Inside Amazon's Secret 'Moonraker' Project: Turning Alexa Into a Full AI Agent

One internal document from earlier this year labeled Moonraker Alexa+'s "highest cost" new initiative. The same document suggested the team consider delaying or scaling back the project due to expense . According to the sources, Amazon's total capital expenditure for 2026 is approximately $200 billion, much of it tied to AI infrastructure .

Projected GPU Cost in 2026

Internal planning documents project more than $100 million in GPU costs in 2026 alone for Moonraker . This figure reflects the immense computational expense of running large language models at scale, particularly for a consumer product that processes millions of requests daily.

The AI Models Behind Moonraker

Moonraker has been tested on Anthropic's Claude models, which Amazon already uses extensively through its deep partnership with Anthropic . Amazon has invested more than $20 billion in Anthropic and secured up to 6 GW of Trainium chip capacity for running Claude models . In June 2026, Amazon restructured its Anthropic billing from compute-hours to a token-based model, a shift some internally disputed as potentially raising costs .

Moonraker in the Broader AI Agent Race

Moonraker is Amazon's attempt to catch up in the increasingly competitive AI agent market, where rivals like Google (Gemini agents), OpenAI (Operator / ChatGPT agents), and Microsoft (Copilot agents) have already launched agentic capabilities . The Verge characterized Moonraker as Amazon "trying to catch up" in the agent race . The industry is moving rapidly from simple chatbots to autonomous agents that can book travel, manage calendars, and execute purchases on behalf of users.

Internal Spending Concerns

The spending concerns are acute and multi-layered:

• Moonraker was flagged internally as a "highest cost" project, with the $100M GPU projection noted as a risk .
• Alexa as a whole has cost Amazon an estimated $25 billion-plus over its lifetime with limited direct revenue return .
• Internal planning documents explicitly suggested the team consider delaying or scaling back Moonraker due to cost .
• The shift from compute-hour billing to token-based pricing with Anthropic has also caused internal friction over cost predictability .

Alexa+ Rollout: A Troubled History

Rollout timeline:

• Alexa+ was unveiled at a Manhattan press conference in February 2025 .
• It entered Early Access beta in late 2025, with over 6,400 Amazon employees providing feedback via an internal Slack channel .
• By January 2026, Amazon automatically enrolled all Prime members into Early Access, even those who hadn't opted in—sparking backlash from users who didn't want the service .
• In June 2026, the broader public rollout continued, with users rapidly reporting issues on social media .

Beta test issues:

Beta testers described the service as "unbearably erratic" . Notable failures included Alexa+ turning off an aquarium power strip and killing a user's fish . Employees reported ignored commands, erratic performance, and harmful errors like shutting down critical home equipment .

Early internal tests in late 2024 scored Alexa+ at only 4.57 out of 7 on satisfaction—well below the 5.5 target . Latency was a "critical issue": response times reached up to 10 seconds on older Echo devices, far above the 2–4 second internal goal . Engineers in Boston and Seattle reportedly worked 80-hour weeks for two years to make Alexa+ conversational and hit an internal target of 97% correct answers .

Jassy's reported usage metrics:

In his April 2026 shareholder letter, CEO Andy Jassy reported that customers are talking to Alexa twice as much as before, for longer durations and across wider topics . Purchases completed on Alexa devices increased 3x, music streaming rose 25%, and smart home functionality usage jumped 50% . Jassy also disclosed that AWS AI revenue hit a $15 billion annual run rate, and that Amazon's internal chip business (Graviton/Trainium) is generating over $20 billion a year .

Industry Tensions: The Cost of AI at Scale

Moonraker's cost concerns are not happening in a vacuum. Several broader industry dynamics are relevant:

• Massive infrastructure spending: Amazon's ~$200B capex in 2026 is one of the largest corporate investment programs in history, driven almost entirely by AI demand .
• Anthropic's $100B AWS commitment: Anthropic has agreed to spend over $100 billion on AWS over 10 years, securing up to 5 GW of computing capacity .
• Custom chip pivot: Amazon is aggressively pushing its own Trainium and Graviton chips as lower-cost alternatives to Nvidia GPUs, with two large customers already asking to buy all of Amazon's available Graviton capacity for 2026 .
• Token-based billing dispute: Amazon shifted Anthropic from compute-hour to token-based pricing in mid-2026, which some internally argued could drive up costs rather than control them .

The Bottom Line

Moonraker represents Amazon's most ambitious bet yet on AI agents—and also its most expensive. With projected GPU costs exceeding $100 million in a single year, and a history of Alexa's cumulative losses of over $25 billion, the project embodies the central tension of the AI industry: staggering infrastructure costs against uncertain consumer monetization. Whether Amazon can afford to build Moonraker—and whether consumers will pay for it—remains an open question.