One Data Center Took Down Grok, Claude, and ChatGPT This Morning

Thursday morning was a mess if you live inside AI tools. Grok went dark for three and a half hours, and not just the web version — the Android app, the iOS app, everything. SpaceXAI blamed a "models outage" at their Memphis compute center and apologized to Grok users plus a bunch of unnamed "compute partners."

Here's the uncomfortable part: the timing lines up almost perfectly with problems at Anthropic and OpenAI too. Claude had elevated errors across multiple models starting around 6:30 AM PT. ChatGPT and Codex followed at 7:30. Anthropic won't say if their issue was related to the Memphis outage, but remember — they signed a deal earlier this year to lease compute from SpaceXAI. You don't need a conspiracy theory to connect those dots.

That's the real story here, and honestly it's the one nobody in the industry wants to talk about. AI infrastructure has quietly become a small-world network. A handful of mega data centers, a handful of compute partners, and one blown transformer in Tennessee can ripple through half the major assistants on the market at once. The cloud had this problem a decade ago with AWS zones; we're relearning the lesson, just slower.

The token price collapse is already here

Meanwhile, the money side of AI is doing something weird. Seeking Alpha's LLM Token Expenditures Index collapsed in Q2 — token prices falling hard while yields climb. The argument from that corner: AI capex might finally stall or get cut, and the "bubble" narrative is less a prediction than a description of what's already in motion.

To be fair, I've read this exact article format a dozen times since 2023, and the doomsday crowd has been wrong more often than right. But the token-price data is hard to wave away. When inference gets this cheap, the per-token economics shift underneath everyone — and that feeds directly into the next story.

The 97% problem: your bill is one model

A dev.to post has been making the rounds, and it's the most practical thing I've read all week. Weio, an AI company, instrumented their production traffic and found one frontier model — gpt-4o — was carrying 77 percent of their calls and 97 percent of their model spend. Read that again. The other 23 percent of traffic, running on cheap models, cost about 3 percent of the bill.

Nothing about their setup was exotic. The demo gets built on the strongest model, the demo becomes the default, the default becomes the architecture. No single call ever feels expensive — a frontier call costs a fraction of a cent — so nobody challenges the monthly aggregate. Blame is asymmetric: nobody gets questioned for using the strong model, but somebody gets questioned when the cheap model fails.

Their fix is a written routing policy: cheap or mid-tier models for routine glue work, file edits with a clear target, research and reading; frontier models reserved for ambiguous requirements, subtle debugging, architecture decisions — anything where being wrong is expensive. The dividing question is simple: is the path already decided?

I've been running something similar for my own workflows and it's the single biggest cost lever most teams are leaving on the table. If you haven't looked at your own token spend by task type, do it this week. The gap between "frontier default" and "routed" is usually not 10 or 20 percent — it's often 60 or 70.

Local AI gets even cheaper

On the opposite end of the spectrum, XDA ran a fun one: a Raspberry Pi that boots straight into a local LLM, no Linux in sight, just a USB drive and some patience. You're not running anything frontier on a Pi, obviously — but that's kind of the point. The loudest voices insist you need a rack of GPUs to do anything useful, and meanwhile people are serving reasonable local models off a $60 single-board computer.

It's a good reminder that the "cheap tier" keeps moving down. Two years ago local meant a MacBook Pro; now it means a Pi with a thumb drive. The token price collapse and the hardware floor dropping are the same trend from two directions.

AI scribes are great until they're not

Last one, and it's a cautionary tale. University of Edinburgh researchers looked at AI note-taking tools for doctors — the scribes that record consultations and write summaries. The tools genuinely help with admin: one clinic in Dudley cut a six-month letter backlog to 14 days. But the study found the AI summaries tend to prioritize clinical info at the expense of the patient's actual story — facial expressions, gestures, emotional state all get flattened out.

There's also a subtler problem they flagged: "cognitive offloading." The tool takes over the mental work of note-taking, which sounds great until clinicians stop recognizing their own notes or remembering the patient next visit. And some patients clam up about sensitive stuff — substance abuse, domestic issues, mental health — when they know an AI is processing the conversation.

I'm not against AI scribes; the backlog numbers are real and the relief for burnt-out GPs matters. But this is the pattern with every AI assistant in a high-stakes room: it captures what's sayable and loses what's visible. Training data won't fix that by itself.

---

Quick add-on note: if you're trying to make sense of your own model spend or just want a clean way to run the numbers on any of this, Math Calculator is a handy little tool for the arithmetic side of things. No affiliation, just a bookmark I keep going back to.

That's the week in AI from my corner. The infrastructure fragility, the cost wake-up call, and the quiet reminder that cheaper doesn't always mean worse — it sometimes just means you should route around the expensive default.