0:00 NVIDIA just put one petaflop of AI performance into a laptop.0:04 That announcement is the capstone on a week that saw the entire AI landscape shift from chaotic growth to calculated power plays.0:14 Last week, in episode 205, we talked about Meta’s open model push as a power shift.0:20 But what NVIDIA just did isn’t a shift.0:23 It’s a consolidation.0:25 They’re not just making the chips anymore.0:28 They’re building the entire stack, from the silicon in your laptop all the way up to a new vision-language model that they claim is five times faster than anything else in its class.0:41 This is NVIDIA drawing a line and daring anyone to cross it.0:46 They are making a direct play for the entire AI ecosystem, and this week, they showed their hand.0:53 The model is called Nemotron 3.5 Super VL.0:56 It’s a 120-billion-parameter vision-language model.1:00 It reads documents, images, video.1:03 It reasons.1:03 It uses tools.1:05 And according to NVIDIA, it does it all five times faster than the next-best competitor.1:11 Partners are already using it.1:13 Terray Therapeutics is using it to extract data for chemical synthesis.1:19 BLACKBOX AI is using it for encrypted inference, hitting over five hundred output tokens per second.1:26 On a single GPU.1:27 And that GPU could now be in your laptop.1:31 The new RTX Spark laptop, from Microsoft, ASUS, Dell, and others, features a 6,144-core Blackwell GPU and a 20-core Grace Arm CPU.1:40 This isn’t just a faster machine.1:43 It’s a statement.1:44 It’s a local, private AI supercomputer designed to sit on your desk.1:49 The era of everything running in the cloud just got a serious competitor.1:55 While NVIDIA was busy trying to own the entire stack, Microsoft was busy building the single biggest piece of it.2:03 They just finished pre-training MAI-Base-2.2:06 That’s a two-point-eight TRILLION parameter model.2:10 For perspective, it has eighty billion active parameters, trained over thirty-eight days on sixteen THOUSAND of NVIDIA's own top-tier GB200 GPUs.2:21 Microsoft says it was trained on organic data, no distillation from other models.2:27 But here's the turn.2:29 They aren't open-sourcing it.2:31 And they aren't keeping it fully proprietary.2:34 They plan to license the weights to "trusted partners." This is a new category.2:40 Not open, not closed.2:42 Controlled.2:42 It’s a private club for AI giants, and Microsoft is handing out the first invitations.2:49 Meanwhile, national borders are becoming the new firewalls for AI.2:54 On October 16th, China’s Ministry of Commerce and Ministry of Science and Technology proposed new export controls.3:02 They cover twenty-three AI-related technologies, including foundation models and AI agents.3:09 The immediate reaction was that China is closing off.3:13 But the details matter.3:15 The proposal explicitly exempts technologies that are already public.3:20 If it’s on an open-source repository or in an academic paper, it’s exempt.3:26 This isn't a blanket ban.3:28 It’s a targeted measure.3:29 It allows China to keep consuming global open-source innovation while controlling the export of its own cutting-edge, proprietary work.3:40 The public has until November fifteenth to comment, but the message is clear: the free-for-all is ending.3:47 The developer experience is also heating up.3:51 It’s becoming a battleground for efficiency.3:54 OpenAI announced that its Codex coding tool is now more efficient.3:59 They didn't give numbers.4:01 They just said they got there in three ways.4:04 One, subagents reuse context instead of rebuilding it.4:08 Two, that context stays cached for longer.4:11 And three, improved request routing means more requests hit that cache.4:17 A poll on X showed eighty-six percent approval, so developers are happy.4:22 But they’re happy without any hard data.4:25 On the exact same day, the team behind Claude Code, a competitor, shipped an update.4:31 They implemented gzip compression.4:34 And they gave a number.4:36 The p90 first-token wait time dropped from four-point-zero-seven seconds to one-point-three-three seconds.4:44 A specific problem, a specific solution, a specific result.4:48 That contrast says everything about the state of the market right now.4:53 This growing power and complexity is now on the radar of global finance.4:59 G7 finance ministers met in Bangkok this week for the IMF annual meeting.5:04 The topic?5:05 AI-driven cyberattacks.5:07 They discussed holding simulated drills and tabletop exercises to prepare for future attacks.5:14 The bankers are worried about what they call "unknown unknowns." When the people who manage the world's money start war-gaming your technology, you know it's become systemically important.5:28 This is no longer just a tech story.5:30 It's a national security and economic stability story.5:35 And beneath all this commercial and geopolitical maneuvering, the deep, fundamental research continues.5:42 At the IWAI2026 conference, Professor Tadahiro Taniguchi of Kyoto University gave a keynote on the "Collective Predictive Coding Hypothesis." He’s exploring how language and cooperative behaviors might emerge in AI systems from first principles.6:01 It’s a reminder that while companies are building trillion-parameter models and governments are drawing up regulations, there are still fundamental questions about intelligence itself that we are just beginning to ask.6:17 So let's go deeper.6:18 Let's start with NVIDIA.6:20 Because this week wasn't just a product launch.6:24 It was a declaration of intent.6:26 For years, the game was simple.6:28 NVIDIA made the best shovels for the AI gold rush.6:32 Everyone, from OpenAI to Google to Meta, bought their GPUs by the tens of thousands.6:39 NVIDIA’s job was to make the hardware faster, and everyone else’s job was to figure out what to do with it.6:47 That's over.6:47 NVIDIA just laid out a complete, vertically integrated vision for AI.6:53 Look at the pieces.6:54 First, the high end.6:56 Nemotron 3.5 Super VL, the 120-billion-parameter model.7:00 It’s a vision-language model, which is the frontier right now.7:04 The ability to understand images and video, not just text, is what separates the latest models from last year's.7:13 And NVIDIA claims a five-times speed advantage.7:16 That's not an incremental improvement.7:19 That’s a leap.7:20 And they’re not just putting it on a shelf.7:23 They’re integrating it.7:25 Partners like Terray Therapeutics are using it for drug discovery.7:30 BLACKBOX AI is using it for high-speed encrypted inference.7:35 This model is already working.7:37 Then, look at the other end of the spectrum.7:40 On October 15th, just before the big launch, they quietly dropped Nemotron 3.5 Bolt.7:46 It's a 74-billion-parameter model, but with only nine billion active parameters.7:52 It’s compressed to 2-bit precision.7:55 What does that mean?7:57 It means this powerful, context-aware model—it supports up to 256,000 context tokens—can fit into 24 gigabytes of memory.8:06 It can run, and run fast, on a single, high-end consumer GPU.8:10 It's designed for tool calling and multi-step agent workflows.8:15 This is the model you build products on top of.8:18 And then, the centerpiece.8:20 The RTX Spark laptop.8:22 A 6,144-core Blackwell RTX GPU.8:24 A 20-core Grace Arm CPU.8:26 Up to 128 gigabytes of unified memory.8:29 And the headline number: one petaflop of FP4 AI performance.8:33 In a laptop.8:34 You can buy it at Best Buy.8:36 Now, connect the dots.8:38 NVIDIA has a flagship model for the cloud, a hyper-efficient model for the edge, and now, the perfect hardware to run it all locally.8:48 They are building the Apple of AI.8:51 A tightly integrated hardware and software ecosystem where every component is designed to work with every other component.9:00 Why rely on the cloud and pay for inference every time you run a query, when you can have a petaflop on your desk?9:08 That's the question they are forcing every developer and every enterprise to ask.9:14 But there’s friction.9:16 PCMag reported that while people lined up for the launch event at a San Francisco Best Buy, few actually walked out with a machine.9:26 The cost is high.9:27 And early reports suggest the battery life, while good, doesn't beat the Snapdragon X Elite machines, even with a larger battery.9:37 This is the core tension in NVIDIA’s strategy.9:40 They’ve built the F-22 of laptops.9:43 But how many people need, or can afford, to fly an F-22 to work?9:48 This isn't a criticism of the technology.9:51 It's a question about the market.9:53 NVIDIA is betting that the demand for local, private, powerful AI will be so strong that customers will pay the premium.10:02 This week was the first test of that bet.10:05 Now, let's talk about the other major shift.10:08 The splintering of the AI world.10:11 For the last couple of years, the debate was simple: open versus closed.10:16 You had the OpenAI, Google, and Anthropic camp, with their proprietary, API-gated models.10:23 And you had the Meta and open-source camp, releasing weights for anyone to use.10:29 It was a philosophical and a business model divide.10:33 This week, two new, powerful factions emerged.10:36 First, you have Microsoft's "controlled release" model.10:40 MAI-Base-2 is a monster.10:42 Training a 2.8 trillion parameter model is a feat of engineering that only a handful of companies on Earth can pull off.10:51 It took 16,000 GPUs running for 38 days.10:54 The cost is astronomical.10:56 And what are they doing with this crown jewel?10:59 They're not locking it in a vault like OpenAI.11:03 But they're not giving it away like Meta.11:06 They are licensing the weights to "trusted partners." So what does it all add up to?11:12 It's a third way.11:14 A feudal system for AI.11:15 Microsoft is the king, and it will grant land and titles—the model weights—to its most powerful and loyal vassals.11:24 This allows them to maintain control, to ensure the model is used in ways that align with their interests, and to capture a different kind of value than pure API sales.11:36 It avoids the chaos of true open source while enabling a deeper level of integration and customization than a black-box API.11:46 It’s a power play designed for a mature ecosystem, where the goal is not just to build a model, but to build an empire around it.11:55 At the same time, we're seeing the rise of the "nationalist" model, exemplified by China's proposed export controls.12:04 On the surface, it looks like a clampdown.12:07 But the exemption for open-source is EVERYTHING.12:11 Think about what this policy actually does.12:14 It allows Chinese researchers and companies to continue to pull in, use, and build upon all the open-source AI innovation happening around the world—from Meta's Llama to models from Mistral in France.12:29 But when a Chinese company builds something truly novel and proprietary on top of that foundation, the government can step in and say, "That, you cannot export without our permission." It's a one-way valve.12:45 Innovation flows in, but cutting-edge intellectual property stays home.12:50 This complicates things massively for Chinese developers who want to participate in the global open-weight community.12:59 Do they release their new model and make it exempt from the rule, or do they keep it private and face potential export restrictions?13:09 It introduces a layer of strategic calculation that developers in other parts of the world don't have to deal with.13:17 It also signals that the Chinese government sees AI not just as a commercial technology, but as a strategic national asset, to be cultivated and protected just like any other critical resource.13:32 So the simple "open vs.13:34 closed" world is gone.13:35 Now you have closed, you have open, you have Microsoft's "trusted partner" feudalism, and you have China's "one-way valve" nationalism.13:45 The landscape just got a lot more complicated.13:49 And a lot more interesting.13:51 This week wasn't about a single breakthrough.13:54 It was about the entire board changing.13:57 The era of pure, unrestrained innovation is maturing into an era of strategic consolidation.14:04 NVIDIA is no longer content to sell the shovels.14:08 They are now building the entire gold mine, the refinery, and the bank.14:13 Their vertical integration, from the Nemotron models down to the RTX Spark laptop, is a direct play to own the entire AI platform.14:23 Microsoft's "trusted partner" model for MAI-Base-2 and China's nuanced export controls show that the very definition of "open" is being contested and reshaped by corporate and national interests.14:37 The simple binary of open versus closed is fracturing into a complex multipolar world.14:44 So, what does this set up?14:46 The battle for the next year is not just about who can train the biggest model.14:52 It's about who can build the most compelling, integrated, and defensible platform.14:58 It’s about controlling the stack.15:00 The question is no longer just "what can AI do?", but "who decides what AI does?" This week, we saw the major powers make their opening moves in that much, much larger game.15:13 The platform wars have begun.
