Latest / Freightvine / Jason Miller, MSU | The Capacity Driven TL Cycle and the ROI of AI

Transcript

Jason Miller, MSU | The Capacity Driven TL Cycle and the ROI of AI · Freightvine

• Chris Caplice (DAT/MIT): Hi Jason, welcome back to the Freightvine Podcast. Hello and welcome to the Freightvine Podcast, your source for all things freight transportation. I'm Chris Kapalas, Chief Scientist at DAT Freight and Analytics, and today I'm joined by Dr. Jason Miller. Hello and welcome to the truckload market update for 25 June 2026. In the dry van market, we saw that the change in active contract rates increased slightly, 0.4%, spot rates increased 1.7%, and the current level of replacement rates for dry van is positive 8.5%, which means on average, rates coming in are about 8.5% higher than rates that are exiting. Jason Miller: Hey, thanks for having me back, Chris. Chris Caplice (DAT/MIT): you're joining an esteemed crowd, Jason. You're my third fifth timer. I've been doing this is my 178th, 176th podcast. ⁓ ⁓ I've ⁓ you joined Bill Cassidy ⁓ Kevin's wire is a five time. So I'm not gonna send you a jacket, but I might send you a coffee mug. So we'll see how that goes. Professor of Supply Chain Management at Michigan State University's Eli Broad College of Business. This is Jason's fifth time on the Freightvine podcast. I last had him on about a year ago to discuss tariffs as they were impacting the market. This time I wanted to talk about the current truckload cycle. ⁓ both agreed that it started in December 2025, but that it took a couple months, really through the first quarter, to make sure it wasn't a false start, like what we saw at the end of 24, beginning of 25. Jason Miller: Hey, I'll take it, man. That's awesome. Chris Caplice (DAT/MIT): Yeah, so I wanted to talk back because ⁓ as you might know that the truckload cycle has changed again. ⁓ and in another cycle that seemed to start and wanted to talk to you about ⁓ about that. And first question is when do you think this cycle ⁓ started? When do you pig ping the date? The market gap between spot and contract is 27 cents a mile, spot above contract, and the year-over-year change in spot rates is about 36% for dry-van, and about contract is about 5%. And Jason notes that this expansion is unique because it's supply-driven, and it's essentially a mirror image of that demand-driven bull market that we saw in 2017-2018. And he notes that while aggregate demand remains flat, we are seeing what he termed a feeding frenzy in specific sectors like data center construction, steel production, mainly due to the rapidly expanding DAT center networks, that ecosystem by hyperscalers. Okay, temp control. We saw the contract rates actually dropped a little, 1.6% over the last two weeks. This is the second period in a row, so we're seeing some leveling off there. Spot rates increased 0.6%. And we see the current level of replacement rates for temp control to be 6.9% positive. Again, which means rates coming in, contract rates are about 7% higher than those they were replacing. Market gap between spot and contract is 16 cents a mile, spot above contract. Jason Miller: I I ping it as December of twenty-five. So if I had to say when did this cycle turn, it was, you know, December twenty-five. Chris Caplice (DAT/MIT): Me too. Me too. What and what is your signal? Everyone I've had students look at this and people have little fine points. How do you define when the market transitions from a contraction to expansion, vice versa? And sticking with the AI theme, we kind of turned the conversation from transportation to the trillion-dollar question of AI's actual return on investment or ROI. Jason shared his skepticism regarding the current hype, discussed the high cost of compute, the issue of model drift, and why we haven't seen a true labor substitute in AI yet. Jason Miller: Yeah, so h the one index that I've used a lot is basically just taking DATs, ⁓ contract rates, spot rates, and if you do contract minus spot over the average of the two, there's seemingly has been this magical ten percent threshold that whenever you sort of dip below that significantly, that's when it seems that a market has has flipped. and so in retrospect that's when it flipped. I don't think any of us really started to feel confident though, saying the market has Chris Caplice (DAT/MIT): Year-over-year growth in spot rates for temp control is 24.4%, contracts about 3%. Mm-hmm. Intermodal, we saw about a 2% increase in contract rates. We saw spot rates drop about 0.4%, but there's very little spot in intermodal. Current level replacement rates is negative 7.9% for intermodal. So we're seeing a drop there, which is a little surprising. I thought we'd see a leveling out or an increase. Market gap is negative 2 cents a mile for spot versus contract, where spot is below contract. Yeah. I am more of an optimist and think that the cost of AI will drop and that it is a long term massive benefit to the workforce. But it's always great to debate these topics with Jason. He always comes with an opinion and with data. So that's always a great debate. Whether you're navigating the routing guide failures that we're seeing now more frequently or trying to separate AI hope from hype, hopefully this episode gives you some data driven clarity because that's what Jason always brings. And need to understand where this current capacity Jason Miller: Had truly flipped until probably I'd say mid-February or even March. And and the reason was is we had you've always got the December holiday tightening. We moved into January and it actually felt like it was loosening a little bit. And then you had some really horrible winter weather. And so that then created the, well, maybe it's the winter weather narrative. And that was late January through like mid-February. And then things just never really started to sort of ⁓ you know lose. Chris Caplice (DAT/MIT): Meet up. And year-over-year change in spot rates for intermodal is about positive 6.4% and about 8.5% for contract. Finally, flatbed. We see the change in active contract rates drop 0.5%, a little bit of a change we've seen recently. Spot rates are still going up, positive 2.4%. And the current level of replacement rates between ⁓ rates that are entering versus rates that are exiting is positive 8% for flatbed. Right. Right, right. Capacity driven cycle is headed. Following my conversation with Jason, I will present the truckload market update. So let's get started. Jason Miller: loosen up again and actually started tightening more and more. And so that's when it felt, you know, certainly by March it was okay. We we are in Chris Caplice (DAT/MIT): Yeah. Yeah, in the g in the January, February time, we're concerned and you know, w is the tightening due to ice or ice? The weather or you know, the border control stuff that that's freaking people out. And it seemed like it was a Market gap between spot and contract is 19 cents a mile, spot above contract. And the year-over-year change in spot rates for flatbed is positive 24%. And for contract rates for flatbeds, a positive 6.7%. Okay, a lot of numbers out there. What are we seeing? Temp control is continuing to drop from a peak in May, ⁓ both ⁓ for ⁓ contract. So I'm wondering if it's kind of leveling out. It's still 24% higher than it was year over year, ⁓ but we're Toin coin toss. But go back to December of 2024. To me, that felt exactly like December 2025, but it didn't have the stickiness. In January, once the ⁓ new administration came in, the the tariff started hitting and it just dropped. But it seemed like we're in the exact same spot, but then it just changed. What do you think happened in 24 that stopped or 25 rather that stopped it from turning to the cycle? Seeing it it drop a little bit for that spot rate, and contract is only about three percent year change ⁓ year over year. ⁓ dry van is staying pretty steady. ⁓ spot rates are up across the board, except for intermodal, which has minimal spot anyway. And comparing year over year, all the modes except intermodal has spot between 24 to 36 percent above year over year and three to eight percent for contract. So we are higher. The question is, how long will that stay? ⁓ replacement rates are going. Jason Miller: Yeah, so you were starting to see sort of a nascent manufacturing recovery, you know, December twenty-four, January twenty-five. And then you just had the tariff, and I think even more so, it was the tariff uncertainty that caused sort of so much chaos. Because at the end of the day, the tariffs have generally proved to be much more bark than actual bite when you look at where effective tariff rates actually settled in at. ⁓ and so I think what you had was sort of a big capex pause by a lot of firms. You didn't see the machinery demand and Chris Caplice (DAT/MIT): Yes. Through the roof, all are in the positive seven to nine percent, except ⁓ intermodal, which did drop. So we are seeing bids coming in are going to be higher than the rates they're replacing. ⁓ similarly, the gap between spot and contract is positive. It's 16 to 27 cents a mile, except intermodal, which is flat. All this is pointing out that the the cycle that started tightening, the the ⁓ expansion that started hit in December of 2025 is continuing, ⁓ not at the same rate that it was. So we'll see if Jason Miller: Things like that in through most of 25 that we would have expected. ⁓ plus, you also had just this continual decline of capacity all through 25, even setting aside any type of English language proficiency, non-domicile CDLs. And I think part of that was a lot of carriers entered 25 thinking this is the year, it's finally going to turn. Chris Caplice (DAT/MIT): how far that continues, probably through 2026 into 2027. We'll see how long that lasts. We'll keep paying attention to the different metrics to see if we see the canary in the coal mine that gives us an indicator when the market will level out or actually contract, because we all know that it will eventually. All right, that's it for the truckload market update for 25 June 2026. Right. Jason Miller: And then when that didn't happen, you started to see operations that were borderline just finally closed up shop. And so, you know, when I look right now, you know, freight volume today isn't much above where it was in 25. But supply is down, you know, several percentage points. And so I think we finally sort of just tipped over that threshold where you start to see more routing guide failures and we have so much more pricing visibility. Chris Caplice (DAT/MIT): Leave that. Jason Miller: today that you know once it became clear that spot prices were higher you started to see a quick adjustment of bidding behavior by carriers and then that starts to you know build the feeding frenzy cycle. Chris Caplice (DAT/MIT): Right, right. Let's ⁓ before we get into the driver exodus, because I want to talk about that. ⁓ this cycle, you made a recent LinkedIn po post that said you thought this looked like the twenty seventeen, twenty eighteen ⁓ cycle, which lasted what, forty-eight months? Is that right? Four years, three and a half, something like that? What what makes you think it look like like this? And what does that imply? Jason Miller: Well. Well, so I'm saying it's the mere image of that because seventeen and eighteen was all about demand growth. We were seeing the one index I co-author ⁓ the Tump Trucking Ten Mile Index, that was growing three to four percent in twenty seventeen, then through most of twenty eighteen. And then it real growth really started to cool down in Q four of that year, and then twenty nineteen was flat. And so when you look at sort of that expansionary cycle, I kind of date that from a pricing standpoint. point from about June of ⁓ tw of seventeen through if you want to call it October of eighteen or December of eighteen, that's kind of your, you know, somewhere between about a fifteen to eighteen ish month ⁓ bull market. But that was almost all demand driven, little bit of supply driven with the ELD mandate, but but not much. Chris Caplice (DAT/MIT): ELD, sure, sure. Jason Miller: This is the exact opposite that we're in, in that this is ninety percent supply driven and maybe ten percent demand driven, but it's not aggregate demand, it's demand for certain goods. So it's the data center build-out, it's the AI ecosystem, because we'll see sectors right now You know, construction steel production is up eight, ten percent from where it was last year, but major appliance production is down six, seven percent from last year. And so you've got a lot of sort of the imbalance issue ⁓ I think, from what carriers were expecting heading into this year. Chris Caplice (DAT/MIT): But ⁓ so that's a really good point. I think we all agree this is a a capacity-driven cycle, but usually ⁓ which one ends faster? Because when the price is going up, you capacit a lot of capacity is just parked, right? A lot has gone out, and we knew we'll talk about drivers in a second, but a lot of that can come back in rather quickly. Do you think that what's the fundamental difference is between a demand-driven cycle and a supply or capacity-driven cycle, in your opinion? Jason Miller: Well, so we've never seen a supply driven cycle like this. Like if you go back and you look, you know, back the last thirty years, every expansionary market was demand driven, whether it was the nineteen nineties, it's ⁓ th you know, ⁓ four through about ⁓ six with the housing boom, you know, the thirteen fourteen fracking boom. ⁓ and so it really comes down to I think assumptions folks make about where capacity can come from. The Chris Caplice (DAT/MIT): Demand. Jason Miller: The view that this is going to be a very long-term expansionary pricing cycle because supply will be constrained seems to operate under a lot of assumptions that a lot of the driver pool is immigrants, they're going to have a harder time getting jobs with this administration and the rules in place, and therefore you're not going to have supply come back as fast. I tend to be a little bit more skeptical of that because when you look at the studies that have actually been published of where do truck drivers come from. Chris Caplice (DAT/MIT): Right. Jason Miller: From using the current population survey, so sort of the the big survey the Census Bureau does to track unemployment, it suggests that there's a lot more nuance to this. That a lot of people that get into trucking are folks that worked on loading docks and they see the drivers interact with them, they learn through whatever mechanism that hey, the market's really hot, it's a good time to start, and they say, Hey, I'd rather be driving a truck than driving a forklift. folks lose jobs in manufacturing, folks lose jobs. Chris Caplice (DAT/MIT): Yeah, yeah. Jason Miller: and management they go become truck drivers. And so when you look at ⁓ there's a good paper by Burks in Monaco that's ⁓ public domain from the Bureau of Labor Statistics from twenty nineteen that really kind of lays out where do drivers come from and you don't see any evidence that it's you know a disproportionately immigrant pool. And so I that's Chris Caplice (DAT/MIT): So do think that's just an inflation of the popular press? 'Cause a lot of the press coming out, you know, you have all the videos come out and it's like we're being swarmed. Is that just an exaggeration in your opinion? Jason Miller: So I think that probably on the margin you had a lot of capacity that grew and ⁓ you know, more recently through the cycle it may have been a little bit more from there. But it's important for everybody to remember, you know, we had you know, because of COVID there was large scale sort of crackdown on immigration, sorta through that same period where trucking capacity was growing a lot. Chris Caplice (DAT/MIT): Mm-hmm. That's a good point. Jason Miller: You know, so so I I am just I to me this is always just supply demand, this is just market dynamics. And this is where I really like to look at sort of class eight orders and where you look right now, I mean, they're high. They are, you know, ⁓ more than double where they were last year, and since December they've been well beyond replacement level of new trucks. Chris Caplice (DAT/MIT): Now how much of that is due to emission standards changing? Okay. Jason Miller: And that's where we don't know. And so that that's the challenge, as we don't know how much of this could be admi at admi admission standards. But my my sense is is, you know, if you look, the correlation between year over year change in line haul spot rates and new class eight orders pre-COVID well now, it's like point nine. We're pretty much back to where that was. And so I I just have a sense that capacity will start to make its way online, maybe Q4 of this year and through next year. Because from a carrier standpoint, I just think there's such a you know, economists have called it an expansion drive. There's this desire to grow. And for carriers that are out there that currently can't service all the demand they're saying, it's hard to tell your customers, hey, I'm I'm sorry, I can only take. 80% of your tenders because what's gonna happen is next year that shipper is gonna say, look, I want 95%, or by God, I am going to find somebody who is going to give me 95 plus percent acceptance. And so you run the risk as a carrier that if you don't expand to meet that demand, you may lose that shipper. Chris Caplice (DAT/MIT): Yeah, you know, this raises another question. It's really a taking a left turn here. routing guide compliance. To me, I've been studying this and doing this stuff since the nineties when I started doing bidding to create routing guides. I think that they're fundamentally changing now. About thwenty percent ⁓ higher of different industries ⁓ doesn't even the the freight doesn't even touch a routing guide. It goes straight that whether it's API or something informal, it's going straight direct to spot or direct to dynamic, call it whatever. So routing guides, since they're getting a lot of the the infrequent freight is not touching the routing guide, the routing guide compliance by nature is gonna be more consistent freight, so it's gonna get better now anyway. So what are your thoughts on how this whole idea of measuring routing guide compliance? To me, that's that's like a partial ⁓ picture of it. Jason Miller: No, I I think you're making a good point. I think it too depends on the extent we're calling compliance shipper to carrier versus shipper to broker. and so you know, 'cause to the extent that we're having brokers turn down loads because maybe they can't find capacity at that what that line haul rate is, that that I think does provide at least a little bit more information. but I mean we're even seeing that now, so ⁓ you know, I'll g I I've always liked Freightwave's ⁓ tender rejection rate. And if you look right now, they're I think 16-ish, 17%. But if you go back and look in 2018, they were 25% when they started publishing the data back in March of that year. And if you look at 21, they were, you know, tender rejection rates were 25 to 30% through much of 2021. And so this is where I, you know, it's tight out there, but I think from a purely I need to find a truck, I do not believe it is. as tight as tight today as it was back in the worst of twenty one or the worst of twenty eighteen. And so I I think it's another you know it's another metric and that we have to put together. I haven't seen what CHR's most recent sort of average tender depth index is at Chris Caplice (DAT/MIT): I agree. See, that's another one that's just because a lot of people, I mean, I've done a lot of work with CHR over the years and it's all out there and it's public. But ⁓ the depth, you they used to go to twelve, thirteen, fourteen. Now they they pull the the ripcord at three. So naturally the long tail's gone away. Because if you're waiting ⁓ the second, third, you're gonna go to a different mechanism now. ⁓ I think technology's gotten so so much smarter in real time. that you don't need to rely on that waterfall method for everything. So I think it's really changing the metrics. But we'll see. We'll see how that changes. Jason Miller: No, and and that's one reason for me I've liked to, you know, rely on DAT's data. ⁓ just doing, you know, contract minus spot, because that's at least pure rate based at that point. So it's hopefully a little bit less sensitive, even though again it's not not maybe a perfect basket, apples to apples. And, you know, there's there's no doubt, like it is it is tight right now. I do, you know, as we're filming this and you know, mid ish to late June, it does seem, even though we're approaching fourth of July, it does seem that we're Chris Caplice (DAT/MIT): Sure. Sure. Jason Miller: starting to asymptote a bit though on rates and yes diesel's diesel's come down a bit so line overall line haul is likely up but we do seem to be running into some resistance right now of don't Chris Caplice (DAT/MIT): Agree. I you know, it's funny, Jason. I just running numbers this morning that comes in, it came in at Saturday as of that. We do it every every week. We just look at the real time, even though we know it's gonna bounce around because stuff comes in. And I'm seeing exactly what you just described, especially for reefer. Reefer tailed out a little more. Now new rates coming in in contracts are going through the roof still, because of exactly what you said in the beginning. ⁓ but let me go back to the spot versus contract, because I also look at spot premium ratio, which is and and when we define spot for DAT, the shippers. Side, it's if we haven't seen the rate more than four days. And so we don't rely on the shippers saying yes, this is spot or not, because they they every company has different policies, so it's more dynamic. And when I look at it as a spot premium ratio, and it's exactly what you said when spot goes above contract, I don't even look at the 10%. And this is all shipper buy, whether from broker or from carrier. But the point I want to make, if I look at the last three cycles, the spot premium ratio peaked out at about 25%. Each cycle. It's pretty starred there. And right now we're around ten, fifteen percent. So it looks like we have some headroom. But as you said, this is a different kind of cycle. So I'm curious how far up do you think it's going to continue to go? Or do you think that spot's gonna start it's gonna start closing? Jason Miller: I mean, so the challenge is like June, this is end of quarter loading right now. We're getting around for fourth of July. We've got a little extra freight moving because of FIFA and the World Cup. ⁓ here right now, 'cause even though folks would not necessarily think that generates a lot of freight, it it it does. I mean, people are eating more, they're drinking at bars, things like that. Yeah, and Chris Caplice (DAT/MIT): Talk about Boston. We were inundated with Scots and they were so much fun. And that whole thing about we ran out of beer, it's true. It was the Sam Adams Brew House and Faniel Hall and Hennessy's right there. They literally had four times the demand for beer than they do on St. Patrick's Day, which is insane. But there were so much fun, so many skirts, so many kilts in Boston this last couple of weeks. It was very fun. Jason Miller: But but so this is the thing, right? Is once you get past July four, you got the auto industry goes into model changeover. So July tends to be a fairly weak month for overall freight demand because of the auto sector being at the lowest capacity it is through the entire year. You start getting past peak summer building season. ⁓ so we're gonna start to see single family starts are gonna decline. So that means less bricks and things like that. And so, you know, my concern right now and is Chris Caplice (DAT/MIT): Yep. Jason Miller: I think there's a decent chance that we're probably towards the spot peak right now at the moment. Because, you know, for up until maybe again a December period, but that's just pure seasonality. Because the challenge is right now is it's difficult to see where there's an influx of freight demand that is going to happen that is sustained. You may see an early peak season for containerized imports because folks are going to try to run stuff in before whatever new set of terrorists. tariffs are gonna kick in. ⁓ but you're not gonna have single family housing dramatically change this year. You're not gonna have discretionary consumer spending on furniture or major appliances or motor vehicles. That's not gonna change this year. Chris Caplice (DAT/MIT): Back to school. That's just a a little blip. Jason Miller: I think that's not that much and you know, you're gonna have I think again Cautious consumer, I mean, look at ⁓ the technology products and how those are begin to become more expensive. You know, you've had Wall Street Journal talking about iPhone eighteen's gonna be, you know, a couple hundred dollars more expensive than what iPhone seventeen was because of the memory shortage. And so you're gonna start seeing, you know, back to school, I think on the electronic side, a lot of Americans are gonna have some sticker shock because tablets are more expensive and all of this stuff. And so, you know. I think the thing is is when we started the year, we were all expecting a couple of interest rate cuts, maybe two to three. Now then most likely movement is gonna be an interest rate hike at the rate we're at in Chris Caplice (DAT/MIT): So tell explain, you're an economist, I'm just a dumb trucking guy. Why are the rates going up? Was it employment? What is causing it to go? Jason Miller: You've had a huge energy price spike and even with, you know, futures prices of crude oil right now I think they're around seventy-five ish as we're filming this. That's still way above where we started the year at. No one was really expecting oil to be around seventy-five. And, you know, you've had a big impact on the petrochemical ecosystem, you know, so you've seen plastic prices go up. This stuff's gotta filter its way through. And again, you have this dramatic ⁓ you know, memory chip shortage. ⁓ that has caused prices to skyrocket. And that's making its way into essentially so many not only consumer electronic goods, but also producer goods. Talking to one man Chris Caplice (DAT/MIT): And and what's driving the memory chip shortage is really being driven by the AI data centers, sucking up all the the other chips? Jason Miller: Yeah, exactly. As you've got, you know, as you're having so many so much GPU demand and things like that, you have, you know essentially what I'm gonna say is a almost price insensitive set of buyers and the hyperscalers that are buying Nvidia GPUs because they need Nvidia GPUs, and therefore Nvidia's having to have Foxconn or and Han High Precision build more of those ⁓ GPUs, and that's just pulling all this memory demand and there's There's basically three memory producers in the world. You got SK Heinex and Samsung, which are South Korean, and you got Micron, which is US. And so you just have essentially this almost insatiable demand that is pulling all of this memory capacity, and it's leading to massive price increases. Chris Caplice (DAT/MIT): Now do you think that's ⁓ short term when when I hear about all the the hyperscaling, you read about all this, ⁓ I have think I think back to 2000 with global crossing and all the fiber buildup where they kind of you know overbuilt because they assumed price would be a certain thing and it collapsed of course. But that capacity was used eventually. So do you think the same thing is gonna happen here? Cause the AI they seem to be overbuilding. Like there's a race and I don't think there's gonna be a dominant AI. I'm curious what your thoughts are, but it seems like the capacity is being way overbuilt. Jason Miller: Yeah, so I mean, it it feels a lot like the late nineties telecom boom, you know, you know, ecosystem boom. ⁓ you know, and a good example of this is production of energy wiring cable, that is actually an industry in the US. Production peaked in ninety nine, two thousand at three times where it ended up being in twenty seventeen. B ⁓ so you can imagine we're sitting so high it collapsed. It ne Never has even remotely gotten back to that level. And so that's not even a global sourcing situation. That's a we just overbuilt. Chris Caplice (DAT/MIT): But but that capacity got consumed. I mean that enabled other things that we didn't see coming, right? Jason Miller: It it did. And and so th this is the question I think in the AI space is what good is a data center for other than running, you know, LLMs? you know, a agentic AI. And this really comes down to the question of can, you know, I mean, l you know, being realistic, there are two ⁓ you know, AI shops out there that really matter. Open AI and anthropic. Like those are the two that matter. No offense to Grok, but no one cares about, you know, that AI. Or ⁓ you know, some of these other more micro ones ⁓ in the U.S. So it's really comes down to open AI. Chris Caplice (DAT/MIT): Wait, wait, wait, wait. You're discounting Google? Just out of the box? Jason Miller: J so I mean Gemini is a decent p platform, but when you look, I mean the this all depends on really can open AI and Anthropic be profitable? That's what we don't have an answer to right now. ⁓ you know, open AI's financials got leaked and they they were burning a lot of money last year. Depend it's it Chris Caplice (DAT/MIT): Yeah. Shocked. I'm shocked. Jason Miller: Yeah. And and I mean we're talking burning money that is exponentially more than what Uber burnt ⁓ at the start of, you know, essentially trying to scale that business and create the network densities and what ⁓ you know Amazon did creating Amazon Web Services. And so the question really is going to be, will customers pay the cost of compute? You know, Wall Street Journal recently had a story that both Anthropic and OpenAI are looking at lowering their prices ⁓ after they had raised them to a lot of ⁓ call them industrial customers through the token based billing because folks found out hey we were spending you know Per engineer $100,000 a month, and somebody goes, ooh, where's the ROI on that right now? And so I think that, you know, that is the right now multi-trillion dollar question. is because that the build-out of the ecosystem to support this is really supporting not just the US economy, but a s a substantially important share of the freight market, hauling the steel, hauling the building materials, holding the tra hauling the transformers. Chris Caplice (DAT/MIT): Yeah, yeah. Yeah, yeah. Jason Miller: And so any slowdown to this would be, I think, catastrophic to the freight market and the overall economy. ⁓ Is Chris Caplice (DAT/MIT): Right. Jason Miller: Do we start to see signs of approaching profitability in a reasonable time period? Chris Caplice (DAT/MIT): That that's that's that's a big question. But two two things that I wanna raise the point. One is When the price goes up, I mean the price per token, the cost per token is going down dramatically because the pressure's on, right? And so there's a lot of we're so early days. Three years ago, we didn't know where we'd be now. And I'm thinking three years from now, but I think the price per token's gonna continue to drop because we're much they're being much more clever. When something's way expensive and is limiting, you're gonna put a lot of time into it. But the second thing, I have a hypothesis, and I'm curious what if you would agree with me or probably don't. ⁓ one of the things that's happening. Happening is the electrical grid can't support what's currently being put in, much less the EV transition for trucks, cars, and all that. My hypothesis is because we have these deep pockets of money right now and they're investing into the electrical infrastructure, 20, 30 years from now we will have cheap, plentiful electricity because of this overbuild. And a lot of the investments gonna be put in that we will then use for other things. So what do you think about that? That the the AI revolution is really funding the future of ⁓ solid solidification of the ⁓ grid in the US. Jason Miller: So what matters is whether those investments in the Greg can be us it can be essentially that electricity can be transferred elsewhere. So I'll pick on the state of Texas, right? Texas is basically its own island when it comes Hey, I mean but No, but you know that so I think that Chris Caplice (DAT/MIT): ⁓ yeah. I got my master's there, be careful. Welcome horn. Welcome horns. Jason Miller: It it will on the electricity side, it will really depend on again the flexibility of being able to move those electrons potentially elsewhere if needed. On the AI side, the big question, you know, the questions I have is one is will the cost come down as use increases? ⁓ that that is I think that is the most important critical thing. Because the one thing for everybody watching to understand is you can't just make one of these models and just it's not like a piece of machinery. Put in a factory, it's installed, and then it just operates. You have to continually maintain this to make sure we don't have something called model drift. That is very different than as you think about maintaining machinery where you know, your machine is not going to drift as it it ingests more information into it. And so this is a shockingly CapEx intensive sector. And so the real question is, is will people pay the amount of money this costs? And so it comes down to is AI a true, you know, truly revolutionary, or is it something kind of just like a different form of software? In which case we should not be pouring trillions of dollars into it. ⁓ you know, I I hate to say it like I'll I'll pick on content I see on LinkedIn. You can tell what's AI generated. Most of it ain't very good. ⁓ you know, you can look at a graphic and say, ⁓ that that's 100% AI. ⁓ as an academic, I can read when people have used AI to write reviews, and it's typically not very good. I can read papers and be like, ⁓ this section was written by AI, and it's not very good. Chris Caplice (DAT/MIT): What are what are your tells or can can you not tell us? Jason Miller: Well, so I mean ⁓ on academic papers, ⁓ v so tends to be certain language cues with how it's written. It'll wrap around itself the logic continually. And so it'll pick certain words out that ⁓ show up a little too much. ⁓ the damn thing'll just hallucinate and create citations that don't exist. ⁓ it Chris Caplice (DAT/MIT): That's that's that's my fear. Yeah. But the word ⁓ one of the words that I've seen is delve. AI loves delve and it just ⁓ looks in Jason Miller: Yep. Well and the thing is too is you know, as an academic when you submit to certain journals, you'll be sure to put in certain citations that you know will appeal to the It won't it won't do that. And so I'll look through a citation list and I'm like, Man, you would never have found this no nobody submitting to X journal would ever have cited this completely unknown work. And so y you're like, Okay, the the the this thing went out and it did a search. Chris Caplice (DAT/MIT): See, but that's that's like sloppy AI use. There's still good AI use. But I have to give you a story. I looked at I was reviewing a paper and they cited me in something with someone I've never heard of. Apparently we wrote a paper together in 2017, and it was like, no, this is wrong. I don't know where this came from. ⁓ but I let it go through because I got the citation, right? Right? That's what I should do. Jason Miller: But But but this though is the the reason. ⁓ you gotta gotta you gotta love it, we we keep track. No, but the here's the thing though, is what you're describing, Chris, shows you use it but then you have to go put so much work into it, you then ask yourself the question, was I better off just doing it myself in the first place? And so that's where I think, you know, when you taught when you see a lot of this, the question is, is will it get progressively better? Or not. 'Cause again, all LLMs are doing is just predicting the next thing that's most likely in a string of text or filling you know, and so That's that's for me sort of my my concern is that I think people are fascinated by this because it's new, it's the exciting thing, right? I mean, this is it but it will it deliver the value. That's where I don't know. And yes, you know, in the in our you know, in a in the trucking space, you're seeing interesting applications, talking to brokers. Okay, we use Agenic AI to contact carriers that we've not talked to for four hours to get an update on a load. It's like okay. Okay, that is a valuable use, but is this something that we as a society should be spending trillions of dollars for to basically have that type of substitute? I don't I that's where I get a little bit more worried. ⁓ I also think what's gonna happen for a lot of companies that think, ⁓ man, we can use AI to replace our entry level workers. It's like, ⁓ so you think entry level workers only do these little things? I can't wait till senior managers actually have to go check this stuff and realize like, ⁓ man, I I I d I don't know about this and I'm not sure that this thing has done it right and now who goes and checks it? That's where you're Chris Caplice (DAT/MIT): You know, okay. So great story ⁓ came out in the Wall Street Journal today. Actually, I saw it. I don't know if you saw it where ⁓ someone wrote an editorial and talked about how actually AI is helping younger employees make a have a bigger impact. ⁓ and the idea was that it enables a young employer doesn't need to wait for to get gain this experience and whatever. They can actually apply some things. They're only limited by their curiosity. And I'd be and also they're much more I you know when you get more senior you have ⁓ which you must have you have p doctoral students ⁓ you have ⁓ other people working with you for this stuff and you've gotten to a level to have that. So someone coming in, th th this is a ⁓ an assistant, if it's used correctly, that get maybe it doesn't do the final product, but gets them to use tools, use things they haven't ⁓ had to have the experience to actually take years to to collect. Jason Miller: I mean my my response to the WSJ on that is why are college students booing people when they talk about it at commencement? Nah, I'm I'm I Nah n nobody no but I d na to quote ⁓ Ed Zitron, who's a fairly ⁓ you know, popular AI skeptic, no one booed the automobile in air conditioning. Chris Caplice (DAT/MIT): 'Cause they're stupid. They boo a lot of things. They boo a ⁓ they boo so many things. Yeah. Did they br did they boo VisiCalc? I don't know. I don't know. You gotta look at these things. Jason Miller: And so so that's where I'm just so that that's where again I'm I'm ⁓ you know I Chris Caplice (DAT/MIT): They also boot anyone to the right of center. Someone, let's get real here. Jason Miller: Nah, well I I guess my point is this, is that Well, what you're saying may be true, that's not how employers are gonna a lot of employers are gonna view this. They're gonna view this as ooh, I can substitute labor for machinery and I I just am gonna guess they're gonna find out that it's not gonna work well. So so here's ⁓ you know, some examples. Some of the law cases where you've had AI clearly being used to write it and it's created phantom citations, to the point where judges are saying, To the federal government, what are you doing? Like you Submitted legal documents with it with phantom citations. Like this is and you know stuff like that can cause you to lose court cases. And so I think that that's gonna be the issue is to the extent like hallucinations are part of this process. They're not part of the normal human decision-making process. Until we figure out how to get around that, when a hallucination could cost you, you know, hundreds of millions of dollars, it's gonna make that substitution more. Chris Caplice (DAT/MIT): But don't you think the hallucinations are a growing pain? I mean, look at it was three years ago. When it first came out, it was crazy what you could do with it. And now it's gotten smarter. And we're smarter too, it because if you have two different models cross-check each other, the probability of hallucination on both is very low. So they're we're learning that stuff. I think something that you said earlier, you see the AI slop in your domain. ⁓ you I do a lot of like AI music because it's fun. I can create songs and I don't play an instrument, but you can tell. And and younger generations have an ear for it. Just like we have an idea, you know. You we we can filter through stuff that our parents or grandparents probably would have been hoodwinked on. So I I I I don't know. To me the hallucinations are growing pains. We're getting over that. Jason Miller: So the question then is again it goes back to cost. If you're now having to run two models to cross check each other to ⁓ how how ⁓ how much does that cost versus again just having a human do it? The other Chris Caplice (DAT/MIT): Yeah, but look at the cost the cost of doing computing in early days was tremendously high. And now that's gone down. So I think the cost of doing this analysis is just gonna drop. It's just gonna continue to drop. Yeah. Jason Miller: And and and that that that's the question is will that will that cost come down to to justify it? And so this is where, you know, again, we we have there's no doubt the infrastructure build out is assuming incredibly wide scale adoption. That then but the only way to have wide scale adoption is for users to eventually put an ROI to this where it makes sense. And you know the challenge is is right now the subsidization that is taking place for the average user. So I think the numbers I heard were for one for chat GPT for the twenty dollar a month subscription, if you have one user maxing out their tokens, it takes thirty-five non-using accounts to basically make up for that. So they're basically Chris Caplice (DAT/MIT): Yep. Jason Miller: subsidizing that user 35x times. That can't continue, but then you have to ask somebody, okay, rather in 20 a month, let's see, I'm just sitting here because I just don't do mental math anymore, 20 times 30. That's $700 a month. That user is not going to pay $700 a month for this. That is almost a car payment nowadays in this country. Chris Caplice (DAT/MIT): Absolutely. Right, right. Jason Miller: And so how do we work make that work out? And yes, again, Uber subsidized on the ride sharing piece. Yeah, you could drive across San Francisco for five bucks. That was not profitable, but that was inherently a network-based system where you had to come to scale on the u on the demand and the supply side to make it work, but the economics were super, super, super clear. Chris Caplice (DAT/MIT): So so look look it this way. I see, I mean, who are the big players right now? They are so flush, right? They're they're they've got a lot of money now. Will they all survive doing this? I don't know. But do you think one, two things. Do you think AI it's just gonna go away? Or do you think it's gonna settle down to one dominant one or some other few? What's gonna happen five, ten years from now, do you think? Jason Miller: I th I so I think right now we are in, you know, if you look at the Gartner sort of hype cycle we're on the trending up piece. I think we'll get to the disillusionment piece within a year or two. I think you're gonna probably have specialty AIs for different things. I mean you'll have one that's more focused on clothing coding, you're gonna have one that's more focused on health and medical. And and you will have that. So it will be a tool, but it will be like computing. know it'll be like industrial machinery and factories. You know, when you look at the US, we effectively automated away most manufacturing jobs that can be automated by I'm gonna say 2010. And the reason I say that is if you look you had these you know, sort of secularly the big recession in nineteen late nineteen seventies, early nineteen eighties. A lot of very inefficient plants got wiped out that were more labor intensive. Two thousand and one recession did the same thing. Two thousand eight nine recession did the same thing. But if you look, that's kind of when manufacturing employment in this country bottoms out and then it starts to climb back upward a little bit. And so, you know, you've na you've had Yeah. You know, essentially you sort of asymptote to the extent that you can reasonably replace, let's say, labor. I think that what we'll find is that AI is not the replacement on the labor side that we thought it was. And because of that, that is gonna affect how much money companies can spend on it. Because if it's not that labor substitute that we think, then clearly we can't be spending, you know there there's no companies that once we're on token based belling that are gonna allow engineers to burn thirty thousand thousand dollars of tokens a month. Chris Caplice (DAT/MIT): I agree, I agree. But to me that's we have a lot of big companies chasing and they think there's gonna be a winner, a couple winners, and they're betting all for it, and there will be a fallout. There there just will. ⁓ but I think at the same time the cost per token, I'm I will make a bet with you, and I don't even know what it is right now. I don't know what the general market cost per token is, or if that's the the the metric that we're gonna use going forward. I I think it'll continue. There's gotta be some correlation to Moore's Law. It's gonna keep dropping because of there's we're we're not even you know, when new things happen, you don't try to be elegant, you just get it done. And now when the focus is on cost, you get more elegant. And there's a lot of different ways we're not doing just stupid searching. You know, it's more selective. You're doing pre selections for these kind of things. So I think you're gonna see a drop dramatically in cost per token at the same time. But I agree with you, there's way overbuild. But is there any silver lining to that overbuild? Jason Miller: Well ⁓ so where I would where I would push back is i on the it will get cheaper is if No, no, no, is it's just if you look, for example, the the producer price indices for stuff like semiconductors that adjusts for quality change. And yes, that plunge down, if you go from the nineteen fifties through about twenty nineteen, it drops and it drops very sharply. Chris Caplice (DAT/MIT): I'd be disappointed if you didn't. Come on. Mm-hmm. Jason Miller: It has been rising ever since 2019. So that is telling you cost increases are over or overwhelming, or I should say price increases are overwhelming any additional quality improvement. You've now seen, because of the memory shortage, this huge exponentiation upwards of that producer price index. And so the challenge is that I see is how does computing become cheaper when you have GPUs, you know, cost. costing more because of memory. And so the challenge is now is we've created such an insatiable demand, we've pulled up the cost of all the material needed for the compute that it's hard to see how that becomes cheaper anytime soon. Plus you're pulling up the cost of electricity. And so that to me is going to be the big question is on you know we've basically seen a sort of a cra, you know, for lack of better term Quality gains, overwhelming price increases, that that has stopped in semiconductors even before this. And so that's where I get more, you know, more questioning on that. It's the same thing the consumers experiencing, right? As we're starting to see those consumer electronics becoming fundamentally more expensive, even holding quality constant because of this. Chris Caplice (DAT/MIT): I think it doesn't this get into okay if the price of any input goes up, then we start using other inputs. We start looking for other things. And I'm not saying we're gonna have a a difference of chips. I say the software, the way we're using it, we're gonna be clever there. Because we're not being clever right now because efficiency is secondary at this point. They're just grabbing market share. But as soon as the focus gets to be, hey, don't run the full model every time. Let's be a little more selective on it. And there's a lot of work going here on that. And I think it's to me it's basic economics as a non-economist. When the price of something goes up, you look to use less of it, you divert it, you try to find a better way to get that end result you want using other inputs. Jason Miller: No, a absolutely, and we will we of course are gonna see efforts to that. The question is really can we? So I mean and even, you know, tying it back transportation, think about right now, you know, intermodal in twenty six is intermodal in twenty eighteen. When you count container and truck trailer, if you look at just container, you know, it's ⁓ it's higher today, but that's because truck trailer ⁓ you know, there's less trailer on flat car today because that doesn't fit precision schedule railroading. Chris Caplice (DAT/MIT): Yeah. Jason Miller: But, you know, if you think about it, modally from a transportation standpoint, you know, that we will see shifts over, but at the end of the day, there's a lot of structural constraints that are on systems that hey, you can't can't lower costs anymore. You can't you can't do it this way. And so that's really the question is how much room is there to change this in that AI space versus are you hitting some type of you know physical physical constraint within the system? You know, take again motor vehicles produced in the United States. We have physical constraints because of our cost of inputs where we are not going to produce as cheap as Japan or as cheap as South Korea. We literally cannot do that. that. And it no amount of creativity can overcome that issue. And so that's where ⁓ I know it it's gonna be interesting 'cause it Chris Caplice (DAT/MIT): I think AI is a little different in this respect. To me, it's a but we'll see. We'll see how it comes out. But I'm not saying we're gonna have suddenly new chips or anything, but there there is a lot of push on changing the way that we can use existing inputs. But let me change the topic totally. I want to hit this before we're almost out of time. You were appointed by Governor Wit Whitmer to be on Michigan's Commission for Logistics and Supply Chain Collaboration through 2029. First, congratulations. ⁓ Jason Miller: Thank you. Chris Caplice (DAT/MIT): And tell me what does that role entail? What will you be doing for Michigan? What is the what's the objective here? Jason Miller: So it's part of a group that is, you know, sort of cross-functional backgrounds. A lot of my role, we just had a meeting ⁓ earlier this month, our quarterly meeting. And so for example, I sort of I g delivered a presentation on sort of the state of manufacturing in Michigan. Because when you think freight generation, it's still mostly manufacturing. And it's really, you know, trying to level set for folks and thinking about okay, where are their opportunities? So one of the conversations that came out about that is Michigan actually has the second most Chris Caplice (DAT/MIT): Really? How do you measure diversity? Where how is that? Okay. Huh. Jason Miller: ⁓ different different types of products grown and things like that. This is something that our our ag folks have quantified. And but the challenge has been is historically Michigan was not that big of a food manufacturer, even though we had all of these products. But if you look over the last decade, food manufacturing has actually been the fastest growing manufacturing sector in our state. And it's the only one that I can point to and say employment today is higher than in 1997. It's the only major manufacturing sector in this state where we can say we employ more people than 1997. And so, you know, conversations like that sort of provide a starting point to say, okay, are there more opportunities to essentially get more food manufacturing in this state by explaining the value ⁓ to different constituencies helping small businesses that are maybe producing in one site, could they expand to a second site? So I view a lot of my role as kind. Chris Caplice (DAT/MIT): Interesting. Jason Miller: of bringing sort of the data piece. ⁓ you know, it same way it also allows me to be networked with folks to say, okay, Gordy Howe bridge, what are the economic implications of that? I'm the person that can provide the trade data to say, look, this is what comes in through the quote unquote port of Detroit. We can start to get a sense of these are the shipper verticals that matter the most on an import side and an export side, and we go from there. So a lot of how I view it is it's bringing you know data data to these conversations and simplifying folks' life otherwise trying to find somebody who could ⁓ where's all the import data coming from? Well I I know how to use the Census Bureau's ⁓ Trade Online website. Chris Caplice (DAT/MIT): Yeah, yeah. No, that's that's great. So is this your first step for political office or is this just something to apply data? Jason Miller: I I I ⁓ my time as department chair was enough politics for me, thank you very much. I did my three years and ran away from Chris Caplice (DAT/MIT): Yeah, yeah. Let me ask a question about the the food manufacturing, because that's really interesting. Do you think that's because there's less automation in food manufacturing than in, say, auto? Or do you think there's something inherent? Jason Miller: I think part of it is certainly the automation piece. ⁓ you know The end of the day, Michigan's ma manufacturing in Michigan, ⁓ you know, there just huge drops after 2000, and most of that is industrial robots because NAFTA's 1994. You can't blame NAFTA for magically hitting seven years later. That that that that would not fly ⁓ from a from an explanation standpoint. So I think that's part of it. And I think it's more, again, more plants, more, you know, diversity of food in the US has been continually increasing. And so I think, you know, you have again manufacturers taking advantage of being close to various, you know, sort of diverse sources of ⁓ of prod of ⁓ inputs. And so I think that, you know, there there's a lot there. ⁓ you know, cross border trade has been immensely beneficial for us. Canada provides a lot of agricultural inputs as well. Chris Caplice (DAT/MIT): Has that how much of a hit has that taken over the last two years? I know bourbon isn't going north anymore. Jason Miller: ⁓ so Yeah, so it really depends on whether it's stuff that's been targeted or not. So, you know, food imports from you know from Canada that are coming in USMCA or, you know, tomatoes, things like that, not as negatively affected. Certainly has been a negative effect on the auto sector. ⁓ and you know, we even see that nationally. Our imports of motor vehicles are down about fifteen percent from where they were in twenty twenty-three, but our exports of motor vehicles are down thirty percent from where. They were in 2023. ⁓ And so this is where I'm going to say there's, you know, the challenge has been with the tariffs, is it's just making us more insular as a nation because we're not only are we seeing less imports in certain categories, but we're also seeing less less exports. Chris Caplice (DAT/MIT): Yeah. So I wanted this is a t unfortunately I didn't get to half the topics I wanted to. We'll have to save that for another time, ⁓ because the whole import export and how that's changing ⁓ with the tariff impacts w is in a whole nother podcast. But Jason, thank you so much for joining me today. I always learn something new. We don't always agree, but I respect your opinion. Jason Miller: Hey, that's good, right? I mean, you know, pure agreement usually just leads to group things. So my academia is awesome. Chris Caplice (DAT/MIT): Yeah, what the ⁓ if two people agree on everything together, then one person's not needed. Yeah. Anyway, thanks a lot, Jason. Appreciate it. Everyone stay tuned for the truckload market update. Jason Miller: Exactly.