Amazon has a well defined growth flywheel. The flywheel connects cost to price to customer experience and selection to customer experience. This flywheel has driven Amazon’s dominance of eCommerce. Is it now building a similar flywheel for generative AI?

In Ibbaka’s AI Monetization in 2025 survey, we are seeing more and more companies partner with infrastructure providers to develop generative AI applications. Of the 336 companies that took the survey, 164 or 49% are partnering with an infrastructure provider like Amazon, Google, Microsoft, or Salesforce. From other data sources we know that AWS is the dominant partner (about 70%), with Microsoft and then Google running well behind.

At the AWS Re:Invent conference (Dec. 2-6 2024) Amazon has been making some interesting announcements. Here is Amazon’s summary of the event. It is promoting its own chips as an alternative to Nvidia and saying these will offer much lower computing costs. Amazon is also taking the position that no one model will dominate and that access to multiple models is critical (this is something Ibbaka has found in its own work, different models deliver different results to the same prompt, and sometimes you want to use one model, sometimes another, and sometimes to combine the results of multiple models).

Amazon CEO Andy Jassy said

“I figured that almost everybody would end up using Anthropic's Claude models...but they're also using Llama models, and they're also using Mistral models, and they're also using some of our own models... This kind of surprised us.”

Is having access to multiple models, so that more and different agents can be built, equivalent to having more selection in eCommerce? We will see this answered real time over the next few years. Note that Hugging Face currently has 1,180,784 different models as this is being written and that there is a lot of talk about the importance of SLMs (Small Language Models). SLMs are likely. to dominate in the agent space and AWS may be the key to deploying agents at scale (setting up competition with Salesforce and its agent-centric vision).

Is this a general pattern for sustainable growth in generative AI?

Amazon is also offering its own foundation models under the name Amazon Nova. Setting aside performance questions, how does the pricing compare with other models?

Google’s Gemini models are priced closest to Amazon, and in the case of Gemini 1.5 Flash-8B Google is significantly cheaper. Google also has its own chips, TPUs (for Tensor Processing Units) which makes it more cost-competitive than relying on Nvidia, but perhaps only for Google models.

It will be interesting to watch the Amazon - Google - Salesforce dynamic play out in 2025 and to see how Microsoft and OpenAI respond. Here are some possibilities:

• The compute cost for agents using SLLs will drop rapidly, while the compute cost for LLMs will grow as the use of reasoning models expands - this will lead to two very different markets with different pricing strategies

• Virtually all companies with an agent packaging model will move to Amazon, Google or Salesforce

• There will be a major fight for mindshare between the Microsoft (and to some extent Google) co-pilot packaging and the Salesforce (and it seems Amazon) agent packaging - both will win and establish their own niches in the larger ecology

All this will complicate packaging and pricing decisions for B2B applications and agent providers. Ibbaka is working to define best practices for the packaging and pricing of B2B generative AI, whether it is delivered as an application, a co-pilot, or an agent.