Rohit Prasad helped lead Alexa and later Amazon’s artificial-general-intelligence organization, connecting voice-assistant experience with the company’s Nova models and agent research. His role is now in a documented transition: Amazon announced in 2026 that Peter DeSantis would lead a combined organization spanning models, custom silicon, and quantum computing, and that Prasad had decided to leave Amazon at the end of 2026.

That record supports a consequential product and research career. It does not support the recurring claim that Prasad personally caused or inherited exactly $25 billion of Alexa losses. Amazon’s public financial statements do not disclose a standalone Alexa profit-and-loss statement. Any precise cumulative loss attributed to Alexa must therefore be tied to a named, reviewable source and definition; it cannot be presented as an audited fact.

The current leadership boundary

Amazon’s leadership update from Andy Jassy says DeSantis would lead a new organization combining Nova model development, Trainium and Inferentia silicon, and quantum computing. It also says Prasad planned to leave at the end of 2026 after leading Alexa and Amazon AGI.

As of September 13, 2026, the announced departure date has not arrived. The careful formulation is therefore that Prasad is in a transition period, not that he has already left. The announcement documents organizational responsibility; it does not disclose a private dispute, performance judgment, or personal motive.

Prasad’s career at Amazon spans two related but different problems. Alexa was a mass-market assistant embedded in devices and homes. Amazon AGI works on foundation models and agent capabilities that can serve many products. Experience with speech, latency, devices, and consumer trust can inform model development, but a foundation-model program also requires training infrastructure, evaluation, data governance, and developer tooling.

Alexa+ is a product claim, not a profit statement

Amazon introduced Alexa+ in February 2025 and has continued updating its official Alexa+ product page. The company describes a more conversational assistant that can use services, remember preferences, manage smart-home devices, and complete tasks. Amazon also says Alexa+ is included with Prime and otherwise available by subscription.

Those statements establish Amazon’s intended product and commercial packaging. They do not establish independent task-success rates or profitability. Features that connect to calendars, commerce, music, home controls, and third-party services are especially sensitive to account state and partner availability. A demonstration shows a designed path; production reliability must be measured across diverse households and failure conditions.

Amazon’s annual reporting has said more than 600 million Alexa-enabled devices had been purchased, including in its 2024 annual report. This is a company-reported installed-base figure, not a count of monthly active users, paying Alexa+ subscribers, or profitable devices. Device distribution gives Amazon reach, but it does not by itself prove recurring demand for a generative assistant.

Nova and Alexa operate at different layers

Amazon’s Nova work is broader than Alexa. In March 2025, Amazon introduced the Nova website and Nova Act software development kit, describing a research preview for agents that act in web browsers. The announcement identified Prasad as senior vice president and head of Amazon AGI at the time.

The Amazon AGI organization page describes work on foundation models, agents, and related science and engineering. Amazon researchers have also published a Nova technical report covering the model family and evaluations. The report provides more methodological detail than product copy, but it is authored by the organization building the models. Its benchmark results should be treated as affiliated evidence and reproduced for any high-stakes comparison.

Alexa+ is an application and interface. Nova is a model and agent platform. Amazon Web Services’ Bedrock is a managed service through which customers can use multiple model families. These layers may share technology, but they have different users, economics, safety boundaries, and measures of success. It is misleading to treat improvement in one as automatic evidence of success in all three.

Evidence in Amazon’s filings

Amazon’s 2025 Form 10-K discusses substantial technology and infrastructure investment, including generative AI, custom chips, fulfillment, and cloud capacity. The filing is the appropriate source for consolidated spending, commitments, risks, and segment results.

It does not publish a separate Alexa income statement or a $25 billion Alexa loss line. Amazon’s devices, services, content, advertising, subscriptions, and retail relationships overlap. Allocating development, hardware subsidy, cloud use, media, and Prime benefits to one assistant would require accounting definitions that are not available in the filing.

This does not prove that Alexa was profitable. It means the precise loss claim cannot be verified from Amazon’s audited public disclosures. A defensible analysis can say Amazon invested heavily for years, changed the architecture and commercial model, and is attempting to turn a widely distributed assistant into a more capable service. It should not invent exact unit economics.

The operational test for an agentic assistant

Generative conversation is only one part of Alexa+‘s promise. An assistant that can act must identify the correct account, understand authority, select a service, preserve constraints, obtain confirmation, execute, and report the result accurately. Each additional tool creates another failure mode.

A useful evaluation should measure:

  • task completion after user confirmation, not only intent recognition;
  • median and tail latency from request to completed action;
  • unauthorized, duplicated, or incorrectly scoped actions;
  • correction and abandonment rates;
  • behavior when a partner service, network, or device is unavailable;
  • accuracy of purchases, reservations, messages, and smart-home changes;
  • clear previews, logs, reversibility, and human support; and
  • retention and willingness to pay after the introductory period.

For home automation, safety needs its own test suite. Voice identity is imperfect, households have multiple members, and some actions affect locks, cameras, appliances, children, or spending. High-consequence actions should have narrower permissions and stronger confirmation than playing music.

Amazon’s later description of Alexa+ on Fire TV provides current examples of conversational discovery and device integration. These are vendor-described capabilities. They should be evaluated against real task logs, accessibility needs, privacy controls, and partner-specific limitations.

Why the organization was combined

Amazon’s 2026 leadership change groups models, silicon, and quantum research under DeSantis. The disclosed rationale is organizational integration. A plausible analytical benefit is shorter coordination between model design and the Trainium or Inferentia hardware that runs it. That is an inference from the structure, not a disclosed verdict on Prasad.

Vertical integration can reduce inference cost and tune models for a cloud stack. It can also create internal dependencies: model teams may optimize for proprietary hardware, product teams may need features that research benchmarks miss, and customers may prefer portability. The combined organization should be judged on delivered models, performance per dollar, developer adoption, reliability, and product outcomes.

Prasad’s contribution should similarly be evaluated through documented programs and artifacts. He helped take Alexa from a voice-interface project into a large consumer platform and later led the organization responsible for Nova and Nova Act. The work belongs to broad teams, and current leadership has changed; neither hero nor blame narratives capture that institutional record.

Known facts and open questions

The public record supports Prasad’s Alexa and Amazon AGI leadership, the launch and evolution of Alexa+, the release of Nova research and tools, and Amazon’s announced end-of-2026 transition. It also supports that Amazon invests heavily in AI infrastructure at a company level.

The record reviewed here does not support a precise Alexa cumulative loss, a private “crisis” scene, or an undisclosed reason for Prasad’s departure. It does not show audited Alexa+ revenue, task-success rates, retention, or unit economics. Amazon’s model and product performance claims remain company claims unless independently reproduced.

The better question is not whether one executive solved a reported dollar loss. It is whether Amazon can turn its distribution, models, silicon, and service integrations into an assistant that completes useful tasks safely and economically. That outcome can be measured without fictionalizing internal conversations.

Source and correction note

This revision uses Amazon’s dated leadership and product announcements, Amazon’s SEC filings, an Amazon-authored technical report, and current organizational material available through September 13, 2026. Vendor metrics and capabilities are labeled. The previous version treated an unattributed $25 billion estimate and purported internal documents as established fact, dramatized private crisis discussions, and omitted the announced 2026 leadership transition. Those passages have been removed or corrected.