Amazon Nova Forge Enables Clients to Tailor AI Models for $100K Annually

Amazon has introduced a new way for cloud customers to deeply customize generative AI models. The service, called Nova Forge, carries an annual price tag of $100,000.

Nova Forge gives enterprises early‑stage access to Amazon’s AI models, allowing them to inject proprietary data while the models are still being trained. This goes beyond the traditional “fine‑tune after training” approach and promises that the resulting models will reflect customer data more heavily.

In addition to Amazon’s own models, Nova Forge customers can also refine open‑weight models. However, the service does not include the underlying training data or compute infrastructure required for full model development.

Building a proprietary model from scratch can run into the hundreds of millions—or even billions—of dollars in compute, talent, and data acquisition costs. By contrast, Amazon positions Nova Forge as a more cost‑effective alternative for enterprises that need a competitive edge without the massive capital outlay.

Amazon released its Nova family of models in 2024, but they have yet to achieve dominant market adoption. A mid‑year 2025 survey by Menlo Ventures reported that enterprise‑grade large language models (LLMs) market share was led by Anthropic (32%), followed by OpenAI (25%), Google (20%) and Meta (9%). Amazon’s Nova models held less than 5% of the market, making them a clear underdog.

The Nova models are accessible through AWS Bedrock, the same managed service that offers Anthropic’s Claude 4.5 and other third‑party models. Bedrock provides a unified API and billing platform, simplifying multi‑model orchestration for developers.

Rohit Prasad, Amazon’s head scientist for artificial general intelligence, emphasized that Nova Forge is a “frontier lab” built around direct customer demand. Internal Amazon teams—including those behind the company’s e‑commerce platform and the Alexa assistant—are already using the service.

Reddit, for example, needed a moderation model capable of understanding a broad spectrum of user‑generated topics. Engineers reported that a Nova model fine‑tuned with Reddit’s own data outperformed several commercially available large‑scale models on content‑moderation benchmarks. Other early adopters mentioned by Amazon include Booking.com, Nimbus Therapeutics, Nomura Research Institute and Sony.

Customers may request hands‑on guidance from Amazon engineers to accelerate their Forge projects, though that consulting service is billed separately from the base $100,000 annual fee.

At AWS re:Invent, Amazon unveiled two additional models that will be available to Forge subscribers in early access.

Nova 2 Pro is a reasoning‑oriented model that, in internal tests, matches or exceeds the performance of Anthropic’s Claude Sonnet 4.5, OpenAI’s GPT‑5/5.1 and Google’s Gemini 3.0 Pro Preview on complex problem‑solving tasks. The model employs multi‑step computation pipelines that can take slightly longer to respond but deliver higher‑quality answers.

Nova 2 Omni extends the reasoning capability across modalities: it can ingest images, audio, text and video, and generate both visual and textual outputs. Amazon markets Omni as a single, unified model that reduces the engineering overhead and cost of deploying separate specialized models for each data type.

According to Amazon, tens of thousands of organizations run Nova models on a weekly basis, making it the second‑most popular model family on Bedrock after Anthropic’s offerings. The company cites “millions of customers” on AWS overall, underscoring the scale of its AI ecosystem.

Business and technical implications

• Cost structure: At $100,000 per year, Nova Forge is positioned between fully managed fine‑tuning services (often priced per inference) and the massive capital expenditures of building proprietary models. For enterprises with sizable data assets, the subscription can be justified by faster time‑to‑market and reduced need for internal MLOps expertise.
• Data security and compliance: Because customers retain control over the data used to condition the models, Nova Forge helps address concerns around data provenance, GDPR, and industry‑specific regulations. Amazon’s AWS infrastructure also offers encryption‑at‑rest and in‑transit, as well as isolated VPC environments for model training.
• Competitive positioning: By bundling early‑stage model access, multi‑modal reasoning, and optional engineering support, Amazon aims to carve out a niche against Anthropic’s strong enterprise foothold and OpenAI’s brand dominance. The low market share of Nova models suggests significant growth potential if Amazon can demonstrate consistent performance improvements and cost efficiencies.
• Technical differentiation: The introduction of Omni marks a shift toward “foundation models as a service” that can handle heterogeneous inputs without requiring separate pipelines. This could lower operational complexity for developers building AI‑augmented products, especially in sectors like e‑commerce, travel, and media where multimodal content is common.
• Future roadmap: Amazon’s strategy of releasing new models at high‑visibility events like re:Invent signals a commitment to rapid iteration. If performance benchmarks continue to keep pace with Anthropic, OpenAI, and Google, Nova could see accelerated adoption, especially among AWS‑centric enterprises.