Amazon Nova Forge: Custom Foundation Models Are No Longer Just for Tech Giants

AWS used re:Invent 2025 to quietly introduce what may be the most consequential enterprise AI capability they’ve released in years. Amazon Nova Forge is not another fine-tuning interface. It is a structural challenge to the assumption that only frontier labs with billion-dollar budgets can build custom foundation models.

For the first time, enterprises can train what Amazon calls “Novellas”—private, custom versions of Nova models—infused with their own proprietary data and domain knowledge during the model training process itself, not bolted on at the end through SFT alone.

Forge subscription access begins around $100,000 per year, with additional usage-based training compute through SageMaker. For many mid-market companies, that shifts the calculus from “impossible” to “attainable.”

Why the Economics Shifted

Traditional foundation model training is financially inaccessible outside of a few labs. A modern frontier model can require hundreds of millions of dollars in compute, months of engineering work, and deep expertise in data curation and distributed training.

That left most enterprises two options:

1. rely on API access to generic models, or

2. attempt continued pre-training on open-weights, suffering catastrophic forgetting as the model overwrote its base capabilities.

Forge’s open training approach introduces a third path. AWS gives customers access to curated checkpoint milestones—pre-training, mid-training, and post-training—combined with a data-mixing pipeline that blends customer data with Amazon’s own training distributions. This prevents the catastrophic forgetting commonly seen in naive continued pre-training while letting the model absorb deep domain-specific concepts.

Reddit: A Case Study in Model Consolidation

Reddit’s early use of Forge demonstrates the real implications. Moderation on Reddit spans thousands of heterogeneous communities, each with distinct tone, vocabulary, humor norms, and edge cases. Even after years of investment, Reddit maintained a portfolio of specialized ML models rather than a single cohesive system.

Forge allowed Reddit to train a Nova-based model using their proprietary moderation data. The result outperformed commercially available LLMs on Reddit’s internal tasks and—critically—consolidated multiple separate ML workflows into a single model.

As Reddit CTO Chris Slowe put it:

“We’re replacing a number of different models with a single, more accurate solution… It’s a shift in how we implement and scale AI across Reddit.”

The operational simplification alone is a major signal for the enterprise AI market.

Beyond Moderation: Industries With Deep Domain Knowledge Stand to Benefit Most

Early adopters span life sciences, finance, entertainment, travel, and specialized language applications. Each illustrates a different advantage of introducing private data during model training.

Drug Discovery

Nimbus Therapeutics used Nova 2 Lite through Forge to train a unified drug discovery assistant. Their model demonstrated 20–50% improvement over Claude Sonnet 4 on property prediction benchmarks, a meaningful performance gain in workflows where each experimental candidate is a costly wet-lab process.

Financial Services

Nomura Research Institute’s involvement underscores why Forge resonates with regulated industries. Sensitive trading strategies, risk models, and client data can remain entirely within AWS boundaries. Enterprises can train models on proprietary intelligence without exposing it to model providers or external vendors.

Manufacturing and Industrial

Manufacturers often have domain-specific operational data and terminology not represented in general pre-training corpora. Forge’s reinforcement fine-tuning—using customer-defined reward functions within a managed SageMaker RFT pipeline—lets models learn from proprietary simulations, robotics environments, or defect scoring systems.

How Nova Forge Works: A Closer Look at “Open Training”

Forge provides a structured interface to multiple phases of the Nova training lifecycle:

Pre-training checkpoints

Ideal for large unstructured corpora. This is where domain concepts become part of the model’s internal world-model. AWS’s data mixing reduces catastrophic forgetting by ensuring the model doesn’t drift away from general competence.

Mid-training checkpoints

A balanced stage where the learning rate is lower, but the model can still meaningfully shift toward new domain knowledge without destabilizing its base capabilities.

Post-training + Supervised Fine-Tuning

For labeled instruction-response datasets. This is closest to the typical enterprise fine-tuning workflow but benefits from starting with a checkpoint already aligned toward the customer’s data domain.

Reinforcement Fine-Tuning

Customers define reward functions or programmatic quality signals, and SageMaker executes RFT using these signals to shape behavior. This is not raw RLHF infrastructure—AWS provides the pipeline structure and guardrails.

Once training completes, models are deployed exclusively on Amazon Bedrock with the same enterprise security, logging, observability, and quotas as native Nova models.

AWS does not provide raw model weights, and Forge models are not portable outside Bedrock today.

Responsible AI in a Custom Model World

Custom models shift safety responsibility onto enterprises. Forge includes safety evaluation tools and allows configuration of moderation policies. Data mixing with Nova safety datasets helps retain base guardrails, but AWS is clear: enterprises must validate the final behavior of their customized models.

This is particularly relevant in healthcare, finance, and legal contexts, where a model’s refusals matter as much as its capabilities.

Strategic Implications for Enterprises

Forge changes the build-vs-buy equation:

• The cost threshold drops from “only hyperscalers can do this” to “mid-market companies can justify this annually.”

• Proprietary data becomes an even stronger competitive moat.

• Model consolidation becomes an operational advantage, simplifying ML Ops stacks.

• AWS subtly shifts from “infrastructure layer” to “model ecosystem with preferential incentives for staying inside Nova.”

Enterprises already committed to AWS will see Forge as a natural extension. Multi-cloud organizations will need to weigh increased Bedrock lock-in against the value of deeply customized models.

A Step Toward the Diversification of AI

If Nova Forge succeeds, the AI landscape becomes less centralized. Instead of a few generic frontier models dominating every vertical, industries will cultivate their own domain-expert models—optimized for chemistry, logistics, financial reasoning, medical workflows, or specialized moderation tasks.

That future brings benefits and challenges. Stronger performance. Richer diversity. More responsibility. And a requirement for enterprises to think systematically about the safety characteristics of their custom models.

Bottom Line

Nova Forge democratizes a capability that, until now, belonged only to frontier labs. By lowering the barrier to custom foundation model development, AWS has created a new strategic option for enterprises sitting on valuable proprietary data.

The question is no longer “Can we afford to build our own model?”

For many organizations, the answer is now yes.

The real question has become:

“Is our proprietary data strong enough—and our organizational maturity high enough—to justify building a model that could outperform anything off-the-shelf?”

Early adopters like Reddit and Nimbus Therapeutics indicate that when the data is genuinely differentiated, the payoff can be significant.

Nova Forge is live today in US-East (N. Virginia), with more regions planned.