The one-screen version: Earnings week delivered the first real crack in the AI buildout. On October 28–29, Microsoft trimmed its full-year capex guidance to ~$170B (from ~$175B) and Alphabet held its $195–205B range instead of raising it — both short of estimates, which CNBC called the buildout "hitting a snag." The cause isn't cooling demand; it's chips: Microsoft's CFO said third-party AI-chip delays push some capacity into the first half of 2027. Meanwhile OpenAI will merge your ChatGPT and Codex usage into one allowance in December, and DeepSeek doubled its raise to ~$15B days after a US federal ban. Three stories, one throughline: the layer you build on just repriced — capacity tightened, your allowance consolidated, the competition got funded. Budget for scarcity, not for last quarter's discounts.

1. The AI buildout paused raising its bets#

For a year, every hyperscaler earnings call ended the same way: capex guidance up. This quarter it stopped. Microsoft trimmed its full-year capital-expenditure outlook to about $170B from ~$175B, Azure grew a still-enormous ~48%, and Alphabet kept its $195–205B guidance rather than raising it. Both landed shy of Street estimates, per CNBC. A week earlier, Meta narrowed its own full-year data-center capex to $135–145B. The combined big-four number is still roughly $720–745B for 2026 — but it is no longer accelerating.

The line that matters isn't the dollar figure; it's why. Microsoft CFO Amy Hood said delayed third-party AI chips will push some data-center capacity into the first half of 2027. This is a supply crack, not a demand one.

What you do Monday: read 2027 capacity as scarce. If Microsoft can't get chips on time, the downstream effects — longer GPU waitlists, firmer rental floors, tighter API rate limits — land on you. Reserve the capacity you know you'll need instead of assuming on-demand will be there, keep the model layer swappable so an availability shock is a config change, and re-run the rent-vs-API math at today's firm floors, not last summer's falling ones. We mapped exactly where those floors sit in the cheapest GPU rental right now.

2. Your ChatGPT and Codex limits become one#

On October 31, OpenAI updated its Help Center to confirm that Plus, Pro and Business plans will share a single token-based usage allowance across ChatGPT and Codex starting in December — instant chat excluded — as ChatGPT Work's agent features fold into the main app, per DEV Community and TokenPost. OpenAI did not publish an exact date or per-plan token amounts.

The practical effect for a team: if the same seats drive both everyday chat and Codex coding sessions, your effective coding headroom shrinks in December without the invoice changing. Inventory who shares a seat now, and decide before the switch whether your heaviest Codex users need their own plan — or whether a portion of that work belongs on an API budget you control.

3. DeepSeek doubled its war chest — days after a US ban#

Around October 22, DeepSeek wrapped a round of about $15B (~100B yuan) led by CATL and Tencent, roughly doubling its original ~$7B target at a reported ~$75B valuation, with a possible 2027 IPO — per TechCrunch and Reuters. The timing is the story: it closed about two days after the US ordered federal civilian agencies to remove DeepSeek over security concerns.

For most founders this isn't "switch to DeepSeek." It's a pricing signal: a lab the US just banned from federal systems still raised a frontier-scale war chest, which means downward pressure on open-weight and API prices is funded well into 2027. Keep a cheap open-weight fallback wired and tested, and don't architect around any single lab's promotional rate.

The week's three stories rhyme: compute got scarcer, your allowance got consolidated, and the competition got richer. None of those moves in your favor on price — so budget for firm floors, keep every layer swappable, and reserve the capacity you'll actually need before someone else does.