AI Capex vs. Headcount | AI Field Notes #39

AnalysisJetBrains shipped TeamCity 2026.1 with Model Context Protocol (MCP, an open standard that lets AI agents access external data and tools) built into the build system, allowing AI agents to query build logs, analyze failures, and suggest fixes without custom webhook integration. A new open-source CLI with more than 60 commands is installable via Homebrew, npm, or Windows Package Manager, and exposes agent-callable skills for checking build status and streaming log output in real time. The release also patches a critical vulnerability, CVE-2026-44413, affecting all prior On-Premises versions, which teams should treat as an immediate update. JetBrains also made the AI Assistant available during Enterprise license trials for the first time, previously an Enterprise-only feature.

AnalysisOpenAI's promotional 2x usage boost for Codex Pro's $100/month tier expires on May 31, today. Teams on that plan have been running at 10x the base Plus capacity since the promo launched; after today they revert to the standard 5x. The $200/month Pro plan is unaffected, keeping its 20x Plus capacity on an ongoing basis. For engineering teams that sized agentic workflows around the promotional limits, the effective compute budget halves overnight without any change to their bill. OpenAI's Codex, the AI coding agent that runs inside ChatGPT and via API, has seen heavy adoption for autonomous coding tasks; the promo was designed to encourage that adoption. The options after today are to absorb the reduction, move to direct API billing, upgrade to the $200 plan, or shift some tasks to a cheaper model.

AnalysisDell's AI server business generated $16.1 billion in revenue last quarter, a 757% increase from the same period a year earlier and the company's strongest growth since its 2018 return to public markets. Dell's stock rose 32% in a single session on the earnings release, its best single day on record. The jump reflects how directly hyperscaler capex commitments translate into hardware purchase orders: when Amazon, Microsoft, Meta, and Google collectively pledge $700 billion in 2026 infrastructure spending, a meaningful share lands as server chassis and rack configurations that Dell builds and ships. The momentum is real, but so is the customer concentration. A handful of hyperscalers account for the bulk of the AI infrastructure buildout, which means Dell's revenue trajectory is tied to their spending decisions.

AnalysisAmazon's custom silicon business, spanning Trainium AI accelerators, Graviton CPUs, and Nitro chips, runs at roughly $20 billion in annual revenue. Trainium2 is largely sold out, Trainium3 launched nearly fully subscribed, and Trainium4 is significantly reserved despite being 18 months from broad availability. Multi-year commitments across Trainium total $225 billion, with Anthropic accounting for up to 5 gigawatts of capacity, the largest disclosed single-customer commitment. OpenAI committed approximately 2 gigawatts. Andy Jassy, Amazon's CEO, estimated that if the chip business were standalone it would generate around $50 billion annually, placing it among the top global data-center silicon providers. The level of advance commitment signals that infrastructure scarcity is real and already priced into supply chains.

AnalysisColorado's comprehensive AI regulation takes effect June 30, applying to high-risk AI systems and creating the kind of state-level compliance requirement the Trump administration has been working to preempt since its December 2025 executive order directing the Justice Department to litigate against state AI measures. A federal judge issued a 14-day stay blocking enforcement of one Colorado AI provision while the court weighs xAI's request for a preliminary injunction, a ruling that may set the precedent for how enforceable state AI rules are in the absence of federal law. Washington, Florida, Virginia, and Utah are advancing their own bills. Without a federal floor, every AI company and developer shipping to US consumers faces a compliance matrix that is expanding with no settled endpoint.

AnalysisAnthropic opened its Seoul office with KiYoung Choi as Representative Director, citing Korean Claude usage running 3.5 times higher than the country's population share would predict. The company briefed South Korea's intelligence, finance, and science ministries on Claude Mythos, the same frontier model used in Project Glasswing for autonomous vulnerability discovery. A second European headquarters opened in Milan this week, designed to address GDPR (Europe's data-protection regulation) data-sovereignty requirements for Italian enterprise and research clients. Taken together, the expansion pattern suggests Anthropic is moving from US-centric API business to direct government relationships in allied nations, with Mythos as the asset that justifies the conversations.