NVIDIA Corporation (NVDA) UBS Global Technology and AI Conference (Transcript)
Call start: 9:35 AM Eastern Standard Time on January 1, 0000 to 10:09 AM
NVIDIA Corporation (NASDAQ: NVDA )
UBS Global Technology Conference Call
9:35 AM (Eastern Standard Time) on December 3, 2024
Company participants
Collette Kress - Executive Vice President and Chief Financial Officer
Phone conference attendees
Tim Archule-UBS
Tim Archule
Good morning. Thank you. I'm Tim Archule. Thank you for coming to NVIDIA. Colette is also here. Thank you for your time, Colette. Let's get started. Colette, these past 2 years have been fantastic. Colette, you are still growing rapidly. Can you talk about how some of the use cases have evolved? Companies are widely adopting AI. From the perspectives of cloud, consumer, internet, and enterprise, how has the demand evolved? Also, could you talk about some very exciting use cases.
Colette Cress
Understood. Let me start with the statement I need to read. Just a reminder, this presentation contains descriptions related to future estimates, and investors are advised to read our reports filed with the SEC to obtain information about the risks and uncertainties our business faces.
It's truly a pleasure to meet all of you. Thank you so much for having us. Let me talk a little about what we've seen. Even just in the past few quarters, it has certainly been a very fast journey. But please remember our business has a 30-year history. However, we are certainly at a very important stage.
Thinking about the stage we are at now and what we are seeing, many of us who have used and seen the computing platforms for 20 to 30 years believe that the computing platforms we have used will undergo transformation over the next several decades. This means there are people focusing on existing computing platforms, such as the transition to general-purpose computing or accelerated computing.
In addition to focusing on accelerated computing, a new element of focusing on AI and Generative AI has also been added. Of course, the new use cases we have seen over the past several quarters mainly involve model size. If you remember, we were talking about large language models just before Generative AI appeared.
The importance of consumer internet companies, recommendation engines, and work related to model size is becoming increasingly important. Currently, work on foundational models is crucial in terms of the ongoing projects. However, due to the emphasis on building various types of foundational models and multimodal models, scaling of these models is also observed.
The next stage of this transition also focuses on the inference phase. After developing large language models, the work on supporting inference models is increasing. All our different types of customers are unique. We collaborate with all types of companies from startups to research. CSP is a major part of launching high-speed computing for end-user companies in need. While we collaborate with consumer internet companies, it is even more important that we are globally engaged in supporting this initiative.
It was something unprecedented in terms of speed and understanding of what is to come. This is part of what we are seeing today.
Tim Archuri
We are all concerned that ultimately we will end up building excess production capacity. And your opinion, I think, suggests that we are still far from it. So, could you talk about to what extent you have a perspective, and also about the demand situation in relation to the available supply?
Colette Crest
Yes, looking at the past quarters and our scaling, our primary goal regarding scaling is to collaborate with all partners, downstream customers, and bring about supply. No, at this point, we do not see a situation indicating a slowdown. However, it is currently confirmed that due to the size of the models and the complexity of inference, the demand continues to increase significantly, and we are preparing for the next architecture, Blackwell.
Regarding the upcoming Blackwell in the current quarter, it appears that we are likely experiencing supply constraints, which are expected to continue until the next fiscal year. Therefore, we do not expect a slowdown. In particular, demand and interest for the new architecture to be introduced in the future continue to be very high.
Tim Archule
Indeed, it was a great quarter. We shipped Blackwell in January and have shipped more Blackwell than we expected three months ago. At the same time, I hear that there is a lot of disruption in the supply chain. Many articles have been written about that. Could you tell us how Blackwell differs from the previous product cycle?
Collette Cress
Yes, our Blackwell architecture is unique. However, what we are doing here is building at the data center scale. Please do not misunderstand. We have been focusing on the platform perspective for many years. It's not just at the chip level or end-to-end scale. But when you look at our Hopper architecture, you can see that our Hopper architecture is for rack scale and the work we have been doing.
Essentially, some of what you see with the Blackwell architecture is fully built within our supply chain. Are you ready? We run it. We remove it. We install it, then turn it on together. And depending on which stage of the lifecycle you are in, we provide more options to our customers.
In other words, data centers are complex, and various things are being done to make them more efficient. Each one is at different stages. Therefore, you can choose from various options we offer. That means liquid cooling is possible. Air cooling can also continue. You can incorporate ARM CPUs and x86 as needed.
We have various network options such as InfiniBand and Ethernet, as well as various switch options. The choice is up to the customer, and there are various configurations available. Considering the current situation of Blackwell this quarter, everything related to the chips is complete.
The chip is doing well. The work we did with the chip is progressing smoothly. Currently, we are setting up configurations for various customers. Everyone will see photos on the internet. The first configuration is excited to start, happy faces preparing to assemble the entire data center and various racks. This is our current situation.
Tim Archule
That's great. There are various product release spaces, B100, B200. There are racks with CDs. And in about 6 months, Blackwell Ultra will be introduced. Are there any risks for customers waiting for products to be released this quickly?
Colette Cress
No, when considering what is needed for data center design, planning is essential. Over the past 5 to 10 years, we have been working with various customers on more projects. Customers plan every 6 months. What's here? What are we building? It is necessary to have computing ready when working on projects. Therefore, considering the shortage of supplies, we continue to offer customers the excellent Hopper 200 configuration. This is an opportunity to start some of the work being done in the HGX system and GPU demand is extremely high, so we have not touched the GPU yet. People understand how to build data centers by working with us. Data centers are often already procured. They just know what needs to be placed inside. So, there are two architectures.
As you know, we will also implement Vera Rubin in the future. We will discuss this with customers as well, but first, let's talk about potential services. How can we assist in understanding what might be needed in the future?
Tim Archule
However, this seems a bit similar to when customers did not have to wait for successful products like H100 and H200. Does this mean there is evidence that customers did not have to wait immediately after the launch of Vera products?
コレット・クレス
顧客は待ち望んでいます。待つというのは強い言葉です。顧客は毎日電話をかけてきて、いつコンピューティングが見られるのか尋ねており、私たちはその新しい供給のために必死に取り組んでいます。しかし、重要なのは、これはおそらく20年かかる旅だということを忘れないことです。誰もが参加するでしょう。私たちのアーキテクチャは、アクセラレーテッドコンピューティングとAIの世界で必要なことを何でもできるように、エンドツーエンドのスケーリングアプローチを可能にします。そして私たちは、インフラストラクチャだけでなくソフトウェアでも顧客を支援するための非常に有力な候補です。
ティム・アーキュリ
ありがとうございます。それで、キャッシュについてお聞きしたいのですが、実は昨夜このことについて話し合いました。過去 4 四半期で約 56 billionドルのキャッシュフローを生み出しました。25 暦年には約 120 billionドルを生み出すと聞いています。年間約 50 billionドルの自社株買いをすると仮定しても、来年末には 100 billionドルのキャッシュが残り、2026 年末には 200 billionドルのキャッシュが残る可能性があります。これは明らかに莫大な数字で、100 billionドルのキャッシュがあった Apple でさえ、最終的にはその額を減らしました。キャッシュをどう使うつもりですか?
コレット・クレス
現金に関する当社の取り組みと現金の使用は、どの企業にとっても最も重要な事項の 1 つです。当社は、戦略とニーズを通じて物事を手助けしてくれる優秀な人材と資金を費やしています。まず検討すべきことは、イノベーションと、当社が行っている拡大に向けたあらゆる支援に必要な現金がいくらかということです。これが最初の部分になります。しかし、当社は継続しており、イノベーションと、R&D で行っているすべてのことが、その重要な部分になることを知っています。
今、私たちは企業から学んだことの両方からそれを実行できます。また、M&A の形で優秀なチームを招き入れるという私たちの仕事の観点からもそれを考えることができます。これは私たちにとって素晴らしい機会であり、私たちはその分野でも引き続き取り組んでいきます。そして、それは AI の新しい分野に追加して重点的に取り組みたい新しいタイプのビジネス モデルについて考えることにつながります。
そして、私たちができるサポートは、ソフトウェアの構築だけではなく、他の企業向けの完全なシステムの構築であり、私たちはそれに投資します。こうした種類の投資を決定した後、私たちの仕事は、株主への還元という点において常にそうであり、それに対する私たちの焦点は、自社株買いと配当の組み合わせであり、皆さんはこれを引き続き目にすることになるでしょうし、これを注意深く見守ることになるでしょう。
We do not prefer excess cash. Therefore, we will carefully monitor it and intend to continue this in the next quarter.
Tim Archule
That's great. Thank you. I briefly talked about gross profit, and as Blackwell's performance expands, it was mentioned that the gross profit will decrease in the first quarter of next year. It was explained that it will rise to the low 70 percent range, then during the conference call, it will increase from 71 percent to 72.5 percent. However, it will rise again afterwards, reaching the mid-70 percent range by the end of the year.
Could you please tell us how confident you are about this point? After Rubin comes into play, some people believe that the gross profit margin will be squeezed again as Rubin begins to grow. Can you talk about the confidence in maintaining the mid-70s in the long term?
Colette Cress
Understood. Blackwell is unique in various aspects. Just this quarter alone, they are preparing various versions to be introduced to the market. It will not be just one version being shipped. Multiple versions are planned for shipment. Therefore, the current shipment volume is very low.
By expanding this throughout the year and embarking on the expansion of various system configurations that the company holds, we will be able to improve the gross profit margin. However, when thinking about the future prospects and Vera Rubin, it is still a bit early. We still need to run an analysis on TCO and what we should be doing.
Therefore, in terms of what we are looking at, we will postpone it a little later. However, when it comes to Blackwell and what we are looking at, we can say that we are in a unique position now.
Tim Archulee
How was it possible to raise gross profit so much? I remember that the gross profit was in the mid-60s when Hopper was first released, but now it's approaching the mid-70s. So, how was the gross profit increased to this extent?
Collette Cress
Understood. When determining the value provided to customers, various factors are involved. It's not just about performance. It's not just about the chip's performance. It's important to find out what can be done to provide an end-to-end solution and the minimum total cost of ownership that customers can decide. This helps to determine how to enter the market and use a specific price for the product.
This total cost of ownership (TCO) value essentially looks at the entire end-to-end view. How to complete the software? How to complete the entire datacenter architecture? Are other teams necessary? It's not just about looking at various components. It's also not a situation of components and cost-plus models.
Therefore, strong performance, strong efficiency, and the best TCO will allow for a complete TCO that incorporates all software provided internally in the system and supporting the overall lifespan of the system.
Tim Archulee
Understood. Can you talk about the networking business? There have been many questions about it. It declined in the previous quarter while an increase was expected. Some people also believe that the networking business's position weakened slightly due to the transition from InfiniBand to Ethernet. Can you discuss the networking business? Do you expect it to grow alongside the computing business?
Collet Kress
Our network business is one of the most important businesses added when transitioning to data center scale. The ability to consider not only the time when data and work are processed but also the use of computing and GPUs is essential when considering the positioning of the network within that data center.
We have two different products. InfiniBand. InfiniBand has been very successful in many of the world's largest supercomputers over the past decades. This was very important in terms of data size and data throughput speed. There were different views on how to handle the traffic that occurs there.
Ethernet is an excellent configuration that has become a standard for many companies. However, Ethernet is not built for AI. Ethernet is only built for networks within data centers. Therefore, we are incorporating some of the excellent points of inter-InfiniBand into an Ethernet for AI. This allows customers to have both options. Customers can build a complete end-to-end system with InfiniBand and choose what to do with Ethernet.
Both of these are growth options for us. There were timing issues for this quarter. However, now, in terms of network continuity, what everyone will see will definitely grow. Our designs for computing and networks have led to some of the most powerful clusters built using our network. The connections we made have been a very important part of the work we have done since the acquisition of Mellanox. People understand and appreciate how our network usage contributes to the entire system.
Tim Archuleta
Great, can you talk about the scaling of these large language models? There have been several articles mentioning that Google and OpenAI are struggling to get better results from these large models. However, on the other hand, Meta attended the earnings call and other companies like Anthropic mentioned that the scaling is progressing smoothly.
So, can you speak about that point? We are aware that there are subtle differences after training, and of course, there are also differences in test time for some of OpenAI's new models. Should scaling issues be considered by investors?
Colette Cres
As we observe the size of the cluster under construction and the tasks many customers are attempting to execute, the scaling laws, particularly with regard to training, are still in effect. We expect to see larger and more complex models in this next generation of Blackwell. In other words, the stage after training will be revived through reinforcement learning, truly seeking the human element, and the stage of fine-tuning the model using synthetic data will be revived.
In other words, training never truly ends and many tasks continue. However, new skill methods focusing on the inference stage have also been established. As a reminder, we are currently the largest inference provider in existence, performing more inferences than any other type of configuration. Why is that? Because it's extremely challenging.
What we are seeing from scaling is a crucial part from the beginnings of generative AI to the present, with an increasing focus on the inference for deep thinking and the time required to do so. This necessitates additional computing power and the ability to execute computing with minimal latency relative to the time spent on inference elements.
Therefore, the laws of scaling remain important, and it is expected that many more new laws will be formulated over the next few decades.
Tim Archule
You are talking about a new aspect of demand, namely government-supported projects, the so-called Sovereigns. You mentioned planning to invest a 1 billion-dollar double-digit amount across these projects this year, could you provide some examples of where this demand is coming from? Also, could you share how large the demand could potentially become? I also think that some of these large-scale projects in the Middle East could eventually match the scale of the USA's BSP.
This has the potential to lead to very significant demand, how do you think we can achieve that?
Collette Cress
Sovereign AI, in terms of generative AI, is a very interesting part that we have seen. In simple terms, what they have seen here is what is seen in the usa, and all countries have an obligation to see it when I desire. And they want models and basic models based on their language, their culture to support their own country. They recognize the importance of what AI will become in the coming decades.
Therefore, the activities in various countries or specific regions that we are collaborating with account for a very large part of the global efforts we are making in the world of AI. This is not just on the west coast of the usa. It is happening worldwide. Not all, but only a part are receiving government funding. Many of them are focused on launching new regional CSPs that can support accelerated computing and have a set of principles. And I think we'll be working with them to build the foundational models to support businesses.
We have already discussed what is happening in Japan. SoftBank is very interested in building very large models in certain areas. In India, many CSPs are working on what to incorporate.
This extends to Europe, spreads to the Nordic countries, and plays a very important role in the Asia-Pacific region. Therefore, thinking about the next generation of supercomputing that we have seen in each country will continue. We will understand what needs to be done for AI and sovereign AI from the GDP.
Tim Archuleta
That's great. Could you talk a little about the software business? You mentioned exceeding an annual $2 billion run rate this year. I'll roughly calculate how many GPUs directly generating revenue are modified.
The attachment rate is about 10%, and about 10% are directly licensed. Could you share your thoughts on the attachment rate and how successful you are in directly licensing GPU software?
Collette Cress
Yes. Our software platform is essential for our many corporate and regional CSPs, future AI factories, and the AI foundation work we are undertaking. Why? As the first step in understanding how to transition to AI and how to start, we have thousands of different applications, CUDA libraries, CUDA operations performed for each industry, and the major workloads within those industries.
Companies need that part not only for the work they need to do themselves, but also for the work needed to support the data center infrastructure. Therefore, we build that part for the companies, and often, the customers of the companies have a strong attachment to that software because they need that part for the work they are doing.
Companies with very large software teams are different from companies that I have built on my own over the decades. However, as you can imagine, we cannot go back to the era of training all software engineers. Therefore, we have spent a lot of high-quality time to support many companies.
We support companies from choosing what kind of computing to do, through model creation, configuring various apps, to all aspects of overall inference, from start to finish. In other words, it's not just the software that's involved. That software comes with true support and services from the companies.
Tim Arkale
So, we looked only at the enterprise market, but is there a possibility that your attachment rate could actually be much higher than that?
Collette Cress
That's right. That's correct.
Tim Arkuri
Understood. That's great. We all talk about datacenters, but inference at the edge will be a bigger theme. Could you tell us about your position at the edge? PCs have a large installed base, which should be quite advantageous for you, and there's also Omniverse. Robotics is a big theme. So, could you talk about what you offer and how to think of yourself as a player at the edge?
Colette Crest
Yes, edge computing and edge AI are very similar and will probably continue to exist in conjunction. That means factories are a thing. Data will be collected in data centers and made available to many edge appliances, such as automobiles. Self-driving cars.
The next phase will probably be about robotics. Data and learning take place in data centers, and our functions support a very large industry taking place inside various devices. This is an important industry. We know that the data center portion is a very large market and crucial for future developments.
This includes the new software we are developing. As you know, we are developing software for self-driving cars that will be launched in the market in the second half of next year. Furthermore, within robotics, tasks that can be performed with that software, tasks that can be executed in many factories using Omniverse, and overall layout of how it functions.
Therefore, these are areas that should receive a lot of focus beyond traditional data centers. However, edge computing will also be an important element.
Tim Archuleta
That's great. Could you please talk a bit about inference and training? You mentioned that inference accounts for about 40% of revenue, how do you think it will evolve in the future?
Colette Cress
It's about 40%. We mentioned starting to think about the use cases of what they're doing. It's evident that a lot of time is being spent on inference. This is before many generative AI applications being in the worst state to launch into the market.
The recommendation engine plays a crucial role in today's inference. Therefore, we may expect to see 40% growth in the future. However, as mentioned earlier, we are still the major player in inference. When considering the Blackwell architecture, especially GB200, and NVL, this is a significant configuration, and the inference performance has improved 30 times compared to our current generation.
This is a very important aspect for many customers. Customers will likely use it when initially building what's necessary for the base model. However, important aspects of advancing inference with Blackwell have been receiving high praise from many customers.
Tim Archuleta
That's great. And then, you mentioned about easing the constraints of Blackwell around the mid-25s, would it be correct to assume that they will begin to relax by mid-25s? Could you talk about those constraints?
Colette Cress
When I think about collaborating with customers on Blackwell's manufacturing, design, and composition, the demand for Cane's Fast and Furious is outstanding. We work with many partners and discuss with suppliers every day to support them. Therefore, at this point, we need to expand the scale to manufacture enough Blackwell to meet the current demand.
And perhaps in the first part of the new fiscal year, supply will be limited. Where will the limitations be? It varies depending on the composition, but part of the challenge is the work on the shared space. All tasks that need to be done for various configurations and the network switch must be done correctly. Depending on the configuration, there may be supply limitations. However, we are planning to ship Blackwell soon in this quarter. Blackwell is progressing well, and we are excited to offer multiple configurations to customers this quarter.
Tim Arcuri
Well, I think next year will definitely be a great year. Anyway, Colette, thank you. I believe I've been appreciated.
Colette Cress
Yes. Understood. See you.
NVIDIA Corporation (NASDAQ: NVDA )
UBS Global Technology Conference Call
9:35 AM (Eastern Standard Time) on December 3, 2024
Company participants
Collette Kress - Executive Vice President and Chief Financial Officer
Phone conference attendees
Tim Archule-UBS
Tim Archule
Good morning. Thank you. I'm Tim Archule. Thank you for coming to NVIDIA. Colette is also here. Thank you for your time, Colette. Let's get started. Colette, these past 2 years have been fantastic. Colette, you are still growing rapidly. Can you talk about how some of the use cases have evolved? Companies are widely adopting AI. From the perspectives of cloud, consumer, internet, and enterprise, how has the demand evolved? Also, could you talk about some very exciting use cases.
Colette Cress
Understood. Let me start with the statement I need to read. Just a reminder, this presentation contains descriptions related to future estimates, and investors are advised to read our reports filed with the SEC to obtain information about the risks and uncertainties our business faces.
It's truly a pleasure to meet all of you. Thank you so much for having us. Let me talk a little about what we've seen. Even just in the past few quarters, it has certainly been a very fast journey. But please remember our business has a 30-year history. However, we are certainly at a very important stage.
Thinking about the stage we are at now and what we are seeing, many of us who have used and seen the computing platforms for 20 to 30 years believe that the computing platforms we have used will undergo transformation over the next several decades. This means there are people focusing on existing computing platforms, such as the transition to general-purpose computing or accelerated computing.
In addition to focusing on accelerated computing, a new element of focusing on AI and Generative AI has also been added. Of course, the new use cases we have seen over the past several quarters mainly involve model size. If you remember, we were talking about large language models just before Generative AI appeared.
The importance of consumer internet companies, recommendation engines, and work related to model size is becoming increasingly important. Currently, work on foundational models is crucial in terms of the ongoing projects. However, due to the emphasis on building various types of foundational models and multimodal models, scaling of these models is also observed.
The next stage of this transition also focuses on the inference phase. After developing large language models, the work on supporting inference models is increasing. All our different types of customers are unique. We collaborate with all types of companies from startups to research. CSP is a major part of launching high-speed computing for end-user companies in need. While we collaborate with consumer internet companies, it is even more important that we are globally engaged in supporting this initiative.
It was something unprecedented in terms of speed and understanding of what is to come. This is part of what we are seeing today.
Tim Archuri
We are all concerned that ultimately we will end up building excess production capacity. And your opinion, I think, suggests that we are still far from it. So, could you talk about to what extent you have a perspective, and also about the demand situation in relation to the available supply?
Colette Crest
Yes, looking at the past quarters and our scaling, our primary goal regarding scaling is to collaborate with all partners, downstream customers, and bring about supply. No, at this point, we do not see a situation indicating a slowdown. However, it is currently confirmed that due to the size of the models and the complexity of inference, the demand continues to increase significantly, and we are preparing for the next architecture, Blackwell.
Regarding the upcoming Blackwell in the current quarter, it appears that we are likely experiencing supply constraints, which are expected to continue until the next fiscal year. Therefore, we do not expect a slowdown. In particular, demand and interest for the new architecture to be introduced in the future continue to be very high.
Tim Archule
Indeed, it was a great quarter. We shipped Blackwell in January and have shipped more Blackwell than we expected three months ago. At the same time, I hear that there is a lot of disruption in the supply chain. Many articles have been written about that. Could you tell us how Blackwell differs from the previous product cycle?
Collette Cress
Yes, our Blackwell architecture is unique. However, what we are doing here is building at the data center scale. Please do not misunderstand. We have been focusing on the platform perspective for many years. It's not just at the chip level or end-to-end scale. But when you look at our Hopper architecture, you can see that our Hopper architecture is for rack scale and the work we have been doing.
Essentially, some of what you see with the Blackwell architecture is fully built within our supply chain. Are you ready? We run it. We remove it. We install it, then turn it on together. And depending on which stage of the lifecycle you are in, we provide more options to our customers.
In other words, data centers are complex, and various things are being done to make them more efficient. Each one is at different stages. Therefore, you can choose from various options we offer. That means liquid cooling is possible. Air cooling can also continue. You can incorporate ARM CPUs and x86 as needed.
We have various network options such as InfiniBand and Ethernet, as well as various switch options. The choice is up to the customer, and there are various configurations available. Considering the current situation of Blackwell this quarter, everything related to the chips is complete.
The chip is doing well. The work we did with the chip is progressing smoothly. Currently, we are setting up configurations for various customers. Everyone will see photos on the internet. The first configuration is excited to start, happy faces preparing to assemble the entire data center and various racks. This is our current situation.
Tim Archule
That's great. There are various product release spaces, B100, B200. There are racks with CDs. And in about 6 months, Blackwell Ultra will be introduced. Are there any risks for customers waiting for products to be released this quickly?
Colette Cress
No, when considering what is needed for data center design, planning is essential. Over the past 5 to 10 years, we have been working with various customers on more projects. Customers plan every 6 months. What's here? What are we building? It is necessary to have computing ready when working on projects. Therefore, considering the shortage of supplies, we continue to offer customers the excellent Hopper 200 configuration. This is an opportunity to start some of the work being done in the HGX system and GPU demand is extremely high, so we have not touched the GPU yet. People understand how to build data centers by working with us. Data centers are often already procured. They just know what needs to be placed inside. So, there are two architectures.
As you know, we will also implement Vera Rubin in the future. We will discuss this with customers as well, but first, let's talk about potential services. How can we assist in understanding what might be needed in the future?
Tim Archule
However, this seems a bit similar to when customers did not have to wait for successful products like H100 and H200. Does this mean there is evidence that customers did not have to wait immediately after the launch of Vera products?
コレット・クレス
顧客は待ち望んでいます。待つというのは強い言葉です。顧客は毎日電話をかけてきて、いつコンピューティングが見られるのか尋ねており、私たちはその新しい供給のために必死に取り組んでいます。しかし、重要なのは、これはおそらく20年かかる旅だということを忘れないことです。誰もが参加するでしょう。私たちのアーキテクチャは、アクセラレーテッドコンピューティングとAIの世界で必要なことを何でもできるように、エンドツーエンドのスケーリングアプローチを可能にします。そして私たちは、インフラストラクチャだけでなくソフトウェアでも顧客を支援するための非常に有力な候補です。
ティム・アーキュリ
ありがとうございます。それで、キャッシュについてお聞きしたいのですが、実は昨夜このことについて話し合いました。過去 4 四半期で約 56 billionドルのキャッシュフローを生み出しました。25 暦年には約 120 billionドルを生み出すと聞いています。年間約 50 billionドルの自社株買いをすると仮定しても、来年末には 100 billionドルのキャッシュが残り、2026 年末には 200 billionドルのキャッシュが残る可能性があります。これは明らかに莫大な数字で、100 billionドルのキャッシュがあった Apple でさえ、最終的にはその額を減らしました。キャッシュをどう使うつもりですか?
コレット・クレス
現金に関する当社の取り組みと現金の使用は、どの企業にとっても最も重要な事項の 1 つです。当社は、戦略とニーズを通じて物事を手助けしてくれる優秀な人材と資金を費やしています。まず検討すべきことは、イノベーションと、当社が行っている拡大に向けたあらゆる支援に必要な現金がいくらかということです。これが最初の部分になります。しかし、当社は継続しており、イノベーションと、R&D で行っているすべてのことが、その重要な部分になることを知っています。
今、私たちは企業から学んだことの両方からそれを実行できます。また、M&A の形で優秀なチームを招き入れるという私たちの仕事の観点からもそれを考えることができます。これは私たちにとって素晴らしい機会であり、私たちはその分野でも引き続き取り組んでいきます。そして、それは AI の新しい分野に追加して重点的に取り組みたい新しいタイプのビジネス モデルについて考えることにつながります。
そして、私たちができるサポートは、ソフトウェアの構築だけではなく、他の企業向けの完全なシステムの構築であり、私たちはそれに投資します。こうした種類の投資を決定した後、私たちの仕事は、株主への還元という点において常にそうであり、それに対する私たちの焦点は、自社株買いと配当の組み合わせであり、皆さんはこれを引き続き目にすることになるでしょうし、これを注意深く見守ることになるでしょう。
We do not prefer excess cash. Therefore, we will carefully monitor it and intend to continue this in the next quarter.
Tim Archule
That's great. Thank you. I briefly talked about gross profit, and as Blackwell's performance expands, it was mentioned that the gross profit will decrease in the first quarter of next year. It was explained that it will rise to the low 70 percent range, then during the conference call, it will increase from 71 percent to 72.5 percent. However, it will rise again afterwards, reaching the mid-70 percent range by the end of the year.
Could you please tell us how confident you are about this point? After Rubin comes into play, some people believe that the gross profit margin will be squeezed again as Rubin begins to grow. Can you talk about the confidence in maintaining the mid-70s in the long term?
Colette Cress
Understood. Blackwell is unique in various aspects. Just this quarter alone, they are preparing various versions to be introduced to the market. It will not be just one version being shipped. Multiple versions are planned for shipment. Therefore, the current shipment volume is very low.
By expanding this throughout the year and embarking on the expansion of various system configurations that the company holds, we will be able to improve the gross profit margin. However, when thinking about the future prospects and Vera Rubin, it is still a bit early. We still need to run an analysis on TCO and what we should be doing.
Therefore, in terms of what we are looking at, we will postpone it a little later. However, when it comes to Blackwell and what we are looking at, we can say that we are in a unique position now.
Tim Archulee
How was it possible to raise gross profit so much? I remember that the gross profit was in the mid-60s when Hopper was first released, but now it's approaching the mid-70s. So, how was the gross profit increased to this extent?
Collette Cress
Understood. When determining the value provided to customers, various factors are involved. It's not just about performance. It's not just about the chip's performance. It's important to find out what can be done to provide an end-to-end solution and the minimum total cost of ownership that customers can decide. This helps to determine how to enter the market and use a specific price for the product.
This total cost of ownership (TCO) value essentially looks at the entire end-to-end view. How to complete the software? How to complete the entire datacenter architecture? Are other teams necessary? It's not just about looking at various components. It's also not a situation of components and cost-plus models.
Therefore, strong performance, strong efficiency, and the best TCO will allow for a complete TCO that incorporates all software provided internally in the system and supporting the overall lifespan of the system.
Tim Archulee
Understood. Can you talk about the networking business? There have been many questions about it. It declined in the previous quarter while an increase was expected. Some people also believe that the networking business's position weakened slightly due to the transition from InfiniBand to Ethernet. Can you discuss the networking business? Do you expect it to grow alongside the computing business?
Collet Kress
Our network business is one of the most important businesses added when transitioning to data center scale. The ability to consider not only the time when data and work are processed but also the use of computing and GPUs is essential when considering the positioning of the network within that data center.
We have two different products. InfiniBand. InfiniBand has been very successful in many of the world's largest supercomputers over the past decades. This was very important in terms of data size and data throughput speed. There were different views on how to handle the traffic that occurs there.
Ethernet is an excellent configuration that has become a standard for many companies. However, Ethernet is not built for AI. Ethernet is only built for networks within data centers. Therefore, we are incorporating some of the excellent points of inter-InfiniBand into an Ethernet for AI. This allows customers to have both options. Customers can build a complete end-to-end system with InfiniBand and choose what to do with Ethernet.
Both of these are growth options for us. There were timing issues for this quarter. However, now, in terms of network continuity, what everyone will see will definitely grow. Our designs for computing and networks have led to some of the most powerful clusters built using our network. The connections we made have been a very important part of the work we have done since the acquisition of Mellanox. People understand and appreciate how our network usage contributes to the entire system.
Tim Archuleta
Great, can you talk about the scaling of these large language models? There have been several articles mentioning that Google and OpenAI are struggling to get better results from these large models. However, on the other hand, Meta attended the earnings call and other companies like Anthropic mentioned that the scaling is progressing smoothly.
So, can you speak about that point? We are aware that there are subtle differences after training, and of course, there are also differences in test time for some of OpenAI's new models. Should scaling issues be considered by investors?
Colette Cres
As we observe the size of the cluster under construction and the tasks many customers are attempting to execute, the scaling laws, particularly with regard to training, are still in effect. We expect to see larger and more complex models in this next generation of Blackwell. In other words, the stage after training will be revived through reinforcement learning, truly seeking the human element, and the stage of fine-tuning the model using synthetic data will be revived.
In other words, training never truly ends and many tasks continue. However, new skill methods focusing on the inference stage have also been established. As a reminder, we are currently the largest inference provider in existence, performing more inferences than any other type of configuration. Why is that? Because it's extremely challenging.
What we are seeing from scaling is a crucial part from the beginnings of generative AI to the present, with an increasing focus on the inference for deep thinking and the time required to do so. This necessitates additional computing power and the ability to execute computing with minimal latency relative to the time spent on inference elements.
Therefore, the laws of scaling remain important, and it is expected that many more new laws will be formulated over the next few decades.
Tim Archule
You are talking about a new aspect of demand, namely government-supported projects, the so-called Sovereigns. You mentioned planning to invest a 1 billion-dollar double-digit amount across these projects this year, could you provide some examples of where this demand is coming from? Also, could you share how large the demand could potentially become? I also think that some of these large-scale projects in the Middle East could eventually match the scale of the USA's BSP.
This has the potential to lead to very significant demand, how do you think we can achieve that?
Collette Cress
Sovereign AI, in terms of generative AI, is a very interesting part that we have seen. In simple terms, what they have seen here is what is seen in the usa, and all countries have an obligation to see it when I desire. And they want models and basic models based on their language, their culture to support their own country. They recognize the importance of what AI will become in the coming decades.
Therefore, the activities in various countries or specific regions that we are collaborating with account for a very large part of the global efforts we are making in the world of AI. This is not just on the west coast of the usa. It is happening worldwide. Not all, but only a part are receiving government funding. Many of them are focused on launching new regional CSPs that can support accelerated computing and have a set of principles. And I think we'll be working with them to build the foundational models to support businesses.
We have already discussed what is happening in Japan. SoftBank is very interested in building very large models in certain areas. In India, many CSPs are working on what to incorporate.
This extends to Europe, spreads to the Nordic countries, and plays a very important role in the Asia-Pacific region. Therefore, thinking about the next generation of supercomputing that we have seen in each country will continue. We will understand what needs to be done for AI and sovereign AI from the GDP.
Tim Archuleta
That's great. Could you talk a little about the software business? You mentioned exceeding an annual $2 billion run rate this year. I'll roughly calculate how many GPUs directly generating revenue are modified.
The attachment rate is about 10%, and about 10% are directly licensed. Could you share your thoughts on the attachment rate and how successful you are in directly licensing GPU software?
Collette Cress
Yes. Our software platform is essential for our many corporate and regional CSPs, future AI factories, and the AI foundation work we are undertaking. Why? As the first step in understanding how to transition to AI and how to start, we have thousands of different applications, CUDA libraries, CUDA operations performed for each industry, and the major workloads within those industries.
Companies need that part not only for the work they need to do themselves, but also for the work needed to support the data center infrastructure. Therefore, we build that part for the companies, and often, the customers of the companies have a strong attachment to that software because they need that part for the work they are doing.
Companies with very large software teams are different from companies that I have built on my own over the decades. However, as you can imagine, we cannot go back to the era of training all software engineers. Therefore, we have spent a lot of high-quality time to support many companies.
We support companies from choosing what kind of computing to do, through model creation, configuring various apps, to all aspects of overall inference, from start to finish. In other words, it's not just the software that's involved. That software comes with true support and services from the companies.
Tim Arkale
So, we looked only at the enterprise market, but is there a possibility that your attachment rate could actually be much higher than that?
Collette Cress
That's right. That's correct.
Tim Arkuri
Understood. That's great. We all talk about datacenters, but inference at the edge will be a bigger theme. Could you tell us about your position at the edge? PCs have a large installed base, which should be quite advantageous for you, and there's also Omniverse. Robotics is a big theme. So, could you talk about what you offer and how to think of yourself as a player at the edge?
Colette Crest
Yes, edge computing and edge AI are very similar and will probably continue to exist in conjunction. That means factories are a thing. Data will be collected in data centers and made available to many edge appliances, such as automobiles. Self-driving cars.
The next phase will probably be about robotics. Data and learning take place in data centers, and our functions support a very large industry taking place inside various devices. This is an important industry. We know that the data center portion is a very large market and crucial for future developments.
This includes the new software we are developing. As you know, we are developing software for self-driving cars that will be launched in the market in the second half of next year. Furthermore, within robotics, tasks that can be performed with that software, tasks that can be executed in many factories using Omniverse, and overall layout of how it functions.
Therefore, these are areas that should receive a lot of focus beyond traditional data centers. However, edge computing will also be an important element.
Tim Archuleta
That's great. Could you please talk a bit about inference and training? You mentioned that inference accounts for about 40% of revenue, how do you think it will evolve in the future?
Colette Cress
It's about 40%. We mentioned starting to think about the use cases of what they're doing. It's evident that a lot of time is being spent on inference. This is before many generative AI applications being in the worst state to launch into the market.
The recommendation engine plays a crucial role in today's inference. Therefore, we may expect to see 40% growth in the future. However, as mentioned earlier, we are still the major player in inference. When considering the Blackwell architecture, especially GB200, and NVL, this is a significant configuration, and the inference performance has improved 30 times compared to our current generation.
This is a very important aspect for many customers. Customers will likely use it when initially building what's necessary for the base model. However, important aspects of advancing inference with Blackwell have been receiving high praise from many customers.
Tim Archuleta
That's great. And then, you mentioned about easing the constraints of Blackwell around the mid-25s, would it be correct to assume that they will begin to relax by mid-25s? Could you talk about those constraints?
Colette Cress
When I think about collaborating with customers on Blackwell's manufacturing, design, and composition, the demand for Cane's Fast and Furious is outstanding. We work with many partners and discuss with suppliers every day to support them. Therefore, at this point, we need to expand the scale to manufacture enough Blackwell to meet the current demand.
And perhaps in the first part of the new fiscal year, supply will be limited. Where will the limitations be? It varies depending on the composition, but part of the challenge is the work on the shared space. All tasks that need to be done for various configurations and the network switch must be done correctly. Depending on the configuration, there may be supply limitations. However, we are planning to ship Blackwell soon in this quarter. Blackwell is progressing well, and we are excited to offer multiple configurations to customers this quarter.
Tim Arcuri
Well, I think next year will definitely be a great year. Anyway, Colette, thank you. I believe I've been appreciated.
Colette Cress
Yes. Understood. See you.
Disclaimer: Community is offered by Moomoo Technologies Inc. and is for educational purposes only.
Read more
Comment
Sign in to post a comment
