Meta’s Watermelon Model, AWS Quotas, and AI Cost Controls

Compact Conversations for 2026-07-03: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Meta’s next AI model, codenamed Watermelon, reportedly matches GPT-5.5

Meta’s superintelligence chief Alexandr Wang told employees the upcoming ‘Watermelon’ model has caught up with OpenAI’s GPT-5.5 on internal benchmarks and uses an order of magnitude more compute than its predecessor.

Why it matters: If accurate, this signals Meta’s massive infrastructure and talent investments are narrowing the performance gap with leading AI labs, which could reshape enterprise model choices and competitive dynamics.

Source: Business Insider

The Feed

AWS raises AgentCore runtime quotas by up to 5x to help enterprises scale AI agents

AWS increased default quotas for Amazon Bedrock AgentCore, raising concurrent session limits to 5,000 in key US regions and boosting token processing rates.

Why it matters: Higher default limits reduce administrative friction for enterprises moving AI agents from pilot to production, supporting more concurrent users and complex workflows without manual quota requests.

Source: InfoWorld

Microsoft launches $2.5 billion ‘Frontier Company’ to embed 6,000 AI engineers inside enterprise clients

Microsoft is investing billions in a new consulting unit that will place thousands of engineers directly with enterprise customers to integrate AI into core business processes.

Why it matters: This massive services push positions Microsoft as a platform-agnostic implementation partner, competing directly with AI labs’ own deployment arms and focusing on measurable enterprise ROI.

Source: The Decoder

Companies are throttling employees’ AI use because it’s too expensive

Leaked internal communications show companies across industries are pleading with employees to use less powerful AI models and cutting off access to some models as monthly costs spiral into the millions.

Why it matters: Unchecked usage-based pricing is leading to unexpected cost overruns, forcing enterprises to implement usage controls and reconsider their AI deployment strategies and vendor contracts.

Source: 404 Media

Crusoe in talks to raise $3 billion in round that may triple firm’s value

The data center startup, which powers AI compute for Meta and Oracle using stranded energy sources, is negotiating a funding round that could value it at around $30 billion.

Why it matters: The massive valuation jump reflects intense investor demand for alternative, cost-effective AI compute infrastructure as the industry grapples with power and cost constraints.

Source: Bloomberg

Visual Studio Code improves tools for agents

VS Code 1.127 makes browser tools for AI agents generally available, adds per-site permissions for the integrated browser, and improves session management and cost transparency.

Why it matters: These updates make it easier for developers to build, test, and manage AI agents that interact with web applications, a key workflow for creating production-ready AI tools.

Source: InfoWorld

One Thing to Try

Kimi K2.7 Code is now generally available as a selectable model in GitHub Copilot. It’s the first open-weight model offered in Copilot, providing a potentially lower-cost option for code generation and refactoring tasks.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: [curious] Today’s lead is about Meta’s AI progress. Business Insider reports that Meta’s superintelligence chief Alexandr Wang told employees in an internal town hall that the company’s next AI model, codenamed Watermelon, has caught up with OpenAI’s flagship GPT-5.5 model.

Host B: Wang said Watermelon is currently in training and uses an order of magnitude more compute than Avocado, Meta’s internal codename for Muse Spark, the model they released in April. He cited closely followed AI model benchmarks, though it’s not clear which ones. The claim is based on internal testing, and Meta hasn’t released public benchmark results for Watermelon yet.

Host B: One number to know today: 5,000. That’s the new default limit for active concurrent AI agent sessions in AWS’s US East and West regions, up from 1,000. It’s a fivefold increase that reflects how enterprises are moving from pilot projects to production deployments.

Host A: [thoughtful] On the infrastructure side, AWS has increased key Amazon Bedrock AgentCore runtime quotas by up to fivefold. The new limits support up to 5,000 active concurrent sessions in US East and West regions, and 2,500 in all other supported regions. [conversational] Analysts say this is AWS’s response to enterprises rapidly shifting AI-agent experiments to production deployments. The higher defaults help reduce operational friction when scaling AI agents from pilot projects to production. Forrester’s Charlie Dai notes the bigger change is the move from single-task copilots to multiple production-grade agents serving larger user populations.

Host B: Microsoft is taking a different approach. The company is investing 2.5 billion dollars in a new unit called Frontier Company that puts 6,000 engineers directly at enterprise customers. The goal is to integrate AI into core processes with measurable ROI, not more experimentation. [with emphasis] The Decoder reports Microsoft is positioning this as a platform-neutral alternative to OpenAI and Anthropic, which push their own models through their own deployment companies.

Host A: In cost control news, 404 Media reports that companies across tech, entertainment, banking, and other industries are throttling employees’ AI use because it’s too expensive. Leaked Slack chats, emails, and internal dashboards show companies pleading with workers to use less powerful models to stop costs from spiraling out of control. Some companies are even cutting off access to certain models altogether. [skeptical] The report cites material from half a dozen companies, including Atlassian, Adobe, and Amazon, and notes that in at least one case, AI spending has tripled to more than 15 million dollars a month.

Host B: On the funding front, Bloomberg reports that Crusoe, the data center upstart with contracts to supply AI computing power for Meta and Oracle, is in active talks to raise about 3 billion dollars. The funding round is expected to value the company in the 30 billion dollar range, up from a 10 billion dollar valuation in October. [curious] Crusoe specializes in using stranded energy, like excess natural gas, to power its data centers, which is seen as a cost-effective approach for AI compute.

Host A: In developer tools, Visual Studio Code 1.127 brings improvements to the Agents window for managing agent sessions and makes browser tools for agents generally available. The update features agents that can build and test web apps in the integrated browser, with safer per-site browsing permissions. [with a small lift] Microsoft says the browser tools let agents open pages, read content and console errors, take screenshots, and select, type, and navigate to verify their own work.

Host A: [conversational] One thing to try is Kimi K2.7 Code, which is now generally available in GitHub Copilot as a selectable option in the model picker. This is the first open-weight model offered in Copilot, giving you more choice and a lower-cost option for coding workflows.

Host B: If you’re on Copilot Pro, Pro+, or Max plans, you can select it in Visual Studio Code version 1.127 or later. [thoughtful] The vendor says this model is particularly efficient for code generation and refactoring tasks, so switching to it could be a simple way to test cost control for your team’s AI-assisted development.

Host A: That’s Compact Conversations for Friday. More AI news tomorrow. Until then, happy prompting.