Token Bill Dwarfs CPU Bill | Ben Pouladian on Meta's Muse, Anthropic S1, and Why All Roads Still Lead to NVIDIA

Jack Farley (0:00)

I'm joined once again by Ben Pouladian of BEP Research. Ben is a Semiconductor, autodidact enthusiast. He's been an investor in Nvidia since 2016 Ben, great to see you again. Welcome back.

Ben Pouladian (0:13)

Glad to be back with you, Jack. Some exciting things definitely have happened in the space, and glad to be here to talk about it and chop it up and share some thoughts about AI and the cycle and the new products that are coming out and where it's being applied and who's winning and who might be losing and where I see things are headed.

Jack Farley (0:34)

Ben, you say who might be winning, who might be losing. A month ago in the public markets, it was clear who was the loser. It would be Meta, among the worst performance of publicly traded stocks, that could be called AI stocks. That has totally reversed upon the release of Muse, which is a customer-facing application that can do lots of different things for people and has currently about 3 million monthly active users. How big of a deal is Muse? What's your take on it? What are the consequences here?

Ben Pouladian (1:05)

Muse is like the first AI assistant or AI app that I actually gave my wife to use. So I've been deep in the weeds on Claude Code, Claude, ChatGPT, Codex, Grok, GrokBot, Gemini. I'm playing with all of them. And whenever the kids needed help of homework, she would call me over and say, can you use your AI? Because I have all the apps. And then finally, I said, you know what, honey, I think this Muse, this is something you could use.

And she got on it and she's like, wow, look, it connects to my Instagram.

It's helping me book appointments. Is this what AI really is? I'm like, yeah, it's useful. So from a everyday user standpoint, I think the team at Meta has really taken the AI that we see that's Michelin star high-end restaurants and they made a McDonald's version for the average American consumer to use useful AI for their everyday lives. And I think they built a great product. It's probably not a frontier model, but it does 90 percent of the tasks what people want it to do.

Jack Farley (2:13)

And Ben, I think that it speaks to a broader point, which is maybe about 24 months ago, we started hearing about agentic AI, agents, agents, agents. Kind of the promise was that AI is going to make your life easier. In the 23 months since, there's been tremendous development in AI for enterprise applications, particularly coding. AI helps me a lot for researching stuff in finance, but it's actually making the hard work easier and allowing you to do more stuff kind of at the frontier. But it actually had not actually saved people's times in terms of booking appointments and stuff, really. It hadn't really done that at all. And I think Muse is the beginning of this.

Ben Pouladian (2:51)

Yeah, people are looking for ways to earn their time back. No one wants to click out of fill out a form for a rebate.

I use Muse for my daughter's AYSO emails and updates for soccer practices and things like that. There's just a lot of ways that you could use it. I use Muse to sign me up to wait in line for the DMV appointments. The whole idea, I think the premise goes back to what Amazon did with Prime and Deliveries, is consumers will pay to get their time back. So if you can figure out ways to use technology to save time on mundane tasks that people hate doing, you can really create a product that is sticky and is a winner for the user.

Jack Farley (3:41)

Ben, at BEP Research, you've been doing a lot of work on how much Meta is charging. And I've got my own sort of forecast that I just asked Claude to rip up of how much they can make from this. Tell us about the work about how much it's going to cost Meta to serve use in terms of the token costs versus the CPU costs.

Ben Pouladian (4:03)

The CPU cost is not as much as people think it is because they use like a AMD virtual server and it's provisioned in a way where people aren't constantly using and hitting these CPUs all the time.

I think that's kind of the misnomer that people have with CPU usage. It's not as big as what they think it is, but obviously it's going to cost a lot of tokens because with an agentic workflow, the whole point is that the CPU will orchestrate with the AI to run a task multiple times until it kind of gets it right. And in that time period, you're going to be blowing through a lot of tokens. So there is going to be a huge token cost behind it. So that's kind of where you see the big usage. And it's basically why you see a company like Meta spending so much on AI capex, building their own data centers from Louisiana all the way up to Canada, and also leasing facilities in other parts of the world, because they have a large install user base, and they're thinking maybe we can convert maybe half of them to Muse users. I mean, that's a lot of compute that's needed to serve the request these people have every day.