AI Agents & Automation
What Amazon's Moonraker Project Says About the State of AI Agents
Internal documents reveal Amazon is spending over $100 million on GPUs to turn Alexa into an AI agent that handles multi-step tasks. The cost, and the rocky rollout so far, tell a bigger story about where agentic AI actually stands for the rest of us.
July 8, 2026 · 5 min read
What changed
Amazon is working on a project codenamed Moonraker to give Alexa agentic AI abilities, according to internal planning documents seen by Business Insider. The goal is an Alexa that handles multi-step tasks from a single voice command — "book me a ride and text my friend" — rather than needing a separate request for each action. The documents project over $100 million in GPU costs for 2026 alone, with plans to deploy hundreds of NVIDIA GPUs. Engineers are reportedly testing with Anthropic's Sonnet model for the reasoning layer. With over a billion Alexa-enabled devices already in homes, Amazon has more distribution than any competitor in the AI agent race.
The honest nuance
The same internal documents floated delaying or scaling back Moonraker to ease cost pressure, and some senior leaders believe the Alexa team has already overspent on the underlying AI models. Alexa+ itself, the generative-AI-powered version that launched earlier this year, has had a rough start: multiple delays before a US launch, reports of hallucinations, and at least one documented incident where a beta tester claimed Alexa turned off a fish-tank filter and killed their fish. Amazon also declined to comment on the Moonraker leak, which means everything here is based on planning documents, not shipped product. If a company with Amazon's resources is wrestling with cost and reliability at this level, that is worth paying attention to.
Why it matters
Amazon spending $100 million to make agents work is not a signal that your small business needs to do the same. It is the opposite. The big players are doing the expensive, failure-prone R&D right now so that by the time agentic AI is reliable enough for everyday use, the core capabilities will be commoditized and available through tools you already have. The pattern that actually matters for a small business hasn't changed: pick one bounded, repeatable task, use a no-code tool to chain a trigger to an AI step, and keep a human review on the output. That approach already works. You don't need Moonraker's budget or Amazon's GPU fleet to get value from it.
What to actually do
Pick one task you or your team does the same way every time — sorting incoming leads, drafting a standard reply, pulling together research notes for a client call — and automate just that one thing. Use Zapier, Make, or n8n. Add an AI step that drafts or sorts the output, then review it yourself before it goes anywhere. Run that single workflow for two weeks before you even think about adding a second one. The businesses getting real value from AI agents right now are not the ones waiting for Amazon or Google to ship a finished product. They are the ones running small, boring automations with a human check at the end.
From Kindloom Labs
The AI Small Business Starter System bundle includes a setup guide for exactly this kind of low-lift automation, scaled for a business without a dedicated ops person.
Sources
Keep reading
Stay in the loop
Get notified about new releases, content, and blog updates from Kindloom. No spam, unsubscribe anytime.