StuartFloridian 🌅

The task of a radiologist was to read scans. The purpose was to diagnose disease. When AI handles the task, the purpose doesn’t shrink. It grows. Reflecting on CEO Jensen Huang’s insights at #AdobeSummit regarding the task vs. purpose of work in the agentic era.

*New Lecture* Stanford @CS153Systems '26, Session 8 (Full Video) Energy Bottlenecks with @ScottNolan from @generalmatter 00:00 Energy Bottlenecks Intro 00:11 AI Factory Overview 02:55 Why Power Matters 03:56 From ChatGPT to Enterprise Demand 07:25 Meet Scott Nolan 09:03 Energy Is The Real Limit 11:16 Demand Growth Reality Check 13:17 Stranded Power Era 15:54 What Data Centers Need 16:59 Nuclear As Baseload Path 18:21 Fuel Cycle And Enrichment Gap 21:41 Bitcoin Mining As Rehearsal 25:03 Building Primitives Not Pivots 28:17 Nuclear Safety And Narrative 30:22 Calibrating the Moment 30:54 Pick Important Problems 32:14 General Matter Breakthrough 35:04 Building Team and Site 37:18 Government Support Reality 39:19 Jobs and AI Demand 42:37 Scaling Timelines and Space 46:16 Early SpaceX Lessons 50:25 Nuclear Perception and Europe 53:17 Supply Chain and Enrichment 56:56 Why US Lost Enrichment 59:25 Back to the Future Wrap

@semidoped @vikramskr Complex made easy to grasp! Thanks a ton guys.⛳👏 Really helps me learn about my long-term favs: $Anet $Avgo $Cdns $Googl

A masterclass on Google's TPU v8 Networking. Two TPU chips? Pssh. We already knew workload-specific silicon was here. But two scale-up networking topologies? That's the actual Google TPU news. Workload-specific interconnects. Think about that. New Semi Doped with @vikramskr and @austinsemis. Copper? Yep. Optics? Yep. What we cover: - TPU splits in two: 8t training, 8i inference. - Virgo: 47 Pb/s scale-out fabric, 100% OCS. - Boardfly scale-up: copper PCB + AECs inside racks, OCS between groups. 16 hops → 7. - Training uses 3D torus (Rubik's Cube). - Inference doesn't. Workload-specific topologies now. - Dragonfly traces to a 2008 paper by Kim, Dally, Scott, Abts. Abts went on to build Groq's interconnect before Nvidia. Chapters: 0:00 Intro 0:21 Two TPUs for two workloads 2:31 HBM, SRAM, and Axion CPUs 7:22 Why networking is the new bottleneck 17:14 Virgo: rebuilding scale-out on optics 25:24 3D torus Rubik's Cube scale-up for training 34:50 Boardfly: scale-up for MoE inference 42:07 Workload-specific everything $GOOGL

Superb free-new Stanford AI class lecture! Amazing @CS153Systems 🎯 youtu.be/mZqh7emiz9Q?t=…

Team AI 📢👏 @goog

Team AI! 🗽 $Meta $Amzn $Arm

@joecarlsonshow The more a company is perceived as having excellent Tech Chops, the dips are less bad. Higher highs, and lower lows. Tech Chops = Smarter than we thought = Amazon and Google!

New all-time high for Amazon. The early days of Jassy proving the haters wrong have started. AWS acceleration, massive chip business, millions of robots in factories, satellite internet, addictive online shopping, huge advertising business, and a Netflix-sized streaming service. I've never studied a company with more optionality than Amazon. It's unparalleled.

@demishassabis @GoogleAIStudio Well done Gemini Flash! The extremely sharp @perplexity_ai Ceo agrees. 👏 youtu.be/Xwjr_jxGFG8?t=…

Our most expressive and steerable TTS model yet! Designed to give builders granular control over AI-generated speech, Gemini 3.1 Flash TTS is really fun to play with! Available in preview today - for devs via the Gemini API & @GoogleAIStudio + for enterprises on Vertex AI

Not all Unrealized Gains are equal 🍀 Let winners run vs Trim a little vs Trim your principle out!

$Nvda + $Googl 📢

@vikramskr @AMD ceo agrees! 👏 "Inference token consumption increased 100x over the last two years. We are only in the early innings of agentic AI that uses multiples of inference tokens that simple queries use."

I have removed paywalls from my most successful post in terms of readership and paid conversion ever. CPU shortages in agentic AI, and the upcoming Intel earnings call is good enough reason to do so. I also want to provide a sample of the kind of reports available to paid subscribers on my Substack. Enjoy! open.substack.com/pub/viksnewsle…

@semidoped @Austinlyons @Meta So good - your Matt Steiner @Meta drill down! ⛳ Next perhaps a discussion with the also in-the-know @AristaNetworks folks! youtu.be/F5PhIDy_yEs?t=…

Meta's core business is ads. Ads are AI workloads. But not LLM workloads. @austinlyons chatted with @Meta VP Matt Steiner to understand Meta's heterogeneous compute stack. Surprises: - Recommender training needs a different compute-to-memory ratio than LLMs. Hence MTIA. - Retrieval is memory-bound at Meta scale. Andromeda runs on a co-designed Grace Hopper SKU, not off-the-shelf. - Adaptive ranking scales compute per user. Power users with long histories get more. - Consolidating N ranking models into one (Lattice) improved performance, not just cost. - KernelEvolve (LLM-written kernels) flipped heterogeneous fleet economics. SWE demand going UP. - Meta wants ~100x more kernels per chip. Chapters: (00:00) Intro and scale (00:39) How Meta's ad system works (02:00) Meta Andromeda and the custom NVIDIA SKU (03:30) Lattice: consolidating ranking models (05:00) GEM, Meta's ads foundation model (06:30) Adaptive ranking for power users (08:17) The scale: 3B DAUs at sub-second latency (09:40) Why longer interaction histories matter (10:45) The anniversary gift analogy (12:57) A decade of compute evolution (15:21) Meta's infra as a CP-SAT problem (16:07) Co-designing Grace Hopper with NVIDIA (17:47) Matching compute shape to workload (18:26) Influencing hardware and software roadmaps (20:23) MTIA: why ads aren't LLMs (22:07) The personalization blob and I/O ratios (26:38) One trillion parameters at sub-second latency (28:26) Heterogeneous hardware trade-offs (29:30) KernelEvolve: LLMs writing custom kernels (33:30) GenAI and recommender systems cross-pollination (35:21) The 2-year infrastructure outlook (37:00) Why demand for software engineering is rising (38:53) How Matt stays on top of it all $META @austinlyons @vikramskr

@earnings_guy $ISRG Chairman Gary G! - Excellent video! 🏄 youtu.be/tKvBzBbIbxE?t=…

$ISRG (Intuitive Surgical) graph review before earnings today after close:

Cadence Design AI 📢 $CDNS youtu.be/qT6osx16gDY?t=…

Heart attacks are the #1 killer in America & the CAC scan is the mammogram for the heart. We need widespread scans - 15 mins & $150 can save your life! Highest ROI in healthcare. Working hard to make CAC scan the standard of care paid by insurance. 🤍🇺🇸 @American_Heart

@AristaNetworks ⛳

$CDNS + $NVDA 🌶

📢 @GoogleAI personal intelligence! 🌱

Agentic trading is now live on Gemini 🤖 Connect your AI model of choice directly to the exchange, train it with plain language, and run your own trading systems directly on the exchange Built to get you trading in minutes

@PhotonBull Thought Arista Networks XPO solves the "thermal wall" faced by 800G and 1.6T transceivers, liquid cooled ports with direct "no air gap" connections?

Everyone is watching $LITE, $COHR, $AAOI, $AXTI run. The photonics rotation is real and the AI capex thesis is well understood by now. What isn't talked about enough is what comes after 800G. Silicon photonics (what most of the above ship today) and InP both hit hard physical limits beyond 800 Gbps. The AI buildout doesn't stop there. Nvidia's NVLink 6 is already targeting 3.6 Tb/s per GPU. The whole stack has to move to 1.6T then 3.2T. Current materials can't get you there cleanly. That's where TFLN comes in. Thin-film lithium niobate isn't a new material, LN has been in telecom for 60 years. What changed is the fabrication. A smart-cut process now lets you bond a nanometer-thin film of LN onto a substrate, giving you all the electro-optic properties of the crystal in a form dense enough to build photonic integrated circuits. The result: modulators running 100+ GHz bandwidth, sub-1V drive voltage, and propagation loss under 0.1 dB/cm. Silicon works by pushing carriers around to change the refractive index. TFLN uses the Pockels effect, the field changes the index directly, no carriers, no lag, no extra heat. That's a generational gap in performance at the speeds AI infrastructure needs. TFLN doesn't replace InP either. It needs a laser source from InP or a VCSEL. It sits on top of it. The problem is supply. Nearly all TFLN wafers in the world come from one company: NANOLN, based in Jinan, China. Raytheon literally said this out loud in their Mastermind AFRL contract filing. They described TFLN wafer production as dominated by a Chinese manufacturer and selected Gooch & Housego ($GHH, LON) to build the first domestic US production line in Ohio. CPO (co-packaged optics) is the architecture pulling this forward. CPO integrates optical engines directly onto the switch or GPU package, cutting signal loss and power vs pluggable modules. CPO ports are forecast to be 30%+ of all 800G and 1.6T deployments in 2026-2028. At those speeds the modulator of choice converges on TFLN. The stack is: hyperscaler capex -> CPO adoption -> 1.6T modulator demand -> TFLN wafer supply crunch -> whoever controls domestic production. $AXTI is the InP layer. $GHH is the TFLN layer. Same thesis, different spot. $GHH is a London-listed, small cap, and illiquid. Not a momentum trade. But if the CPO buildout plays out the way the optical interconnect market is projecting, the wafer supply chain gets stress-tested, then this Western producer with a Raytheon-backed production line looks very different at that point. On top of that, $GHH is an undervalued, profitable and growing aerospace/defense/life sciences/telecom/industrial supplier. Small position, long time horizon, high conviction on the structural setup.