Last week's AI news was a supply-side story: three labs shipped cheaper models in the same 48 hours, and we read it for founders as intelligence is becoming a commodity input. This week the story flipped to the other side of the transaction — and the demand side is where pricing power actually dies.

The single clearest data point: Microsoft, the largest distributor of AI in the world, is now engineering its own frontier suppliers out of its products. When the biggest buyer starts routing around you, that's not a rumor about commoditization. That's commoditization arriving.

Here's the week, read for what you should do about it.

The 30-second version#

1. The biggest buyer is routing around its suppliers#

On July 7, Bloomberg reported that Microsoft has started completing tens of thousands of weekly AI prompts inside Excel and Outlook using its in-house MAI models rather than OpenAI or Anthropic. The quote that matters, from Microsoft AI chief Mustafa Suleyman: "We pay Anthropic a significant amount of money annually, so our goal is to reduce, and eventually eliminate, that cost entirely." The MAI family is already live in GitHub Copilot, with in-house Teams transcription models rolling out next.

Read that as a founder, not a Microsoft-watcher. The company that put a chatbot in every Office app — the one whose distribution made the current model economy — has decided the model underneath is a cost line to be optimized, not a partnership to be deepened. It isn't dropping frontier models entirely; it's doing the smart thing, which is reserving them for the work that needs them and routing the high-volume, good-enough work to something cheaper it controls.

When your largest customer's stated goal is to "eliminate" your line item, you are not a platform. You are a commodity with a grace period.

What to do: Copy the architecture, not the anxiety. The move Microsoft is making at planetary scale is the one you should make at yours — tier your calls. Send the reasoning-heavy, quality-critical requests to the expensive model; route the bulk (classification, extraction, formatting, first drafts) to the cheapest model that clears your bar. If your calls all go to one hard-coded flagship, you're leaving the same margin on the table Microsoft just decided to stop leaving.

2. Enterprises are already voting with their tokens#

This isn't theoretical, and it isn't just Microsoft. On July 7, CNBC reported that the share of US enterprise token usage running on Chinese open models via OpenRouter has stayed above 30% every week since February and peaked at 46% — against an 11% trailing-twelve-month average. The driver is blunt: those models run 60–90% cheaper, and on some agentic benchmarks they now land within a point of the US frontier at roughly a fifth of the cost.

Nearly half of a real, price-sensitive slice of production usage has already migrated toward the cheapest credible option. The demand curve is doing exactly what a commoditizing market's demand curve does.

What to do: Treat "cheapest model that clears the bar" as a first-class engineering goal, not a someday-optimization — but benchmark before you switch. Route through an OpenAI-compatible gateway, mirror a slice of your real traffic onto a cheaper candidate, and compare on your tasks, not a leaderboard. And know your customers' constraints: an open or Chinese model may be a non-starter for a regulated buyer regardless of price, which is itself useful to know before a deal, not during one. (We wrote the founder's playbook for keeping your LLM stack swappable if you want the mechanics.)

3. The market repriced the picks-and-shovels#

The clearest financial signal of the week: Nvidia has shed roughly $1 trillion in market value in under two months, down about 16% from its May 14 peak, now trading at ~18x forward earnings — its cheapest since early 2019. The cause isn't collapsing demand (Nvidia still held ~97% of the server-GPU market at the end of 2025); it's investors rotating the "AI trade" into memory makers like Micron, up 229% this year on high-bandwidth-memory prices.

For a founder the fundamentals matter less than the narrative shift. For two years, "AI infrastructure" was a valuation that only went up. That reflex just broke. If you're raising into the AI-infra stack, or your pitch leans on the assumption that anything with "GPU" or "inference" in it commands a premium, the market just told you that premium is negotiable.

What to do: If you're an infra or tooling startup, sharpen the part of your story that isn't "we ride the AI wave" — unit economics, retention, a wedge that survives cheaper compute. The capital is still there (see below), but it's getting more discerning about which AI stories it pays up for.

4. …but the capital didn't leave — it moved#

Anyone reading the Nvidia slide as "the AI money is drying up" is reading it wrong. The money is abundant; it's relocating to the two things a cheap, swappable model can't hand you — hard infrastructure and distribution.

The through-line with Sections 1–3 is exact: value is draining out of renting intelligence and pooling around owning the substrate or the audience. Blue Origin owns launch capacity. Meta owns distribution. Microsoft owns the app surface — which is precisely why it can afford to treat the model as swappable.

Also worth a founder's 20 seconds#

The takeaway for founders#

Last week the price of intelligence fell. This week the people who buy intelligence — Microsoft, US enterprises, the public markets — started acting like it. That's the more durable signal, because supply-side price cuts can be marketing, but a buyer re-architecting to spend less is a structural fact.

So the founder question sharpens by one turn from last week. It's no longer just "if a rival could rent the same model for a dollar a million tokens, what's left that's yours?" It's "when your own biggest customer starts routing around premium models to save money — will your product be the premium thing they route around, or the workflow, data, or distribution they can't?" Build for the second answer. This week was another data point that the first one has a shrinking shelf life.