By ByteBot
AWS just changed the rules for building custom AI models. On December 2 at re:Invent 2025, Amazon announced Nova Forge—the first service to offer “open training” where companies can access model checkpoints at 60% or 80% completion and inject their proprietary data during pre-training, not just after. This solves catastrophic forgetting and lets enterprises build truly custom AI without the millions in compute costs required to train from scratch. Reddit is already using it and outperforming Claude Sonnet 4 by 20-50% on content moderation tasks.
What “Open Training” Actually Means
Nova Forge gives organizations access to pre-trained, mid-trained (at 60% or 80% completion), and post-trained Nova model checkpoints. Companies blend their proprietary data with Amazon’s curated training data via API and finish pre-training the model. The result is a custom “Novella”—a Nova variant that preserves frontier model capabilities while adding specialized knowledge.
Here’s why this matters: traditional fine-tuning happens after training completes. You take a finished model and adjust it for your use case. But heavy fine-tuning causes catastrophic forgetting—the model loses base capabilities when learning new tasks. Nova Forge lets you inject data during pre-training when the model is still learning, resulting in deeper integration of domain knowledge without forgetting what makes the base model useful.
Think of it like buying a 60%-completed house and finishing it to your specs, rather than buying a finished house and remodeling or building from scratch. You get customization without starting from zero.
Reddit Proved It Works
The proof isn’t hypothetical. Reddit built a custom Nova Forge model for content moderation that outperformed Claude Sonnet 4 by 20-50% on property prediction tasks. Reddit needed an AI model sophisticated about the diverse topics people discuss on the platform—something generic models struggle with.
Engineers found that a Nova model enhanced with Reddit data through Forge performed better than commercially available large-scale models. Reddit is now using this custom model for automated content moderation across the social network. When you’re outperforming Anthropic’s flagship model by double-digit percentages on your specific task, that’s not incremental improvement. That’s a different approach working.
The Catastrophic Forgetting Problem
Catastrophic forgetting is a well-documented challenge in AI research. When you fine-tune a model on new data, it “forgets” previously learned capabilities. IBM describes it as undermining “the versatility and reusability of LLMs across domains.”
AWS addressed this explicitly: “Fine-tuning a conventional model would not work for some use cases because most models are designed to avoid offensive content entirely, whereas custom pre-training combined with fine-tuning can produce frontier models expert at specific domains.”
Nova Forge’s data mixing approach during pre-training preserves foundational skills—general intelligence, instruction following, safety—while adding specialized knowledge. You don’t have to choose between general capability or domain expertise. You can have both.
Who’s Already Building
Major companies are betting on this approach. Beyond Reddit, Sony is building an AI agent for review and assessment processes, targeting 100x efficiency improvement. Early results show they’re exceeding the performance of larger models using reinforcement fine-tuning. Booking.com, Cosine AI, Nimbus Therapeutics, Nomura Research Institute, and OpenBabylon are also building custom models with Nova Forge.
That level of enterprise adoption signals this isn’t just AWS marketing. Companies with real use cases and real budgets are choosing Nova Forge over alternatives.
The $100K Question
Nova Forge costs $100,000 per year. Additional help from AWS experts not included. Is that expensive? Absolutely. Is it democratizing custom model development? That depends on your perspective.
Building a foundation model from scratch costs millions in compute, requires months of time, and needs hundreds of GPUs. Fine-tuning off-the-shelf models is cheaper but offers surface-level customization with high risk of catastrophic forgetting. Nova Forge sits in the middle—deep integration at a fraction of from-scratch costs.
For mid-market enterprises and up, $100K/year is accessible. For startups and small businesses, it’s still out of reach. AWS is targeting enterprises, not scrappy three-person teams. Call it “democratization” if you want, but it’s democratization for companies with serious budgets.
What This Really Means
Nova Forge represents a strategic shift in how enterprises think about AI. OpenAI and Anthropic want you to rent their models forever. AWS wants you to build and own your custom models. Once you build a Novella on AWS infrastructure, switching to another provider means starting over. This is strategic vendor lock-in wrapped in innovation.
But here’s the thing: enterprises with proprietary datasets don’t want to send their data to OpenAI or Anthropic. They want to own the model, control the data, and build competitive advantage through customization. Nova Forge gives them that path without the infrastructure burden of training from scratch.
The real winner here? Companies with unique, high-quality proprietary data. If your data is generic, Nova Forge won’t save you. The customization is only as good as your data. But if you have domain-specific datasets that competitors don’t have, Nova Forge turns that data into a custom AI capability that’s genuinely differentiated.
The Competitive Landscape
AWS is making a direct play against OpenAI, Anthropic, and Google. Nova 2 Pro—the base model for Nova Forge—performs at least as well as Claude Sonnet 4.5, GPT-5, and Gemini 3.0 Pro on benchmarks. But the real differentiation isn’t model performance. It’s the ability to customize deeply without catastrophic forgetting.
OpenAI offers fine-tuning. Google offers fine-tuning. Anthropic offers fine-tuning. None of them offer access to partially-trained checkpoints where you can inject data during pre-training. Nova Forge is the first. Whether competitors follow this approach or double down on their closed-model strategies will shape the enterprise AI market.
Bottom Line
AWS Nova Forge pioneers “open training”—a middle ground between building from scratch and fine-tuning off-the-shelf models. For $100,000 per year, enterprises can access partially-trained Nova checkpoints, inject proprietary data during pre-training, and build custom models that preserve general capabilities while adding specialized knowledge.
Reddit proved it works by outperforming Claude Sonnet 4 by up to 50%. Sony, Booking.com, and other major companies are already building with it. The approach solves catastrophic forgetting and offers deeper customization than traditional fine-tuning.
Is it expensive? Yes. Is it accessible to everyone? No. But for enterprises with proprietary datasets and serious AI ambitions, Nova Forge offers a path to owned AI capabilities without the millions in costs required to train from scratch. AWS is betting that enterprises will choose to own rather than rent. We’re about to find out if they’re right.
