Jensen Huang – Will Nvidia's moat persist?
Jensen Huang argues Nvidia's real moat isn't chip specs but ecosystem lock-in through CUDA, and makes the contrarian case that export controls are backfiring by forcing China to build a competing AI stack that will become the global standard.
Read OriginalMy Notes (5)
Do as much as necessary and as little as possible
- Nvidia's job: turn electrons into tokens. Do as much as necessary and as little as possible to make that happen.
- Whatever Nvidia doesn't need to do itself, it partners out and folds into the ecosystem.
- Result: the largest partner ecosystem in the industry, upstream and downstream.
- The test is simple: if we don't do it, does it get done?
- NVLink, the full stack, 20 years of CUDA at a loss, CUDA-X libraries, cuLitho for computational lithography — none of this existed without Nvidia building it.
- If nobody else will build it, Nvidia commits fully. Whole company, all the might.
- Clouds already exist. If Nvidia didn't run one, somebody would.
- So Nvidia doesn't become a cloud. It doesn't become a financier either. There are people in the financing business already.
- The business model stays simple on purpose.
AI is a five-layer cake
- AI isn't just a model. It's a five-layer stack.
- The bottom layer is energy. Above that sits chips, then systems, then models, then applications.
- Abundant energy makes up for weaker chips. Abundant chips make up for weak energy.
- The US is scarce on energy, so Nvidia has to push architecture and extreme co-design to get max throughput per watt.
- If watts are free, you don't care about perf per watt. You can run old chips. 7nm chips are essentially Hopper.
- This is why China having cheap abundant energy matters so much for their AI position.
Every layer has to win, not just one
- Applications matter most because that's where AI diffuses into society and drives the industrial revolution.
- But every layer has to succeed. You can't sacrifice one to protect another.
- Conceding the chip layer in China to protect the model layer is picking a fight with your own stack.
- Long-term tech leadership means winning all five layers, not trading them off.
Conceding China is a loser's premise
- Nvidia's share is growing, not shrinking. The assumption that they'd lose China anyway is wrong on the facts.
- Jensen didn't wake up a loser. The US isn't a loser. The industry isn't a loser.
- Walking away from the second largest market in the world because you assume you'd lose it makes no sense.
The policy argument relies on extremes
- The export-control case depends on absolutes: any compute at all means we lose everything.
- Real life isn't absolute. You can keep the best tech at home and first, and still compete abroad.
- Both things can be true at the same time. It takes nuance and maturity, not all-or-nothing thinking.
- Conceding China hands away ~40% of the world's tech industry for the benefit of one layer of the stack.
- That's a disservice to the country, to national security, and to American tech leadership.
- The goal should be: best tech here first, and compete to win everywhere else.
"This is one of the concerns that I have about the doomers describing the end of work and killing of jobs. If we discourage people from being software engineers, we're going to run out of software engineers. The same prediction happened ten years ago. Some of the doomers were telling people, 'Whatever you do, don't be a radiologist.' You might hear some of those videos still on the web saying radiology is going to be the first career to go and the world is not going to need any more radiologists. Guess what we're short of? Radiologists."
Why jobs and tasks are not the same thing
- A job is the full role. A task is one thing inside it.
- A radiologist's job is patient care. Reading a scan is just one task.
- AI automating the task doesn't eliminate the job. It frees the person for the rest of the role.
- Confusing the two is how you end up with a shortage of radiologists and worse healthcare.
- Moore's Law now delivers about 25% per year. That's it.
- Between Hopper and Blackwell, the transistors themselves got roughly 75% better over three years.
- But Blackwell is 50x Hopper. Lithography can't explain that gap.
- Jensen originally announced Blackwell as 35x more energy efficient than Hopper. Nobody believed it. Dylan later wrote that Jensen sandbagged and it's actually 50x.
Where the 10x–100x leaps actually come from
- The lever is computer science, not transistors.
- Great algorithm work adds another ~10x on top of Moore's Law each year.
- MoE is the clearest example. New attention mechanisms also cut compute dramatically.
- Most of AI's recent progress came from algorithm advances, not raw hardware.
Why Nvidia can capture those algorithm gains and TPUs can't
- CUDA is programmable enough to invent new architectures on, whether that's MoE, diffusion, disaggregated inference, or hybrid SSMs.
- Nvidia is an extreme co-design company. They can change the processor, system, fabric, libraries, and algorithm at the same time.
- They push computation off the chip into the fabric via NVLink, or into the network via Spectrum-X.
- Without CUDA underneath, Jensen says he wouldn't even know where to start doing this.
Why process node isn't the whole story
- There's no 10x gap between 5nm and 7nm. Lithography is a small part of the picture.
- Architecture matters. Networking matters, which is why Nvidia bought Mellanox. Energy matters.
- Semiconductor physics still counts, but computer science and the full stack above the transistor drive most of AI's impact.
Summary used for search
TLDR
• Nvidia's 50x Hopper→Blackwell improvement came from architecture/co-design, not transistors (only 75% better) - Moore's Law is dead, computer science is the lever
• Supply chain bottlenecks (CoWoS, HBM, even plumbers) are temporary 2-3 year problems that get swarmed once identified - real constraint is energy policy
• Export controls are strategic mistake: China has energy abundance to gang up older chips, 50% of AI researchers, and US is forcing them to optimize for non-US hardware that will become open source standard
• Nvidia's moat is ecosystem (CUDA install base, programmability, richness) not hardware specs - TPUs/ASICs can't match because they lack flexibility for new algorithms
• Investment philosophy shifted: now willing to make $30B+ bets on OpenAI after initially missing that foundation labs needed supplier capital, not just VC money
In Detail
Jensen Huang defends Nvidia's position by reframing what their actual moat is. It's not the GDS2 file they send to TSMC - it's the full stack of CUDA ecosystem, install base across every cloud, and architectural programmability that enables rapid algorithm innovation. He points to Blackwell being 50x better than Hopper despite only 75% transistor improvement over three years as proof that computer science matters more than process nodes. The real advances come from new algorithms (MoEs, attention mechanisms, hybrid architectures) that require a flexible, programmable platform to implement quickly.
On supply chain constraints, Jensen argues every bottleneck is a 2-3 year problem maximum once the industry swarms it - whether CoWoS packaging, HBM memory, or EUV machines. Nvidia shapes the ecosystem years in advance through direct investments, technology licensing (like COUPE patents), and convincing CEOs across the supply chain about the scale of AI's future. The real bottleneck is energy policy preventing new datacenter builds, not chip manufacturing capacity. He notes China has massive energy abundance and can simply gang up more 7nm chips to compensate for being behind on process nodes.
The most striking part is his defense of selling to China. Jensen argues export controls are a strategic blunder that's accelerating exactly what the US fears: China developing their own chip ecosystem and AI stack. With 50% of the world's AI researchers, abundant energy, and now forced to optimize everything for domestic chips, China will build the dominant open source AI ecosystem. When those models diffuse globally, the world will standardize on Chinese hardware instead of the American tech stack. He frames it as the US "conceding the second largest market" and repeating the telecom industry mistake. His test: in a few years, when the US wants to export AI technology to India, Middle East, and Africa, those regions will already be locked into Chinese standards.
On competition, he dismisses TPUs as limited to one customer (Anthropic) and notes Nvidia runs everywhere while maintaining the best performance-per-TCO. He acknowledges missing the initial Anthropic investment because he didn't realize foundation labs needed multi-billion supplier investments, not VC funding - a mistake he won't repeat. The company philosophy is "do as much as needed, as little as possible" - only build what won't exist otherwise, partner for everything else. This means investing in neoclouds like CoreWeave but not becoming a hyperscaler themselves, since clouds will exist regardless.