Amazon isn't walking away from AI. It is cutting the messy middle: older AWS tools, overlapping agent products, and some of the research labor behind customization.
Amazon's AI reset is less dramatic than the headline version and more useful for you if you buy cloud software. AWS has not killed every AI service it launched in the last two years. It has moved a long list of services and SageMaker AI features into maintenance, and the sharpest names on that list are Amazon Kendra, Amazon Q Business, and Bedrock Agents, now called Bedrock Agents Classic.
That distinction matters. AWS said on June 30 that services moving to maintenance will no longer be open to new customers starting July 30, while existing customers can keep using them with support. So this isn't a switch-off. It's a warning label. If your team was about to standardize on Kendra or Bedrock Agents Classic, you now have to ask why AWS is pointing new work somewhere else.
Forbes wrote on July 23 that AWS was retiring the first generation of several AI services from new customer growth, including Q Business, Kendra and Bedrock Agents. Some of the timing is brutal. Bedrock Agents launched in November 2023, just as companies were trying to turn chatbot demos into internal tools that could search documents, call APIs and get things done. Less than three years later, AWS is telling new customers to look at AgentCore instead.
The job cuts tell the same story from inside Amazon. Reuters reported on July 22 that Amazon cut roles in its artificial general intelligence group, including employees under Adeeb Shanaa, vice president of AGI data services, and Vishal Sharma, vice president of AGI information. Amazon did not disclose the number of affected employees. A spokesperson told Reuters that the company was "sharpening our focus on the initiatives that matter most for customers, so we can move faster on what counts."
That statement is bland. The product moves are not.
AWS wants fewer doors into the same stack
AWS is steering customers toward broader platforms. Kendra users are being directed toward Bedrock Managed Knowledge Base, which AWS announced on June 17 with native data connectors, hybrid search and an agentic retriever for enterprise RAG work. Q Business customers are being pointed toward Amazon Quick, the broader Quick Suite workspace that Amazon launched in October 2025 after evolving QuickSight into a bigger analytics and automation product. Bedrock Agents Classic customers are being sent to Bedrock AgentCore, the agent platform AWS first put into preview in July 2025 and expanded heavily in 2026.
You can see the sales logic. A narrow search product and a workplace assistant each need their own positioning, pricing conversation and migration path - and so does an agent builder. A larger platform gives AWS one place to land the customer. It also makes the customer harder to lose once the workflows, indexes and permissions are inside the same system.
Amazon's AI budget is not shrinking. The Information reported that Amazon expects capital spending to reach about $200 billion in 2026, up from $131.8 billion in 2025, with spending focused on AI, chips, robotics and its Leo satellite internet service. Amazon is still buying the infrastructure. It is being more selective about the products and people wrapped around it.
Nova Forge sits right in that calculation. Amazon introduced Nova Forge in December 2025 as a way for organizations to build customized Nova model variants by using checkpoints across stages of training and blending their own data with Amazon-curated data. The roles Reuters described - data services, model work and customization-adjacent functions - sit close to the kind of capability Amazon now wants to package for customers. If you're selling customization as a service, you don't keep every part of that work as bespoke internal labor.
The cleanup is industry-wide
Amazon isn't alone in pruning AI sprawl. Microsoft announced a Copilot leadership update on March 17, bringing consumer and commercial Copilot work together while Mustafa Suleyman shifted toward frontier model and superintelligence work. Google has used voluntary buyouts and smaller cuts across several units this year, including search, core engineering and research, according to reporting from The Information and CNBC-linked coverage. The whole sector overbuilt menus in 2023 and 2024. Now the menus are getting shorter.
Frankly, that part is healthy. Enterprise buyers don't need 12 overlapping ways to build an internal assistant. They need a product that survives budget season, passes security review and still has a roadmap after the champion who bought it leaves the company.
The risk is trust. If your company built workflows on Bedrock Agents Classic, a migration to AgentCore may be sensible, but it still means engineering work. If you chose Q Business for internal knowledge search, Amazon Quick may be the future, but it is not the same procurement decision you made before. AWS says existing users will keep support, security patches and bug fixes, but no new features. That's a runway, not comfort.
The harder question is whether Amazon can cut AGI roles and still credibly chase frontier models. Business Insider reported on July 28 that Amazon is also winding down several in-house Nova models, including Premier, Omni, Reel and Canvas, while shifting resources toward a new Frontier Model Research effort led by Pieter Abbeel. That makes the strategy clearer. Fewer models, fewer tools - and a sharper bet on what remains.
If Amazon wants to be the infrastructure layer for models from Anthropic, OpenAI, its own Nova line and whatever comes next, this cleanup makes sense. If it wants to win the frontier model race outright, cutting around post-training and customization is a harder call. You can't sell depth forever by narrowing the bench that builds it.
Also read: Tesla spent nearly $2 billion buying an AI hardware company and told almost no one • Dario Amodei wants a harder China chip ban that would only speed up the loss he is trying to prevent • Snappyit Brings AI Ghost Mannequin Shots To Fashion Sellers Without The Studio Bill
