Plexo Capital's Lo Toney: AI 'inference economics' important to watch in 2026

Jon Fortt (Anchor) 00:00.000

Well, let's dive deeper into that AI trade and what could be in store next year. Check out the top performers in the Nasdaq 100 for 2025. Western Digital, Micron, Seagate, all names on the hardware side. On the other hand, major underperformers include software names like Adobe,

Jon Fortt (Anchor) 00:17.080

Workday, DataDog with investors skeptical about AI monetization. So, should investors expect a catch-up trade on the software side in 2026? Joining me now is Low Tony. He's founding managing partner at Plexo capital and a CNBC contributor. Low, a lot of people cautious about

Jon Fortt (Anchor) 00:34.760

2026 being that great for software here. What will it take?

Lo Toney (Founding Managing Partner) 00:40.320

Well, first, thanks for having me. And to piggyback on what Steve said, I think I just want to mention that what's important about Apple is that even if their AI execution stumble slightly, their economics still protect them. And that's very different from a company like Meta.

Lo Toney (Founding Managing Partner) 00:56.440

And that difference, I think, is is the real story if we dive in and in particular to the Manus acquisition. Maybe we can even walk through a few definitions that I think were important to frame the context and to look for things in 2026. First, let's think about training.

Lo Toney (Founding Managing Partner) 01:12.080

That's a one-time build cost, but where we're moving is to inference, which is the ongoing operating cost of actually running AI in the real world. Or you can think about it this way. Training creates capability, inference determines profitability. And inference economics, which

Lo Toney (Founding Managing Partner) 01:29.800

is is a concept we've been writing about extensively this year is the study of really who pays for AI every time it runs and whether that revenue scales faster or slower than the cost. And this is good context to talk about Apple, Meta and inference economics I think are going

Lo Toney (Founding Managing Partner) 01:48.320

to be important to watch for 2026.

Jon Fortt (Anchor) 01:50.480

But look, I I think there's a potential bottleneck here that we don't talk enough about and that's workflow and human culture. There's certain ways that people are used to getting things done, and there are certain ways that people have organized their data centers and their

Jon Fortt (Anchor) 02:06.440

data. I'm not sure that enough companies are in position to scale their AI use just because of how they're structured and how they do things enough in 2026 to capture the full expectation of the market in what this software is going to do.

Lo Toney (Founding Managing Partner) 02:23.520

It's a really important point, John, and I appreciate bringing that up because when we talk about the cost to build out infrastructure typically, we're talking about the cost for training infrastructure, which is very GPU dependent. Again, we can go back to Apple. And what I

Lo Toney (Founding Managing Partner) 02:40.800

look at with Apple is the fact that even if they stumble, they really have minimal exposure to this concept around this massive build out because they're not really spending money on large-scale build out. It's really happening by the third-party apps. So AI can almost feel free

Lo Toney (Founding Managing Partner) 02:59.000

to the users because Apple's not exposed to those marginal inference costs, right? So really Apple benefits from AI without being on the hook for the inference costs. Microsoft, their exposure is a little bit lower, but with the model from Microsoft, inference is pretty much

Lo Toney (Founding Managing Partner) 03:17.400

just built into their existing model of seats, licenses, and contracts. So those absorb the inference costs. Now, Google is really interesting because of their vertical integration we've talked a lot about their TPU chips. And their TPU chips are within one data center that

Lo Toney (Founding Managing Partner) 03:38.840

manages all of the services that handle billions of users, handles Google Cloud, it even handles the developers that use it. And then as well, it also does a dual purpose with AI, both handling a lot of the training for the models as well as the inference. And so when we think

Lo Toney (Founding Managing Partner) 03:56.440

about Google's business model, they don't subsidize inference. They tax it. Now, if we were to think about Meta, that's where there's some exposure because Meta really monetizes attention. And so, the inference costs can rise faster than monetization unless their pricing power

Lo Toney (Founding Managing Partner) 04:16.000

improves. And I think that's why one of the reasons at least that they made the Manus acquisition. And and what Meta is trying to do here with Manus is Manus is good at making AI into tasks and work workflows. So it kind of helps Meta to fix their balance sheet in a few or fix

Lo Toney (Founding Managing Partner) 04:36.000

overall financials in a few ways, Right they fix the balance sheet with the Blue Owl transaction which took 30 billion off their balance sheet. This Manus acquisition kind of helps the P&L by making some incremental revenue. And look Meta has been working on their own ship.

Lo Toney (Founding Managing Partner) 04:52.960

They've also been talking with Google about potentially using TPU's. Yeah. So that could help cash flow