Amazon rarely wins the AI news cycle, and its Nova 2 family — launched at AWS re:Invent in December 2025 — is a good example of why: no superlatives, no AGI talk, just a competent fleet of models slotted into Bedrock and priced to move. But the strategy underneath is worth understanding, because it's distinctly Amazon.

Four models, four jobs

Nova 2 is a fleet, not a hero model:

• Nova 2 Lite — a fast, cost-effective reasoning model for everyday workloads, handling text, images and video to generate text.
• Nova 2 Pro — Amazon's most intelligent reasoning model, spanning text, image, video and speech, aimed at hard agentic coding, long-range planning and complex problem-solving.
• Nova 2 Omni — multimodal in (image, text, video, speech) and out (text and images).
• Nova 2 Sonic — a speech-to-speech model unifying speech understanding and generation for real-time, human-like conversation, with expanded multilingual support and a 1M-token context window.

The segmentation is deliberate: most enterprise workloads don't need a frontier model, they need the right model at the right price point. Amazon is selling a menu, not a flagship — Lite for volume, Pro for the hard stuff, Omni and Sonic for multimodal and voice. It's the AWS playbook applied to models: give customers tiers and let them optimise spend.

Nova Forge: rent the recipe

The more interesting announcement is Nova Forge — a service letting AWS customers build their own frontier version of Nova, branded 'Novellas', for $100,000 a year. Crucially, enterprises can access pre-trained, mid-trained or post-trained checkpoints, intervening at whichever stage suits them.

That's a genuinely differentiated offering. Most labs sell you either a fixed API model or open weights you have to train yourself. Nova Forge sells the middle: a managed path to a custom frontier model without standing up your own training infrastructure. For a large enterprise with proprietary data and the budget, that's an appealing third option — more control than an API, far less heavy lifting than DIY pre-training.

Why it matters

Amazon isn't trying to have the smartest model; it's trying to be the place you run and customise models, whoever built them. Bedrock already hosts Anthropic, Meta and others alongside Nova. Nova 2 and Nova Forge round that out with a competent first-party fleet and a managed customisation pipeline. It's an unglamorous, infrastructure-first bet — and given how much AI ultimately runs on cloud rails, an entirely rational one.