The AI and Tech Weekly: June 21, 2026

Mouad Belghiti·2026-06-2110 min read

The AI and Tech Weekly: June 21, 2026

Top market news, developer tips, open source picks, and startup ideas from the community

Market News

SpaceX Acquires Cursor for $60 Billion

Four days after the largest IPO in history closed at a $2.5 trillion valuation, SpaceX announced it is buying Cursor, the AI coding startup, for $60 billion in an all-stock deal. The merger is expected to close in Q3, pending regulatory approval. Cursor was founded in 2022 by Michael Truell and three MIT classmates who built an AI-native code editor that lets developers write, edit, and refactor code with natural-language instructions.

The numbers are real. Cursor crossed $1 billion in annualized revenue in November 2025 and hit $4 billion in ARR by June 2026, with roughly $2.6 billion from enterprise customers. One survey found Cursor deployed inside 64% of the Fortune 500. The strategic logic is clear: Grok, the chatbot that came with SpaceX's acquisition of xAI in February, still trails behind models from Anthropic, OpenAI, and Google on coding. Cursor solves that problem immediately.

For developers, the acquisition reshapes the AI coding market. Cursor's independence made it a neutral platform. As a SpaceX subsidiary, that changes. Expect competitors like Windsurf, Copilot, and open alternatives to pick up users who want a vendor not tied to a single hardware and defense company.

Trump Administration Bans Anthropic's Claude Fable 5 and Mythos 5

On June 13, Anthropic received a government directive ordering it to suspend access to Claude Fable 5 and Mythos 5 for all foreign nationals, including Anthropic's own foreign national employees. The administration cited national security authorities but offered no specific justification. Anthropic said it had to "abruptly disable" access across all customer accounts to comply.

Anthropic launched Fable 5 on June 9. The model is a Mythos-class release with exceptional performance on software engineering and scientific research benchmarks. A limited group of cybersecurity providers got access to the less restricted Mythos 5 through a program called Project Glasswing. The government told Anthropic it had learned of a potential jailbreak method, but Anthropic disputed that this was a valid reason for a full recall. More than 150 cybersecurity experts signed an open letter calling for the ban to be lifted.

For developers, this is a real infrastructure risk. Fable 5 pricing was $10 per million input tokens and $50 per million output tokens, less than half the price of the previous Mythos Preview. The abrupt removal shows that AI model availability is now subject to government action with short or no notice. Building on closed, high-capability models carries supply risk that open-weight models do not.

OpenAI's Leaked Financials: $13B Revenue, $21B Operating Loss

Audited 2025 financial statements obtained by blogger Ed Zitron and verified by the Financial Times show OpenAI earned $13.07 billion in revenue last year, more than triple its 2024 figures. Total costs hit $34 billion. Operating loss came to roughly $20.9 billion, excluding one-time accounting charges from the company's conversion from a nonprofit structure.

The headline loss figure of $38.5 billion includes a $41.55 billion non-cash charge from fair value adjustments on convertible interests and warrants. Strip that out and the underlying business is still burning $21 billion a year. The ratio of costs to revenue has improved: in 2024, OpenAI spent $2.37 for every $1 earned. In 2025, that ratio fell to $1.60. OpenAI has filed for an IPO expected later this year.

The context matters for any developer building on OpenAI's APIs. The company is burning money at scale and has not shown a path to profitability. It is also raising prices and cutting API deals more selectively. Developers who need stable, predictable costs should treat pricing from any top-tier AI provider as subject to change.

Midjourney Pivots to Health Hardware With a Full-Body Ultrasound Scanner

Midjourney, best known for generating images from text, announced the Midjourney Scanner on June 17. It is a full-body ultrasound device that scans a person in roughly 60 seconds without radiation or magnetic fields. The hardware works by lowering a user on a platform into a shallow pool of water ringed with half a million ultrasonic sensors. Soundwaves fire from every angle, generating terabytes of data per second, and a compute cluster reconstructs the data into 3D cross-sectional images using a method the company calls Ultrasonic CT.

The hardware runs on technology licensed from Butterfly Network, which disclosed up to $74 million in expected payments over five years. The device is not yet FDA-cleared for diagnostic use. Midjourney will start with body composition maps, which do not require clearance, and plans to add FDA approvals in layers. The first deployment will be a Midjourney Spa opening in San Francisco in 2027.

The pivot signals something broader. AI software companies with strong cash positions are starting to move into physical products where the margins can be higher and the competition is less developed. Notably, the Scanner runs no AI at all. Midjourney says the image reconstruction is purely computational. This is a bet on hardware design and sensor coverage, not model performance.

Tips and Tricks from the Community

Attach reasoning traces to PRs, not just the diff — When an AI agent writes code, it produces a chain of thought before outputting the change. That reasoning is almost always discarded. Teams that attach it as a PR comment cut review time significantly because reviewers reconstruct intent instead of guessing at it. Set up your agent to append a short reasoning summary to each PR description automatically. (addyosmani.com/blog/agentic-code-review)
Design agent loops for durability, not just speed — Most agent loops break because they have no durable execution layer underneath them. A single network timeout or process restart loses all state. Use a durable orchestration system so that when the agent loop fails, it resumes from the last checkpoint rather than restarting from zero. The TLDR newsletter's June 19 edition linked a detailed breakdown of this architecture. (tldrnewsletter.com)
Store agent-operating knowledge close to your code — The most effective teams put project rules, coding conventions, and repeatable workflows directly in the repo as skills or instruction files that agents can read. This gives every agent session the same starting context without relying on prompts pasted at runtime. Claude Code, Cursor, and similar tools all support project-level instruction files. Put yours in version control. (developersdigest.tech)
Use AWS S3 annotations for queryable object metadata — AWS shipped a new S3 annotations feature now available in all regions. You can attach up to 1,000 named annotations per object, each up to 1 MB, in JSON, XML, YAML, or plain text. Annotations are queryable and can be modified without rewriting the object. This is useful for tagging ML training data, documents, or any binary asset with rich context that needs to be searched later without moving the data. (aws.amazon.com)

Open Source This Week

OpenClaw — A personal AI assistant that runs entirely on your own devices and connects to 50+ integrations including WhatsApp, Telegram, and Slack. Surged from 9k to 210k stars in early 2026. github.com/openclaw/openclaw
Bumblebee — A read-only supply chain scanner from Perplexity AI that checks your dependencies, MCP servers, and VS Code extensions for suspicious packages. Still at 2.6k stars but growing fast since the v0.1.1 release. github.com/perplexity-ai/bumblebee
Nanochat — Andrej Karpathy's minimal LLM training stack in a single cohesive repository, 55k stars. The goal is to make every part of the training loop readable and reproducible with no hidden abstractions. github.com/karpathy/nanochat
Clear — A programming language where the specification and implementation live in the same file. No drift between docs and behavior. Built so agents can both read and execute it, and compiles to any target without changing the spec. github.com/clear-lang/clear

Startup Ideas for Inspiration

Drawn from YC's Requests for Startups and other leading accelerators

Software for Agents

Right now, almost every developer tool, API, and SaaS product was designed for humans. Agents interact with these tools the same way a person does: they read text, click buttons, and parse HTML that was never meant to be parsed. This creates enormous friction. Agents need APIs with machine-readable documentation, granular permission models, stateless CLIs, and versioned outputs. Nobody has built a well-designed suite of services specifically for agent consumers.

YC's framing is direct: "the next trillion users on the internet won't be people, they'll be AI agents." Every service that wants to capture agent traffic needs to be rebuilt from the ground up. The starting point is picking a single vertical, like legal document retrieval or financial data, and building a clean API with MCP support, detailed capability declarations, and structured outputs that an agent can use without fallback logic.

AI-Native Service Companies

Most AI startups from 2023 to 2025 sold software. The next generation sells the work itself. A law firm that uses AI to complete contract reviews, a recruiting firm that delivers vetted candidates, a bookkeeping shop where the entire back office runs on agents. The customer does not buy a product; they buy the outcome. Headcount stays flat while capacity grows.

The reason this works now and not two years ago is that AI agents can handle multi-step tasks reliably enough to complete whole workflows, not just assist with steps. The starting point is identifying a service business with a well-defined, repeatable workflow that currently relies on 10 to 50 people doing mostly the same work every day. Build the agent stack to do that work, deliver the service, and price it at what the market pays today for a human team.

Company Brain

Every engineering team has tribal knowledge that lives nowhere: which services talk to which, why a particular library was chosen, what broke during the last incident, and which PRs touched a given module. Today this lives in Slack threads, Notion pages, and the heads of your senior engineers. When someone leaves, it walks out the door.

YC calls this "Company Brain" and compares it to a dependency manager for organizational knowledge. The idea is a system that ingests code history, PR descriptions, incident reports, documentation, and conversations, then builds a queryable graph that agents and engineers can use to understand any part of the codebase in seconds. Tools like Unblocked are early examples, but none have become the default. The starting point is building the ingestion and indexing layer, then focusing on one specific query type that engineers do more than five times a day.

SaaS Challengers

Most SaaS tools charge a flat monthly fee for software that helps a human do a job faster. The same job can now be done entirely by an agent at a fraction of the cost. This creates an opening to compete with any established SaaS product by offering the service, not the software, at a lower price.

YC is actively looking for founders targeting specific SaaS incumbents. The criteria are simple: pick a product where the core workflow is well-defined, the output is measurable, and the existing customers are unhappy with the price. Build an agent that produces the same output and undercut the subscription price by 70 to 80 percent. The starting point is finding a SaaS product with 100,000 to 500,000 paying customers, mapping its core workflow, and testing whether an agent loop can complete it end-to-end before writing a single line of product code.

※