Amazon just dropped a game-changer that could reshape enterprise AI. The company unveiled Nova Forge, a new tool that lets customers build their own frontier AI models by plugging into Amazon's base model training process - something previously reserved for tech giants. This isn't just another fine-tuning service; it's custom pre-training that could give businesses the specialized AI they've been waiting for.
Amazon is making its biggest AI play yet, and it's not what you'd expect. While everyone's been focused on who has the smartest chatbot, Amazon's betting on something potentially more valuable: letting customers build their own frontier models.
The company announced Nova Forge alongside its second-generation Nova AI models at re:Invent in Las Vegas. But here's what's really interesting - Nova Forge doesn't just let you fine-tune existing models like everyone else. It gives customers access to custom pre-training, the foundational stage where base models are built. That's territory normally reserved for companies like OpenAI and Google.
"Everyone is looking for a frontier model that's an expert in their domain," Rohit Prasad, who leads Amazon's AI efforts, told WIRED. "This is essentially a new open training paradigm."
The proof is already in the pudding. Reddit used Nova Forge to create a custom model specifically designed to identify rule-breaking content. Traditional fine-tuning wouldn't work because most AI models are built to avoid offensive content entirely - they'd refuse to analyze the very materials Reddit needs to moderate.
"Other LLMs understand Reddit as a concept, and how Reddit works, but they're not down in the weeds," Reddit CTO Chris Slowe explained to WIRED. "We really built a Reddit expert model."
This hits at a massive enterprise problem. According to Bain's November survey, about three-quarters of US companies see AI as a high priority, but they're struggling with a lack of expertise and resources to build truly useful custom models.
The AI model landscape has traditionally been binary: closed models you access through APIs, or open models you download and run yourself. Nova Forge creates a third path - what Amazon calls "mid-stage training access" that's locked into AWS infrastructure but gives unprecedented customization power.
Amazon's timing here is strategic. While OpenAI and Google battle over general-purpose model supremacy, Amazon's going after the enterprise market that wants specialized AI but lacks the hundreds of millions needed to train models from scratch. Prasad says Nova Forge models should be "significantly cheaper" than building from scratch, though Amazon isn't sharing specifics yet.
The new Nova 2 lineup itself is impressive. Nova 2 Pro allegedly matches or beats OpenAI's GPT-5, Google's Gemini Pro 2.5, and Anthropic's Sonnet 4.5 across benchmarks. But the real standout is Nova 2 Omni, which Amazon claims is the first fully multimodal reasoning model that processes images, audio, video, and text while performing simulated reasoning.
Early Nova Forge customers include Booking.com, Sony, and biotech firm Nimbus Therapeutics - a diverse mix that suggests broad applicability across industries.
This move also positions Amazon strategically in the broader cloud wars. While Microsoft has its OpenAI partnership and Google pushes Gemini integration, Amazon's betting that customization trumps raw model performance. It's a smart hedge given that Amazon was relatively late to the frontier model game.
The company's also making infrastructure bets to support this vision, investing billions in AI hardware while trying to challenge Nvidia's dominance with custom Trainium chips. Anthropic's latest models already run on Amazon's hardware thanks to the company's $8 billion investment.
What makes Nova Forge particularly intriguing is how it could democratize specialized AI development. Instead of every company needing a massive AI research team, they could plug their domain expertise into Amazon's training infrastructure and get models that truly understand their business.
Reddit's experience suggests this isn't just theoretical. Slowe says their custom model will likely automate content moderation tasks that would be impossible with general-purpose AI. "For a large set of situations, it will be substantially better than what we get off-the-shelf," he told WIRED.
The big question is whether Amazon can execute at scale while maintaining the performance advantages that make custom training worthwhile.
Amazon's Nova Forge represents a fundamental shift in how companies might approach AI development. Instead of choosing between expensive custom development or limiting off-the-shelf solutions, businesses could soon have a middle path that combines the power of frontier models with domain-specific training. If Amazon can deliver on the performance and cost promises, this could be the breakthrough that finally makes AI genuinely useful for the enterprise customers who've been waiting on the sidelines. The real test will come as more companies beyond Reddit put Nova Forge through its paces in production environments.
