share_log

英伟达电话会全文来了!Blackwell产能爬坡顺利,Q4收入将超预期,Scaling Law没放缓

nvidia conference call full text is here! Blackwell's production capacity is ramping up smoothly, Q4 revenue will exceed expectations, Scaling Law has not slowed down.

wallstreetcn ·  Nov 21, 2024 04:28

During the conference call regarding nvidia's financial report, it was stated that at the initial launch of Blackwell, the gross margin will drop to a low of 70%, specifically between 71-72.5%; revenue for Blackwell in the fourth quarter is expected to exceed previous estimates by several billion dollars, and demand for Hopper will continue into next year, at least for the first few quarters.

AI chip giant Nvidia's revenue growth continued to slow in the third quarter, and is expected to further slow down this quarter. Although last quarter's revenue exceeded analysts' expectations, the mid-range of this quarter's guidance fell short of the high-end expectations of buy-side analysts. From the stock price performance, investors seem to think that such guidance is not explosive enough, causing Nvidia's stock price to fall after hours.

In the conference call after the financial report announcement, Nvidia released its expected guidance. In the fourth quarter, Nvidia Hopper demand will continue, and Blackwell will achieve initial growth, with revenue expected to exceed the previous estimate by tens of billions of dollars. With the expansion of Blackwell product scale, it is expected that the gross margin will slow down to below 70%.

During the Q&A session, analysts focused their 'fire' on Blackwell, raising a series of questions about the progress of this chip. Will potential customers drive greater demand for Blackwell? Are there supply limitations for Blackwell? Is the fourth quarter the most challenging period for Blackwell in terms of gross margin pressure?

Here are the key points of this conference call:

1. Gross margin situation: Due to the launch of Blackwell this quarter, increased costs will lead to a reduction in gross margin. At the initial stage of chip launch, the gross margin will drop to a low point of 70%, specifically 71-72.5%. By the second half of the 2025 fiscal year, it will reach a mid-point of over 70%, around 75%.

2. Blackwell demand situation: Blackwell is planned to start shipping this quarter, accelerating pace in the coming year, with demand expected to exceed supply by the 2026 fiscal year. The continuous increase in inferred demand will drive continuous growth in chip demand. CEO Huang Renxun stated that Blackwell's delivery volume in the next quarter will exceed the company's previous expectations.

3. Blackwell roadmap and supply constraints: Nvidia will continue to execute the roadmap proposed at GTC, launching Ultra next year and transitioning to Rubin in 2026. Nvidia's execution work is progressing well, with a vast supply chain network including Taiwan Semiconductor, Amphenol, Vertiv, SK Hynix, Micron, Achronix, KYEC, Foxconn, Quanta, Wistron, Dell, HP, Supermicro, Lenovo, among others. Progress on Blackwell's production capacity ramp-up is good.

AI demand will continue to grow long-term, projected to reach trillions of dollars for global data center spending by 2030. The first point is to modernize data centers from coding to machine learning. The second point is generative AI, establishing AI factories, creating a new industry, and a new niche market that has never existed in the world before.

Hopper demand growth will continue, at least through next year, likely exceeding the shipments of this quarter next quarter.

Scaling Law has not slowed down: There are now three training methods, pre-training will continue, this is an empirical law not a physical one. In addition, there is post-training and inference scaling law. The industry is evolving in the crucial aspects of pre-training, post-training, and the increasingly important inference time.

Below is the full text of the NVIDIA conference call

Colette Kress, Chief Financial Officer:

It's another record quarter in Q3, continuing our incredible growth. Revenue reached $35.1 billion, up 17% quarter-over-quarter, 94% year-over-year, well above our expected $32.5 billion. Driven by NVIDIA's accelerated computing and AI, all market platforms achieved strong growth both sequentially and year-over-year.

Starting from the data center, the data center achieved another record high. Revenue reached $30.8 billion, up 17% quarter-over-quarter, 112% year-over-year. NVIDIA Hopper's demand was exceptionally strong, with H200 sales growing significantly, reaching sales in the billions, making it the fastest growing product in the company's history. H200's inference performance has doubled, and TCL has increased by about 50%. Cloud service providers account for about half of our data center revenue, with revenue more than doubling year-over-year. CSPs have deployed NVIDIA H200 infrastructure and high-speed networking, with tens of thousands of GPU installations to drive business growth, meeting the rapid increase in demand for AI training and inference workloads. AWS and Microsoft Azure now offer NVIDIA H200-driven cloud instances, while Google Cloud and OCI are also due to be launched. GPU cloud revenue in North America, Europe, the Middle East, Africa, and Asia-Pacific has also doubled year-over-year due to the increasing deployment of NVIDIA cloud scenarios and sovereign clouds.

Consumer internet revenue more than doubled year-over-year, enterprises are expanding the scale of NVIDIA Hopper infrastructure to support next-generation AI models, training, multimodal and agent AI, deep learning recommendation engines, as well as generative AI inference and content creation workloads. NVIDIA's Ampere and Hopper architectures are driving customer's inference revenue growth. NVIDIA is the world's largest inference platform, with substantial installations and a rich software ecosystem encouraging developers to optimize for NVIDIA, continually improving performance and TCL. The rapid development of NVIDIA's software algorithms has increased Hopper's inference throughput by an astonishing 5 times and reduced time by 5 times within a year.

The upcoming NVIDIA NIM will further improve the performance of Hopper Inference by 2.4 times. Continuous performance optimization is a major feature of Nvidia, bringing more and more economic returns to the entire company's installed user base. After successfully implementing large-scale changes, Blackwell has been fully deployed for production. We delivered 13,000 GPU samples to customers in the third quarter, including the first batch of open artificial intelligence Blackwell DGX engineering samples. Blackwell is a full-stack, fully infrastructure artificial intelligence datacenter-scale system with customizable configurations to meet the diverse and growing demands of the artificial intelligence market, ranging from X86 to ARM, training to enhanced GPU, InfiniBand to Ethernet switches, and liquid cooling to air cooling, etc.

Every customer is rushing into the market, and Blackwell is now in the hands of all major partners who are working hard to enhance their data centers. We are integrating Blackwell systems into various data center configurations for customers. The demand for Blackwell is astonishing, and we are rapidly expanding supply to meet the huge demand from customers. Customers are preparing for large-scale Blackwell deployments. Oracle has announced the launch of the world's first artificial intelligence cloud computing cluster, scalable to over 131,000 Blackwell GPUs, to help enterprises train and deploy some of the most demanding next-generation artificial intelligence models. Yesterday, Microsoft announced that they will be the first to offer CSP-based Blackwell cloud instances in private preview featuring Nvidia GB200 and Quantum InfiniBand. Last week, Blackwell made its debut in the latest round of training results, winning in GPU benchmark tests, with performance 2.2 times that of Hopper.

These results also demonstrate our relentless pursuit of reducing computing costs. Compared to 256 H100s, running GPT-3 benchmarks only requires 64 Blackwell GPUs, reducing costs by four times. Nvidia's Blackwell architecture switch can increase inference performance by 30 times, elevate inference throughput and response time to a new level, ideal for running new inference applications like OpenAI's o1 model. Each new platform transition spawns a wave of new startups. Hundreds of native artificial intelligence companies have already achieved great success in providing AI services. Leaders like Anthropic, Perplexity, Mistral, Adobe Firefly, Runway, MidJourney, Lightrix, Harvey, Codium, Cursor, as well as Google, Meta, Microsoft, and OpenAI, are paving the way, with thousands of native AI startups building new services.

The next wave of artificial intelligence is enterprise AI and industrial AI, with enterprise AI in full swing. Nvidia's AI Enterprise Edition includes NeMo and NIM microservices, serving as an agent artificial intelligence operation platform. Industry leaders are using Nvidia AI to build collaborative driving and agents. Companies partnering with Nvidia like Cadence, Cohesity, NetApp, Nutanix, Salesforce, SAP, and ServiceNow are accelerating the development pace of these applications, possibly deploying billions of agents internally in the next few years. Consulting industry leaders like Accenture and Deloitte are introducing Nvidia AI to global enterprises. Accenture has established a new business unit with 30,000 professionals trained in Nvidia AI technology to help drive this global initiative. Additionally, Accenture, with over 0.77 million employees, is internally leveraging Nvidia-supported Agentic AI applications, with one case reducing manual steps in marketing activities by 25% to 35%.

Nearly 1000 companies are using NVIDIA NIM, highlighting the monetization speed of NVIDIA AI Enterprise. We expect the annual revenue of AI Enterprise Edition to grow by more than 2 times compared to last year, and our sales channels are still expanding. Overall, the annual growth of our software, services, and support revenue is $1.5 billion, with a projected increase of over $2 billion this year, as industrial AI and robotics technology accelerates.

This is spurred by the breakthroughs in physics-based artificial intelligence, the fundamental models capable of understanding the physical world. Similar to NVIDIA NeMo for enterprise AI agents, we have built NVIDIA Omniverse for developers to build, train, and operate industrial artificial intelligence and robotics. Some of the world's largest industrial manufacturers are adopting NVIDIA Omniverse to accelerate business development, automate workflows, and elevate operational efficiency to new levels. The world's largest electronics manufacturer, Foxconn, is utilizing NVIDIA Omniverse-based digital twins and industrial artificial intelligence to speed up the construction of its Blackwells factory and increase efficiency to new levels. In the Mexican factory alone, Foxconn expects annual kilowatt-hours usage to decrease by 330%. Looking regionally, our data center revenue in China has seen continuous growth, thanks to industry shipments of copper products meeting export standards, although the proportion in total data center revenue is still far below pre-export control levels. We anticipate intense competition in the future Chinese market. We will continue to comply with regulations while providing service to our customers.

As countries adopt NVIDIA's accelerated computing technologies to drive a new round of AI-driven industrial revolution, our Sovereign Artificial Intelligence (AI) initiative continues to build momentum. Leading Indian CSPs, including Product Communications and Zoda Data Services, are building artificial intelligence factories with tens of thousands of Nvidia GPUs. By the end of the year, they will increase the deployment of Nvidia GPUs in India by nearly 10 times. TCS and Wipro are adopting NVIDIA AI Enterprise, training nearly 0.5 million developers and consultants to help customers build and run AI agents on our platform. In Japan, SoftBank is using Nvidia DGX Blackwell and Quantum InfiniBand to build the country's most powerful AI supercomputer. SoftBank is also partnering with Nvidia to transform telecommunications networks into distributed AI networks with Nvidia AI Aerial and ARAN platforms, which can process 5G RAN and AI on CUDA.

We are working with T-Mobile in the USA to launch the same products. Leading Japanese companies such as NEC and NTT are adopting the NVIDIA AI Enterprise, while large consulting companies including EY, Strategy, and Consulting will help introduce NVIDIA AI technology to various industries in Japan. Network business revenue increased by 20% year-on-year. Areas of continuous revenue growth include InfiniBand and Ethernet switches, SmartNex and BlueField DPU. Although network revenue has been continuously declining, network demand is strong and continuously growing. We expect network revenue to continue to grow in the fourth quarter.

CSPs and supercomputing centers are using and adopting the NVIDIA InfiniBand platform to power the new H200 cluster. Revenue from NVIDIA Spectrum-X Ethernet for AI has more than tripled year-on-year, and our product line continues to expand. Several CSPs and consumer internet companies are planning to deploy large clusters. Traditional Ethernet was not designed for artificial intelligence. NVIDIA Spectrum-X leverages previously InfiniBand-exclusive technology to enable customers to achieve large-scale GPU computing. With Spectrum-X, the Colossus 100,000 Hopper supercomputer from xAI will achieve zero application latency and maintain 95% data pass-through output, while traditional Ethernet only achieves 60%.

Now let's talk about gaming and AIPC. Gaming revenue is $3.3 billion, an increase of 14% month-on-month and 15% year-on-year. The third quarter was a quarter of great harvest for the gaming business, with revenue growth in laptop, gaming console, and desktop categories both month-on-month and year-on-year. The strong momentum in back-to-school sales is driving the end-demand for RTX as consumers continue to choose GeForce RTX GPUs and devices to support gaming, creativity, and AI applications. Channel inventory remains healthy, and we are preparing for the holiday season. We are starting to ship new GeForce RTX AI PCs from ASUS and MSI, with the ability to accommodate up to 321 AI vertices, with an expected inclusion of Microsoft's Copilot+ feature in the fourth quarter.

These machines leverage the powerful capabilities of RTX ray tracing and AI technology to provide superpowers for gaming, photo and video editing, image generation, and encoding. Last quarter, we celebrated the 25th anniversary of the world's first GPU, the GeForce 256. NVIDIA GPUs have been a driving force behind some of the most influential technologies of our time, from transforming graphics computing to sparking the AI revolution. Looking at ProViz, revenue is $0.486 billion, an increase of 7% month-on-month and 17% year-on-year. NVIDIA RTX workstations continue to be the preferred choice for professional graphics, design, and engineering workloads.

Moreover, AI is becoming a powerful demand driver, including autonomous vehicle simulations, generative AI model prototypes for productivity use cases, and generative AI content creation in media and entertainment. Automotive industry revenue reached a record $0.449 billion, with a 30% month-on-month increase and 72% year-on-year growth. The strong growth is mainly attributed to the advancements in AI Orin autonomous driving technology and the strong demand for NAVs in the end-market.

Automakers are launching all-electric SUVs based on NVIDIA Orin and DriveOS. Now let's look at other parts of the income statement. GAAP gross margin is 74.6%, non-GAAP gross margin is 75%, with a decrease mainly due to the transfer of H100 systems in data centers to more complex and higher-cost systems. GAAP and non-GAAP operating expenses increased by 9% due to higher costs for computing, infrastructure, and engineering development required for launching new products. In the third quarter, we returned $11.2 billion to shareholders in the form of share buybacks and cash dividends.

Now, let me talk about the outlook for the fourth quarter. Total revenue is expected to be $37.5 billion, with a growth range of plus or minus 2%, including continued demand for the Hopper Architecture and initial growth of Blackwell products. Although demand far exceeds supply, as our understanding of the supply situation increases, we are poised to exceed Blackwell's previous forecasted billion-dollar revenue. In the gaming sector, despite strong sales in the third quarter, we expect revenue to decline in the fourth quarter due to supply constraints. Gross margin is expected to be 73% and 73.5% under Generally Accepted Accounting Principles (GAAP) and non-GAAP, respectively, with a range of plus or minus 50 basis points. Blackwell is a customizable AI infrastructure with various NVIDIA in-house chips, multiple network task options, suitable for both air-cooled and liquid-cooled data centers. Our current focus is on meeting strong demand, improving system availability, and providing customers with the best configuration options.

With the expansion of Blackwell scale, we expect the gross margin to slow down to below 70%, with Blackwell's gross margin around 70%. Operating expenses under Generally Accepted Accounting Principles (GAAP) and non-GAAP are expected to be approximately $4.8 billion and $3.4 billion, respectively. We are a data center-scale AI infrastructure company, and our investments include building data centers, developing our hardware and software stack, and supporting the launch of new products. Excluding non-related investment income and losses, other income and expenses under Generally Accepted Accounting Principles (GAAP) and non-GAAP are expected to be around $0.4 billion. The tax rate under GAAP and non-GAAP is expected to be 16.5% plus or minus 1%, excluding any discrete items.

For more financial details, please refer to the Chief Financial Officer's comments and other information on our investor relations website. Finally, let me highlight the upcoming financial industry events. We will be attending the UBS Global Technology and AI Conference in Scottsdale on December 3. Jensen will deliver a keynote address at CES in Las Vegas on January 6, and the next day (January 7) we will host a Q&A session for financial analysts. We will hold an earnings conference call on February 26, 2025, to discuss the performance of the fourth quarter of the 2025 fiscal year. We are now ready to take questions.

Q&A Session

Q1, Cantor Fitzgerald analyst C.J. Muse: I just want to ask you one question, it's about the scaling debate of large language models. Obviously, it may still be early, but I'd love to hear your thoughts on this. How are you helping clients address these issues? Clearly, we're talking about clusters that have not yet benefited from Blackwell, which could drive greater demand for Blackwell. Thank you.

Responder:

The base model before training is intact and ongoing. As we all know, this is an empirical rule, not a fundamental physical law, but there is evidence to suggest that training is still ongoing. However, what we are learning is that this is not enough, and we have now discovered two other training methods, one is post-training, and the other is post-training with reinforcement learning human feedback. We now have reinforcement learning artificial intelligence feedback and various forms of synthetic data generation to aid in expanding post-training.

One of the major and most exciting developments is the raspberry, namely OpenAI's o1 model, which performs reasoning time expansion, the so-called reasoning time. The longer it thinks, the better and higher quality the answers are, considering methods such as thinking chains, multi-path planning, and various necessary reflection technologies. In a sense, this is somewhat like thinking in our minds before answering a question.

Therefore, we now have three training methods, and we have seen all three. Hence, there is a huge demand for our infrastructure. You can now see that the capacity of the previous generation base models was about 0.1 million Hopper chips, while the next generation starts from 0.1 million Blackwell chips. You can see the industry's development direction in pre-training, post-training, and the now crucial reasoning time expansion.

So, for all these reasons, there is indeed a high demand. But please remember, at the same time, our company's reasoning capacity is also constantly strengthening. We are the largest reasoning platform in the world today because our installed capacity is very large, and all reasoning is trained on and with jumpers, which is incredible.

As we switch to using Blackwells to train basic models, it also brings a large amount of installed infrastructure for inference. As a result, we see a continuous increase in demand for inference. We see the inference time continuously lengthening. We see a continuous growth in the number of native AI companies. Of course, we also begin to see enterprises adopting Agentic AI, which is indeed the recent trend. We see a large demand from different places.

Q2, Goldman Sachs' Toshiya Hari: Jensen, earlier this year you implemented a large-scale reform, and there were some reports of heating issues last weekend. In this context, do you have the ability to execute the roadmap presented at GTC this year, launching Ultra next year and transitioning to Rubin in 2026? Can you talk about this issue? Some investors have doubts about this. It would be very helpful if you could talk about your ability to execute on time. Then, regarding the Blackwell part, concerning supply constraints, is it various components causing this situation? Or specifically HBM? Is it supply constraints? Have the supply constraints improved? Are they deteriorating? Any information on this would be very useful. Thank you.

Responder:

Yes, thank you, so let's take a look back at the first question. Blackwell's production is now in full swing. In fact, as Colette mentioned earlier, this quarter we will deliver more Blackwells than previously expected. Therefore, the supply chain team is working excellently with our supply partners to increase the output of Blackwells. Next year, we will continue to strive to increase the output of Blackwells.

The current situation is one of supply not meeting demand, which is expected since, as we all know, we are in the early stages of the artificial intelligence revolution. We are in the early stages of the next generation of basic models that can infer, that can think long-term, and of course, one of the truly exciting areas is physical artificial intelligence, meaning AI that can now understand the structure of the physical world.

Therefore, the demand for Blackwell is very strong. Our execution work is progressing smoothly. Obviously, we are conducting a large amount of engineering design globally. What you see now are systems being installed by Dell and Core Weave. I think you also see Oracle's systems being put into use. And Microsoft's systems, they are about to preview the Grace Blackwell system.

Google has its own system, and all these CSPs are in a race. As you know, the engineering design we cooperate with them on is quite complex. The reason for this is that even though we have built a full stack and full infrastructure, we break down all artificial supercomputers and integrate them into all custom data centers around the world.

We have been doing this integration process for several generations, we are very good at it, but there is still a lot of engineering work to be done. But as you can see from all the systems being installed, the situation with Blackwell is very good. As mentioned earlier, the supply volume and planned shipment volume for this quarter exceeded our previous expectations.

About the supply chain, there are seven different kinds of chips, seven custom chips, we have built-in ordering, so we can provide the Blackwell system. The Blackwell system adopts air-cooled or liquid-cooled, NVLink 8 or NVLink 72 or NVLink 8, NVLink 36, NVLink 72, we have x86 or GRACE, integrating all these systems into global data centers is considered a miracle.

Therefore, to achieve such scale of growth, the component supply chain is essential. You know, you have to look back at how much our Blackwell shipments were in the last quarter, which was zero. While this quarter, the total shipment of Blackwell systems will be in billions, with an incredible growth rate. Almost all companies in the world are involved in our supply chain, we have very good partners, including Taiwan Semiconductor and connector companies Amphenol, Vertiv, SK Hynix, Micron, Amkor, KYEC, Foxconn, Quanta, WeWin, Dell, HP, Supermicro, Lenovo, etc., the number of companies involved is incredible.

I am sure I have also missed partners involved in the expansion of Blackwell production, for which I am very grateful. Overall, I think the current situation with capacity ramp-up is very good.

Finally, regarding the issue of our execution roadmap, we are executing the annual roadmap, and we hope to continue doing so. Of course, by doing this, we can improve the platform's performance, but equally important is that as we are able to improve performance and increase by a factor of X, we reduce training costs, reduce inference costs, reduce the costs of artificial intelligence, making artificial intelligence more accessible.

However, another very important factor is that when data centers reach a certain fixed scale, it may have been 10 megawatts in the past, but now most data centers are 100 megawatts to several hundred megawatts data centers, and gigawatt data centers are planned. The size of a data center is not actually important, power is limited. In data centers with limited power, the highest performance per watt will directly translate into the highest revenue for our partners. Therefore, on the one hand, our annual roadmap reduces costs. But on the other hand, because our performance per watt is better than any other product, we can create as much revenue for customers as possible. Therefore, this annual rhythm is very important for us, and we are fully committed to continuing to do so. As far as I know, everything is going according to plan.

Q3, UBS's Timothy Arcuri: I would like to know if you can talk about the current trajectory of Blackwall. It was indeed mentioned that Blackwall is doing better than the tens of billions of dollars mentioned in January, it sounds like the company has more to do. But I believe in recent months, you have also mentioned that Blackwall will surpass Hopper in the fourth quarter. So I have two questions, first, will Blackwall surpass Hopper in the fourth quarter, is that correct? Then, Colette, you mentioned that the gross margin of Blackwall will drop to below 70%. So I wonder, if the fourth quarter is the crossover period, is that the time when the gross margin pressure is the most severe? The gross margin will drop to below 70% in April. I just want to know if you can describe that for us. Thank you.

Colette Kress, Chief Financial Officer:

Of course, let me first answer your question, thank you for your question about our gross margin. We have discussed our gross margin because right from the start, we have been enhancing the performance of Blackwall, and we will be introducing many different configurations, many different chips to the market, we will focus on ensuring that our customers get the best experience at installation.

Our gross margin will start to increase, but we believe that in the first phase of growth, the gross margin will be around 70%. So, you're right, in the next few quarters, we will start to increase the gross margin, and we hope to quickly reach 70%.

Jensen Huang, Chief Executive Officer:

Hopper's demand will continue until next year, definitely in the first few quarters of next year. Meanwhile, the shipment volume next quarter will exceed this quarter. Our Blackwells shipment volume next quarter will exceed the first quarter. So, from this perspective, we are at the beginning of two significant fundamental changes in the computing field. The first transformation is from encoding on CPUs to creating neural network machine learning on GPUs.

Currently, the fundamental shift from encoding to machine learning is very common. No companies are not engaged in machine learning. Therefore, machine learning is also the basis for generative AI. One the one hand, the first thing happening is the modernization of $1 trillion worth of computing systems and datacenters worldwide for machine learning. On the other hand, I believe that based on these systems, we will create a new type of AI capability.

When we talk about "generative AI," we are essentially saying that these datacenters are actually AI factories. They are producing something. Just like we produce electrical power, we are now also producing AI. If we have a large number of customers, like a large number of electricity consumers, these generators will be running around the clock. Today, many AI services are running 24/7, just like AI factories. Therefore, we will see the launch of this new type of system, which I call AI factory, because it is very close to its essence. It is different from traditional datacenters. So, these two basic trends are just beginning. Therefore, we expect this growth - this modernization and creation of new industries to continue for several years.

Q4, UBS Group Analyst Vivek Arya: Colete, I'd like to ask, do you think that NVIDIA's gross margin in the second half of '25 can return to around 70%? Is this assumption reasonable? Then, my main question is, historically, when we look at hardware deployment cycles, there will inevitably be some digestion process. When do you think we will enter this phase, or is it too early to discuss this now because Blackwells has just started? How many quarters of shipment volume do you think will meet the first wave of demand? Can the growth continue until '26? How should we prepare for the long-term, secular digestion period for hardware deployment, right?

Colete Kres, Chief Financial Officer:

Let me clarify your question about the gross margin. Can our gross margin reach 70% in the second half of next year? I think this is a reasonable assumption and a goal we want to achieve, but we have to see how the climb mix goes, but it is definitely possible.

Jensen Huang, CEO:

I believe that until we modernize the $1 trillion datacenter, it is not digestible. Therefore, looking at datacenters globally, the vast majority were built in the era of manually coding applications and running on CPUs. It is now unreasonable to do so. If every company's capital expenditure, if they are preparing to build a datacenter tomorrow, they should build it for future machine learning and generative AI, right?

Because they have many outdated datacenters. So what will happen in the next X years? Let's assume that in 4 years, as we evolve in the IT field, datacenters worldwide will all be modernized. Everyone knows that IT's annual growth rate is about 20-30%. For example, by 2030, the global datacenters dedicated to computing will reach trillions of dollars. We must adapt to this growth. From coding to machine learning, we must achieve modernization of datacenters. That's the first point. The second part is generative AI. We are now creating a new capability in the world that has never existed before, a new sub-market in the world that has never existed before.

If you look at OpenAI, it doesn't replace anything. It's something completely new. In many ways, it's like when the iPhone was born, it's completely new. It doesn't replace anything. So, we will see more and more companies like this. They will create and generate inherently intelligent things from their services, some of which will be digital artist intelligence.

Some are basic intelligence, like OpenAI. Some are legal intelligence, like Harvey Digital Marketing Intelligence, like writers, and so on. These companies number in the hundreds and are called AI-native companies, almost every platform is changing, otherwise, as you remember, there were internet companies, cloud-first companies, mobile-first companies, now they are AI-native companies. They were mobile-first companies, and now they are AI-native companies. These companies are born because people see the shift in platforms, see the new opportunities to do new things.

Therefore, my feeling is that we will continue to build modern information technology, first modernized computing, and then create AI factories, shaping new industries for AI production.

Q5, Bernstein Research Analyst Stacy Rasgon: To clarify, when you say the gross margin drops to 70%, is 73.5% considered dropping to 70%? Do you have any other thoughts? As for my question, you are guiding total revenue, so what I mean is, the total revenue of the next quarter in the datacenter will definitely increase by several billion dollars, but it sounds like Blackwells' growth is more than that now. But you also mentioned that Hopper is still strong. Will Hopper decline continuously in the next quarter? If so, why? Is it because of supply constraints? The Chinese market has always been strong, but after entering the fourth quarter, the Chinese market will decline. So, it would be very helpful if you could tell us about Blackwells' growth and the performance of Blackwells and Hopper in the fourth quarter. Thank you.

Colette Kress, Chief Financial Officer:

Starting with your first question, regarding our gross margin and the defined low point. Of course, our gross margin may be below the median, for example, our gross margin may be 71%, or even 72%, 72.5%, we will be within this range. We may also exceed this. We need to see the results. We do hope that in the remaining time of this year, we can ensure that our output and products continue to increase and improve, and by then we will reach around 75%.

The second question is about our Hopper, the H200 has seen significant growth, not only in terms of orders, but also in the speed of project approvals. This is an amazing product and one of the fastest-growing products we have seen.

We will continue to sell the Hopper in this quarter and the fourth quarter. This includes all our different configurations, which involve measures we might take in China. But please remember, people are also looking to establish their Blackwell at the same time.

Therefore, we may see both situations in the fourth quarter. It is possible that the Hopper could experience growth between the third and fourth quarters, but we can only wait and see.

Q6, Morgan Stanley Joseph Moore: I would like you to talk about what you see in the inference market. You have already mentioned the impact of Strawberries and longer extension inference projects. But you also mentioned that as some clusters of Hopper age, you may use some potential Hopper chips for inference. Do you expect inference to surpass training in the next 12 months?

Jensen Huang, Chief Executive Officer:

Our hope and dream is that one day, there will be a massive amount of inference worldwide. By then, artificial intelligence will truly be successful, right? By then, every company's marketing department, forecasting department, supply chain department, legal department, engineering department, and of course, coding department, will be doing inference internally in the company. We hope that every company will be able to do inference 24/7, with a large number of native AI startups and thousands of native AI startups generating tokens and artificial intelligence in every aspect of your computer experience, from using Outlook to making PowerPoint, or sitting there using Excel, all constantly generating tokens.

Every time you read a PDF, open a PDF, it generates a lot of tokens. One of my favorite apps is NotebookLM, an app launched by Google. I enjoy using it so much, just because it's fun, you know. I put in every PDF, every archive file just to listen and scan. So I think our goal is to train these models for people to use. Now, AI is entering a whole new era, a new branch of AI called physical AI, these large language models can understand human language and our thought processes. Physical AI understands the physical world, understands the meaning of structure, understands what is reasonable, what is unreasonable, what might happen, what won't happen, it not only understands but also predicts and deduces a brief future.

This capability is incredibly valuable for industrial artificial intelligence and robotics technology. Therefore, many native artificial intelligence companies, robotics companies, and physical artificial intelligence companies have become popular because of this, and you may have heard of these companies. This is also the true reason why we are building Omniverse. Omniverse allows these artificial intelligence to create and learn in Omniverse, and learn from synthetic data generation and reinforcement learning of physical feedback, instead of learning from human feedback.

To have these capabilities, we created Omniverse to achieve physical artificial intelligence. Therefore, our goal is to generate tokens. Our goal is reasoning, and we have already begun to see this growth. So I am very excited about this. Now let me say one more thing. Reasoning is super difficult. The reason why reasoning is super difficult is because on the one hand you need high accuracy, and on the other hand, you need high throughput. You need high throughput so that costs can be as low as possible, but you also need low latency. Computers that have both high throughput and low latency are very difficult to manufacture. These applications have longer context lengths because they want to understand, they want to be able to reason with an understanding of the context. Therefore, the context length is increasing.

On the other hand, the models are getting larger and more multimodal. It is incredible how many dimensions of innovation are involved in reasoning. This innovation speed is the greatness of the Nvidia architecture, because our ecosystem is outstanding. Everyone knows that if they innovate on Nvidia's CUDA architecture, they can innovate faster, and they know that everything should be feasible.

If something happens, it is most likely because of their code and not ours. Therefore, we have the ability to innovate in all directions at the same time, with a huge installed base, so that whatever you create can be implemented on Nvidia computers and widely deployed to various data centers around the world, all the way to the edge of robot systems. This capability is truly amazing.

Q7, Wells Fargo analyst Aaron Rakers: When we look at the cycle of the data center business, I would like to ask you a question. When I saw the performance of the last quarter, Collette, you mentioned that the networking business had clearly declined by 15% consecutively, but your comment was that you saw very strong demand.

You also mentioned that in these large cluster areas, you have won multiple cloud CSP designs. So I would like to know if you can explain the development of the networking business, and in which areas you see some constraints, and whether you are confident in the development speed of Spectrum-X to reach the scale of billions of dollars you mentioned before.

Colette Kress, Chief Financial Officer:

Starting with the network first, the year-on-year growth is huge. Since the acquisition of Mellanox, our focus has been on integrating the work we have done in the data center. Networking is a crucial part of it. The ability to sell our network together with many systems we have done in the data center is continuously growing, and performing quite well.

As a result, the performance of this quarter has only slightly declined, and our growth momentum will soon recover. They are preparing for Blackwells and more and more systems, which will not only use our existing networks, but also the networks we provide to these large systems.

Q8, Citi Bank Atif Malik: I have two questions for Colete. Colete, in the last earnings conference call, you mentioned that sovereign demand is in the low double digits. Can you provide the latest update on this? Then, can you explain the situation of tight gaming supply? Is this because you are redirecting supply to data centers?

Colete Cress, Chief Financial Officer:

Starting with sovereign artificial intelligence, it is an important part of growth, with the emergence of generative artificial intelligence and the establishment of models in various countries around the world, this has truly come to fruition. We have seen many such companies, and we also discussed many of these companies and their work on today's conference call. Therefore, our 'Sovereign AI' and our future pipeline remain absolutely intact, as these people are working hard to build these foundational models in their own language and culture, and are working within companies in these countries. I believe this will continue to be a growth opportunity, and you may see our regional clouds storing and/or many parts of artificial intelligence factories focusing on sovereign artificial intelligence.

Growth is not only in Europe, but also in the Asia-Pacific region. Let me address the second question you raised about gaming. In terms of supply, we are now busy ensuring that all our different products can be produced smoothly. In this situation, our game console supply is progressing quite rapidly from the sales we see. The challenge we face now is how to quickly supply these products to the market this quarter. Don't worry, I think in the new year, we will have more suppliers back on track. It's just that the supply in this quarter will be relatively tight.

Q9, Melius Research analyst Ben Reitzes: The continuous strong growth this quarter, with your revenue guidance of around 7%. Does your comments on Blackwells imply that our growth rate will pick up again as the supply increases? Just in the first half, there seems to be some catch-up. So, I would like to know how much guidance you have in this regard.

Jensen Huang, Chief Executive Officer:

We guide one quarter at a time.

Colete-Cress, Chief Financial Officer:

Currently, we are working on this quarter and establishing the transportation system we need at Blackwells company. All suppliers worldwide are working closely with us. Once we enter the next quarter, we will help everyone understand the next phase and the subsequent production growth.

Q10, Analyst Pierre Ferragu from New Street Research: You mentioned pre-training, actual language models, and reinforcement learning in your comments. Reinforcement learning is becoming increasingly important in training and inference. Then there's inference itself. I would like to know if you have a high-level typical sense of the whole artificial intelligence ecosystem, such as one of your clients or a large model. How much computation is there per computing unit? How much is used for pre-training, reinforcement, and inference today? Do you know how the computation is allocated?

Jensen Huang, CEO:

Now, the work of the pre-training base model is very important because everyone knows that the new technologies for post-training have just come online. Whatever you do in pre-training and post-training, you will try hard to minimize the inference cost for everyone. However, there are limits to what you can prioritize. Therefore, you always need to think on-site, think in context, and reflect. So, I think in our case, the scale of these three aspects is actually very reasonable.

In the field of basic models, we now have multimodal base models, and the volume of videos these base models need to be trained on has reached an incredible PB level. My expectation is that in the foreseeable future, we will continuously expand the scale of pre-training and post-training, as well as the scale of inference time. This is also the reason why I believe we need more and more computation, we must maximize performance as much as possible, increase by X times each time, so that we can continue to reduce costs, increase revenue, and drive the AI revolution. Thank you.

Now, please have Jensen Huang deliver the closing remarks.

Jensen Huang, CEO:

Thank you, the tremendous growth of nvidia's business is benefiting from two fundamental trends, which are driving global adoption of nvidia computing. First, the computing stack is undergoing a reshaping, transitioning from coding to machine learning platforms. From executing code on the CPU to processing neural networks on the GPU. The $1 trillion installation base of traditional data center infrastructure is being rebuilt for software 2.0, which will apply machine learning to generate artificial intelligence. Secondly, the era of artificial intelligence is fully approaching. Generative ai is not only a new software capability, but also a new industry where digital intelligence is manufactured by ai factories, marking a new industrial revolution that can create a trillion-dollar ai industry. The demand for Hopper and the expectations for Blackwell (which is now in full production) are incredible, for the following reasons. There are more infrastructure model manufacturers than a year ago, and the scale of computation for early-stage and late-stage training continues to grow exponentially.

There are more artificial intelligence native startups than ever before, with a growing number of successful reasoning services. With the launch of ChatGPT o1 and OpenAI o1, a new scaling rule called test-time scaling has emerged. All of these require a significant amount of computation. Artificial intelligence is changing every industry, company, and country. Enterprises are adopting delegative ai to fundamentally transform workflows. Over time, artificial intelligence colleagues will assist employees in completing work faster and better. Due to the breakthrough in physical artificial intelligence, investments in industrial robots are skyrocketing.

As researchers train world-scale models on megabytes of video and synthetic data across the universe, it is driving the demand for new training infrastructure. The era of robots is approaching. Countries around the world recognize the fundamental trends of artificial intelligence we are witnessing and the importance of developing their own artificial intelligence infrastructure. The era of artificial intelligence has arrived, with a large scale and a wide variety. Nvidia's expertise, scale, and the ability to provide full-stack and full infrastructure capabilities enable us to serve the future trillion-dollar value of the artificial intelligence and robot opportunities.

From every hyperscale cloud, enterprise private cloud to sovereign region artificial intelligence cloud, on-premises to industrial edge and robot technology. Thank you for attending today's meeting, see you next time.

Editor/Lambor

The translation is provided by third-party software.


The above content is for informational or educational purposes only and does not constitute any investment advice related to Futu. Although we strive to ensure the truthfulness, accuracy, and originality of all such content, we cannot guarantee it.
50
Comment Comment 2 · Views 142.7k

Comment(2)

Recommended

Write a comment
2

Statement

This page is machine-translated. Futubull tries to improve but does not guarantee the accuracy and reliability of the translation, and will not be liable for any loss or damage caused by any inaccuracy or omission of the translation.