After Wall Street closed the markets for the day and Nvidia announced its financial results for the second quarter of fiscal 2025, we had the opportunity to speak with Colette Kress, CFO of the accelerator computing giant.
We wanted to better understand how the Blackwell GPU delays are affecting Nvidia’s finances, how the company’s software business differs from that of other platform providers, how Nvidia is managing exponential growth, and how the generative AI revolution could expand the world’s IT budgets, but also cannibalize some of them.
Timothy Prickett Morgan: I have limited time for you, so I want to jump right in. How much did the Blackwell redesign displace the Blackwell ramp and how much revenue, if any, shifted from fiscal 2025 to fiscal 2026? My guess is that at some level it doesn’t matter because Hopper can be made at a higher margin and people have to buy two of them for every Blackwell, and they’ll buy what they can get today.
Colette Kress: In reality, nothing has changed or impacted anything because the demand for systems and GPUs is still there to continue the work they do to build generative AI and accelerated computing.
We haven’t reached all the companies we need to reach. If you keep going to your top cloud service providers, you’re not going to find anything you can get. They’ve sold everything. Everyone else is using it. That’s why this transition from Hopper to Blackwell is important, because Blackwell is great, but they’re still buying because if they don’t, they’re behind. People still have to make that journey, and that’s why Hopper will continue.
TPMs: One more clarification. When you refer to the CFO’s commentary on “several billions” of revenue generated by Blackwell in the fourth quarter of the fiscal year, I assume that by “several” you mean certainly more than $2 billion and at least $3 billion. . . .
Colette Kress: I have this rule. When I say single, that’s one. And a couple is two, like when you got married or you have two cookies. Several is more, and you got that right.
TPMs: When I spoke to Jensen a few years ago, he reminded me, or admonished me, that Nvidia is not really a hardware company. It’s a software company, and 75 percent of the people at Nvidia work on software. Of course, there are some people in management, but the rest work on hardware.
I annoyed him about it then, and I’m going to annoy you about it now, because if you had, that would be your job.
Can you extract the software value from the platforms and more accurately reflect the inputs and outputs of your software efforts? I know you’re building a software business separately from that, but you know that in the relational database world, for example, the hardware costs x and the database and middleware license costs 10x. Nvidia does the opposite.
They’ve been giving away software as part of the hardware for a long time and now they’re building this additional revenue stream. Will Nvidia eventually turn into a software company that also makes hardware? Will it eventually be a 50:50 split or something like that?
Colette Kress: That’s a really good summary. And let me add a little more color to the situation.
The transition to accelerated computing and AI is not a situation where you turn on the hardware, load your stuff on it, and it’s running. The software we’re building should be paving the way for companies to accelerate computing as opposed to their initial general-purpose computing. They need a different path. The majority of our software – when we say we’ve given it away – should be supporting that transition to accelerated computing.
The reason we do so much software development is because they have to redesign workloads to move to accelerated computing. That’s why we have so many that stitch that together for customers. That’s why it’s so difficult to compete. We have a competitor that says, “Here’s my chip,” and a company will say, “What am I supposed to do with it? I don’t have anything around it. How do I even bootstrap this?”
Our software is so important because you have to adapt to the changes in the workflow. If it was going directly to the CPU, you have to redirect it to use the GPU.
Our process now is that we continue to work on helping every industry – top workloads, top applications – move to accelerated computing. Now we have a different situation: customers are working on generative AI, they have a model and they need help to evolve that model to make it work perfectly, and they need to deploy services to make sure it has the right security, to make sure it has the right level of permission, the right level of overall operation.
Software is going to be a big part of Nvidia, whether we sell it or give it away for free to help them transition. Companies don’t necessarily write their own software. There are some big companies that do that, but most buy it from others. In order for them to use those GPUs and AI in mission-critical applications, they need to make sure they have someone managing that software. That’s us. We need to manage it. We need to keep it secure. We need to continually update it.
TPMs: Exponential growth can’t last forever. We know that, but it can last a long time. When I look at Nvidia’s growth over the last few years, driven by accelerated computing in general and generative AI in particular, it’s easy to get excited about exponential curves, but there’s also a desire to have realistic ideas about how long it can last.
How do you plan as a CFO in this kind of climate? Do you plan exponentially at this point and hope for the best? Because it’s very difficult to predict how this is all going to play out. This is a unique situation.
Colette Kress: Most of the planning for a company this size is planning our capital to design more products and build more products for all the different areas of our business.
You’re right, I’m always looking ahead to make sure our planning assumes the best of what I think is possible without going in the other direction – and to move fast enough. Are we OK? We don’t waste a day, we don’t waste a penny. And those two things are really important for us: our agility to move fast, but our ability not to waste anything along the way.
We’re still in the early stages of this journey to accelerate computing. We’re not going to do all of this in two years. These transitions require two, three decades of work to move to accelerated computing and integrate AI into everything we do.
Do we know what’s going to happen every day? Is everything going to grow at the same beautiful speed? No. But we know it’s going to be with us for decades. And if you hold onto that understanding and that vision and just keep planning to make sure you can do it fast and meet the expectations of how fast things need to move, then that’s all we’re going to do in terms of planning. And we’re planning a lot.
TPMs: This is a related question. When I got into this IT business nearly four decades ago, most companies’ IT spending was 1 to 2 percent of revenue. It fluctuated a bit by industry and company size — larger companies spent more, different industries like computer services and financial services spent more. During and after the dot-com boom, average IT spending rose to 4 to 6 percent of company revenue.
As I reflect on this GenAI wave, I wonder if this new workload will force companies to spend 8 or 10 percent of their revenue on IT budgets, thereby increasing overall IT spending faster than expected, or will it simply cannibalize current spending?
Colette Kress: So it’s a bit of both: you make the investment now, even though you know it’s not an instant solution.
There’s work involved in getting the generative AI application ready and getting it to where it’s going to be used. Every day, more data is coming in and they’re going to work on it. So you’re going to make sure you move faster because if you don’t, you’re going to get left behind. That’s one part.
But the question is: do I just do more? Yes, but you’ll find that this is your most important work and you’ll invest less in the other areas that you invest in. I can invest in that, but that won’t give me the productivity boost that AI can do for me.
For example, general purpose X86 CPU computing is not necessarily going to give me a great return on investment, OK? So do I need to refresh that or do I just leave them plugged in and they’ll probably work fine for a few more years and focus my allocation of how I invest my time and money on what’s coming in the next two decades rather than what’s already in the past?
So there will be two things: more, because it will be more productive and efficient for them, and on the other hand, you will equally stop doing things that do not contribute to productivity.
