원문 보기
Jensen, this is intriguing. I did n't even know you'd be here, but I don't think you'd miss such an event for any reason. This is incredible. This is the epicenter of world trade. Yes. Commerce. Why not? Do you know? Well, let me ask you something. Creating opportunities for the world. I...before I saw you, I went to my trusts and estates lawyer and said, " Listen, if I know I'm going to die in 2030, is there anything I can do to help my children?" You question it .
Will we all die in 2030? Maybe it won't matter. We're not...we're not going to die in 2030. There are so many people in the world who will build AI properly. We will have all kinds of protection, we will invent a bunch of technologies for security. All this is happening all over the world. It's not just two companies developing something.
Thousands of companies are engaged in development. So we will create AI that will be incredibly useful. It will be extremely helpful and incredibly safe. We can get it all at the same time. Because this is an engineering task. This is an engineering task. Testing is an engineering task. We must create products and test them properly.
And if they're not ready for release, just hold them, keep testing and improving them until they are . This is common sense. I... I believe in it. Two years ago, you told me that companies would be able to buy something from you and get four times more than they expected. It was, as always, conservative, right?
You lowered expectations because they're making money now. They just don't like to talk about it . Well, the most incredible thing is that in the last 6 months, AI has become extremely useful . And because it's so useful, people are building these computing systems, Nvidia's "AI factories," like crazy.
Each gigawatt of Nvidia's "AI factories" costs about $50-60 billion . OK. However, the rent for this 1- gigawatt "AI factory" is about $50 billion per year. For a year. Good. So yes, the return on investment is now about 1 year. And they can use it because the flexibility and durability of Nvidia systems are so high that they will use them for many years, much longer than five or six.
Therefore, the return on this invested capital is simply incredible. That's why he's building it. No, it's difficult because the closed models, so to speak, Frontier, they can probably do three times as much. But at the same time, we need both open and closed models, right? You should rent and use the best available models as a product or service, these closed models.
However, every country, every company, every industry must also create its own intelligence. How can you completely outsource your intelligence to someone else? And that's why all these companies are here. This is the reason for our partnership with Salesforce: Salesforce will use Nvidia NeMo Triton and our agent technologies to build its own agents running on Salesforce.
Also, because these agents will be extremely fast and there will be so many of them, the data processing layer in Salesforce sits on top of it—it's in Salesforce, it's in Databricks, it's in Snowflake. We are accelerating data processing in all of this. So you get lightning- fast data processing. You will have many agents to help you complete tasks.
And all this is based on the Nvidia NeMo Triton. So Salesforce is now also an AI company. Well, then it will be the same as with all the other industrial revolutions. In the beginning, everyone said, " Listen, the loom will put a lot of seamstresses out of work." And it was a difficult situation for them.
True, but after 4 years, there were ten times more jobs. You know, this narrative is annoying, unnecessary, and, frankly, just plain wrong. The fact is that as a company becomes more productive and uses artificial intelligence, it becomes more ambitious. Their ambition allows them to imagine and want to achieve more, and that's why, if you look at Salesforce, they were the data layer and the application layer.
Now they are also a level of agents on top of that. They are much more ambitious about the future of Salesforce today than they have ever been when I've seen Mark, and he's always enthusiastic . And today he is simply bursting with ideas. So, AI has helped his company raise its ambitions, and as a result, they will be able to reach greater heights.
This will happen not just at Salesforce, but across all Salesforce customers, across every single industry. It is completely illogical to think that engineering will become unnecessary if we have agents to write software code. I grew up in a generation where engineers didn't have any software. We used all sorts of other tools and created prototypes .
So, engineering existed before software. Engineering will exist even after the advent of AI. no sales, so Yeah, it just doesn't make sense that Yeah, you're watching all this on a popular podcast. You got a call in the middle of it . That was a little strange. Was this planned? No. The president just dialed.
I was on stage and backstage I saw out of the corner of my eye my team waving their phones in the air, and I was trying to figure out what was going on. As they got closer, they said, "Mr. President." Uh, and of course, you can't help but answer this call. Yes, yes. It is wise to accept it. When the president calls you, answer.
In short, he just wanted to talk to me about some tweets he posted. I told him I was just on stage with my friends and he said, “Turn me on. "Let me talk to everyone." So I plugged it in. Here's what happened. OK. Good. That was ...okay. So it was completely unplanned. He used the term " mystification".
A lot of people picked up on this , but the consensus of that panel was that some things are fictional. You would agree with this term, "fictional." Are we...How to stop this? AI safety is a real thing. OK. Uh, designing products to be safe for the world—that's a real thing. Companies should innovate as quickly as possible, but should never do so quickly that they release unsafe products.
This is a real thing. However, the fact that we need new laws, new antitrust laws, or new regulations to allow these companies to do fundamental engineering and do it properly before releasing products is completely unnecessary. And I think that's exactly what the president means—that we don't need new laws.
We don't need...yes, we don't need new ones...we have enough laws. We have many rules that govern the reliability and functionality of products. And, you know, there are a lot of them . Yesterday was a strange day . So your stock drops five and a half percent because everyone says OpenAI won't go public .
So, you will not fulfill the quarterly plan. Blah-blah -blah. When we understand that you are selling at a multiple of 14 to the earnings of the 28th fiscal year, which begins in January, and that you have so many prospects, what is the best investment in the world? Nvidia. Well , let's start the battle.
Let's buy billions. Jim, listen, this is incredible. We are the only platform that launches every closed model. That's right . Absolutely every closed model. And we are increasing our share in the closed model segment. We are the only platform that runs every open model, including Nvidia with our own NeMo models.
Demand for token generation has increased 25 times in less than a year. And the number of public tokens generated increased from 30% to almost 70%. So the demand for AI, because it's useful now, is just growing exponentially. Nothing, nothing that happened this week will slow that down. In fact, what's likely to happen is this : we used to have training systems and output systems, huge systems.
Now we will have a third system. This third system is essentially a testing lab for testing the environment. So these testing systems are going to be extremely large so that we can have products that are functional, capable, safe and secure. And this requires massive data centers for testing them. So, we just created another new demand for AI.
This is brilliant. What about whether we have enough computing power? And when will we force Wall Street financiers to start trading computing power? Well, Jim, this is a big breakthrough in recent years. People now understand that Nvidia computing is the only system that is interchangeable and durable.
The reason we are interchangeable is that we run everything. We run every model, closed model, open models. We run every stage of the AI lifecycle: from data processing to training, to post-training, and, of course, testing. Every company uses us, and as a result, it's now the world's first computing system that can be used as a collateralized asset.
So you can use computers as collateral , and we're seeing more and more evidence of that. If so, this will be a huge revelation for all companies that do not have the ability to finance such a scale, as the asset itself will be collateral. So this is a moment in time that we will all remember as the moment when everyone who hired you and bought your product experienced this tipping point.
And the profitability, especially given the service life of 6, 7, 8, 9 years, will be extraordinary. Can you help people explain this? I'm tired of talking about it . The strange thing is that many of the first tenants signed long-term leases, long-term contracts. And now all these cloud providers are regretting it a little bit, because the spot prices, because of demand, the rents are just incredible right now .
You can now rent Nvidia Grace Blackwell for $16 per hour of GPU work. And a year ago, some of them were rented for $5. So now that these contracts are ending, everyone is just raising their prices; they raise revenue forecasts. The return on investment is increasing; they buy more equipment; they want to create more infrastructure.
And so, I think this flywheel is really, really spinning right now. Well , I'll tell you something. Maybe you've heard of it . Nvidia. Hold it . Don't trade it. Hold, don't trade . That's what I do. Thank you very much, Jensen. Nice to see you, CEO, co-founder and all . Everyone, Jim Cramer . Jensen Huang!
. Tuesday was a "red" day for most of the market as oil prices and bond yields rose. We saw oil confidently cross the $100 per barrel mark and 10-year bonds rise above 5%, all in the lead-up to the FOMC meeting on Wednesday . The Fed's rate decision and SEP report are due to be released on Wednesday at 2:00 PM ET , followed by a press conference at 2:30 PM .
On Tuesday at the AI Infra Summit, Nvidia shared performance results from Vera Rubin and AI Factory, as well as impressive energy efficiency figures. Nvidia reported that in SemiAnalysis' Agent X benchmark for generative inference , Nvidia Vera Rubin and NVL72 achieved up to 30x higher throughput per megawatt than GB200 and NVL72 on DeepC V4 Pro, and up to 45x lower cost per million tokens.
Nvidia noted that with limited power consumption, throughput per megawatt determines how much revenue an AI factory can generate , and the cost per million tokens determines the margin of that revenue. As I have explained many times, Nvidia purposefully lowers the cost of tokens with each new generation architecture, and they benefit greatly from this not only from the sale of hardware that lowers the cost of tokens, but also from the continued growth in usage and demand for tokens that results from lower costs.
Nvidia also announced that Lambda provided the first validation of Nvidia DSX Max LPS on Nvidia Blackwell servers. DSX Max LPS continuously monitors GPU and rack power consumption, dynamically directing available power to where it is needed most. Lambda launched 19 nodes within a power budget typically allocated for 16 full-power nodes, increasing the cluster’s token throughput by 24 % and performance per watt by 23%.
DSX Max LPS dynamically reallocates power between racks based on demand, allowing up to 40% more GPUs to be used within the same facility power consumption and increasing token throughput by 35%. Nvidia also noted that Nvidia Grock 3 LPX adds ultra-low latency deterministic inference to Vera Rubin and complements SXM LP S.
The platform delivers 35 times higher token throughput per megawatt than GB200 NVL72 for models with parameters over 4.2 trillion with long context. Higher token throughput and ultra-low latency inference are extremely important for agent workloads. Additionally, NVIDIA and Salesforce released Co-pilot, the first Salesforce model specifically designed for CRM considerations.
Co-pilot is based on the NVIDIA NeMo Megatron 3 Super, trained on synthetic data derived from Salesforce's approximately 27 years of CRM experience. It is now available to select Agent Force customers, with general availability in the US expected this winter. Salesforce is also implementing NVIDIA models and accelerated computing in Mission Force for government customers.
Meanwhile, KC Rajkumar of Lynx Equity Strategies said his checks did not reveal any slowdown in GPU deployments, cancellations of memory or TSMC orders, changes in supply schedules at hardware companies, or component outages following calls last weekend to slow the development of advanced AI. As I said in Monday's video, I don't expect the events of last weekend to have a .
Even if leading-edge companies slow down the release of new versions, there are still usage growth, agent workloads, corporate implementations, cybersecurity, etc. to consider. Today, there is still not enough computing power to meet the current demand for modern advanced models. Furthermore, we must understand that there is a difference between words and actions.
Since Dario's essay last weekend, the only concrete change so far seems to be that Anthropic and OpenAI have committed to bringing in third-party evaluators with employee-like access to the systems. Everything else at this stage is just speculation and suggestions. Furthermore, regarding the costs of AI infrastructure, I believe that companies' increased focus on security could lead to significant computing resources being directed towards model evaluation, monitoring, and tuning.
If so, this could be a potentially large new source of computing demand that many market participants are not even considering yet. Furthermore , we don't know what leading companies will actually do, as opposed to what they say. Words and actions are two very different things. And, as I said in the video on Monday night, pay attention to what they do.
There are no signs of a real slowdown yet. And even if there is a slowdown in the development of advanced models, this does not automatically mean a slowdown in overall spending on AI infrastructure. The most likely scenario is that the additional computing power that will emerge will be directed towards data inference workloads rather than training.
And again, that's if the companies developing the advanced models actually deliver on what they propose. Also on Tuesday, we got a new GTC announcement from Nvidia. Jensen Huang is scheduled to speak at GTC in Washington, D.C., on December 1st. So mark it on your calendars. And before that, we also have GTC in Berlin scheduled with Jensen performing on October 21st.
On Tuesday, Micron announced the demonstration of the world's first 512GB DDR5 RDIMM module. The module has been demonstrated on several server platforms, and Micron claims that it consumes 60% less power than 4 modules of 128 GB RDIMMs. Serial production is planned for the second half of 2027. This is regular server DRAM memory, not HBM, but it's clearly aimed at AI rendering, in-memory databases, and other resource-intensive data center tasks.
In other news: Micron's Taiwanese workers' union said Tuesday it may resort to a strike if Micron does not agree to a permanent profit-sharing system. The union reiterated its demand that Micron consistently allocate 15% of operating profits to employees . The union leader reportedly said: "If there is no concrete offer on September 18 and 21, and the workers feel that the company is simply dragging out the process, we will declare the negotiations a failure and proceed to a vote for a strike." Taiwan is Micron's main manufacturing hub, so it's important to monitor this situation.
Micron's earnings report is expected on September 30, but judging by comments from the union president, the strike vote could take place before the report is released. So keep that in mind if you plan to trade Micron shares during the reporting period . If a strike is called, it will likely result in short-term negative stock movement.
Nevertheless, I expect this situation to be resolved eventually. Looking ahead, we expect Micron to report on September 30th after the market closes. And then Jensen is scheduled to perform again at GTC in our long ago , he quite seriously suggested that Nvidia could become likely that Nvidia's market capitalization will what Jensen is hinting at when he talks about the possibility of turning Nvidia into a $3 trillion revenue company.
Apparently, the question is what multiplier the market will be willing to provide Nvidia at that moment. much what computing resource capacity they can find. Today, there are no GPUs sitting idle due to lack of demand, as was the case with fiber optics during the peak of the dot-com boom. Back then, companies laid out networks, hoping that use cases and demand would emerge over time.
Today we observe a completely opposite situation. oom, they ignore the fact that this time the development of new ways to use it at meeting demand. New capacities begin to be used immediately after their launch. Hyperscalers monetize capacity as soon as it becomes available. Each of the hyperscalers mentioned supply constraints during are expanding their infrastructure to meet already builoom.
2026 will be a turning point compani corporation inference. tipping point in the demand for inference. That is why we are seeing a surge in revenues for leading AI labs. I wish Anthropic and OpenAI were public companies so everyone could see their revenue growth. I think that the rapid growth of revenues of leading laboratories may be the first evidence that will attract the attention of the market and make it understand that investments in AI infrastructure have a clear payback.
I think the growth in revenue from industry leaders will also help reassure investors about the sustainability of Nvidia's growth, as these revenueat there is not enough shortage of computing resources until at least the first half of 2028, and possibly longer. So, regardless of short-term events, , remain forward-looking, computing demand just the company.
has taken a position that will allow it also sell hardware for real-time local inference via world even without a connection to a data center. Note that NVIDIA takes a holistic approach to physical AI, becoming solutions based on the NVIDIA . As for production, Blackwell Ultra has already been launched and demand for it remains high.
Vera Rubin begins shipping to customers . NVIDIA Grock 3 LPX is now in full-scale production. Subsequently, we expect the release of Rubin Ultra in 2027, and after that, Feynman in 2028 . We have a clear product roadmap for data centers through 2028, and Jensen believes annual spending on AI infrastructure will reach $3-4 trillion close to anything resembling a bubble bursting event.
With that in mind, growth opportunity ahead. And I think that in the future this company will be worth much mor closing: . Sometimes the same montageat you need .
Comments 0