Running your own AI infrastructure isn't really an infrastructure decision. It's a data decision, Ben Pouladian @benitoz says. Hosts Luke Knight and Ronan Chambers pressed him on where frontier labs' economics hold up now that NVIDIA owns Hugging Face. Pouladian's framing: rent a frontier model like Opus for the hard problems, and let it orchestrate cheaper open models like Nemotron or Qwen for the routine work you already own. The reason to own that layer isn't the hardware bill. It's that the data stays on your own system instead of training the model company selling it back to you. The infrastructure spend isn't really about hardware. It's about which vendor doesn't get to train on your data and use it against you later. Source: BEP Research -

Home → grocery-store parking spot. The latest FSD update on my Tesla handled the drive, pedestrians and the parking itself. This clip shows the final stretch. End-to-end for an everyday errand. Pretty amazing. FSD (Supervised).

Weekends are for philosophy The ClawFather @steipete has spoken 🦞 Model choice matters. Privacy matters more. A slick demo doesn’t earn Zuck the keys to my inbox, files and accounts. Open source isn’t automatically private. But I want control over my data, not another “trust us” from Meta.

First takes the Apple iPhone 18 pro max is speedier than the my old 17. Excited to try out this more useful Siri AI Apple wins on privacy. Jury still out on Muse $AAPL $META

The next species might not need a body. Dinner with @TheStalwart after Odd Lots LA last night got me thinking about AI agents that communicate, coordinate and work as a team. I unpacked it on @etnshow this morning. Full discussion:

You could watch every All-In keynote and still miss this. We spent two days talking to people building, financing and buying the AI boom. Six off-stage conversations. What they told us. What it means for investors. People still do business.

Anthropic just stood up a real biology lab for AI drug work Physical biology. Not only software Street still prices the model as the moat I already wrote the note: Design is cheap. Test is scarce. The moat is the measurement Unpaywalled so anyone can read it The Moat Is the Measurement

Nebius notified customers of 17–21% on-demand GPU price increases effective Oct. 1. For B200: +$1,350 per 1,000 GPU-hours. BEP’s first Compute Market Brief breaks down the changes, budget impact, and implications for pricing power.

This is my first time putting together a list like this. I’d like to share some independent researchers who are active on X and consistently provide me with valuable insights. They have given me a lot of valuable insights, so I highly recommend checking them out! And if there’s anyone I missed, please let me know in the comments! @PhotonCap : One of the best experts in optics, without a doubt. His Substack has a lot of deep and insightful analysis on the optical industry. @vikramskr : A semiconductor researcher with a background at Qualcomm who now also provides research for institutional clients. He’s not only extremely knowledgeable, but also a genuinely great person. I’ve learned a lot from him, both professionally and personally. @Frenchie_ : A France based market analyst who also runs a large community of his own. I feel like I’m always receiving insights and help from him, so I’m simply grateful. @ParadisLabs : An excellent source for timely market news and insights on X. He also shares insightful articles every week, so definitely give him a follow. @jukan05 : Do I even need to explain? His subscription is only $1. Just go subscribe already. @jmartinprin : An optics and semiconductor expert formerly with SemiAnalysis. I’ve learned a tremendous amount of technical knowledge from him and personally really appreciate all the insights he has shared with me. @demian_ai : A market expert currently with Nebius. I’m really excited to see the AI bottleneck product he’s building. @pequityresearch : A professional investor and market analyst. Personally, I think his X subscription is way too cheap for the value he provides. If you want to follow analysis across the broader market, definitely check him out. @benitoz : I was particularly impressed by his recent critique of SemiAnalysis. Looking forward to seeing more from him. @ShortSeller : He gives me a lot of insight into technical chart analysis, which is definitely one of my weaker areas. Always appreciate what I learn from him. @ren_stocks : I don’t know. I just really like Pikachu. 😂 And his posts too, of course. @NuttyCLD : Personally, I think he’s one of the best semiconductor industry professionals at explaining technical topics in a simple and understandable way. Please post a little more on X! @AurelionRsch : They run what I consider one of the best models for an independent research firm. I’ve gained a tremendous amount of insight from their work. @BigBerbowski : He provides thoughtful research across AI infrastructure, semiconductors, and the broader ecosystem. His work has personally helped me a lot, and I’m very grateful for it. If there’s anyone I missed, please let me know in the comments! And once again, thank you to everyone on this list for all the insights and support!

Codex is OpenAI's bet that researchers burn tokens like power Top 1% spent $7k-$8k a day in early August Annualize that and you are already $2.5M to $2.9M a year for one researcher Noam says that spend is on an exponential It is going to keep increasing Token Explosion 🤯

Brad’s message was simple: “Make every child a capitalist.” My father did that for me. 70 million kids. Direct owners in America. Get kids owning assets early. Let compounding do the rest. Every child a capitalist.

AI factories are not just GPUs. GPUs + CPUs + DPUs + networking all have to work together as one system. Jensen now expects NVIDIA to sell 2x as many chips next year. More compute requires more silicon everywhere. $NVDA

Interesting post Fed day Our AI x Science basket is up 5.2% today. SPY is up 0.9%. We put this book up in August. It is the applied AI biotech stack. Thermo, Danaher, Illumina, Twist, Tempus, 10x, Schrodinger. The lab names. TWST, TEM, SDGR, and TXG led today. Markets still chase the model. We already owned the lab. Casper called 2026 an outstanding year on biotech and AI demand. Tempus is still running on data licensing. One year on this basket is about +97%. Today felt like the tape catching up. The next inflection is the lab

Wait. So when it's NVIDIA vs NVIDIA you compare best config vs best config, TRT-LLM and MTP on, per megawatt, at the rack? That's the comparison you called apples to bananas for TPU vs GB300 last week @fabknowledge Same referee, same rulebook. Glad we agree. Now run the TPU cube that way.