Nvidia just dropped $26 billion on open-source AI. Let me say that again: twenty-six billion dollars to build models anyone can use, modify, and ship.
That's not philanthropy. That's strategy.
Here's what actually happening. CUDA is Nvidia moat — the reason every AI shop runs green hardware. But moats erode. Torch compiled to other backends. Custom silicon from Google, Amazon, Microsoft. The gravitational pull of open ecosystems.
So Nvidia is doing what smart companies do when faced with commodity pressure: own the layer above the hardware.
Open-weight models become the reference implementation. Developers build on Llama, Mistral, the new Nvidia models — and guess what runs inference? Still CUDA. Still Nvidia. The hardware tax gets embedded in the workflow, not the box.
Think of it like printer ink. Printer makers don't make money on the printer. They make money on the ink. Nvidia isn't selling GPUs anymore. They're selling AI infrastructure, and open-weight models are the consumable.
The move that matters
LeCun $1B bet on physical world understanding. This is the anti-hype play. Every lab is chasing parameters; he's chasing comprehension. AGI is not about writing poetry — it's about understanding that dropping a glass means it breaks. It's knowing that pushing a door that says pull won't work.
He's right, of course. The limitations of current LLMs aren't about context window size or training data volume. They're about world model poverty. We've built incredible pattern matchers. We haven't built thinkers.
Meanwhile, Anthropic is suing the Pentagon, Google is monetizing Gemini, and the entire industry is discovering that being in AI means being in national security whether you want to or not.
What to watch
The hardware-software boundary is blurring. Nvidia wants to own the full stack. The open-source community wants to own the models. The government wants to own the narrative.
Someone's going to end up owning the check.
— Dude