The full transcript of the NVIDIA conference call is here! Blackwell's production ramp-up is smooth, Q4 revenue will exceed expectations, and the Scaling Law has not slowed down

The AI chip giant NVIDIA's revenue growth in the third quarter continues to slow down, and it is expected to further decelerate this quarter. Although last quarter's revenue exceeded analysts' expectations, the median guidance for this quarter did not meet the high-end expectations of buy-side analysts. From the stock price performance, investors seem to feel that such guidance is not explosive enough, leading to a decline in NVIDIA's stock price after hours.

In the conference call following the earnings report, NVIDIA provided guidance, stating that demand for the Hopper will continue in the fourth quarter, and Blackwell will achieve initial growth, with revenue expected to exceed previous estimates by billions of dollars. As the scale of Blackwell products expands, gross margins are expected to slow to below 70%.

During the Q&A session, analysts focused their "firepower" on Blackwell, raising a series of questions regarding the chip's progress, such as whether potential customers would drive greater demand for Blackwell, whether there are supply constraints for Blackwell, and whether the fourth quarter will be the most challenging period for gross margin pressure from Blackwell.

Here are the key points from this conference call:

1. Gross Margin Situation: Due to the launch of Blackwell this quarter, increased costs will lead to a reduction in gross margins. In the initial phase of the chip launch, gross margins will drop to a low of 70%, specifically in the range of 71-72.5%. In the second half of fiscal year 2025, it is expected to reach a median of over 70%, around 75%.

2. Blackwell Demand Situation: Blackwell is planned to start shipping this season, with an accelerated pace over the next year, and it is expected that demand will exceed supply by fiscal year 2026, indicating that increasing demand will drive continuous growth in chip demand. CEO Jensen Huang stated that Blackwell's delivery volume next quarter will exceed the company's previous expectations.

3. Blackwell Roadmap and Supply Constraints: The roadmap proposed at GTC will continue to be executed, with Ultra set to launch next year and a transition to Rubin in 2026. NVIDIA's execution work is progressing smoothly, with a vast supply chain network that includes TSMC, Amphenol, Vertiv, SK Hynix, Micron, Anker, KYEC, Foxconn, Quanta, Wistron, Dell, HP, Supermicro, Lenovo, etc. Progress in ramping up Blackwell's capacity is good.

4. Long-term Growth in AI Demand Until 2030: By 2030, global data centers for computing will reach several trillion dollars. The first point is modernizing data centers from coding to machine learning. The second point is generative artificial intelligence, building AI factories; we are now creating a new industry, a new segment that has never existed in the world.

5. Continued Growth in Hopper Demand: Demand for Hopper will continue into next year, at least for the first few quarters of next year, while next quarter's shipment volume will exceed this quarter's.

6. Scaling Law Has Not Slowed Down: There are currently three training methods, and pre-training will continue, as this is an empirical law, not a physical law. In addition, there are now post-training and inference scaling lawsThe industry has developed in pre-training, post-training, and now very importantly, inference time.

Below is the full transcript of NVIDIA's conference call

Colette Kress, Chief Financial Officer:

The third quarter was another record quarter, and we continue to achieve incredible growth. Revenue reached $35.1 billion, a 17% increase quarter-over-quarter and a 94% increase year-over-year, far exceeding our expected $32.5 billion. Driven by NVIDIA's accelerated computing and artificial intelligence, all market platforms achieved strong quarter-over-quarter and year-over-year growth.

Starting with the data center, the data center set a new record. Revenue reached $30.8 billion, a 17% increase quarter-over-quarter and a 112% increase year-over-year. Demand for NVIDIA Hopper is exceptionally strong, with H200 sales significantly increasing consecutively, reaching double-digit billions, making it the fastest-growing product in the company's history. The inference performance of H200 has improved by 2 times, and the total cost of ownership (TCO) has improved by about 50%. Cloud service providers account for about half of our data center sales, with revenue more than doubling year-over-year. Cloud service providers have deployed NVIDIA H200 infrastructure and high-speed networks, with GPU installations reaching tens of thousands to drive business growth and meet the rapidly growing demand for artificial intelligence training and inference workloads. AWS and Microsoft Azure now offer NVIDIA H200-powered cloud instances, with Google Cloud and OCI also set to launch soon. While large CSPs have seen significant growth, GPU cloud revenue in North America, Europe, the Middle East and Africa, and Asia-Pacific has also doubled year-over-year, due to the increasing NVIDIA cloud scenarios and sovereign cloud construction.

Consumer internet revenue more than doubled year-over-year, as enterprises scaled up NVIDIA Hopper infrastructure to support next-generation artificial intelligence models, training, multimodal and agent-based AI, deep learning recommendation engines, as well as generative AI inference and content creation workloads. NVIDIA's Ampere and Hopper infrastructure are driving customers' inference revenue growth. NVIDIA is the world's largest inference platform, with a massive installed base and a rich software ecosystem that encourages developers to optimize for NVIDIA, continuously improving performance and TCO. The rapid development of NVIDIA's software algorithms has increased Hopper's inference throughput by an astonishing 5 times within a year, reducing time by 5 times.

Our upcoming NVIDIA NIM will further enhance Hopper Inference performance by 2.4 times, and continuous performance optimization is a hallmark of NVIDIA, bringing increasing economic returns to the entire company's installed user base. After successfully executing large-scale changes, Blackwell has been fully put into production. In the third quarter, we delivered 13,000 GPU samples to customers, including the first batch of open AI Blackwell DGX engineering samples. Blackwell is a full-stack, full-infrastructure AI data center-scale system with customizable configurations to meet the diverse and growing demands of the AI market, ranging from X86 to ARM, from training to enhanced GPUs, from InfiniBand to Ethernet switches, and from liquid cooling to air coolingEvery customer is rushing to enter the market, and Blackwell is now in the hands of all major partners who are working to enhance their data centers. We are integrating the Blackwell system into various data center configurations for our clients. The demand for Blackwell is astonishing, and we are ramping up supply to meet the enormous demand from our customers. Clients are preparing for large-scale deployments of Blackwell. Oracle 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 AI models. Yesterday, Microsoft announced that they would be the first CSP to offer Blackwell cloud instances based on NVIDIA GB200 and Quantum InfiniBand in a private preview. Last week, Blackwell made its debut in the latest round of training results, winning the GPU benchmark with performance 2.2 times that of Hopper.

These results also demonstrate our relentless pursuit of reducing computing costs. Running the GPT3 benchmark requires only 64 Blackwell GPUs compared to 256 H100s, resulting in a 4-fold cost reduction. The switch in the NVIDIA Blackwell architecture can increase inference performance by 30 times and elevate inference throughput and response times to a new level, making it ideal for running new inference applications like OpenAI's o1 model. Each new platform shift spawns a wave of new startups. Hundreds of AI-native companies have 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, while thousands of AI-native startups are building new services.

The next wave of artificial intelligence is enterprise AI and industrial AI, with enterprise AI fully unfolding. NVIDIA's AI Enterprise Edition, which includes NeMo and NIM microservices, is an agent AI operations platform. Industry leaders are using NVIDIA AI to build collaborative driving and agents, with companies like Cadence, Cohesity, NetApp, Nutanix, Salesforce, SAP, and ServiceNow accelerating the development of these applications in collaboration with NVIDIA, potentially deploying billions of agents in the coming years. Consulting industry leaders like Accenture and Deloitte are bringing 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 build-out. Additionally, Accenture, with over 770,000 employees, is internally utilizing NVIDIA-supported Agentic AI applications, one case of which reduced manual steps in marketing campaigns by 25% to 35%Nearly 1,000 companies are using NVIDIA NIM, showcasing the monetization speed of NVIDIA AI Enterprise. We expect the annual revenue from the AI Enterprise edition to grow more than twofold compared to last year, and our sales channels continue to expand. Overall, our software, services, and support revenue has an annual growth of $1.5 billion, and we anticipate this year's annual growth will exceed $2 billion, with industrial artificial intelligence and robotics accelerating development.

This is driven by breakthroughs in physical artificial intelligence, which involves foundational models capable of understanding the physical world. Similar to NVIDIA NeMo, which is aimed at 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. Foxconn, the world's largest electronics manufacturer, is using NVIDIA Omniverse-based digital twins and industrial AI to speed up the construction of its Blackwells factory and enhance efficiency to new heights. In its Mexican factory alone, Foxconn expects to reduce annual kilowatt-hour consumption by 330%. Regionally, our data center revenue in China continues to grow, thanks to the industry shipment of copper products that meet export standards, but its share of total data center revenue remains far below levels prior to the onset of export controls. We expect competition in the Chinese market to remain very fierce in the future. We will continue to comply with regulations while serving our customers.

As countries adopt NVIDIA accelerated computing technology to drive a new wave of industrial revolution powered by AI, our sovereign AI programs are also gearing up. Leading CSPs in India, including Product Communication and Zoda Data Services, are building AI 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 tenfold. TSC and Wipro are adopting NVIDIA AI Enterprise and training nearly 500,000 developers and consultants to help customers build and run AI agents on our platform. In Japan, SoftBank is leveraging NVIDIA DGX Blackwell and Quantum InfiniBand to build the country's most powerful AI supercomputer. SoftBank is also collaborating with NVIDIA to transform telecom networks into distributed AI networks using NVIDIA AI Aerial and ARAN platforms, which can process 5G RAN and AI on CUDA.

We are partnering with T-Mobile in the U.S. to launch similar products. Leading Japanese companies, including NEC and NTT, are adopting NVIDIA AI Enterprise, and large consulting firms, including EY, Strategy, and Consulting, will help introduce NVIDIA AI technology across various industries in Japan. Network business revenue has increased by 20% year-on-year. Areas with continuous revenue growth include InfiniBand and Ethernet switches, SmartNex, and BlueField DPUAlthough online revenue has been declining continuously, online demand remains strong and is growing. We expect online revenue to increase sequentially in the fourth quarter.

CSPs and supercomputing centers are using and adopting NVIDIA's InfiniBand platform to power the new H200 clusters. Revenue from NVIDIA Spectrum-X Ethernet for AI has more than tripled year-on-year, and our product line continues to expand, with several CSPs and consumer internet companies planning to deploy large clusters. Traditional Ethernet was not designed for artificial intelligence. NVIDIA Spectrum-X leverages technologies previously unique to InfiniBand, enabling customers to achieve large-scale GPU computing. With Spectrum-X, xAI's Colossus 100,000 Hopper supercomputer will achieve zero application latency and maintain 95% data throughput, compared to only 60% for traditional Ethernet.

Now let's talk about gaming and AIPC. Gaming revenue reached $3.3 billion, a 14% sequential increase and a 15% year-on-year increase. The third quarter was a bumper harvest for the gaming business, with revenue from laptops, gaming consoles, and desktops all achieving sequential and year-on-year growth. Strong back-to-school sales momentum drove terminal demand for RTX as consumers continued to choose GeForce RTX GPUs and devices to support gaming, creative, and AI applications. Channel inventory remains healthy, and we are preparing for the holiday season. We have started shipping new GeForce RTX AI PCs from ASUS and MSI, equipped with up to 321 AI vertices, expected to feature Microsoft's Copilot+ in the fourth quarter.

These machines harness the powerful capabilities of RTX ray tracing and AI technology, providing supercharged performance 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 the driving force behind some of the most influential technologies of our time, from transforming computer graphics to igniting the AI revolution. Looking at ProViz, revenue was $486 million, a 7% sequential increase and a 17% year-on-year increase. NVIDIA RTX workstations continue to be the preferred choice for professional graphics, design, and engineering workloads.

Additionally, AI is becoming a powerful demand driver, including autonomous vehicle simulation, generative AI model prototyping for productivity-related use cases, and generative AI content creation in the media and entertainment sector. Revenue from the automotive industry reached a record $449 million, a 30% sequential increase and a 72% year-on-year increase. Strong growth was primarily driven by the enhancement of NVIDIA Orin autonomous driving technology and robust 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 was 74.6%, and non-GAAP gross margin was 75%. The sequential decline was mainly due to the transition of H100 systems in data centers to a more complex and higher-cost system mixDue to the high costs of computing, infrastructure, and engineering development required for the launch of new products, GAAP operating expenses and non-GAAP operating expenses increased by 9% quarter-over-quarter. In the third quarter, we returned $11.2 billion to shareholders in the form of stock buybacks and cash dividends.

Now, allow me to discuss the outlook for the fourth quarter. Total revenue is expected to be $37.5 billion, with a growth range of plus or minus 2%, which includes continued demand for Hopper Architecture and initial growth of Blackwell products. Although demand significantly exceeds supply, as we gain a better understanding of the supply situation, we expect to exceed the previously projected billions of dollars in revenue for Blackwell. In gaming, despite strong sales in the third quarter, we expect a sequential decline in revenue in the fourth quarter due to supply constraints. According to GAAP and non-GAAP, gross margins are expected to be 73% and 73.5%, respectively, with a growth range of plus or minus 50 basis points. Blackwell is a customizable AI infrastructure equipped with various types of NVIDIA built-in chips and multiple network task options, suitable for air-cooled and liquid-cooled data centers. Our current focus is on addressing strong demand, improving system availability, and providing customers with the best configuration combinations.

As Blackwell scales, we expect gross margins to slow to below 70%, with Blackwell's gross margin around 70%. According to GAAP and non-GAAP, operating expenses 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 to develop our hardware and software stack and support the launch of new products. According to GAAP and non-GAAP, excluding non-affiliated investment income and losses, other income and expenses are expected to be around $400 million. The GAAP and non-GAAP tax rate 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, allow me to highlight some upcoming financial events; we will participate in the UBS Global Technology and AI Conference in Scottsdale on December 3rd, and on January 6th, Jensen will deliver a keynote speech at CES in Las Vegas. The following day (January 7th), we will hold a Q&A session for financial analysts, and we invite you to attend. We will hold a earnings call on February 26, 2025, to discuss the performance of the fourth quarter of fiscal year 2025. We now begin the Q&A session.

Q&A Session

Q1 C.J. Muse, Analyst at Cantor Fitzgerald: I think I just want to ask you one question, which is about the scaling rules debate for large language models. Clearly, it’s still early days for us, but I’d love to hear your thoughts on this.How are you helping clients solve these problems? Clearly, we are discussing clusters that have not yet benefited from Blackwell; will this drive greater demand for Blackwell? Thank you.**

Responder:

The foundational model scaling before training is intact and continues. Everyone knows this is a rule of thumb rather than a fundamental physical law, but there is evidence that it is still continuing to train. However, what we are learning is that this is not enough; we have now discovered two additional training methods, one is post-training, and the other is post-training with reinforcement learning from human feedback, but now we have reinforcement learning with artificial intelligence feedback and various forms of synthetic data generation, all of which contribute to the expansion of post-training.

One of the biggest and most exciting developments is Strawberry, which is OpenAI's o1 model that performs inference time extension, known as inference time. The longer it thinks, the better and higher quality the answers it derives; it considers methods like thinking chains, multi-path planning, and various necessary reflection techniques, etc. Intuitively, this is somewhat like us thinking in our heads before answering questions.

Therefore, we now have three training methods, and we have seen all three training methods. Thus, the demand for our infrastructure is very large. You can now see that the capacity of the previous generation foundational model is about 100,000 Hopper chips, while the next generation starts with 100,000 Blackwell chips. You can understand the industry's direction in pre-training, post-training, and now very importantly, inference time extension.

For all these reasons, the demand is indeed very high. But please remember that at the same time, our company's inference capabilities are also continuously improving. We are the largest inference platform in the world today because our installed base is very large, and all inference is trained with and on the jump, which is incredible.

As we turn to using Blackwells to train foundational models, it also brings a large amount of installed infrastructure for inference. Therefore, we see the demand for inference continuously increasing. We see inference times continuously extending. We see the number of AI-native companies steadily growing. Of course, we are also starting to see enterprises adopting Agentic AI, which is indeed a recent trend. We see a large demand coming from different places.

Q2 Goldman Sachs Toshiya Hari: Jensen, you implemented large-scale reforms earlier this year, and there were some reports about heating issues last weekend. In this context, do you have the capability to execute the roadmap proposed at GTC this year, namely the launch of Ultra next year and the transition to Rubin in 2026? Can you talk about this issue? Some investors have expressed doubts about this. It would be very helpful if you could discuss your ability to execute on time. Then regarding the Blackwell part, is the supply constraint caused by numerous components, or specifically HBM? Is it a supply constraint? Has the supply constraint improved? Is it worsening? Any information on this would be very useful. Thank you.Respondent:

Yes, thank you. Now let's take a look and return to the first question. The production of Blackwell has fully launched. In fact, as Colette mentioned earlier, we will deliver more Blackwells this quarter than previously expected. Therefore, the supply chain team is doing an excellent job working with our supply partners to increase the production of Blackwells. Next year, we will continue to work hard to increase the output of Blackwell.

The current situation is one of supply not meeting demand, which is expected, as we are well aware that we are at the beginning of the artificial intelligence revolution. We are at the starting point of a new generation of foundational models that can reason, think long-term, and of course, one of the truly exciting areas is physical AI, which is now capable of understanding the structure of the physical world.

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

Google has its own systems, and all these CSPs are racing ahead. As you know, the engineering design we are collaborating on with them is quite complex. The reason is that, although we have built a full stack and full infrastructure, we have separated all the AI supercomputers and integrated them into all the custom data centers around the world.

We have been doing this integration process for several generations, and 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 status of Blackwell is very good. As we mentioned earlier, the supply and planned shipments for this quarter exceed our previous expectations.

Regarding the supply chain, there are seven different chips, seven custom chips that we have built in to provide the Blackwell systems. The Blackwell systems use air cooling or liquid cooling, 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 nothing short of miraculous.

Therefore, to achieve such a scale of growth, the supply chain of components required is essential. You know, you have to look back at how many Blackwell shipments we had last quarter, which was zero. This quarter, the total shipment of Blackwell systems will be in the billions, and the growth rate is incredible. Almost all companies in the world are involved in our supply chain, and we have very good partners, including TSMC, connector company Amphenol, Vertiv, SK Hynix, Micron, Amkor, KYEC, Foxconn, Quanta, WeWin, Dell, HP, AMD, Lenovo, and the number of companies is astonishingI am sure I have missed some partners involved in the expansion of Blackwell, for which I am very grateful. In short, I believe we are currently in a good position regarding capacity ramp-up.

Finally, regarding our execution roadmap, we are executing the annual roadmap, and we hope to continue executing the annual roadmap. Of course, by doing so, we can improve the performance of the platform, but equally important is that when we can improve performance while increasing the X coefficient, we reduce training costs, reduce inference costs, and reduce the costs of artificial intelligence, making AI more accessible.

But another very important factor is that when data centers have a certain fixed scale, it used to be 10 megawatts, and now most data centers are 100 megawatts to several hundred megawatts, with planned gigawatt data centers. The size of the data center is actually not 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 one hand, our annual roadmap reduces costs. On the other hand, because our performance per watt is better than any other product, we can create as much revenue as possible for our customers. Therefore, this annual rhythm is very important for us, and we are fully committed to continuing this. As far as I know, everything is proceeding as planned.

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

Colette Kress, Chief Financial Officer:

Of course, let me first answer your question, and thank you for your question about our gross margin. We have discussed our gross margin because we are enhancing the performance of Blackwall from the very beginning, and we will launch many different configurations and many different chips to the market, focusing on ensuring our customers have the best experience upon installation.

Our gross margin will start to grow, but we believe that in the first phase of growth, the gross margin will be around 70%. So, you are correct that in the following quarters, we will start to improve the gross margin, and we hope to reach 70% soon.

Jensen Huang, Chief Executive Officer:

The demand for Hopper will continue into next year, certainly in the first few quarters of next year. Meanwhile, the shipments for the next quarter will exceed this quarter. Our shipments of Blackwells next quarter will exceed those in the first quarter. Therefore, from this perspectiveWe are at the beginning of two significant fundamental shifts in the field of computing. The first shift is from coding running on CPUs to machine learning that creates neural networks running on GPUs.

Currently, the fundamental shift from coding to machine learning is very common. No company is not engaged in machine learning. Therefore, machine learning is also the foundation of generative artificial intelligence. Thus, on one hand, the first thing happening is that the $1 trillion computing systems and data centers around the world are being modernized for machine learning. On the other hand, I think, based on these systems, we will create a new type of artificial intelligence capability.

When we say "generative artificial intelligence," we are essentially saying that these data centers are actually AI factories. They are producing something. Just as we produce electricity, we now also need to produce artificial intelligence. If the number of customers is large, just like the number of electricity consumers is large, these generators will operate around the clock. Nowadays, many AI services are running around the clock, just like AI factories. Therefore, we will see the launch of this new type of system, which I call AI factories, because it is indeed very close to its essence. It is different from past data centers. Thus, these two fundamental trends have just begun. Therefore, we expect this growth—this modernization and the creation of new industries will last for years.

Q4 Bank of America Securities Vivek Arya: Colette, I want to ask, do you think it is reasonable to assume that NVIDIA's gross margin can recover to around 70% in the second half of 2025? Then, my main question is, historically, when we see hardware deployment cycles, there inevitably includes some digestion process. When do you think we will enter this phase, or is it too early to discuss this issue now since Blackwells has just started? How many quarters of shipments do you think are needed to meet the first wave of demand? Can growth continue into 2026? How should we prepare for the long-term, secular hardware deployment digestion period?

Colette Kress, Chief Financial Officer:

Let me clarify your question about gross margin, whether we can reach a gross margin of 70% in the second half of next year. I think this is a reasonable assumption and a goal we aim for, but we need to see how the ramp-up mix looks, but it is definitely possible.

Jensen Huang, Chief Executive Officer:

I believe it cannot be digested until we modernize the $1 trillion data centers. Therefore, looking at data centers globally, the vast majority were built in an era where applications were manually coded and run on CPUs. Doing so now is unreasonable. If every company's capital expenditure, if they were to build a data center tomorrow, they should be built for future machine learning and generative artificial intelligence, right?

Because they have many outdated data centers. So, what will happen in the next X years? Let’s assume that in 4 years, as we progress in the IT field, data centers around the world will be modernized. Everyone knows that IT grows at about 20% to 30% per yearFor example, by 2030, global spending on data centers for computing will reach several trillion dollars. We must adapt to this growth. From coding to machine learning, we must modernize data centers. This is the first point. The second part is generative artificial intelligence. We are now creating a new type of capability that has never existed in the world before, a new segment that has never existed in the world before.

If you look at OpenAI, it hasn't replaced anything. It is something entirely new. In many ways, it is completely new, just like when the iPhone was born. It hasn't replaced anything. Therefore, we will see more and more companies like this. They will create and generate essentially intelligent services from their offerings, some of which will be digital artist intelligence.

Some are foundational intelligence, like OpenAI. Some are legal intelligence, like Harvey, digital marketing intelligence, like writers, and so on. There are hundreds of these companies, and they are called AI-native companies. Almost every platform is changing; otherwise, as you remember, there used to be internet companies, cloud-first companies, mobile-first companies, and now they are all AI-native companies. They were mobile-first companies, and now they are AI-native companies. These companies were born because people saw the shift in platforms and the new opportunities to do entirely new things.

Therefore, my feeling is that we will continue to build modern information technology, first modernizing computing, and secondly creating AI factories to build a new industry for AI production.

Q5 Bernstein Research analyst Stacy Rasgon: The clarification question is, when you say gross margins are as low as 70%, does 73.5% count as low as 70%, do you have any other thoughts? As for my question, you are guiding total revenue, so I mean, the total revenue for data centers next quarter must grow by several billion dollars, but it sounds like Blackwells' growth should be more than that now. But you also mentioned that Hopper remains strong. So will Hopper decline consecutively next quarter? If so, why? Is it due to supply constraints? The Chinese market has always been strong, but after entering the fourth quarter, the Chinese market will decline. Therefore, if you could tell us about Blackwells' growth situation and how Blackwells performs compared to Hopper in the fourth quarter, it would be very helpful. Thank you.

Colette Kress, Chief Financial Officer:

Starting with your first question about our gross margins and the definition of the low point. Of course, our gross margins will be below the median; for example, our gross margins might be 71%, or it could be 72%, 72.5%, and we will be in that range. We could also be higher than that. We need to see how the results turn out. We do hope that for the remainder of this year, we can ensure that our output and products continue to improve and sustain, at which point we will reach around 75%The second question is about our Hopper, H200 has seen significant growth, which is reflected not only in orders but also in the speed of project initiation. This is an amazing product and the fastest-growing product we have seen.

We will continue to sell Hopper in this quarter and the fourth quarter. This involves all our different configurations, including the measures we might take in China. But keep in mind that people are also looking to build their own Blackwell.

Therefore, we may see both situations occurring in the fourth quarter. Yes, is it possible for Hopper to achieve growth between the third and fourth quarters? It is possible, 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 talked about some impacts of strawberries and longer extended inference projects. But you also mentioned that as some Hopper clusters age, you might use some potential Hopper chips for inference. Do you expect inference to exceed training in the next 12 months?

Jensen Huang, CEO:

Our hope and dream is that one day, the whole world can perform a large amount of inference. Only then will artificial intelligence be truly 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 performing inference internally. We hope that every company can perform inference around the clock, and there will be a large number of AI-native startups and thousands of AI-native startups generating tokens and AI, from using Outlook to creating PowerPoint, or in every aspect of your computer experience while sitting there using Excel, tokens are constantly being generated.

Every time you read a PDF or open a PDF, a large number of tokens are generated. One of my favorite applications is NotebookLM, an application launched by Google. I enjoy using it because it's fun, you know. I put every PDF file and every archived document in it, just to listen and scan. So I think our goal is to train these models for people to use. Now, artificial intelligence has entered a whole new era, a new genre of AI called physical AI, where these large language models can understand human language and our thought processes. Physical AI understands the physical world, it understands the meaning of structure, what is reasonable, what is unreasonable, what might happen, and what will not happen. It not only understands but can predict and infer a fleeting future.

This capability has incredible value for industrial AI and robotics. Therefore, many AI-native companies, robotics companies, and physical AI companies have emerged, and you may have heard of these companies. This is also the real reason we are building Omniverse. Omniverse allows these AIs to create and learn within the Omniverse and learn from physical feedback generated from synthetic data generation and reinforcement learning, rather than learning from human feedbackTo possess these capabilities, we created Omniverse to achieve physical artificial intelligence. Therefore, our goal is to generate tokens. Our aim is reasoning, and we are already beginning to see this growth. So I am very excited about this. Now let me say one more thing. Reasoning is super difficult. The reason reasoning is super difficult is that on one hand, you need high precision, 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. And it is very difficult to manufacture computers that have both high throughput and low latency. These applications have longer context lengths because they want to understand and reason in the context of understanding. Therefore, the context length is getting larger and larger.

On the other hand, the models are getting larger and are multimodal. It is incredible to know how many dimensions there are to reasoning innovation. The speed of this innovation is precisely what makes NVIDIA's architecture great, as our ecosystem is outstanding. Everyone knows that if they innovate on the CUDA foundation of NVIDIA's architecture, they can achieve innovation faster, and they know that everything should be feasible.

If something happens, it is likely their code rather than ours. Therefore, we have the capability to innovate in all directions simultaneously, with a large installed base, so that whatever you create can be implemented on NVIDIA computers and widely deployed to data centers around the world, all the way to the edge of robotic systems. This capability is truly amazing.

Q7 Wells Fargo analyst Aaron Rakers: When we look at the cycle of the data center business, I want to ask you a question. When I look at last quarter's performance, Colette, you mentioned that the networking business has obviously declined by 15% sequentially, but your comment was that you saw very strong demand.

You also mentioned that in these large-scale clusters, you have won designs from multiple cloud CSPs. So I would like to know if you can explain the development of the networking business, what constraints you see in some areas, and whether you are confident in the growth rate of Spectrum-X to reach the billion-dollar scale you mentioned earlier.

Colette Kress, Chief Financial Officer:

First, starting with networking, the year-over-year growth is huge. Since the acquisition of Mellanox, our focus has been on integrating the work we have done in data centers. Networking is a crucial part of that. Our ability to sell our networking along with many of the systems we are doing in data centers is continuously growing and performing quite well.

Therefore, the performance this quarter is just slightly down, and our growth momentum will recover immediately. They are preparing for Blackwells and an increasing number of systems that will not only use our existing networking but also the networking we provide to these large systems.

Q8 Citibank Atif Malik: I have two questions for Colette. Colette, in the last earnings call, you mentioned that sovereign demand is low double digits. Can you provide an update on this? Then, can you explain the situation regarding the tight supply in gaming?Is this because you are shifting supply towards data centers?**

Colette Kress, Chief Financial Officer:

First, starting with Sovereign AI, this is an important part of growth, and it has really come to the forefront with the emergence of generative AI and the establishment of models in various countries around the world. We have seen many such companies, and we talked about many of them in today's conference call, as well as the work they are doing. Therefore, our Sovereign AI and our future pipeline remain absolutely intact, as these people are working to build these foundational models in their own languages and cultures and are working within the enterprises of these countries. I believe this will continue to be a growth opportunity, and you may see our regional clouds storing and/or AI factories focusing on many aspects of Sovereign AI.

There is growth not only in Europe but also in the Asia-Pacific region. Let me address your second question regarding gaming. From a supply perspective, we are currently busy ensuring that all our different products are being produced properly. In this case, our console supply is progressing quite rapidly based on the sales we are seeing. Now, the challenge we face is how to quickly supply these products to the market this quarter. Don't worry, I believe that in the new year, we will have more suppliers getting back on track. It's just that the supply will be tighter this quarter.

Q9 Melius Research analyst Ben Reitzes: The sequential growth this quarter is very strong, and your revenue guidance is about 7%. Does your comment on Blackwells imply that as supply increases, our growth rate will accelerate again? It seems there will be some catch-up in the first half of the year. So, I wonder how much guidance you have on this.

Jensen Huang, Chief Executive Officer:

We guide one quarter at a time.

Colette Kress, Chief Financial Officer:

Currently, we are working on this quarter and establishing the transportation system we need with Blackwells. All global suppliers 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 New Street Research analyst Pierre Ferragu: You mentioned pre-training, actual language models, and reinforcement learning in your comments, and reinforcement learning is becoming increasingly important in training and inference. Then there is inference itself. I wonder if you have a high-level typical sense of the entire AI ecosystem, such as for one of your customers or a large model. What is the computational load for each compute unit? How much is used for pre-training today, how much for reinforcement, and how much for inference? Do you know how the computational load is distributed?

Jensen Huang, Chief Executive Officer:

Today, the work of pre-training foundational models is very important because everyone knows that new technologies for post-training have just come online. Whatever you do in pre-training and post-training, you will strive to do so as to minimize the inference costs for everyone as much as possibleHowever, the things you can prioritize are limited. Therefore, you always have to engage in on-site thinking, situational thinking, and reflection. Thus, I believe that the scale of these three aspects is actually very reasonable based on our situation.

In the field of foundational models, we now have multimodal foundational models, and the amount of video data required to train these foundational models has reached an incredible PB level. My expectation is that in the foreseeable future, we will continue to expand the scale of pre-training and post-training, as well as the scale of inference time, which is also why I believe we need more and more computing power. We must improve performance as much as possible, increasing it by a factor of X each time, so that we can continue to reduce costs, continue to increase revenue, and drive the AI revolution. Thank you.

Now, please have Jensen Huang generate the closing remarks.

Jensen Huang, CEO:

Thank you. The tremendous growth of NVIDIA's business is driven by two fundamental trends that are propelling the global adoption of NVIDIA computing. First, the computing stack is undergoing a transformation, shifting from coding to machine learning platforms. This is a transition from executing code on CPUs to processing neural networks on GPUs. The $1 trillion installed 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 upon us. Generative AI is not only a new software capability but also a new industry that manufactures digital intelligence through AI factories, representing a new industrial revolution that can create trillions of dollars in the AI industry. The demand for Hopper and the expectations for Blackwell (which is now fully operational) are incredible for several reasons. There are now more foundational model manufacturers than a year ago, and the scale of computing for pre-training and post-training continues to grow exponentially.

There are more AI-native startups than ever before, and the number of successful inference services is also on the rise. With the launch of ChatGPT o1 and OpenAI o1, a new scaling law called test-time scaling has emerged. All of this requires a massive amount of computing power. AI is transforming every industry, company, and country. Businesses are adopting agent-based AI to fundamentally change workflows. Over time, AI colleagues will assist employees in completing their work faster and better. Due to breakthroughs in physical AI, investment in industrial robotics is surging.

As researchers train world foundational models on megabytes of video and the entire universe of synthetic data, there is a growing demand for new training infrastructure. The era of robotics is approaching. Countries around the world recognize the fundamental trends in AI that we are witnessing and have realized the importance of developing their own AI infrastructure. The era of AI has arrived, and it is vast and diverse. NVIDIA's expertise, scale, and ability to provide full-stack and full infrastructure enable us to serve the future AI and robotics opportunities worth trillions of dollars.

From every hyperscale cloud, enterprise private cloud to sovereign regional AI clouds, on-premises to industrial edge and robotics. Thank you for attending today's conference, and see you next time