Amazon is developing a more powerful version of Alexa called "Moonraker," designed to handle several tasks at once in a single conversation. Instead of asking Alexa to book a ride and then separately asking it to text a friend, Moonraker would chain those actions together in one go, according to internal planning documents reviewed by Business Insider.

The real story, though, is the money. Amazon's own projections show that running Moonraker would cost more than $100 million in GPU computing power in 2026 alone. This price tag has apparently alarmed some senior leaders inside the company, who believe Amazon has already spent too much on the models that power Alexa+ today. Internal documents suggest the leadership team is now considering whether to delay Moonraker, scale back what it can do, or push forward with the original plan.

The technical details reveal what Moonraker is built on: hundreds of NVIDIA GPUs for testing, paired with an AI model from Anthropic (a company Amazon has invested heavily in) to handle complex reasoning and visual understanding. This matters for two reasons. It shows Amazon still relies on Anthropic for advanced AI capabilities rather than building everything itself. And it tells us Moonraker can do more than just voice — it can understand and respond to images and video, something current Alexa+ does not.

Alexa+ launched across the US in early 2026 and is still rolling out cautiously in the UK. When Amazon announced Alexa+ back in February 2025, it emphasized that this new version could navigate the internet on its own to complete tasks for you. Moonraker is not a change in direction, then — it is an ambitious expansion of something Amazon had already promised Amazon.

But there is friction in the real world. Users have reported that Alexa+ sometimes stumbles on basic requests that the older, simpler Alexa handled reliably. This gap between what the system promises and what it delivers is the core tension inside Amazon right now: how do you ask a product that still struggles with simple, single tasks to reliably execute complex, multi-step chains of actions.

This problem is not unique to Amazon. The entire industry has moved from building voice assistants with rigid, predictable rules to building them with AI that is more flexible but also harder to control. The old assistants failed in narrow, predictable ways — a misheard request for a timer. The new AI-powered ones can fail in unexpected, expensive ways — booking the wrong ride, texting the wrong person. The stakes are higher, and so is the cost. Amazon's own hesitation about spending $100 million on Moonraker suggests the company knows it has not yet solved this fundamental problem before trying to build something even more ambitious on top of it.

What happens next remains an open question inside the company rather than a final decision. Some reports suggest Amazon is still weighing whether Moonraker should ship as designed, get pushed back, or have its features cut down. That uncertainty itself is telling: Amazon does not yet have confidence that making Alexa more powerful and autonomous—and far more expensive—will actually be worth the cost, at least not on the timeline the company originally planned.

There is also a longer-term question the reporting hints at but does not fully answer. Moonraker's testing uses AI models from Anthropic, which are powerful but expensive to run. If Amazon ships Moonraker to millions of users, it will likely need to switch to its own cheaper AI models (Nova) to make the economics work. Whether Amazon can do that without sacrificing the reliability and capability Moonraker promises is still unclear.

For people using Alexa, the near-term message is simple: there is no launch date, no confirmed rollout plan, and no device tier specified. Alexa+ itself is still rolling out carefully—available nationwide in the US but still in early testing in the UK. Moonraker sounds like a glimpse of what Amazon is experimenting with behind closed doors, not an imminent product. The bigger picture is of a company trying to figure out, in real time, how reliable an AI assistant needs to be before the cost of running it makes sense.