundbreaking platform designed to simplify the creation of custom AI models from scratch. This service allows enterprises to train frontier models using their own proprietary data, starting from early model checkpoints provided by AWS's Nova foundation models. Key features include streamlined workflows that integrate seamlessly with Amazon SageMaker for training jobs, support for multimodal data (text, images, and more), and optimization for AWS's custom silicon like Trainium3 chips for cost-effective scaling. Nova Forge addresses longstanding barriers in custom AI development, such as data preparation and hyperparameter tuning, by providing pre-configured pipelines that accelerate the process from weeks to days. The announcement highlights its general availability, with early access programs for select customers to build optimized variants of the new Nova 2 models now in Amazon Bedrock [source](https://aws.amazon.com/blogs/aws/introducing-amazon-nova-forge-build-your-own-frontier-models-using-nova/). Press coverage from TechCrunch notes it as a response to growing demand for enterprise-grade customization amid the AI boom [source](https://techcrunch.com/2025/12/02/all-the-biggest-news-from-aws-big-tech-show-reinvent-2025/), while official docs emphasize its role in early-stage training infusion for superior model performance [source](https://docs.aws.amazon.com/sagemaker/latest/dg/nova-forge.html).

Why This Matters

For developers and engineers, Nova Forge democratizes access to high-end AI training, reducing the need for specialized ML teams by automating complex orchestration on AWS infrastructure—this means faster iteration on domain-specific models without vendor lock-in risks. Technically, it leverages Nova's architecture for multimodal capabilities, enabling robust handling of diverse datasets and deployment at petabyte scale via Elastic Fabric Adapter networking, potentially cutting costs by up to 50% compared to traditional GPU clusters [source](https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-nova-forge-frontier-models-nova/). Business-wise, technical buyers in enterprises gain a strategic tool to accelerate AI adoption, fostering innovation in regulated sectors like finance and pharma where data privacy is paramount. By embedding proprietary knowledge early, organizations can achieve defensible AI moats, boosting ROI on cloud investments and outpacing competitors reliant on off-the-shelf models [source](https://www.aboutamazon.com/news/aws/aws-re-invent-2025-ai-news-updates).

Technical Deep-Dive

Amazon Nova Forge, unveiled at AWS re:Invent 2025, is a managed service integrated with Amazon SageMaker AI that enables developers to build custom "frontier" AI models—dubbed "Novellas"—by leveraging early checkpoints from Amazon's Nova 2.0 foundation models. Key features include access to pre-training, mid-training, and post-training checkpoints, allowing users to infuse proprietary data early in the training pipeline for domain-specific reasoning. This addresses limitations in traditional fine-tuning, where models often lose foundational intelligence when adapting to niche datasets. Capabilities extend to advanced data mixing, blending user data with Nova's curated corpora, and Reinforcement Fine-Tuning (RFT) using custom reward functions executed in the user's environment. Additional controls for custom content moderation ensure compliance and safety tailoring.

Technically, Nova Forge builds on SageMaker's distributed training infrastructure, supporting continued pre-training (CPT) beyond standard methods. Developers initiate jobs via SageMaker Training Jobs, specifying checkpoints like Nova 2 Lite or Pro as starting points. The process involves data ingestion through SageMaker's processing pipelines, where proprietary datasets (e.g., enterprise documents, codebases) are mixed at ratios defined by users—up to 50% custom data in early stages to embed domain ontology without catastrophic forgetting. Implementation uses Trainium and Inferentia chips for cost-efficient scaling, with automatic sharding for datasets exceeding petabyte scale. Evaluation is handled via YAML-configured "recipes" that define metrics like perplexity, BLEU scores, or custom RLHF alignments. For instance, a basic CPT job can be launched with the AWS SDK for Python (Boto3):

import boto3
sagemaker = boto3.client('sagemaker')
response = sagemaker.create_training_job(
 TrainingJobName='nova-forge-cpt-job',
 AlgorithmSpecification={
 'TrainingImage': '123456789012.dkr.ecr.us-west-2.amazonaws.com/nova-forge:latest',
 'TrainingInputMode': 'File'
 },
 InputDataConfig=[{
 'ChannelName': 'training',
 'DataSource': {'S3DataSource': {'S3Uri': 's3://my-bucket/custom-data/'}},
 'ContentType': 'text/csv'
 }],
 OutputDataConfig={'S3OutputPath': 's3://my-bucket/output/'},
 ResourceConfig={
 'InstanceType': 'ml.trn2.48xlarge', # Trainium2 instances
 'InstanceCount': 8,
 'VolumeSizeInGB': 1024
 },
 EnableManagedSpotTraining=True,
 CheckpointConfig={
 'LocalPath': '/opt/ml/checkpoints',
 'S3Uri': 's3://my-bucket/checkpoints/'
 },
 StoppingCondition={'MaxRuntimeInSeconds': 86400}
)
print(response)

This code provisions a distributed training job, checkpointing intermediates for iterative refinement. Benchmarks for base Nova 2 models show strong performance: Nova 2 Pro outperforms Claude Sonnet 4.5 on 10/16 standard evals (e.g., MMLU, GSM8K), while Nova 2 Lite matches or exceeds Claude Haiku 4.5 on 13/15, with 2-5x lower latency on Inferentia. Custom Novellas inherit this, with reported 20-40% gains in domain-specific tasks like legal reasoning after data blending, though exact metrics vary by dataset quality.

API availability is immediate via AWS SDKs (Python, Java, JS) and CLI, with full documentation in the AWS Nova User Guide, including request/response schemas for endpoints like CreateNovaForgeJob and DescribeCustomModel. RESTful APIs support asynchronous job management, with WebSocket streaming for real-time logs. Integration with Bedrock for inference and SageMaker JumpStart for one-click deployment simplifies pipelines. Enterprise options include private VPC endpoints, SOC 2 compliance, and scalable quotas up to 1,000 concurrent jobs.

Pricing is usage-based: $0.50-$2.00 per Trainium hour (depending on instance), plus $0.10/GB for data processing—up to 70% cheaper than GPU alternatives. A $100K annual commitment tier unlocks priority access to checkpoints and dedicated support. Developer reactions highlight the "10-50x productivity" potential for proprietary model building, though some note initial setup complexity for non-AWS natives [source](https://aws.amazon.com/blogs/aws/introducing-amazon-nova-forge-build-your-own-frontier-models-using-nova/) [source](https://docs.aws.amazon.com/nova/latest/nova2-userguide/nova-forge.html) [source](https://www.aboutamazon.com/news/aws/aws-agentic-ai-amazon-bedrock-nova-models).

Developer & Community Reactions ▼

Developer & Community Reactions

What Developers Are Saying

Developers and technical users in the AI community have expressed enthusiasm for AWS Nova Forge, viewing it as a significant step toward customizable frontier models. Vikram Shenoy, a Senior Software Developer at Amazon AGI, shared his pride in the project: "Nova Forge is now available! Extremely proud to have contributed to this and grateful to have worked with some incredible engineers, managers and product teams on this one. Can’t wait to see what you build with Nova today!" [source](https://x.com/vikramshenoy97/status/1995926620776136939). Similarly, Fardin Abdi from Amazon AGI labs highlighted its enterprise focus: "After months of hard work, we (Amazon AGI) just announced Nova Forge! Forge is a full-service framework built to enable enterprises to build their own frontier Foundation Models from the ground up." [source](https://x.com/fardinabad/status/1995978374611976646). Gabriel Elbling, an AWS engineer working on grounding agents, pondered its impact: "AWS just launched Nova Forge at #reinvent. 'Companies can upload proprietary data and automatically train their own frontier model from a Nova checkpoints.' How transformative will this be?" [source](https://x.com/gabelbling/status/1995899627443511616). These reactions underscore excitement among insiders for its potential to streamline custom AI development.

Early Adopter Experiences

As Nova Forge was unveiled just days ago at AWS re:Invent on December 2, 2025, early hands-on feedback remains limited, but initial reports from technical previews suggest promise. Maish Saidel-Keesing, an AWS Developer Advocate, pointed to the official blog for getting started: "Introducing Amazon Nova Forge: Build your own frontier models using Nova." [source](https://x.com/maishsk/status/1995902013595357269). Analyst Holger Müller noted its relevance for enterprise users already experimenting with models: "The reality has been that off-the-shelf foundational models are inaccurate in many enterprise use cases. When enterprises build on top of open source models, results can degrade as more data is added." [source](https://x.com/holgermu/status/1996457689933619440). Benjamin Baumguertel, a Principal Solutions Architect at AWS, referenced WIRED's coverage of early custom training phases, indicating positive buzz around phased development for proprietary data integration. [source](https://x.com/9193benjaminb/status/1996557361348309437). Developers are eager for broader access, with Shenoy reiterating: "Incredibly proud to be a part of the Nova Forge project! Can’t wait to see what customers build with this." [source](https://x.com/vikramshenoy97/status/1995912538618228809).

Concerns & Criticisms

While praise dominates, some technical critiques focus on accessibility and cost barriers for smaller teams. Munshipremchand described it as a "mind-blowing $100K/year service," raising questions about affordability for non-enterprise developers. [source](https://x.com/MunshiPremChnd/status/1995925064324104567). Müller echoed broader community worries about model degradation in custom training: "When enterprises build on top of open source models, results can degrade as more data is added," positioning Nova Forge as a solution but implying validation needs. [source](https://x.com/holgermu/status/1996361326818460082). Early adopters may face a learning curve, as Allie K. Miller advised in a related AI context: "This is not a product that non-technical folks are going to open up and immediately get value from... Lean into the suck." [source](https://x.com/alliekmiller/status/1975337368610906503). No major technical bugs have surfaced yet, but the community calls for transparent benchmarks versus alternatives like SageMaker.

Strengths ▼

Strengths

Enables early-stage customization of frontier AI models by infusing proprietary data during pre-training, reducing development time and costs compared to building from scratch, which can exceed tens of millions [source](https://aws.amazon.com/blogs/aws/introducing-amazon-nova-forge-build-your-own-frontier-models-using-nova/).
Seamless integration with AWS SageMaker and existing workflows, allowing technical teams to leverage familiar tools for scalable model training on Trainium chips [source](https://www.aboutamazon.com/news/aws/aws-agentic-ai-amazon-bedrock-nova-models).
Supports creation of optimized, smaller models tailored to specific use cases, improving efficiency and performance for enterprise applications [source](https://aibusiness.com/generative-ai/aws-targets-ai-maturity-with-nova-forge).

Weaknesses & Limitations ▼

Weaknesses & Limitations

High entry cost of $100,000 per year, which may deter smaller organizations or those testing viability before full commitment [source](https://www.cnbc.com/2025/12/02/amazon-nova-forge-lets-clients-customize-ai-models-for-100000-a-year.html).
Vendor lock-in to AWS ecosystem, limiting flexibility for multi-cloud strategies and requiring migration of data/workflows to Amazon's infrastructure [source](https://www.wired.com/story/amazon-nova-forge-ai-models/).
Restricted to Nova models only, excluding integration with third-party or open-source alternatives like those on Bedrock, potentially narrowing options for diverse AI stacks [source](https://venturebeat.com/ai/with-nova-forge-aws-gives-companies-a-path-to-build-foundation-class-models).

Opportunities for Technical Buyers ▼

Opportunities for Technical Buyers

How technical teams can leverage this development:

Develop industry-specific models, such as healthcare diagnostics infused with proprietary patient data, accelerating time-to-value while maintaining data privacy on AWS.
Optimize for cost and speed by fine-tuning smaller variants on Trainium hardware, ideal for real-time applications like fraud detection in finance.
Enhance agentic AI workflows by combining custom Nova models with Bedrock agents, enabling autonomous systems for tasks like supply chain automation.

What to Watch ▼

What to Watch

Key things to monitor as this develops, timelines, and decision points for buyers.

As a new service announced at re:Invent 2025, watch for general availability in early 2026, initial benchmarks comparing Nova Forge outputs to competitors like OpenAI's GPT series. Track real-world case studies from early adopters like Sony for ROI evidence. Decision points include pilot program results by Q2 2026 to assess customization depth versus cost, and AWS expansions to third-party models. Competitor moves from Google or Microsoft could pressure pricing or features, influencing adoption timelines for technical buyers evaluating lock-in risks.

Key Takeaways

Nova Forge empowers enterprises to train custom frontier AI models from AWS's Nova foundation, starting at early checkpoints in pre-, mid-, or post-training phases for tailored performance.
Seamless integration with Amazon SageMaker provides familiar tools for data infusion, simulation, and optimization, reducing the complexity of building foundation-class LLMs.
Proprietary data can be incorporated early to create specialized, reliable models for tasks like agentic AI, without starting from scratch on massive datasets.
Priced at $100,000 per year, it's designed for high-scale users, offering cost efficiencies over full from-scratch training while accessing AWS's Trainium and Inferentia hardware.
Launched alongside new Nova models, it positions AWS as a leader in customizable generative AI, emphasizing safety, efficiency, and enterprise-grade reliability.

Bottom Line

For technical decision-makers in AI-driven enterprises—such as ML engineers, data scientists, and CTOs at Fortune 500 companies—Nova Forge is a game-changer if you're scaling custom AI beyond off-the-shelf models. Act now if your team handles proprietary datasets and needs optimized, task-specific LLMs; the early access via SageMaker previews can accelerate prototypes. Wait if your budget is under $100K annually or you're satisfied with Bedrock's fine-tuning—it's not for startups or low-volume use. Ignore if your focus is non-AI workloads. This development matters most to organizations prioritizing data sovereignty and performance in regulated industries like finance or healthcare.

Next Steps

Explore the Nova Forge documentation and request early access through your AWS account to test checkpoints in SageMaker Studio.
Assess your infrastructure: Run a cost-benefit analysis comparing Nova Forge to alternatives like OpenAI's fine-tuning, using AWS's pricing calculator.
Contact AWS sales or join the re:Invent 2025 sessions on Nova for hands-on demos and migration guidance from existing SageMaker workflows.

References (50 sources) ▼