Jensen Huang Just Revealed: Why Nvidia Will Hit $1 Trillion Soon!

[00:00:00]I see through 2027 at least $1 trillion.

[00:00:09]>> $1 trillion to be and Jensen Huang just revealed exactly why Nvidia will hit it soon.

[00:00:15]>> People have heard me say I believe that computing demand has increased by 1 million times in the last 2 years.

[00:00:24]>> So, let's break down the real reasons why $1 trillion isn't a bold prediction.

[00:00:28]It's already a done deal.

[00:00:30]Reason one, the inference explosion nobody saw coming.

[00:00:34]For most of artificial intelligence's history, the expensive part was training. You'd pour enormous resources into building and training a model.

[00:00:42]That was the big computation cost.

[00:00:44]Actually using the model afterward, generating responses, answering questions, that part was considered cheap, almost an afterthought.

[00:00:52]Infrastructure teams planned around it accordingly. That entire logic is now obsolete.

[00:00:57]The first thing that killed it was ChatGPT, which launched the generative artificial intelligence era and multiplied inference demand overnight.

[00:01:06]But, that was just the beginning. Then came reasoning models, systems that don't just answer a question, but actually think through it step-by-step checking their own work before responding.

[00:01:17]Every extra reasoning step burns more compute. The models grew larger, the context windows grew longer, and the computation needed for a single response exploded by roughly 10,000 times compared to earlier systems. Then came the third wave, agentic artificial intelligence, systems that don't just answer you, they do things. They read your files, write your code, run your tests, spot the errors, go back, fix them, and loop until the job is done. And every single one of those actions, every tool call, every file read, every iteration, requires inference. The compute doesn't run once, >> [music] >> it runs continuously for as long as the agent is working, 10,000 times more compute per task, 100 times more users.

[00:02:04]That's where the 1 million figure comes from, and demand is still [music] climbing.

[00:02:09]Reason two, a moat that took 20 years to build. Here's what most coverage gets wrong. They focus on the chip, the chip [music] speed, the chip price, which competitor is catching up. But the chip is almost the least important part of this story.

[00:02:24]What actually makes Nvidia nearly impossible to displace is something called CUDA, a software platform that has been building quietly for 20 years. And by the time most people noticed it, it was already too late to compete with it.

[00:02:37]Every major artificial intelligence breakthrough of the last decade, deep learning, large language models, the entire modern artificial intelligence stack, was built on CUDA.

[00:02:47]Every researcher, every framework, every cloud provider, PyTorch runs on CUDA.

[00:02:55]Most of the world's artificial intelligence training infrastructure runs on CUDA. Hundreds of millions of Nvidia chips are already deployed across every major cloud and every major research institution in the world. That installed base is not a sales number, it's a gravitational force.

[00:03:12]>> The single hardest thing to achieve is the thing on the bottom, installed base.

[00:03:18]It has taken us 20 years to now have built up hundreds of millions of GPUs.

[00:03:25]>> More developers build on CUDA, which produces more breakthroughs, which attracts more industries, which drives more demand for Nvidia hardware, which funds more software improvements. The flywheel doesn't just spin, it accelerates. And the deeper it goes, the harder it becomes for any competitor to pull developers away, because switching means abandoning every tool, every library, and every optimization those developers have built up over years. A competitor can tape out a faster chip. They cannot rebuild 20 years of ecosystem trust in any reasonable timeframe.

[00:03:57]That moat is what gives Jensen the confidence to call $1 trillion by 2027 because the purchase orders were already on the books before he said it out loud.

[00:04:07]But we still haven't answered the real question. Is Jensen actually right or is this the kind of confidence that looks brilliant [music] when it works and delusional when it doesn't? Reason three.

[00:04:18]The last time Jensen made a bold claim, he undersold it.

[00:04:21]>> I said last year at this time that Nvidia's Grace Blackwell in being 72 was 35 times perf per watt.

[00:04:32]Nobody believed me.

[00:04:34]>> 35 times better performance per watt than the previous generation. The audience was skeptical. Then an independent analyst [music] ran the actual benchmarks.

[00:04:43]>> He accused me of sandbagging. He says, "Jensen sandbagged. It's actually 50 times."

[00:04:49]>> The real number was higher than what he claimed on stage. That's what happens when you design chips, [music] software, networking, and cooling as one integrated system rather than just shipping faster silicon and hoping the rest keeps up. And with the new Vera Rubin platform, token output per megawatt increases by 350 times compared to just two years ago. That's what makes the revenue math work. And the math itself is simpler than it sounds. Reason four. The token factory turns every data center into a revenue machine. Jensen spent a significant portion of the keynote explaining something he called the token factory. The idea is straightforward. A data center used to store files. Now it manufactures tokens.

[00:05:34]The outputs that artificial intelligence systems generate every time they think, write, respond, or act. And tokens now have a market price ranging from a few dollars per a at the free tier all the way up to $150 per million for premium high-speed inference.

[00:05:51]>> At every single tier, at every single tier, at every single tier, we increased the throughput.

[00:05:57]And at the tier that where your highest ASP and your most valuable segment, we increased it by 10x.

[00:06:05]>> Each new generation of NVIDIA hardware doesn't just make the same tokens cheaper. It unlocks entirely new service tiers that the previous generation physically couldn't reach.

[00:06:15]Grace Blackwell created throughput levels that enabled premium pricing the older Hopper architecture simply couldn't support. Vera Rubin does the same thing again, one level higher.

[00:06:25]Jensen showed a simple model. Take a 1 gigawatt data center, split the power across four pricing tiers, and the jump from Hopper to Grace Blackwell alone multiplied potential revenue by five times. Vera Rubin multiplies it again.

[00:06:39]That's not a hardware upgrade cycle.

[00:06:41]That's a compounding revenue multiplier baked into every infrastructure investment the hyperscalers are making, which is exactly why Microsoft, Google, and Amazon keep committing to larger and larger builds. Reason five, one open-source project just made all of this permanent. A few weeks before Jensen took the stage, an open-source project called Open Claw appeared on the internet and became the fastest growing open-source project in the history of software development.

[00:07:09]>> Open Claw is the number one, is the most popular open-source project in the history of humanity, and it did so in just a few weeks.

[00:07:20]>> [applause] >> It exceeded It exceeded what Linux did in 30 years.

[00:07:27]>> What Open Claw actually is is an operating system for artificial intelligence agents. It's the layer that lets agents manage files, call tools, run code, spawn other agents, and operate autonomously inside a company's environment. Jensen compared it to what Windows did for personal computers, except instead of unlocking personal computers, it's unlocking personal agents for every employee at every company.

[00:07:53]And here's the direct connection to Nvidia's revenue outlook. Every company that adopts an agentic workflow needs tokens, not as a one-time project cost, as an ongoing operational expense, permanent, growing, [music] and directly tied to how much those agents are being used. Like electricity, like cloud storage, except the demand curve for tokens is still pointing [music] straight up.

[00:08:16]>> I could totally imagine in the future every single engineer in our company will need an annual token budget.

[00:08:24]They're going to make a few hundred thousand dollars a year their base pay.

[00:08:29]I'm going to give them probably half of that on top of it as tokens.

[00:08:34]>> If that becomes standard practice across the enterprise world, and the adoption curve of Open Claw suggests it's already well underway, then demand for compute doesn't plateau.