AI Infrastructure • Compute Economics • Semiconductors • Models • Power • $NVDA since 2016
Los Angeles, CA
Joined April 2009
- Tweets34.8K
- Following1.3K
- Followers25.7K
- Likes75.1K
Pinned Tweet
Always a great time to hang out with @JensenHuang at @NVIDIAGTC Berlin!
I’m not ambitious enough! 😄
🇺🇸🇩🇪🚀🤖💚
Jensen’s ~$2,000 per GPU per year in enterprise software deserves more attention.
Enterprises deploying AI agents need security, governance and someone accountable for keeping the stack running. My bet: OpenShell becomes the standard security layer for enterprise AI, driving adoption of NVIDIA’s paid software and support.
Consider the math. Assume:
• 10M eligible enterprise GPUs shipped in a year
• 50% paid software attach
• $2,000 per GPU annually
That’s $10B in annual recurring revenue potential. Keep those subscriptions for five years and you get $50B from a single year’s GPU shipments.
Each new year of GPU sales adds another potential stream of software revenue. $NVDA
"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
Jensen to the CPO boys
It’s coming.
It took 5 years to invent the process technology. TSMC will deliver
Volume is about to ramp
Laser go pew pew
Nvidia GTC Berlin
$NVDA
Jensen stops by one of the leading companies in Germany, @Siemens to show some GTC Love 💚
Every company will be an Ai company
6pm standing room only for Corweaves $CRWV Richard Ahlfeld at @NVIDIAGTC Berlin
How do you architect for robotics?
Here is the Microduck up close
You can program it work DGX spark
This baby duck represents the future of physical Ai
it relies on reinforcement learning to transition AI-trained movements from digital simulations into the real world
Small duck now large robots later 🐥🤖💚
Jensen handled it with grace and class
He’s truly a generational leader and entrepreneur 💚
Jensen’s Berlin keynote briefly interrupted by a protester. He handled it like a pro.
It was a small moment that captures the broader tension around AI’s impact, both in the room and around the world.$NVDA
Remove the bottleneck
An unplugged GPU produces zero tokens.
NVIDIA says DSX POD cuts installation from up to 6 months to 14 days after site readiness. The modular approach makes time to first token faster and less on site labor.
$NVDA: engineer the whole factory and get it producing sooner.
AI keeps running into atoms.
NVIDIA targets SPIL CPO optical-engine 200GBps made by TSMC Coupe
NVIDIA's side of the Optical Cold War: secure the optical supply behind the GPUs. The moat reaches into the supply chain. $NVDA
The factory is the product.
Compute. Networking. Storage. Software. Deployment.
Our $NVDA thesis: the advantage compounds across the system. A rival has to beat the cost of getting useful work done across that entire stack.
The calm before the storm at GTC Berlin.
@dee_bosa and @GavinSBaker, getting the pregame started. It’s cold outside, but things are about to heat up. 🔥
In Jensen math that’s $1T of Ai capex Alone
But Elon and @SpaceXSI don’t build luxury data centers so it’s a little less call it $800B 😄
$SPCX
ELON:
- WE COULD HAVE 20GW OF SI CAPACITY ONLINE BY NEXT YEAR, SMALL CHANCE BUT IT IS POSSIBLE
- I WOULD BE SURPRISED IF WE BROUGHT ON LESS THAN 13GW OF COMPUTE BY THE END OF NEXT YEAR
If Elon brings 20GW online next year...
If you’ve followed our $NVDA work, tomorrow should connect a few dots.
The factory is the product.
The moat reaches well beyond the GPU. Watch how far Jensen takes that at Nvidia GTC Berlin.
More tomorrow.
🤖💚
Made it to Berlin for @NVIDIAGTC 🇩🇪
Excited to see how Germany’s Mittelstand is embracing physical AI and what’s ahead for robotics, manufacturing, and industrial automation.
It’s time to build 🤖💚
Big thanks to NVIDIA for hosting me at GTC Berlin 🍻
As I head to Europe for GTC Berlin, I keep coming back to a question: what good is building the future if our children don’t have a stake in it?
As a father of three, this is personal. I want my kids to grow up knowing they can help build something and own a piece of it. But that opportunity shouldn’t depend on whether their parents already have money.
Watching the unrest in France, I worry about what happens when people lose faith that the system has a place for them. Ownership won’t solve every grievance. But I believe capitalism is stronger when more people participate in its upside, rather than watch it from the outside.
That’s why @altcap work with Invest America, @MichaelDell and Susan Dell’s gifts, and Elon’s pledge here deserve real credit. Different efforts, united by a powerful idea: give children a stake in the future before they’re old enough to feel excluded from it.
A first investment can teach something a lecture cannot. You own a small piece of a business. You can watch it grow, weather setbacks and learn what patience means. Compounding becomes something you experience, not just something you’re told about.
Everyone wants equity. Not just a seat at the table, but a share in what gets built.
Let’s start the kids young.
👶💪🚀🇺🇸
Buying GPUs gets the headlines. Making them productive is the work.
Meta reports 99% fewer training interruptions at Prometheus from the initial peak, with NVIDIA engineers helping tackle firmware, memory and reliability issues.
Meta also reports over 85% pre-training goodput and over 90% for reinforcement learning on Muse Wonder. Goodput means training time that actually moves the model forward.
Our takeaway: the value is in the whole system. Chips, networking, software and recovery all determine how much useful work your compute delivers.
This is what making compute count looks like.
