Wow, didn’t expect my Sunday to go like this. There’s a lot of hysteria on X right now. Stacy Rasgon @Srasgon has covered NVIDIA for years and understands the company. He’s a voice of reason amid a lot of armchair commentary from people who don’t. In his new note from today, “NVIDIA: Stirring the (honey)pot?”, he says the situation could create political and regulatory noise, and the overhang may stick around. But he doesn’t expect it to materially change demand for NVIDIA’s products. A headline can create uncertainty without changing what customers are buying or the long-term demand for compute. Bernstein maintained its Outperform rating and $400 price target.

Timeline cleanse I’ve been writing about this: the next inflection is the lab. I lost both my parents to cancer and my father in law to Alzheimer’s Jensen lays out the vision here. Connect AI agents to instruments and robots, run experiments, and feed the results back into models. I’ve seen a piece of this firsthand at Terray, where I’m an investor. The moat is the measurement. More experiments. More data. More tokens. 🧬🤖 We need faster cures

What is a GPU worth years after it ships? At NVIDIA GTC Berlin, I sat down with Thomas Galbraith of Barkr on resale prices, rental rates, useful life and the valuations behind compute financing. Full interview below. Chapters 00:00 Opening highlights 00:26 Thomas Galbraith at GTC Berlin 01:03 What Barkr does 02:18 How do you value compute? 03:06 Valuations from resale transactions 03:48 Rental rates versus resale prices 04:37 CPU comparisons and forward valuations 06:07 Methodology, transparency and public comps 06:53 NVIDIA, AMD and other hardware 07:46 The aviation analogy 09:25 Useful life and market expectations 10:22 Financing and the data gap 12:08 Pricing risk and closing thoughts

The AI race is becoming a race for capital. Today on @theallinpod, @chamath argued that AI borrowing is crowding out appetite for Treasuries. Eight days earlier, I filmed Jensen at GTC Berlin explaining how demand for AI drives demand for compute, then capital, putting pressure on financing costs. I put the clips together because the connection matters: strong AI demand can coexist with expensive capital. The winners still have to finance the buildout and earn the return.

You can’t run an AI factory on keynote slides. It needs power. Cooling. Construction. And capacity someone can actually deliver. At GTC Berlin 2026, I sat down with NVIDIA’s @Dion2730 on how @nvidia is tackling the bottlenecks behind the AI boom: Groq LPX, agent guardrails, DSX POD and capacity allocation. Full interview below. 👇 Chapters 00:00 Opening highlights 00:44 Dion Harris at GTC Berlin 01:29 Groq LPX and the inference roadmap 03:13 Premium inference: latency versus throughput 04:35 Vera, OpenShell and BlueField agent security 07:49 DSX and Dynamo: the AI factory stack 10:33 DSX POD, deployment and physical bottlenecks 14:56 Closed-loop cooling and water use 17:04 Power constraints and grid coordination 20:18 Annual hardware cadence and allocation 22:18 Planning the factory and its supply chain 24:36 Throughput, capacity and competing systems 26:09 What comes next for agentic AI 28:21 Closing thoughts

Everyone’s chasing the new vintage. Meanwhile, H100 rental rates rose 10.8% over the month, ahead of B200 and B300 in Silicon Data’s snapshot. GPUs aren’t fine wine, but apparently nobody told customers the older vintage was supposed to be worthless. $NVDA 🍷

Here is the Microduck up close You can program it work DGX spark This baby duck represents the future of physical Ai it relies on reinforcement learning to transition AI-trained movements from digital simulations into the real world Small duck now large robots later 🐥🤖💚

Everyone's watching GPUs. Jensen's watching substrate. "Substrates are getting larger and there are more chips... it's going to be hard to close that gap for a few years." Morgan Stanley's Oct 14 ABF note: undersupply hits 9.3% in CY27 and 13.7% in CY28, prices rise 25%+, and new capacity doesn't arrive until 2028 to 2030. Aunty Lisa took her trips to Taiwan too late. The real bottleneck is the thing the chips sit on anon $NVDA $AMD

"One of these days people might not categorize us as a chip company anymore." - Jensen Most investors still model Nvidia as a GPU company. Meanwhile he's laying out a software licensing business: a couple thousand dollars per GPU per year, across millions of GPUs a year. I wonder how many investors know what CUDA-X is. Or Dynamo. Or Nemotron. CUDA-X is the library layer whole industries build on. Dynamo is the open source inference stack that runs entire GPU fleets. Nemotron is Nvidia's own family of open models. The chip is the wedge. The stack is the business. We'll dig deeper into this at BEP Research and @NVIDIAGTC Berlin. $NVDA

2,000 sandboxes. One CPU. Agents need somewhere to get their hands dirty 🛠️ We wrote about Vera and the agentic CPU opportunity at BEP in July. Now @ivanburazin and @daytonaio are putting numbers behind it: 12.7 jobs/sec for Vera vs 10.9 for AMD EPYC 9755 in their test. Tools, code, tests, retries. Every step hits the bill. We’re digging deeper into this in our upcoming BEP CPU series. Love what you’re building, @ivanburazin 👏 You and the @daytonaio team are making this real. We had some fun visualizing it.

This is the NY–SF disconnect in one debate. One side sees a falling AI bill. The other sees another workflow becoming economical. I’m with @GavinSBaker on the mechanism. Cheaper intelligence can expand the market. We laid out the investment question at BEP after DevDay: does the extra work outrun the efficiency gains? That’s what connects agents to infrastructure demand. We’ve been building the task-cost tools to test it.

Some AI bubble. Lumentum CEO in April: “basically sold out” through 2027. Now: “almost sold out for 2029.” Under-shipping demand by 30% on some products into 2028; ~70% on others in 2027. Demand keeps outrunning the fabs. Laser go pew pew. $LITE

The “NVIDIA killer” business is turning into the NVIDIA partner program. The Information reports NVIDIA plans to invest in d-Matrix, which is already joining NVLink Fusion. Compete on the accelerator, plug into Jensen’s infrastructure. That’s the beauty of owning the platform. Even rival chips can expand your footprint. All roads lead to NVIDIA. $NVDA

The Financial Times having to clarify gross vs. net revenue is quite something for a publication with “Financial” in the name. $70B gross. $50B net. Both can be true. Before framing a $20B shortfall, maybe check whether you’re comparing the same number. Accounting 101.

A wafer can work and still miss the performance spec. That’s the parametric-yield question we’ve been digging into at BEP. For Cerebras, speed has to survive manufacturing, repair and testing at scale. Useful work from @damnang2 on those trade-offs. Modeled scenarios, not disclosed production yields.

. @JensenHuang came to America as a child and worked his way to the top. Now, the National Medal of Science, with his parents there to share the moment. As a first-generation American whose parents emigrated from Iran for a better life, this hits home. They believed in America as a land of opportunity. Imagine how proud his parents must be. 🇺🇸 $NVDA

FT’s ~$50B OpenAI run rate is a September snapshot. @FundaAI has it at $48.8B, close to that headline. Dots launched Sept 29. @thsottiaux now reports 40M active users across Codex + Work. BEP’s hypothesis: dots, new models and tiers, plus coding and enterprise adoption, could drive $100B by year-end. That requires ~26% monthly growth on the same revenue basis. Ambitious, not confirmed. As @altcap argued at All-In, downstream revenue must justify upstream capex.

Big week ahead in SF and San Jose ✈️ DigitalOcean’s Open Intelligence Summit, SEMICON West and dinner, then OCP. AI agents, chips and the infrastructure behind it all. Looking forward to getting out from behind the screen. Around next week? DM me. Let’s meet up 🤝