联想CFO谈AI资本配置：工程师月耗1亿美元Token

[music]

Bloomberg Audio Studios. Podcasts, radio, news.

[music] [music]

Hello and welcome to another episode of the Odd Lots podcast. I'm Tracy Alloway.

And I'm Joe Weisenthal.

So Joe, we are here in Hong Kong still. And we're learning a lot of different things, a lot of interesting things. One of the most interesting things we heard actually came yesterday from the Baidu CFO, where we were just talking casually ahead of our interview, but he was saying that the word token

Yeah.

has now officially been added to the Chinese Mandarin dictionary.

And that the characters that actually make up the Chinese word for token are something like word currency.

I think that's so fascinating. You know, I'm fascinated by the etymology of the word token specifically. So I was thrilled to hear that. But it is, you know, it's like we know the word token in the monetary context. We know Chucky Cheese tokens. We know [laughter] crypto tokens. But we also know that since the middle of the 20th century, linguists have been using token to describe more or less a word. And then obviously with LLMs, we talk about these linguistic tokens a lot.

So to see that in Chinese, the formal term is a merger of these two concepts is I find a very intellectually satisfying thing to learn about.

Bringing up Chucky Cheese tokens is really a way to make sure everyone knows that you're a millennial, Joe.

Yeah, that's right. That's right. Arcade I I should say arcade token.

That's right.

Yeah.

Okay. And I think what's really interesting about word currency itself is it implies that it's connected to spending, right?

And so when you think about the big companies that are spending all this money on tokens and I guess the AI infrastructure build out more broadly, CFOs become really important, right?

Totally. So, we, you know, the headlines by and large are about uh CapEx, right? And that's going to continue to be because of the data center build out. We're going to talk about that. But a lot of it is also going to be OpEx and figuring out how to within a company allocating token permissions and caps and so forth and I doubt anyone has figured out the final answer. But if two different people can get different value out of using AI models, then there is no way that it makes sense for them to have the same token budgets.

By the way, Tracy, can I ask you a personal question that I've never asked you before?

Oh, okay.

Outside of the work context, are you a Mac or a PC person?

Oh, I mhm, I only have I only have work computers at the moment.

Okay. All right, before that, but

Before that, before that, definitely PC.

Okay, good. Me, too.

Yeah.

And in fact,

it's not it's not a choice, is it? Like, I Sorry, I I really don't like Macs.

I don't, either. And in particular, I am long been a fan of the what used to be the IBM ThinkPad laptop with the famous red button, which is now owned by Lenovo.

That's right. You've talked about that computer before, so I can say that this is categorically true. Joe likes that computer. But what's interesting about Lenovo is like, okay, it's famous for the computer with the little red button in the middle, but it's now making an AI play. I mean, everyone's making an AI play, but it's doing it from a different perspective. So, AI integrating into the actual computer, the hardware, but it's also doing cloud, right?

So, this is a really good opportunity to, I guess, take the temperature on AI, the AI build out, the AI spend from a bunch of different perspectives. So, we do in fact have the perfect guest. We're going to be speaking with Winston Shang. He is the CFO of Lenovo. So, Winston, thanks so much for coming on Odd Lots.

Hey, thank you, Tracy, and thank you, Joe.

Why don't you go ahead and describe what Lenovo is and what the AI play is and how it actually fits into the existing business?

Yeah, Lenovo is a global AI infrastructure company that provides pocket to cloud AI infrastructure for the consumer and the enterprise. I think that's really in a nutshell. And so, today we're able to provide this to hyperscalers, which are doing a lot of the spending, are driven by training demand today. And then, given our IBM x86 heritage, which we also acquired the server from the IBM actually. So, we actually are very strong in CPU compute as well, given that IBM was a dominant player in the x86 architecture.

So, from that perspective, we're well positioned for inferencing needs of the enterprise and also the hyperscalers in terms of in the cloud as well. So, I think from the perspective of then, you talked about tokens. And I think from a token perspective, really today people are in the early stage of how much I'm really paying for. I heard someone saying that there was an engineer at a particular company, which I will not mention, that apparently spent $100 million in a month on tokens.

And so, as a CFO, I would have concerns cuz that was clearly not in the budget, right? Not saying that that was from Lenovo. So, we cannot have that happen. And I think we need to be able to drive the productivity or efficiencies as it relates to that budgeting of the token generation. And I think a lot of that will happen on device, where you may just pay a higher price for a device, but you know what you're going to be able to do on compute for the security of the data and for the privacy that you want to interact with your AI agent.

So, just to be clear, for the computers themselves, they can do some inference, right? But where it makes sense, you route it to the cloud. Is that right?

That that is our goal. So, the Lenovo Agentic AI today we what we are good at is really integrating and maximizing the compute capabilities on a device. And from that perspective, given the various needs in terms of various operating systems, agents that may not want to sit on top of a device, I think our agent aims to orchestrate the various LLMs. We'll take the compressed versions of these LLMs. We will, depending on the partnership, be able to do the on-device compressed LLMs versions and do that as a local compute, but in certain queries, allow it to go on the cloud.

And therefore, probably would allow the user probably to spend in terms of the token generation.

It's interesting this word orchestrate because you so you have the AI agent orchestrating maybe a bunch of different sub-agents to complete some task, and that is something that perhaps is done best on a CPU. A company that builds servers is also an orchestrator of the supply chain and acquiring the different components that go into a server, etc. I want to go back to something you said in your first answer because I think this actually will get to the core of this new era.

You said, "Okay, an engineer spends $100 million in a month on tokens." And it's like, that would not make you happy as a CFO. But it could make you happy, right? What if you had a multi-year database migration plan that you think, "Oh, this would be a $500 million job," and the engineer does it in a month for $100 million via tokens? Don't you have to at least be open to the possibility that that was money well spent?

Absolutely, Joe. I think everything is about the return, right? And the planning. So, we're not afraid to invest. As a CFO, you are there to allocate capital. You're not there to constrain capital. You have to allocate, but you have to be clear in terms of that return. And I think in that case, it's probably one where they weren't sure in terms of that the that what they were doing in terms of that spending and wasn't in budgeting.

And that that goes to the point of what is happening today. I think most enterprises were at the early stages of how people are changing from a subscription-based model to a token usage model in terms of the compute capabilities. And so, that is at the beginning and and enterprises are starting to figure out, how do I really track that s

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。