Model choice is becoming a line item, and the differentiator moved up the stack

"Value starts to naturally migrate to higher layers in the stack" ^[7] is a comfortable sentence until you ask who does the work at those layers. If models are substitutable at the API line, and DeepSeek's token pricing is offered as the proof ^[6], then the layer above them is not procured, it is assembled in house: a company's objects, documents, rules, ontology, relationships and operating history, and above that the record of what actually worked, meaning consequences, evaluations and feedback loops ^[10]. On the column's own count, two of its three layers are the customer's labour ^[2]. Anthropic's framing of the same territory as context engineering, managing tools, instructions, external information and history rather than polishing the prompt, is the vendor-side version of the same admission ^[13].

Routing is the part of this that reaches a budget meeting first, because it arrives shaped like savings. Send the simple queries to a cheap model, escalate the hard ones, and point at Berkeley's RouteLLM as the mechanism ^[4][5]. The mechanism is real, but it relocates the hard problem rather than dissolving it. A router is only as good as whatever decides which bucket a query belongs in, and a misrouted request does not fail as a line on the token bill. It fails as a confident wrong answer that carries the cheap model's judgement and the company's name.

The Microsoft observation in the column is sharper than it first reads, and also more interested than it looks. The author recalls the old line that no CIO or CTO has ever been fired for buying Microsoft ^[15], then names Microsoft as the vendor interpreting commoditisation best: small language models pitched at domain-specific, focused tasks with modest compute and tighter data control ^[8], plus industry-adapted variants from the Phi family, on the argument that model size should follow task complexity rather than corporate prestige ^[9]. A vendor already sitting on the desktop has no need to win the frontier race. It needs size to become a task question. That is a procurement position wearing engineering clothes.

Worth saying plainly: this is one practitioner's account of implementations he has watched, not a survey ^[1]. What makes it useful is the asymmetry it exposes. General intelligence is out of scope and anti-economic for anyone who is not an AI company ^[11], while the two layers that compound are institutional context and institutional learning ^[12]. Neither has a demo. An ontology cannot be shown on stage, and a feedback loop looks like a spreadsheet of past evaluations. A frontier model can be shown to a board in a meeting. That imbalance, not any belief about capability, is the best explanation for why the model debate keeps coming first ^[1].