The AI stack got repriced from both ends this week

Cursor shipped Composer 2 at $0.50 per million input tokens and beat Anthropic's Opus 4.6 on Terminal-Bench 2.0 — 61.7 to 58.0. A team of roughly forty people. Five months. Three model generations. They did it with continued pretraining feeding into long-horizon reinforcement learning across three or four GPU clusters. That is the whole story of the week, and everything else is a consequence of it.

Because if a forty-person application company can produce frontier-competitive coding performance at a twentieth of the price, the assumption underneath every frontier-model valuation — that scale and capital are the moat — needs a hard second look. The same day, Alibaba's Qwen3.5-9B beat OpenAI's 120B open-weights model on most language benchmarks, running on a laptop, Apache 2.0, ten cents per million tokens on the Flash tier. A 150M-parameter ColBERT retriever hit 90% on BrowseComp-Plus and outperformed retrieval systems fifty-four times its size. Two independent teams reported 10x data-efficiency gains in the same week. Architecture and algorithm choice now dominate parameter count across three separate layers of the stack.

Yes, but — Cursor still runs on somebody else's frontier model for the hard cases, Qwen's tech lead just resigned with four teammates citing burnout, and Alibaba quietly reclassified Qwen-Image-2.0 from open to closed. The commoditization story is real; the durability of any specific commoditizer is not. The take still holds because the pattern is what matters. Once a forty-person team proves the playbook, the playbook exists.

The labs saw this coming and moved first

On March 19, OpenAI acquired Astral — the team behind uv, ruff, and ty. That completes a nine-month pattern: Google DeepMind bought Antigravity in July, Anthropic bought Bun in December, and now OpenAI owns the Python toolchain. Every frontier lab now controls foundational developer tooling in a major language ecosystem. The Astral team is joining the Codex group specifically, and uv is becoming the package manager inside Codex sandboxes. Codex itself is at 2M+ users, up 3x since January. Fidji Simo told the org internally that OpenAI was "spreading efforts across too many apps and stacks" and is consolidating ChatGPT, Codex, and Atlas into a single desktop superapp.

The strategic read is not subtle. Model APIs alone are not a moat. When Cursor can produce a competitive model for a fraction of the cost, the labs need a different kind of lock-in — the daily tools developers touch. If your CI runs uv today, your package manager is now owned by a company that competes with your coding assistant vendor, sells you inference, and would very much like you to standardize on Codex. That is three counterparty relationships stacked into one dependency.

Anthropic is counter-positioning on protocol. Claude Code Channels runs on MCP, event-driven and persistent, and Google's Stitch adopted MCP independently. Two competing labs converging on the same protocol is how a standard forms. If you build agent-facing surfaces, add MCP support to the roadmap this quarter — the interop layer is being decided now, and later is expensive.

The other flank is atoms

Jeff Bezos is raising $100B from Singaporean and Middle Eastern sovereign wealth to acquire chipmakers, defense firms, and aerospace manufacturers and optimize them with world-model AI. That is more than all US venture capital raised in 2025. Prometheus has already raised $12.2B. Travis Kalanick emerged from eight years of stealth with Atoms — food automation, autonomous mining via Pronto AI, modular transport — and an explicitly anti-humanoid thesis backed by operational data from CloudKitchens. OpenAI and Anthropic have both cut multi-billion-dollar JVs with PE firms (TPG, Bain, Blackstone) to reach portfolio companies directly. Samsung committed $73B to AI chips.

The distribution channel implication matters more than the fund headlines. When a PE firm can hand its 400 portfolio companies overnight AI capabilities through an OpenAI JV, your competitive advantage of being "early to AI" gets neutralized by a wire transfer. The middle of the value chain — standalone AI features bolted onto SaaS, generalist chat tools sold per seat — is where the compression lands. Workday's embedded Sana AI hit 90% adoption in 40 days at one customer and displaced 400 ChatGPT licenses. That is the shape of the compression, not a curiosity.

What to do this week

Re-run your AI feature COGS model against a $0.50-per-million floor for single-domain tasks and a $0.10 floor for anything Qwen-class handles. Features you killed last quarter on unit economics may now be margin-positive. If you use uv, ruff, or Bun in production, pin versions and document a fallback path — the trust model on those tools changed this week and pretending it didn't is a choice. If any of your production workloads sit in AWS me-south-1 or me-central-1, migrate. Amazon itself is telling customers to leave after Iranian drones destroyed three Gulf data centers on March 1. And if you run Microsoft Intune, enable Multi Admin Approval for wipe and reset actions before the weekend — Iran's Handala group used exactly that gap to wipe 200,000+ Stryker devices, and CISA's emergency guidance from March 18 says the misconfiguration is industry-wide, not unique.

The SharePoint RCE Microsoft called "low likelihood" in January is on CISA's KEV with a Saturday deadline. The Cisco FMC unauth RCE has a Sunday deadline and a 37-day exploitation window Interlock ransomware has already used. Patch both before Monday.

The middle is being squeezed. Pick a flank.