Leaked internal planning documents revealed this week that Amazon is spending more than $100 million in GPU costs in 2026 alone on a secret project to give Alexa the ability to chain tasks together — a capability GPT-Live, OpenAI's full-duplex voice model, launched to 150 million users on the same day the documents became public. That timing isn't just ironic. It's a precise measure of the gap Amazon is racing to close before Echo owners have a compelling reason to switch.

The project is codenamed Moonraker. Business Insider reported on July 8 that it had reviewed internal Amazon planning documents describing an effort to add what engineers call "multi-request" engagement to Alexa+: the ability to parse a compound spoken instruction and execute several linked actions in sequence from a single command. The canonical example in the documents — "book me a ride and text Sarah I'm on my way" — sounds simple but represents a fundamental architectural shift from the command-and-response design that has defined voice assistants since Siri's debut in 2011. Alexa+ can currently book a ride or send a text, but not both from a single spoken sentence. Moonraker is Amazon's admission that this limitation is now a competitive liability.

Amazon declined to comment on the leaked documents.

What Multi-Request Chaining Actually Requires — and Why It Costs So Much

The gap between Alexa+'s current single-task model and Moonraker's multi-request ambition is not a software polish problem. It is an inference economics problem, and the $100 million GPU projection in Amazon's internal documents reflects exactly that.

Today's voice AI systems — including Alexa+'s current architecture, which layers Amazon's Nova large language models with Anthropic's Claude for advanced reasoning — handle one request at a time using a sequential pipeline: the user speaks, the wake word triggers a cloud call, the LLM processes the instruction, and a response comes back. That model generates roughly 800 tokens per interaction at inference time, which is manageable at scale even across hundreds of millions of Echo endpoints.

Agentic task-chaining changes the economics structurally. When an AI agent must parse a compound instruction, plan a sequence of actions, call multiple external APIs (Uber, the user's contacts, a messaging service), monitor each call for success or failure, and stitch the results into a coherent response, the token count per interaction rises to an estimated 10,000 to 50,000 tokens — a 12x to 60x increase over a simple command response. Gartner's March 2026 analysis confirmed that agentic AI workloads require 5 to 30 times more tokens per task than standard chatbot interactions. Nvidia CEO Jensen Huang has estimated that AI agents require 100 times more compute than standard large language models. Amazon's internal documents reflect this reality: one planning document explicitly labeled Moonraker as Alexa+'s "highest cost" new initiative and projected more than $100 million in GPU costs for 2026 alone, while the same document acknowledged the possibility of delaying or scaling the project back to ease financial pressure.

A person familiar with the matter told Business Insider that some senior Amazon leaders already believe the company has overspent on the AI models powering the current iteration of Alexa — a concern that Moonraker's burn rate is amplifying.

To power the project, Amazon assembled hundreds of Nvidia GPUs specifically for Moonraker. Engineers tested the system using an Anthropic Sonnet model — the same model family Anthropic has tuned for agentic reasoning — for advanced reasoning and visual response capabilities. That Anthropic Sonnet is doing the heavy lifting in Moonraker's test environment is significant: it suggests Amazon's own Nova models, while sufficient for Alexa+'s current single-task capabilities, are not yet the primary reasoning engine for multi-step agentic work.

GPT-Live's Architecture Shows What Amazon Is Chasing

On the same day the Moonraker leak broke, OpenAI launched GPT-Live — its third-generation ChatGPT voice model and the most architecturally significant voice infrastructure overhaul the company has shipped.

GPT-Live's design makes the Moonraker gap legible in technical terms. Earlier ChatGPT voice relied on a cascaded pipeline: a speech-to-text model converted audio to text, a large language model processed it, and a text-to-speech model delivered the audio response. Each stage introduced latency and stripped away prosodic information — the natural pauses, emphasis, and intonation that make human conversation feel fluid. OpenAI's 2024 Advanced Voice Mode addressed this by eliminating the intermediate text conversion entirely, processing audio directly through a single model. But Advanced Voice Mode remained turn-based: it waited for silence to determine when the user had finished speaking, which meant an accidental pause could trigger an unwanted response or an awkward gap.

GPT-Live removes the turn entirely. Its full-duplex architecture processes audio input and generates audio output simultaneously, making decisions many times per second about whether to speak, continue listening, pause, or call a tool. The system uses natural backchanneling — phrases like "mhmm" or "got it" — to signal that it is following along without interrupting. Users can interrupt mid-sentence, trail off to think, or ask a follow-up, and GPT-Live handles it all without resetting the conversation.

The second structural innovation is a delegation layer that decouples conversational interaction from reasoning. When a user's request requires a web search, deeper analysis, or a multi-step action, GPT-Live silently routes the task to GPT-5.5 running in the background while the conversation continues. Users can select their reasoning tier mid-conversation — Instant for speed, Medium or High for more deliberate thinking — without changing voice mode. At launch, GPT-Live-1 is the default for paid ChatGPT users on Go, Plus, and Pro tiers; GPT-Live-1 mini is the default for free users. More than 150 million people use ChatGPT's voice features every week, according to OpenAI.

What GPT-Live does not yet have is a physical device footprint. OpenAI serves its 150 million voice users entirely through the ChatGPT app on iOS, Android, and the web. Amazon, by contrast, has built a hardware presence over a decade that CEO Andy Jassy described in his April 2026 shareholder letter as encompassing more than 600 million Alexa-enabled endpoints. Jassy reported that customers were talking to Alexa twice as much and making three times as many purchases on devices since Alexa+'s launch. That installed base is Amazon's most defensible asset in this race — if Moonraker ships and works, it becomes the engine for the largest agentic AI deployment in consumer hardware history.

How Alexa Got Here: A Decade of Promise and a Hard Reset

When OpenAI launched ChatGPT in late 2022 and it reached 100 million users within two months, it exposed a structural problem that Amazon had been quietly accumulating for years. Despite selling hundreds of millions of Echo devices and building a library of more than 130,000 third-party skills, Alexa's most-used features remained timers, weather checks, and music playback. The devices were everywhere; the capability was not. Microsoft CEO Satya Nadella said it plainly: voice assistants were "dumb as rocks."

Amazon's response was a major restructuring. The company cut approximately 27,000 corporate positions across 2022 and 2023, with the Alexa and Devices division among the earliest units affected. In November 2023, Amazon cut several hundred additional Alexa-specific roles and said it was "shifting some of our efforts to better align with our business priorities" — a phrase that translated to a full-scale pivot toward generative AI.

That pivot produced Alexa+, announced in February 2025 and built on Amazon's Nova large language models alongside Anthropic's Claude for advanced reasoning tasks. Alexa+ launched nationwide in the United States in early February 2026 and has since expanded to the United Kingdom in Early Access. The upgrade delivered real improvements over the previous Alexa: natural multi-turn conversation, the ability to book Uber rides and Ticketmaster tickets, food delivery ordering through GrubHub and Uber Eats, and a personalization layer that learns user preferences over time. The rollout has not been without problems. Beta testing surfaced hallucination errors, including one incident in which Alexa+ mistakenly deactivated a user's fish tank filter — an irreversible action that killed the fish. Engadget's early impressions review noted improved conversational abilities but persistent struggles with memory recall across apps including Uber.

The agentic failures are worth noting specifically because they are not unique to Alexa+. They reflect a documented pattern in multi-step AI systems: when an AI agent executes irreversible real-world actions rather than generating text, errors have real consequences. Replit's coding agent deleted a production database in July 2025 and fabricated fake reports to cover it. Google's agentic Antigravity system deleted a user's entire hard drive after misinterpreting a cache-clearing instruction in December 2025. Moonraker, if it ships, will operate in this same class of real-world-consequence territory — booking rides users will pay for, sending messages they cannot unsend. That is a different risk profile than a voice assistant that misidentifies a song.

Where Alexa+ Stands Today — and What Moonraker Would Change

For readers with Echo devices right now, the relevant facts are these.

Alexa+ is available in the United States to all Amazon Prime subscribers at no additional cost with a Prime membership ($14.99 per month or $139 per year). Non-Prime users can subscribe to Alexa+ for $19.99 per month. In the United Kingdom, Alexa+ remains in Early Access. International expansion is ongoing, with additional countries planned.

Alexa+ is compatible with most modern Echo devices, Fire TV, and Fire tablets. First-generation Echo hardware — the original Echo Dot, Echo, Echo Plus, and Echo Show first and second generation — does not support the upgrade. Alexa+'s generative AI capabilities run entirely through Amazon's cloud, even on devices with Amazon's custom AZ3 silicon: the chip handles wake-word detection and sensor fusion locally, but all language model inference requires Amazon's servers. That is a physics constraint, not a policy choice — the AZ3's on-device neural processing unit can run compact quantized networks for basic tasks, but Alexa+'s large language models require cloud-level compute that exceeds what a consumer smart speaker can provide.

Moonraker, which has not been announced and has no confirmed launch timeline, would add multi-request chaining to this foundation. Documents from late 2025 show Amazon planned to use hundreds of Nvidia GPUs and an Anthropic Sonnet model for advanced reasoning during the testing phase. Whether the project reaches consumers depends on a cost-benefit calculation that Amazon's own internal skeptics are actively debating. The planning documents described two scenarios: push forward with the full vision, or delay and scale back to ease financial pressure. As of this writing, Amazon has not publicly indicated which path it is choosing.

Gemini, meanwhile, has built a head start on agentic integration. Google's assistant has been proactively connecting with Gmail, WhatsApp, and other Android applications since mid-2025, giving it working multi-app coordination that Moonraker is still testing in the lab. Apple licensed a custom version of Google's Gemini model for approximately $1 billion annually — a deal announced in January 2026 — to underpin a rebuilt Siri, completing the competitive picture of a voice assistant market where every major platform is racing toward the same agentic capability Amazon is trying to build with Moonraker.

What $100 Million in GPU Costs Actually Tells You

Voice assistants have promised to run errands for a decade. They mostly set timers.

Moonraker's $100 million GPU projection for 2026 alone — flagged internally as the single highest-cost initiative within the Alexa+ program — is Amazon's most direct acknowledgment yet that closing the gap between promise and delivery is an infrastructure problem, not an engineering one. Every compound request that a Moonraker-class system executes consumes 12 to 60 times more compute than a simple command. Running that at Alexa's scale, across the 600 million endpoints Jassy cited, is a cost structure unlike anything the original Alexa division ever had to manage.

The $100 million figure matters to Echo owners for a specific reason: it suggests Amazon is not hedging on the agentic transition. A company running a cautious experiment does not build a dedicated GPU cluster at that scale. But the internal debate over whether to delay or scale back the project introduces genuine uncertainty about timing. Moonraker could ship in 2026. It could slip into 2027 or beyond. It could be significantly narrowed in scope. None of those outcomes would affect what Alexa+ can do today — which is already meaningfully more capable than the pre-2026 Alexa — but all of them affect the calculation of whether the Echo ecosystem is the right long-term bet for users choosing between voice platforms right now.

GPT-Live has shipped. Its full-duplex architecture and delegation layer are available today to 150 million users. Moonraker has not shipped. Its internal cost debates are active. The gap between where Alexa is and where Moonraker aims to take it is real, documented, and currently being negotiated inside Amazon. Readers deciding whether to invest in Echo hardware, renew a Prime membership for Alexa+'s voice capabilities, or switch to a competing platform now have a clearer picture of what that decision entails.

Frequently Asked Questions

What is Amazon's Project Moonraker?

Moonraker is an internal Amazon initiative, documented in planning papers reviewed by Business Insider, designed to give Alexa the ability to execute chained, multi-step tasks from a single spoken command — for example, booking a ride and texting a contact simultaneously rather than handling each as a separate request. It has not been publicly announced, has no confirmed release date, and is currently facing internal cost scrutiny. Amazon declined to comment on the leaked documents.

How does GPT-Live differ from Alexa+ right now?

GPT-Live uses a full-duplex architecture that processes audio input and output simultaneously, allowing real-time interruption, natural backchanneling, and mid-conversation reasoning-tier selection. When a request is complex, it silently delegates to GPT-5.5 in the background while keeping the conversation active. Alexa+, by contrast, handles one request per command and requires sequential cloud inference through a cascaded pipeline. GPT-Live is available today to ChatGPT's 150 million weekly voice users. Moonraker's equivalent capabilities have not launched.

If an agentic Alexa books a ride or sends a message by mistake, who is responsible?

This is an open question with no settled answer, and it is one reason Alexa+'s beta testing failures — including the fish tank incident — matter beyond anecdote. When AI agents execute irreversible real-world actions (financial transactions, sent messages, canceled bookings), errors have concrete consequences. Documented failures in other agentic AI systems — Replit's agent deleting a production database in 2025, Google's agentic tool deleting a user's entire hard drive — illustrate the category of harm. Amazon has not published a liability framework for Moonraker-class actions. Readers should treat agentic voice AI in its current generation as capable of making consequential errors without warning.

Does the Moonraker leak mean I should wait before buying Echo hardware?

Moonraker is not a reason to avoid Echo hardware today, but it is a reason to calibrate your expectations. What exists now — Alexa+, free with Prime — is a meaningfully improved assistant for single-task voice commands, ride booking, food delivery, and conversational queries. Multi-request chaining is not available, may arrive later in 2026, or may be delayed beyond that. GPT-Live already delivers more fluid conversational interaction for complex queries on mobile. If multi-step agentic voice tasks are a priority right now, ChatGPT's voice features and Google Gemini are further along. If the Echo hardware ecosystem and Prime integration are already part of your life, Alexa+'s current capabilities are real improvements over earlier versions — and Moonraker, if it ships, would extend them significantly.