Good morning, {{first_name|there}}. Three of the biggest AI earnings reports of the year landed yesterday. Every one of them beat — and the chip index still fell 5%.
Read time: 3 minutes. Same time, every weekday — rate today's issue at the bottom.
🚀 The Big Story: The AI trade beat the numbers and lost the tape
Microsoft, Meta and Arm all reported blowout AI quarters on July 30. Semis sold off anyway — because the spending side finally got specific.
Microsoft printed $90.0B in revenue, up 18%, with net income of $35.8B (+31%). Azure crossed a $100B annual run rate, Copilot passed 30 million seats, and remaining performance obligations hit $678B — up 84% year over year.
Meta did $60.8B, up 28% — and raised AI capex guidance to $130–145B. Operating expenses grew 55% to $42B and Reality Labs lost another $4.62B. Arm posted a record $1.29B (+22%) with data-center royalties more than doubling.
Then the cost side spoke. Dwarkesh Patel's widely-shared argument puts real compute cost 10–15x above the headline framing; spot GPU prices are up roughly 40% since February, and Google is reportedly paying SpaceX around $900M a month for 110,000 orbital GPUs. Nvidia closed down 5%, the Philadelphia Semiconductor Index down 5%.
Jason's take: The market didn't doubt AI demand yesterday — it repriced who pays for it. When Meta's capex range widens by $15B in a single sentence and Arm's data-center royalties double, that money comes out of someone's gross margin, and increasingly that someone is the person buying tokens. If your product's unit economics assume this year's API prices hold, you're carrying an input cost you have never modeled. The 20-minute audit below is the cheapest insurance available.
⚡ Quick Hits
Kentucky landed a $100B data-center campus. Brookfield and NextEra are building 1.2 GW by 2032, scaling toward 1.8 GW and roughly 8,000 jobs — the largest AI infrastructure commitment the state has seen.
Europe opened bidding on seven AI gigafactories. €10B public against about €30B total, 100,000+ chips per site, bids close November 12. Sovereign compute stops being a slogan and becomes a procurement document.
A critical MCP flaw scored a perfect 10.0. CVE-2026-59726 in Ruflo exposed 233 tools on port 3001 with no authentication, patched in v3.16.3 within 24 hours. If you self-host an MCP server, go check your port.
AWS signed a $400M multiyear compute deal with Recursive Superintelligence — more evidence that labs are locking supply years out instead of buying spot.
OpenAI is giving about 100,000 academic researchers free frontier access through 2027. Cheap goodwill, enormous distribution, and a well-timed hedge against the open-weights argument.
Zhongji Innolight raised roughly $6.8B in Hong Kong and closed 2% below issue. The optical-interconnect supplier is a clean read on how much AI-infrastructure appetite public markets have left.
📡 Trending on X
The ARC-AGI-3 scoring fight is the argument of the week. OpenAI's model logged 7.8% on the official harness but 38.3% with tuned API settings — above Claude Opus 5's 30.2%. Everyone agrees both numbers are real; nobody agrees which one counts.
Anthropic is publicly denying it pushed for an open-weights ban, proposing mandatory pre-release safety testing instead. The replies split cleanly between "reasonable floor" and "regulatory moat."
The White House's AI-generated Rick and Morty parody went badly viral. Whatever your politics, it's the clearest signal yet that generative video has reached official comms — and that audiences still punish it.
Stripe's solo-operator data is getting quoted everywhere: businesses clearing $1M+ with zero employees have nearly tripled since 2023. The optimistic read on all of this is one person with good tooling.
🛠 The Workflow: The 20-minute token bill audit
Input costs are moving and most people have never looked at where their AI spend actually goes. Twenty minutes:
Pull your last three invoices from every AI vendor — model APIs, your IDE, your writing tool, whatever is on the card. Put the totals in one row so you see the real number.
Find your top job by volume. Nearly every setup has one task eating 60–80% of tokens. It is usually summarization, classification, or a retry loop nobody capped.
Downgrade that one job to a smaller model and run 20 real examples side by side. If you cannot tell the outputs apart, you just cut your largest line item by most of it.
Cache the repeated prefix. If the same system prompt or document rides along on every call, prompt caching is a config change, not a rewrite.
Set a hard monthly spend cap on every key today — not to save money, but so a runaway loop can't spend your quarter over a weekend.
Reply with the word "TOKENS" and I'll send you the audit worksheet I use.
🧰 Trending Tools
Reclaim AI v4 — calendar automation that renegotiates and reschedules lower-priority meetings on its own. For anyone whose deep work keeps losing to other people's invites.
Canva AI Design Suite — turns a rough sketch or screenshot into an on-brand layout with matching copy. For small teams shipping marketing without a designer.
Kling 3 AI — cinematic video from a text prompt or a single image. For creators who need motion assets faster than a shoot allows.
Groundcover — eBPF observability that just raised a $100M Series C. For anyone who genuinely can't tell which service is burning the infrastructure bill.
GitHub Copilot Workspace — plans and previews multi-file changes before touching the repo. For solo builders making repo-wide edits they can't afford to get wrong.
📣 Put your brand here. The AI Innovator reaches AI-first operators, creators, and marketers every weekday. Primary sponsorships are now booking.
That's a wrap
Monday: whether Amazon and Apple's numbers confirm the capex story — and what Alphabet's cloud margin says about who is really absorbing the compute bill.
Enjoying the format? Share your link — every referral keeps this free:
How was today’s email?
You're reading the 3-minute AI Innovator — same time, every weekday. Hit reply and tell me what you want more of. I read every one.
