The week the AI value chain split at both ends

Cursor's Composer 2 landed this week at $0.50 per million input tokens, scored 61.7 on Terminal-Bench 2.0, and beat Anthropic's Opus 4.6. A roughly 40-person team, three to four GPU clusters, continued pretraining followed by long-horizon RL. Five months from 38% to 61.3% on their internal bench across three model generations. The pricing isn't the headline. The architecture is: an application company built a frontier-competitive model on its own data, on its own timeline, and is now selling inference at one-twentieth what it pays its upstream supplier.

The same week, Alibaba shipped Qwen3.5-9B under Apache 2.0. It runs on a laptop and outperforms OpenAI's gpt-oss-120B on most language benchmarks. The flagship 397B-A17B MoE wins 28 of 44 vision benchmarks against GPT-5.2, Claude 4.5 Opus, and Gemini-3 Pro. Flash tier API: ten cents per million input tokens.

And Bezos is raising $100 billion from sovereign wealth in Singapore and the Gulf to buy chipmakers, defense companies, and aerospace manufacturers and run AI world models across them. More than total US VC raised in 2025. Kalanick surfaced from eight years of stealth with Atoms — food automation, autonomous mining, transport. OpenAI cut a $10B JV with TPG and Bain. Anthropic is talking to Blackstone.

If you're building anything where the differentiation is "we wrap a frontier API," both flanks of your moat just collapsed in the same seven days.

What Cursor actually proved

The interesting thing about Composer 2 isn't that it's cheap. It's that the playbook is now legible. Continued pretraining on domain data, then RL with long-horizon rewards, distributed across a handful of clusters. The Finetuner's Fallacy research that surfaced this week argues base model selection dominates finetuning data choice — early pretraining leaves durable imprints finetuning can't undo. Translation: if you have proprietary data and a clear vertical, the bar to a defensible model has dropped from "raise a billion dollars" to "raise enough to do continued pretraining on the right base."

This breaks the pricing assumption underneath every AI feature shipped in the last eighteen months. If your COGS model used $5/M input as the floor, you're modeling against last quarter. The new floor for single-domain workloads is somewhere between $0.10 and $0.50, depending on how much you're willing to self-host.

Meanwhile every frontier lab spent the same week buying developer toolchains. OpenAI acquired Astral — uv, ruff, ty — the Python toolchain that quietly replaced pip and Black across most of the ecosystem in eighteen months. Anthropic owns Bun. DeepMind has Antigravity. The model API isn't enough of a moat, so the labs are buying the workflow. If your build pipeline runs uv, your package manager is now operated by a company whose strategic interests will diverge from yours within a year.

What Bezos is actually doing

The $100B fund is the part that gets attention. The $10B PE-AI joint ventures are the part that matters this quarter. When OpenAI partners with TPG and Bain, every PE portfolio company across healthcare, financial services, and industrials gets an AI capability injection without ever building an AI team. If your B2B SaaS pitch was "we'll bring AI to your industry," your moat is a phone call between a managing partner and an account executive.

The Bezos move is structurally different and more patient. Buy the factory, run a world model on it, capture the operational data, use that data to improve the model, deploy to the next factory. It's Berkshire with a transformation layer. Whether it works at $100B scale is unknowable. That it will reprice acquisition multiples in chipmaking, defense, and aerospace within twelve months is not.

Kalanick's anti-humanoid thesis — task-specific wheeled robots over bipedal ones, backed by eight years of CloudKitchens operational data — is the contrarian read on the $10B+ pouring into humanoid robotics. He may be wrong. He has more deployment hours than anyone arguing the other side.

What this means for whoever has to ship next quarter

Three concrete moves, in order of urgency.

First, this week: re-run your AI feature unit economics with $0.50/M as the input floor and $0.10/M as the aspirational floor. Features you killed last quarter on COGS may be margin-positive now. Don't wait for the planning cycle.

Second, this sprint: audit your developer toolchain for vendor-becomes-competitor risk. If you depend on uv, ruff, Bun, or any tool now owned by a frontier lab, pin versions, document fallbacks, and treat the tool the way you'd treat any vendor who just acquired your largest competitor — because that's the relationship now. Same audit for your Gulf cloud footprint after Iranian drones hit three AWS data centers on March 1 and Amazon told customers to leave the region. And clear the calendar for March 24 (Node.js patches nine CVEs across all maintained versions) and March 31 (O365 Connectors die, taking your Argo CD and Teams alerting pipelines with them if you haven't migrated).

Third, this quarter: pick a side. The middle of the AI value chain — selling someone else's model with a thin wrapper — is getting compressed from PE-distributed enterprise sales above and Cursor-style vertical model builds below. You either move up into proprietary model capability on your data, or you move out into workflow-embedded distribution where the model is a commodity and the integration is the product. Workday/Sana hit 90% adoption in 40 days at one customer and retired 400 ChatGPT seats. That's the embedded-AI playbook. Cursor at $50B on $2B ARR is the vertical-model playbook. Both are defensible. The middle isn't.

The instinct in a week like this is to wait for the dust to settle. The dust isn't settling. Composer 2 took five months. Astral was acquired in nine. Bezos's fund is being raised right now. The companies that come out of this cycle in good shape are the ones who picked a flank by Q3.