Amazon to let cloud clients customize AI models midway through training for $100,000 a year

Amazon has found a way to let cloud clients extensively customize generative AI models. The catch is that the system costs $100,000 per year.

The Nova Forge offering from Amazon Web Services gives organizations access to Amazon’s AI models in various stages of training so they can incorporate their own data earlier in the process.

Already, companies can fine-tune large language models after they’ve been trained. The results with Nova Forge will lean more heavily on the data that customers supply. Nova Forge customers will also have the option to refine open-weight models, but training data and computing infrastructure are not included.

Organizations that assemble their own models might end up spending hundreds of millions or billions of dollars, which means using Nova Forge is more affordable, Amazon said.

AWS released its own models under the Nova brand in 2024, but they aren’t the first choice for most software developers. A July survey from Menlo Ventures said that by the middle of this year, Amazon-backed Anthropic controlled 32% of the market for enterprise LLMs, followed by OpenAI with 25%, Google with 20% and Meta with 9% — Amazon Nova had a less than 5% share, a Menlo spokesperson said.

The Nova models are available through AWS’ Bedrock service for running models on Amazon cloud infrastructure, as are Anthropic’s Claude 4.5 models.

“We are a frontier lab that has focused on customers,” Rohit Prasad, Amazon head scientist for artificial general intelligence, told CNBC in an interview. “Our customers wanted it. We have invented on their behalf to make this happen.”

Nova Forge is also in use by internal Amazon customers, including teams that work on the company’s stores and the Alexa AI assistant, Prasad said.

Reddit needed an AI model for moderating content that would be sophisticated about the many subjects people discuss on the social network. Engineers found that a Nova model enhanced with Reddit data through Forge performed better than commercially available large-scale models, Prasad said. Booking.com, Nimbus Therapeutics, the Nomura Research Institute and Sony are also building models with Forge, Amazon said.

Organizations can request that Amazon engineers help them build their Forge models, but that assistance is not included in the new service’s $100,000 annual fee.

AWS is also introducing new models for developers at its Reinvent conference in Las Vegas this week.

Nova 2 Pro is a reasoning model whose tests show it performs at least as well as Anthropic’s Claude Sonnet 4.5, OpenAI’s GPT-5 and GPT-5.1, and Google’s Gemini 3.0 Pro Preview, Amazon said. Reasoning involves running a series of computations that might take extra time in response to requests to produce better answers. Nova 2 Pro will be available in early access to AWS customers with Forge subscriptions, Prasad said. That means Forge customers and Amazon engineers will be able to try Nova 2 Pro at the same time.

Nova 2 Omni is another reasoning model that can process incoming images, speech, text and videos, and it generates images and text. It’s the first reasoning model with that range of capability, Amazon said. Amazon hopes that, by delivering a multifaceted model, it can lower the cost and complexity of incorporating AI models into applications.

Tens of thousands of organizations are using Nova models each week, Prasad said. AWS has said it has millions of customers. Nova is the second-most popular family of models in Bedrock, Prasad said. The top group of models are from Anthropic.

WATCH: Amazon set to kick off AI conference next week: Maxim’s Tom Forte on what to expect