Summary: Amazon introduced a new family of frontier AI models at re:Invent in Las Vegas and opened a path for customers to build their own frontier models. The announcements include Nova Lite, Nova Pro, Nova Sonic, and the experimental Nova Omni, plus a tool called Nova Forge that lets customers add their own training data during multiple stages of model training — including base-model training, a process normally reserved for large AI labs. Amazon positions this as a move to give cloud customers domain-expert models that can be far more useful than off-the-shelf general models. How will companies use this? What trade-offs should they expect? Read on for a practical breakdown and a checklist for teams ready to test-customize their next frontier model.
Interrupt / Engage: Why you should stop skimming and think about this
Quick trigger: Amazon is not just shipping another API. It handed cloud customers a way to shape base models with their data — custom pretraining. Custom pretraining. That phrase deserves attention because it flips a common assumption: building a domain expert model is no longer only for labs with billion-dollar budgets. What would a model that truly knows your internal data let you automate, flag, or route differently?
What Amazon announced at re:Invent
At re:Invent, Amazon unveiled four new Nova family models and one tooling story. The models are Nova Lite and Nova Pro (updated large language models), Nova Sonic (real-time voice), and Nova Omni (experimental multimodal model that accepts text, images, audio, and video). These models rolled out to a limited set of customers immediately. The tooling piece, Nova Forge, is the strategic move: it allows customers to inject proprietary data at multiple stages of training, including the pretraining phase that normally builds the base capabilities.
Why Nova Forge matters: custom pretraining explained
Most companies either call a vendor and use closed models through APIs, or they pick an open model, download it, and fine-tune on top. Fine-tuning is useful, but it usually sits on a generalist base that resists some domain-specific behaviors. Nova Forge lets customers add data during different training stages — including the base stage. That means you get a model that starts life closer to your domain. Custom pretraining. Custom pretraining. This is the practical difference between "knows the concept of X" and "is an X expert in the details."
Case study: Reddit built a Reddit expert
Reddit tested Nova Forge to build a moderation-focused model. Reddit’s CTO, Chris Slowe, said a standard fine-tune wouldn’t work because mainstream models are designed to avoid offensive or violent content and often refuse analysis. By using custom pretraining plus fine-tuning, Reddit produced a model that can examine problematic content without refusing to process it — a model that understands Reddit in detail. That model will likely be used to automate parts of content moderation. The takeaway: for some tasks, you need a model that’s comfortable working inside the messy, specialized boundaries of your business.
Other customer signals: Booking.com, Sony, Nimbus Therapeutics
Amazon isn’t testing this in isolation. Other companies — Booking.com, Sony, and Nimbus Therapeutics — are experimenting with Nova Forge. That list gives social proof: travel, entertainment, and biotech firms see value in building domain experts. When different industries test the same tooling, it suggests a broader applicability beyond a single vertical.
What Nova Omni and Nova Sonic add to the story
Nova Omni is Amazon’s multimodal experiment: text, images, audio, video input, with simulated reasoning across those media. Amazon claims no other company has released a fully multimodal reasoning model like this. Nova Sonic targets real-time voice tasks. Put them together and Amazon is signaling an intent to handle multimodal agentic workflows — not just chat, but actions and decisions across media types.
Benchmarks and competition claims — read the fine print
Amazon claims Nova 2 Pro matches or exceeds GPT-5 and GPT-5.1 from OpenAI, Gemini Pro 2.5 and 3.0 Pro from Google, and Anthropic’s Sonnet 4.5 on several benchmarks. Nova 2 Lite is compared to Claude 4.5 Haiku, GPT-5 Mini, and Gemini Flash 2.5. Those are strong claims. Check them against the benchmarks you care about: instruction following, reasoning traces, hallucination rates, and tool use. The vendor’s benchmarks are evidence of capability, but your domain tests are what prove value.
Cost, lock-in, and the "cheaper frontier model" claim
Amazon says a frontier model built with Nova Forge should cost far less than building from scratch, but provided no hard numbers. That’s credible — using shared infrastructure and partially trained models lowers marginal cost. But an important trade-off exists: Nova Forge is tied to Amazon’s cloud. No vendor lock-in? No — it’s AWS-only. Ask this: will the value of a domain-expert model outweigh the cost of cloud lock-in for your organization? Are you ready to commit resources to an AWS-centric AI stack?
Open models versus closed models: where Nova Forge sits
Most open models can be downloaded and run on private hardware; many companies pick open models because experimenting is cheaper and more flexible. However, open models often ship without disclosed training data, making true tuning opaque. Nova Forge sits between closed APIs and raw open models: you get AWS-managed training pipelines and the ability to influence base-model training, but that influence lives on Amazon infrastructure. That’s a practical trade: faster, supported, and powerful — but within AWS.
Amazon’s strategic positioning: infrastructure bets and partnerships
Amazon has built a broad AI portfolio quietly: generative features in shopping (Rufus), investments in Anthropic ($8 billion), and claims that Anthropic’s latest models were trained on AWS Trainium chips. Amazon is competing with Google and Microsoft for cloud AI workloads while OpenAI builds its own stack. Amazon also wants to challenge GPU dominance. The company is betting billions that demand for model training and hosting will keep growing. For customers, that means more tooling choices and an environment that supports large-scale model development.
Risk checklist for teams considering Nova Forge
Before you start, answer these questions honestly. A good negotiation begins with clarity; so does a model build.
- What precise business problem will a domain-expert model solve? How will success be measured?
- Do we have the data quality and volume needed for effective pretraining and fine-tuning?
- Can we commit to an AWS-hosted training and serving pipeline? If not, what’s the migration plan?
- Who will own safety, auditability, and labeling standards inside the organization?
- What budget and governance structure will we set to avoid cost overruns?
These are open questions. How would you answer them for your product or team?
Practical steps for a pilot with Nova Forge
If your answers lean toward a trial, use this staged approach. Start small, keep control, and test assumptions.
- Define the narrow use case. Think moderation, document classification, or an internal assistant task.
- Inventory the data. Label samples, audit quality, and set retention/security rules.
- Design evaluation metrics that capture real business impact — not just benchmark numbers.
- Run a controlled pilot: use custom pretraining on a small base, then fine-tune and evaluate.
- Measure cost per useful inference and compare to existing automation or human costs.
- Iterate or stop. Saying No is a strategy: if the pilot fails to show ROI, stop and reallocate resources.
Ethics, safety, and the role of 'No'
Build safety gates early. Custom pretraining lets a model learn to handle problematic content without refusing to analyze it. That power is useful, but dangerous without guardrails. Who says No when a model’s outputs risk harm? Make that role explicit. Who audits the pretraining data? Who approves deployment? Saying No protects reputation and customers; it enables focused, responsible experimentation.
How to evaluate claims rigorously
Vendors will publish benchmarks and strong language. Treat those as starting points, not verdicts. Ask for reproducible tests, raw evaluation sets, and access to trial environments. Mirror their claims back to them: “You claim Nova 2 Pro matches GPT-5 on X — can we run our benchmarks on a trial instance?” That simple mirroring buys clarity and forces specifics.
Opportunities by industry
Here are quick hypotheses about where custom pretraining adds measurable value:
- Community platforms: moderation and context-aware routing (Reddit example).
- Travel and hospitality: personalized recommendation engines that respect private signals (Booking.com testing suggests this).
- Media and entertainment: content tagging and voice-driven workflows (Sony testing suggests this).
- Biotech and pharma: domain-aware literature synthesis and hypothesis generation (Nimbus Therapeutics testing suggests this).
What would a domain-expert model change in your org’s daily workflow?
Questions to ask Amazon (or any vendor) before committing
Don’t accept marketing slides. Ask open questions to force concrete answers.
- What visibility will we have into the pretraining data pipeline?
- How are provenance, versioning, and audit logs handled for custom pretraining runs?
- What are the rollback and mitigation paths for a model that behaves unexpectedly?
- How will costs scale with model size and training iterations?
- What SLAs and support do you provide for production inference?
If they hedge, mirror that: “You’re hedging on audit logs — can you show a compliance example?” That approach extracts specifics fast.
A pragmatic view: where this fits in your AI roadmap
Use Nova Forge when the marginal value of domain expertise exceeds the marginal cost of cloud lock-in and governance overhead. For exploratory AI teams, run pilots to test the hypothesis. For regulated industries, treat pretraining as you would a clinical trial: careful design, audit trails, and third-party review. The promise is clear: models that act like domain experts rather than generalists. The work is governance, measurement, and cost control.
Final thoughts and an invitation to dialogue
Amazon’s move shifts a big piece of model customization from a closed lab problem into a cloud service. That shift creates opportunity and responsibility. You can build models that truly know your data, but you must decide where to place your bets. What will you build first if you had a domain-expert model tomorrow? What would you say No to, to keep focus and control? I’ll leave that question open — take a moment and answer it for your team. If you want, outline one use case in a paragraph and watch how the decision path clarifies.
#AmazonNova #NovaForge #AIforEnterprise #AWS #MultimodalAI #CustomPretraining #Reddit #Rufus #ReInvent
Featured Image courtesy of Unsplash and Taylor Vick (M5tzZtFCOfs)