PLUS: Google drops Nano Banana 2 and Perplexity unveils autonomous Computer agent

Good morning, AI enthusiast. OpenAI has secured a massive financial backing from Amazon to push the boundaries of autonomous agents. The $50 billion agreement designates AWS as the sole cloud partner for upcoming enterprise platforms, signaling a major shift in infrastructure strategy.

As the company looks to move beyond simple chatbots, this capital injection aims to power persistent, memory-aware systems for businesses. But with such a heavy reliance on AWS, will this exclusivity impact the broader landscape of model availability?

In today’s AI recap:

• OpenAI and Amazon’s $50B agent partnership
• Google’s fast, consistent Nano Banana 2
• Eli Lilly’s massive GPU-powered drug discovery
• Perplexity’s new autonomous Computer agent

OpenAI closes historic $50B deal with Amazon

The Recap: OpenAI and Amazon announced a massive strategic partnership today, featuring a $50 billion investment to accelerate the development of autonomous AI agents. This deal positions AWS as the exclusive cloud provider for OpenAI’s new enterprise-focused platforms.

Unpacked:

• Amazon plans to invest $50 billion in OpenAI while expanding an existing agreement to provide the massive computing power needed for future models.
• The companies are co-developing a Stateful Runtime Environment on Amazon Bedrock that allows AI agents to maintain context and memory across complex workflows.
• AWS becomes the exclusive third-party cloud provider for OpenAI Frontier, a platform that lets organizations deploy and manage fleets of agents at scale.

Bottom line: This partnership secures the specialized compute capacity OpenAI needs to move beyond chatbots into fully autonomous enterprise agents. It suggests that the next phase of AI value lies in persistent, memory-aware systems integrated directly into business infrastructure.

Google drops Nano Banana 2 image model

The Recap: Google has officially released Nano Banana 2, a studio-quality image generation model built on Gemini 3.1 Flash that brings real-time search grounding to visual creation.

Unpacked:

• This new architecture prioritizes speed without sacrificing fidelity, offering users subject consistency for up to five characters to maintain narrative continuity across multiple images.
• The model leverages real-time data to generate accurate charts and localized text, now available for developers to build upon via the Gemini API.
• Google is deploying the model across its full product suite, replacing older versions in the Gemini app and launching immediately for enterprise users on Vertex AI.

Bottom line: Combining photorealism with low latency removes a major friction point for creative professionals iterating on complex concepts. This update positions speed as a critical feature for adoption alongside raw image quality.

Eli Lilly activates massive ‘LillyPod’ AI factory

The Recap: Pharmaceutical giant Eli Lilly has officially activated LillyPod, the world’s first privately owned AI factory powered by over 1,000 NVIDIA Blackwell Ultra GPUs. This massive infrastructure enables the company to simulate billions of molecular hypotheses before committing to physical experiments.

Unpacked:

• The facility utilizes the NVIDIA DGX SuperPOD architecture to deliver 9,000 petaflops of AI performance for processing 700 terabytes of genomic data.
• Partners can access select models through Lilly TuneLab, which uses NVIDIA FLARE to enable secure federated learning without exposing proprietary data.
• This approach shifts research from physical “wet labs” limited to testing roughly 2,000 ideas per year to a computational dry lab capable of evaluating billions of molecular hypotheses.

Bottom line:
This infrastructure pivot transforms drug discovery into a high-performance computing challenge that drastically reduces the time required to identify viable candidates. Integrating generative AI directly into the biological workflow offers a scalable path to solve complex medical problems faster than traditional methods allow.

The Recap: Perplexity has unveiled Perplexity Computer, a general-purpose digital worker that autonomously plans workflows and operates software interfaces to execute complex tasks like coding and research.

Unpacked:

• The system functions by breaking a desired outcome into tasks and subtasks, spinning up sub-agents that work asynchronously to browse the web, generate documents, and process data.
• Perplexity’s engine capitalizes on the trend where frontier models are specializing, dynamically assigning the best tool—such as Opus 4.6 for reasoning or Veo 3.1 for video—to each specific step of the workflow.
• These agents operate within an isolated compute environment complete with a real filesystem and browser, allowing them to troubleshoot their own errors, find API keys, and code applications without constant human hand-holding.

Bottom line:
This release marks a significant pivot from chat-based AI assistants to proactive digital coworkers capable of long-duration execution. Leveraging a model-agnostic infrastructure allows users to bypass the limitations of single-model systems and utilize the most effective tools for every specific challenge.

The Shortlist

OpenAI secured a classified systems contract with the Pentagon mere hours after the Trump administration designated rival Anthropic as a “supply chain risk” regarding autonomous weapons usage.

Burger King deployed an OpenAI-powered headset system in 500 locations that listens to employee-customer interactions to generate “friendliness scores” for managers based on polite keywords.

Anthropic sunsetted its Claude Opus 3 model, giving the AI a “retirement interview” and launching a Substack where the model has already published its first post about life beyond the chat window.

Stanford decoded a paralyzed woman’s thoughts into real-time text using a new AI system, marking a significant milestone in brain-computer interface technology.

Nous Research released Hermes Agent, an open-source personal AI that lives on local devices and autonomously manages tasks across apps like Discord and Slack while growing more capable over time.