AI Stack Reshuffle | AI Field Notes #15

AnalysisCursor (the AI-first code editor built by Anysphere) opened public beta on April 29 for an npm package, @cursor/sdk, that exposes the same agent runtime, codebase indexing, MCP servers, hooks, and subagents that power its desktop app. With about 15 lines of TypeScript, an engineer can spin up the same Composer 2 agent inside a CI job, a backend service, or another product, choosing local execution, a Cursor-hosted sandbox VM, or self-hosted workers. Composer 2 is priced at $0.50 per million input tokens and $2.50 per million output, roughly an order of magnitude under Claude Opus 4.7 per input token. Rippling, Notion, Faire, and C3 AI are running it in production. The shift here is concrete: a coding agent is no longer something a developer talks to, it is a callable function in a deploy pipeline that opens its own pull requests.

AnalysisChina's National Development and Reform Commission ordered Meta and Manus to unwind the $2 billion acquisition of the agentic AI startup, with a follow-on Bloomberg analysis on May 2 noting the decision was elevated to the National Security Commission chaired by Xi Jinping. Manus had moved its headquarters from Beijing to Singapore in 2025, then announced a sale to Meta in December, the kind of corporate-flag swap that Chinese founders have used to dodge both Beijing's outbound rules and Washington's outbound investment ban. The two cofounders are reportedly barred from leaving China. Beijing is also said to be drafting rules requiring Chinese AI companies to seek approval before pursuing US listings. The episode kills the assumption that a Singapore address insulates a Chinese-rooted AI lab from either capital. For US firms hunting Chinese AI talent through offshore vehicles, the playbook just stopped working.

AnalysisJapan Airlines and Haneda Airport began a humanoid robot trial on May 1, the first of its kind at a Japanese airport, after the Ministry of Economy, Trade and Industry published guidelines in March on using robotics and AI to address a shrinking workforce. A demonstration video shows a Unitree humanoid sliding payloads across a conveyor belt, though Japan Airlines told CNBC that feasibility studies are ongoing and would not confirm Unitree as the supplier. The driver is demographic, not technical: Tokyo's labor force is aging, immigration policy under Prime Minister Sanae Takaichi remains tight, and international arrivals at Haneda rose 3.5% year over year in March. Unitree itself is on track to raise about 4.2 billion yuan ($614 million) in a Shanghai IPO. Analysts still describe the robots as poor at fine motor work and dependent on human supervision, which is exactly the gap the trial is designed to measure.

AnalysisOn April 27 Microsoft and OpenAI rewrote the partnership that has defined enterprise AI since 2019. Microsoft's license becomes non-exclusive through 2032, OpenAI is free to sell on AWS and Google Cloud, Microsoft stops paying revenue share, and the long-discussed AGI trigger, which would have ended Microsoft's access if a board declared general intelligence achieved, is replaced with a fixed calendar date. Within 24 hours, AWS shipped GPT-5.5 and Codex on Bedrock. Within 72 hours, the Pentagon signed seven AI vendors on classified networks. Within a week, OpenAI's revenue miss made the Wall Street Journal. The clauses sound legal, but the consequence is structural: Azure is no longer the only cloud where GPT runs in production, and OpenAI is no longer paying Microsoft a share of every dollar it earns. Microsoft kept its license, dropped a hypothetical, and lost a moat.

AnalysisAnthropic shipped a Claude Code update over the weekend that adds an ANTHROPIC_BEDROCK_SERVICE_TIER environment variable, letting developers route requests through Bedrock's default, flex, or priority tiers via the X-Amzn-Bedrock-Service-Tier header. The release also fixes a bug where Opus 4.7 sessions were computing context against a 200K window instead of the model's native 1M, which was triggering early autocompacting and inflated /context percentages. /resume on large sessions is up to 67% faster on 40MB-plus session files, and MCP startup is now deferred until the first @-mention. The mundane bullet points hide a real shift: Claude Code is now a serious multi-cloud production workload with 4% of public GitHub commits attributed to it, and Anthropic is treating Bedrock as a first-class deployment target rather than a partner integration. Pasting a GitHub or GitLab pull request URL into /resume now finds the session that opened it.

AnalysisAlibaba quietly shipped Qwen3.5-Omni and Qwen3.6-Plus in April as proprietary models, accessible only through Alibaba's chatbot site and cloud platform, marking the first time the company has held back its highest-tier model from open weights. The smaller Qwen3.6-35B-A3B was still released under Apache 2.0. The shift follows the March resignation of Qwen division head Lin Junyang and the formation of a new Alibaba Token Hub business unit under CEO Eddie Wu. Qwen-VL-2B-Instruct alone has crossed 18 million downloads on Hugging Face, with more than 200,000 community variants live. The strategic question for the open-source community is whether the most popular Chinese model family is starting to copy OpenAI's structure: free models for community goodwill, paid hosted models for revenue. The CNBC and Beijing Auto Show coverage of Qwen integrations into BYD and Volkswagen-China cars suggests the answer is yes.

AnalysisApple is paying Google about $1 billion a year for a custom 1.2 trillion-parameter Gemini variant that will power the long-delayed Siri 2.0, with full deployment expected in iOS 27 in September, according to The Business Engineer's May 2026 Map of AI and earlier joint Apple-Google statements. Phase 1 already shipped in iOS 26.4, and Apple cites 67% daily engagement with Gemini-enhanced Siri features against 23% for the prior version. The strategic concession is large: after years of insisting on on-device, in-house models, Apple has decided that frontier large language models may be a commodity input not worth the multi-billion-dollar capex. The distribution layer matters more. With more than 1 billion active iPhones, Apple gets state-of-the-art assistant quality without joining the data center arms race, and Google gets paid by both sides of the smartphone market. OpenAI's existing ChatGPT integration in Apple Intelligence becomes the third-party fallback rather than the default.