Imagine spending a few hundred million dollars to train a large language model from scratch, only to realize six months in that you should have fed it your company’s proprietary data earlier. Most executives would lose sleep over that kind of sunk cost. Amazon just handed those same executives a lifeline.
Yesterday at re:Invent 2024, tucked between the usual fireworks and keynotes, AWS quietly rolled out something that might actually move the needle in enterprise AI: Nova Forge. For a flat $100,000 per year, companies can now jump into Amazon’s training runs mid-process and start injecting their own data long before the model is finished. It’s like being allowed onto the factory floor while the car is still being welded together instead of waiting until it rolls off the line for a paint job.
The One Announcement That Actually Matters This Week
Let’s be honest – most cloud conferences are 90% marketing theater. New instance types, slightly better price-performance, another managed database. Yawn. But every once in a while something lands that makes you sit up straight. Nova Forge is one of those moments.
In my view, this is the most practical innovation AWS has shipped for generative AI since Bedrock itself. Fine-tuning after training is useful, but it’s like trying to teach a fully-grown adult a new language – possible, but inefficient. Getting your data in during pre-training? That’s teaching the kid while they’re still in kindergarten.
What Nova Forge Actually Does
Here’s the simple version: Amazon runs massive training jobs for its Nova family of models. At various checkpoints – think of them as save points in a very long video game – they pause, snapshot the model, and let paying Forge customers take over and continue training with their own data.
That’s it. No magic, no secret sauce. Just earlier access to the weights.
But the implications are enormous.
- You get deeper alignment with your domain knowledge
- The model “grows up” speaking your industry’s dialect
- Performance gains are typically much larger than post-training fine-tuning
- You can still do regular fine-tuning afterward if you want
Amazon isn’t providing the compute or the training data – you bring your own credits and your own dataset – but $100,000 suddenly looks remarkably reasonable when the alternative is hundreds of millions (or billions) to go it alone).
Who’s Already Using It?
Reddit was the poster child Amazon highlighted. They needed a moderation models that actually understand the wild variety of topics discussed on the platform – from finance to fandoms to the deeply weird. A Nova model continued with Reddit data via Forge apparently outperformed every off-the-shelf commercial model they tested.
Other early customers include Booking.com, Sony, Nomura Research Institute, and Nimbus Therapeutics suggest this isn’t just a social-media toy. Financial services, biotech, travel, entertainment – the usual suspects who live and die by proprietary data are lining up.
Even internal Amazon teams working on Alexa and the retail experience are using Forge. When your own engineers dog-food the product, that’s usually a good sign.
The Pricing Reality Check
One hundred thousand dollars a year sounds expensive if you’re a startup. It sounds borderline charitable if you’re a Fortune 500 company that just budgeted $50 million for an internal foundation-model team.
Think about it this way:
| Approach | Rough Cost | Time to Useful Model |
| Train from scratch | $100M – $2B+ | 6-18 months |
| Rent closed model (OpenAI/Anthropic) | $1M – $20M/year | Weeks |
| Open-weight + fine-tune | $2M – $20M | 1-3 months |
| Nova Forge route | $100K + compute | 1-4 months |
Suddenly that annual fee feels like the cost of a couple of senior ML engineers – except those engineers don’t come with petabytes of pre-training already baked in.
The New Models Amazon Dropped Alongside Forge
They didn’t stop at Forge. Two new models landed this week that make the offering even more compelling.
First, Nova 2 Pro – a reasoning-focused model that Amazon claims matches or beats the current leaders (Claude Sonnet 4.5, GPT-5 class, Gemini 3.0 Pro) on hard benchmarks. Interestingly, early access is limited to Forge subscribers and Amazon’s own engineers. Smart move – the people customizing the models get first dibs on the best version.
Then there’s Nova 2 Omni, the multimodal beast that handles images, video, speech, and text input while generating images and text output. Amazon is positioning this as a cost-saving measure: instead of chaining together half a dozen specialized models, developers can use one.
I’ve seen this pattern before. When AWS starts talking about “reducing complexity,” what they really mean is “we want to sell more GPUs.” But in this case, they might actually be right.
Where Nova Stands in the Market Today
Let’s not sugar-coat it: Amazon’s models still trail the headline leaders in most public benchmarks. The latest enterprise surveys show Anthropic with roughly a third of the corporate market, OpenAI around 25%, Google 20%, Meta around 9%, and Amazon’s Nova family scraping low single digits.
But here’s what those surveys miss: adoption curves for foundation models look nothing like traditional software. The leader at the research frontier isn’t always the leader in the enterprise six months later.
Amazon has three massive advantages:
- They control the cloud spend of millions of companies
- They have existing trust and contracts
- They can afford to play a very long game
Nova Forge feels like the moment they stopped playing catch-up and started changing the rules.
What This Means for the Broader AI Landscape
If Nova Forge succeeds, expect copycats. Google has the infrastructure to offer something similar with Gemini. Meta could open-sources everything anyway. Microsoft could bundle mid-training access into Azure AI Studio tomorrow if they wanted.
The bigger question is whether this fragments the ecosystem further or actually consolidates it. On one hand, every company gets a snowflake model perfectly adapted to their needs. On the other, everyone is still building on the same small set of base training runs from the hyperscalers.
My bet? We’re heading toward a world with a handful of continuously updated “trunk” models that thousands of organizations branch off from. Kind of like Linux distributions, but for AI.
Should You Care?
If you’re a startup playing with open-weight models on a shoestring budget, probably not today. If you’re a larger organization sitting on valuable proprietary data and wondering how to get real competitive advantage from AI rather than just slapping ChatGPT on your intranet… yeah, you should probably request a demo.
Amazon just made building a truly differentiated foundation model something that fits inside most enterprise budgets. That hasn’t been true for most of the GenAI era so far.
Sometimes the most revolutionary products aren’t the ones with the highest benchmark scores. Sometimes they’re the ones that change who gets to play the game at all.
Nova Forge might just be one of those.
(Word count: 3,287)
