AI Coding Shakeout | AI Field Notes #52

AnalysisAnyone who scripted Google's Gemini CLI into a build pipeline has a migration on their hands as of June 18, when Google retired that command set and replaced it with the Antigravity CLI. Antigravity shipped at Google I/O in May and is built for the Gemini 3.5 agentic model family, where the tool plans and runs multi-step tasks rather than answering one prompt at a time. The old flags and output formats change, so wrapper scripts and CI jobs that parsed Gemini CLI responses will break until they are rewritten. Google is moving its developer surface from chat to agent and giving teams a few weeks to keep up.

AnalysisThree days after the largest IPO in history, SpaceX spent $60 billion of its fresh stock to buy Anysphere, the maker of Cursor, the AI coding editor. The all-stock deal, confirmed in an SEC filing on June 16, closes a year in which Cursor's share of the coding-tool market slid from 41 percent to 26 percent while it burned through $2.3 billion. The buyer matters more than the price. Every serious AI coding assistant now sits inside a tech giant: Copilot at Microsoft, Claude Code at Anthropic, Codex and Windsurf at OpenAI, Cursor and Grok Build at SpaceX. The independent lane just closed.

AnalysisThe company at the center of the AI boom lost $38.5 billion last year, a figure leaked on June 16 and independently verified by the Financial Times against revenue of $13.1 billion. Most of that loss is a one-time, non-cash charge from OpenAI's 2024 restructuring into a for-profit; strip it out and the operating loss was closer to $8 billion. Either number is large for a company this celebrated, and total spending hit $34 billion, with $19 billion going to research. Revenue is growing fast, yet the gap between running frontier models and what users will pay has not closed, and OpenAI is filing toward a trillion-dollar IPO straight into that gap.

AnalysisGitHub's own infrastructure could not keep up with the code its users' AI agents are writing, so Microsoft began routing GitHub traffic onto Amazon Web Services, the cloud of its biggest rival, confirmed on June 16. Availability fell to 88.4 percent in June, well short of the 99.9 percent enterprise contracts assume, after nine service-degrading incidents in May alone. The cause is volume: AI workloads now eat more than 60 percent of GitHub's compute, up from 15 percent in early 2023, and Claude Code alone pushes 2.6 million commits a week, a 25-fold jump in six months. The machines write faster than the platform can absorb.

AnalysisThe heads of the three biggest AI labs sat down to lunch with the leaders of the world's richest democracies on June 17, as the G7 summit in Evian, France, closed with an unusual joint session on AI governance. Sam Altman of OpenAI, Dario Amodei of Anthropic, and Demis Hassabis of Google DeepMind joined the table, alongside Europe's Arthur Mensch of Mistral and Aidan Gomez of Cohere. The optics are the story: the people building the technology now help write the rules meant to constrain it. Talk centered on youth safety and frontier risk, with no binding commitments announced.

AnalysisTwo deadlines are about to break production code for anyone building on Google's Gemini API: the image-preview models go dark on June 25, and several video generation endpoints follow on June 30. Apps still calling the old model IDs will start returning errors instead of images or clips, with no grace period after the date. Google wants developers on Gemini 3.5 Flash and the Veo 3.1 video models instead, which means updating model strings, re-testing prompts, and budgeting for different pricing. The frontier labs now retire models on a fixed schedule, and the price of building on a fast-moving API is the upkeep it demands.

AnalysisEvery AI rollout eventually runs into the same wall: the factories that make the chips cannot keep up. At TSMC's shareholder meeting on June 4, chief executive C.C. Wei said global supply will lag AI-driven demand for years and called this year's growth insane. TSMC, the Taiwanese foundry that manufactures the advanced chips Nvidia and others design, is sold out of its most advanced production through at least 2027, with demand running 25 to 30 percent above what it can make. That shortage sets the ceiling on how fast anyone can add compute, and it keeps GPU and cloud prices high. The constraint is physical, and it clears slowly.

AnalysisFor the first time in the internet's history, machines make more web requests than people do. Cloudflare, which sits in front of a large share of the world's websites, reported on June 5 that automated systems now account for 57.5 percent of HTTP requests against 42.5 percent from humans. Chief executive Matthew Prince had expected the crossover in 2027 and said agentic traffic arrived faster than he thought. The driver is AI agents browsing on a user's behalf: one assistant shopping for a camera may hit thousands of sites where a person would visit five. The web is increasingly built for readers who are software.

AnalysisCompanies bracing for Europe's strictest AI rules just got more time and less clarity. Under the Digital Omnibus, a package of amendments that negotiators provisionally agreed on May 7, the EU AI Act's obligations for high-risk systems, the Annex III category covering hiring, credit, and other consequential uses, slip from August 2026 to December 2027. The requirement for each member state to run a regulatory sandbox moves back a year too. Brussels frames it as simplification; critics read it as the first sign the law's timeline bends under industry pressure. For a team building a hiring or lending tool, the deadline just moved sixteen months, though the rules themselves did not soften.