Welcome, humans.

NVIDIA is worth over $4 trillion today; it’s the most valuable company in the world. Why? First of all, because they make the best AI chips (GPUs) on the planet.

Now, that narrative has been shaky lately, as endless ink has been spilled over how Google’s successful launch of Gemini 3.0 turned their new Ironoowd TPU AI chips into a genuine competitor to NVIDIA (here, here, here, this thread, this video, etc). But great chips are not the only reason developers can't quit NVIDIA.

The real lock-in? A 17-year-old software platform called CUDA. And this company, Modular, just raised $250M to break it.

In our latest podcast episode, Tim Davis—co-founder of Modular and ex-Google Brain—explains why AI infrastructure is fundamentally broken, and what his team is building to fix it. We also get into some spicy territory: Is scaling LLMs even the right path to superintelligence? Tim doesn't think so.

Here's what we cover:

• (9:00) Hardware vs. software: What's actually constraining AGI? Tim's answer might surprise you.

• (12:49) Are LLMs actually intelligent? Tim argues autoregressive models trained on human text aren't the path to superintelligence, and explains what might be.

• (21:00) The autonomous car problem: Why intelligence needs speed, and why CPUs can't get us there.

• (26:49) Introducing Mojo: The new programming language that lets you write code once and run it on any GPU (not just NVIDIA).

• (32:00) The “hypervisor for AI compute”: Tim's vision for a unified platform where you just set your requirements and the software figures out the best hardware.

• (35:28) Case study: How InWorld (good text to speech model!) got 4x performance and 70% cost reduction using Modular on NVIDIA Blackwell machines.

• (42:30) Breaking the CUDA lock-in: Why even NVIDIA partners with Modular, and what more silicon competition could mean.

• (50:00) The bridge analogy: We're deploying AI at massive scale… without understanding how it actually works. Tim thinks that's a problem. Ya think??

Bottom line: While everyone debates which model is smartest, there's a quiet infrastructure revolution happening underneath. Modular is betting that the future of AI isn't about being locked into one chip vendor, it's about software that makes all hardware work together.

Watch / Listen now on: YouTube | Spotify | Apple Podcasts

Keep scrolling for more info an upcoming live with OpenAI that’ll teach you how to code with AI, and three other fantastic videos we just released!

THIS EPISODE WAS MADE POSSIBLE BY OUR PARTNER…

Dell Technologies Is First to Ship NVIDIA's Next-Gen AI Chip—And You Can Buy It Now

Remember waiting months for GPUs during the AI boom? While most companies are still waiting to get their hands on NVIDIA's Blackwell architecture, Dell Technologies just started shipping it.

The Dell Pro Max with GB10 is the first desktop to pack NVIDIA's GB10 Grace Blackwell Superchip—the same next-gen architecture powering the AI labs building tomorrow's models. It's not a server rack. It's a desktop that sits on your desk.

Here's what you get:

• 128GB of LPDDR5X memory for running large models locally.

• 4TB SSD storage for massive datasets.

• NVIDIA DGX OS pre-installed—the same software stack used by AI research teams.

• 20 CPU cores (10 Cortex-X925 + 10 Cortex-A725) optimized for AI workloads.

This isn't for casual ChatGPT users. It's for teams training custom models, running inference at scale, or doing serious data science work who are tired of waiting for cloud compute.

If you burn through thousands of GPU cloud credits in a few months, you'll definitely want to check this out. It’s a surprisingly affordable way to bring Blackwell performance in-house. Check out the Dell Pro Max with GB10 here.

🔴 THIS FRIDAY: Going from Zero to Agentic Coding Wizard with OpenAI’s Codex Agent

Friday, Dec 5 | LIVE with Alexander Embiricos, OpenAI Codex Product Lead 10:30am PST | 12:30pm CST | 1:30pm EST

GPT-5.1 Codex Max just dropped, and its OpenAI's most advanced agentic coding model yet. It can work independently for hours, handle entire PRs and refactors, and run multi-hour agent loops without losing context.

We're doing a 2-hour hands-on deep dive with the product lead himself. The goal? Take you from absolute beginner to confidently managing Codex agents for your next coding project. We'll set it up together, build real agents, and push Codex Max to its limits—plus plenty of time for audience Q&A.

Three More New Videos You Should Watch

Here are three more videos we just released that we think you’ll love.

1. The Future of Windows in an AI World with Pavan Davuluri, CVP of Windows + Devices at Microsoft

Windows is now 40 years old, and it's getting its biggest upgrade yet: AI baked into everything from the taskbar to the file system. Pavan Davuluri (23 years at Microsoft, helped create Surface) walks us through what's changing—and what's staying the same.

• (3:24) How Copilot is integrating into the Windows taskbar—and why voice is the next big interaction layer.

• (5:16) The hybrid compute vision: what runs locally vs. in the cloud, and why it matters for privacy.

• (8:53) “Click to Do“: how local AI turns a static PDF into an Excel sheet you can manipulate with agents.

• (18:47) Addressing “AI fatigue”: how Microsoft is making AI features opt-in, not forced.

• (28:47) Where computing is headed in the next 5 years.

2. Securing AI (With AI) with Vasu Jakkal, CVP of Microsoft Security

Here's a wild stat: Microsoft sees 7,000 password attacks per second. And now attackers are using AI to create self-generating malware. Vasu Jakkal oversees Microsoft's entire security business and explains how they're fighting back—including 6 new AI agents that just launched in Security Copilot.

• (9:26) How a phishing triage agent detects 6.5x more threats per minute (and is 77% accurate).

• (13:28) The 12+ new security agents Microsoft just announced—and the 350+ agents customers are building.

• (15:54) “Trust but verify“: how agents show their work so humans can audit every decision.

• (21:52) Securing AI systems themselves, not just the data they access.

• (23:50) The big question: Do AI agents need their own identities and security controls? (Spoiler: yes).

3. 80% of the World's Data Is Trapped in Video with Danny Nicolopoulos, TwelveLabs

Here's something we didn't know: 80% of the world's data is already in video format…and almost none of it is going into AI models. TwelveLabs built video-native AI that understands video across time, space, and all modalities. Not just frame-by-frame image recognition.

• (1:24) The untapped world of video data: 4.5 billion people recording on smartphones, 300 million video calls per day.

• (2:53) Real use cases: from AI-enhanced mechanics in automotive to reducing insurance paperwork.

• (4:22) How TwelveLabs processes video differently than “language models that tried to modify themselves.“

• (6:41) Launching Marengo 3: 6x smaller embeddings, global language support, now on Amazon Bedrock.

Dive deeper with these resources:

• Try Modular for free

• Check out the Mojo programming language

• Read Richard Sutton's “The Bitter Lesson“

• Check out Tim's blog: “What We Owe the Minds We Create”

• Anthropic's mechanistic interpretability research

Stay curious,

The Neuron Team

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.