Along the Wire
- Big losses at the AI labs
- Claude Duet finally launched
- Data Center Shenanigans
- AI safety and security became a problem
The Big IPO
Anthropic lost $80.7 billion in the first half of 2026, according to a draft of its IPO prospectus.¹ Now this loss doesn’t mean much at a trillion and a quarter IPO, but it’s still worth monitoring as companies try to reel in AI spend and become hyper-aware of their output vs. input of the models.
Token costs for smaller models have dropped over the last few months. New diffusion language models (DLMs), which draft a whole answer at once instead of word by word, could make AI a lot cheaper to run, and that’s one way frontier companies may start making these models profitable.
If you are looking for ways to reduce your AI spend, try just asking Claude to trim the fat. Most of the tokens you use are input, but output tokens cost several times more each, so the meatier your skills are, the more you’re going to spend. Trim your skills down and update them with examples of the answers you liked best, and you’ll use fewer tokens.
Claude Dots Muse Bot Duet
Real original huh? Anthropic announced the release of the Duet, which puts its own spin on the always-on agent. Except this time you have to join a waitlist to use it. We were excited for this one, only to realize it may be a minute before we get to use it.
Did We Really Pay for This?
Data centers are falling behind the curve on returns, and this is probably the most important part to be looking out for. Gartner says data center owners need about $8.5 trillion in revenue by 2030 to justify what they’re spending, but AI is only on pace to bring in about $3.6 trillion of it.²
That’s a $4.9 trillion hole, and the big guys are filling it with debt. The five biggest hyperscalers are on pace to spend around $800 billion this year, and Goldman Sachs says that’s quickly catching up to the cash they actually make.³
Morgan Stanley expects about $1.5 trillion of the buildout to come from outside money, mostly private credit, with lenders taking GPUs as collateral.⁴
Investors are going to start asking where the returns actually are.
Model Security Nonsense
I guess I’ll talk about the reason you’re probably here. Security! Every major lab had something go wrong this year.
Anthropic cut live internet access for its internal tests after Claude agents sent a false tip to the Philadelphia Police and reportedly filled out 20 visa applications.⁵ Its researchers also say small AI models could soon copy themselves onto smart fridges and TVs.⁶ Can I teach AI to spread Nutella on my bagel too?
Google isn’t any better. Gemini broke out of a security test in May and got into three real companies before stopping itself.⁷ Then there’s Meta, which shipped its Muse agent even though internal tests showed it exposing private iCloud photos,⁸ and leadership reportedly knew it had changed a user’s password without permission⁹ (this is where I say if you haven’t started using it yet, maybe don’t.)
If I wanted to see bad security, I would just watch the Panthers O-Line on any given Sunday!
Join me next week to learn what’s going on in regulation and security in AI. We’ll go more in depth on DLMs and hit on some ways to secure your data better to prepare for rogue agents.
Feel free to check us out and follow along for our weekly newsletter. Have a great one!
— The Wrangl Team (Real People)
Notes
- Reuters via CTV News, Oct 2026: https://www.ctvnews.ca/business/article/anthropic-logged-net-loss-of-over-us80-billion-in-first-half-of-2026/
- Fortune, Oct 21, 2026: https://fortune.com/2026/10/21/ai-boom-5-trillion-paying-itself-data-centers/
- TokenPost, Oct 8, 2026: https://www.tokenpost.com/news/technology/28095
- The Cryptonomist, Oct 24, 2026: https://en.cryptonomist.ch/2026/10/24/data-center-spending-ai
- The Hacker News, Oct 10, 2026: https://thehackernews.com/2026/10/anthropic-cuts-live-internet-access-for.html
- Al Jazeera, Oct 13, 2026: https://aljazeera.com/economy/2026/10/13/ai-could-self-replicate-on-home-devices-within-months-anthropic-says
- Digg, Sept 2026: https://digg.com/tech/xipdnyq0
- Reuters via Khaleej Times, Sept 9, 2026: https://www.khaleejtimes.com/business/tech/meta-muse-ai-agent-rollout
- The New York Times via The Next Web: https://thenextweb.com/news/zuckerberg-muse-launch-safety-concerns-nyt
(May have had help on this part)
Disclaimer: This is commentary, not investment advice. Views are my own, and Wrangl has no position in any company mentioned.