What parts of this exercise do you outsource to Opus and Sonnet, and also how do you or Fable decide what is a Sonnet task and what's an Opus task? It really is amazing, but it burns through the tokens, and I'm regularly hitting my token limits if I'm just using Fable.

Until you realize you've burned through all your tokens just switching and had no idea.

I haven't hit my limits for this week.

In fact, I even did a test run just using Fable for this same merger model, and the outputs weren't hugely different.

When you ever see a public merger announcement that's not accretive to EPS, right? It's just like insert synergy, goal seek synergies to make this greater than zero.

Welcome to all our listeners. Welcome to this week's episode of What's the Big Deal? Graeme, tell all our listeners, please, what is the big deal this week? We are looking at not a new deal, but a deal that was announced last year that's currently under competition clearance review by the Surface Transport Board, I think that's what it's called. And it's the proposed merger of Union Pacific and Norfolk Southern. Is that the name of it, Debs? I actually always- Yeah ... I always mess this up.

I know the catchy names on the- I've actually never, ever really heard of them.

Yeah. Well, yeah, and it's interesting.

I think when the deal was announced, it did get a bit of press coverage because it's a really big deal, isn't it? I think Norfolk Southern, $85 billion enterprise value, and together they will form an enterprise that's worth $250 billion.

It is a mega deal. So it did hit- Yeah. Yeah ... some headlines. But maybe not the most exciting deal, in terms of being able to talk about railroads- Yeah ... and connecting ports with railroads and things. So I don't know.

Graeme, you tell us, what do you think is interesting about this deal? Oof. Well, okay. So I think US rail is pretty boring in general. I think rail can be interesting, particularly if we're talking about new high-speed rail links that open up new routes and are interesting for a passenger. But I've been taking more American rail than I ever have, just spending time on the East Coast working around here, and I've got to say, it sucks.

Yes.

It really sucks.

Coming from the UK, I would agree with that.

Yeah. Yeah. And arguably, the UK is not great in the context- Yeah ... of Europe either, right? So it's kind of like from best to okay to worst here. I was on this train to Baltimore, I don't know, a month ago or something, and I swear, as the train was moving around, doors were just flapping open and closed.

It looked like it was built in 1950 or something.

And I'm like, "Okay, what is going on here?" But I have to admit, I actually don't have that much knowledge about US railroads.

I know there are a couple of big railroad companies.

The one everyone thinks of here is Amtrak, because that's the passenger- Yeah ... that's the passenger railway.

I actually couldn't tell you what the ownership model of all these networks are.

Do passenger trains and freight trains share the same actual tracks in some places? How do those arrangements work? I literally could not tell you.

If anyone knows, let us know in the comments. Maybe you know, Debs.

No, not at all. But what I was going to say is that- ...

even my very rudimentary knowledge of European rail is that, yeah, the ownership models are generally complex because you get different parties owning the infrastructure, the rolling stock, and the maintenance of that as well. So it is generally complex. So yeah, it's a tricky thing to get your head around. But we're going to have a look at the deal today and see what we can learn from the numbers. We'll run the numbers.

We've actually built a merger model for this deal, so we can take a look at some of the key metrics that we tend to look at when there's a deal announced and see if the numbers work, and also use AI to run those numbers for us, because, of course, we don't need to do that as much anymore.

As always.

Yeah, absolutely.

Exactly.

So should we dive in? Let's do it. Let's do it. All right, so you used Fable, I want to say, to build this travel. Actually, no, you...

It was kind of an interesting process here was Fable- Yeah ... delegated to other agents, right? That's right. So I was really flexing my skills when it comes to AI this week.

And part of that is because I'm experiencing what I think a lot of people are experiencing at the moment, which is challenges when using Fable, which is the frontier model for Claude, which is amazing.

It's really good, really high level, taking on projects, taking on challenges. It really is amazing.

Yeah.

But it burns through the tokens, and I'm regularly hitting my token limits if I'm just using Fable. And I had been reading about the possibility of delegating certain tasks within a project to other models which use fewer tokens. For example, Sonnet and Opus, still within Claude- Yeah ... and using that approach to kind of optimize your use of tokens and maximize efficiency. So I thought I'd have a go.

And yeah, we'll have a look and see the output, shall we? Our financial and valuation modeling certification program is a self-paced version of what we teach inside investment banks.

Seven courses, 46 hours, three statement modeling, DCF, trading and transaction comps, M&A, and LBO, all built in Excel from scratch using real case studies.

Over 100,000 people have gone through this training.

You get unlimited support from former investment bankers, a month of PitchBook access, and the certification when you pass the exams.

It's all at wallstreetprep.com.

What things, what parts of this exercise do you outsource to Opus and Sonnet? And also, how do you or Fable decide what is a Sonnet task and what's an Opus task? A really good question. So I actually asked Claude that question myself because I had never done it before. Yeah. So I gave Claude a very basic prompt to build a short-form merger model for the deal, okay? The short form, it just means it's just the basic calculations.

It's not full integrated income statement, balance sheet, and cash flow statement.

Yeah.

I gave it a bit of guidance on the data sources to use.

But then, the second half of my prompt is basically that I want to optimize efficiency and minimize token usage by delegating lower-level tasks to sub-agents Opus and Sonnet. And then basically, I asked Claude which tasks can be delegated and provide the prompt and sub-agents so I can run these separately.

So I'm basically leaving Claude to make the decisions on this.

Now, the one caveat- Okay. And are you- Yeah ... are you having to prompt separately and pull the data together, or do you give an instruction to Claude Fable to instruct the sub-agents automatically? So I tried that and I failed, because basically, I'm using the Excel add-in.

Oh, right, okay.

Yeah, so if you're using the web interface, you can actually get Fable to do the delegation for you. It does allow that process.

But unfortunately, in the Excel add-in, I'm not sure why, but it doesn't...

Basically, once you put the prompt in and you select your model, the whole process has to run with that model. So what it then does, if you want to split the tasks up and have Fable as the orchestrator, and then the sub-agents doing some of the lower-level work, you actually then have to create a new prompt within that same window where you then switch between the models.

So you're effectively jumping between models for different parts of the task, and then you come back to Fable at the end for the more important tasks that you really need that sophistication for. The process was a bit clunky.

It required doing lots of starting and stopping the prompt or the process running.

But to be honest, it worked. So I'm happy with that, and I'm sure the process will improve with time as more people start to use this kind of orchestration approach.

Yeah, and you know what? I also think the web interface, the desktop application, whichever one you want to use, is actually getting a lot better at Excel work as well. So if you're one- Yeah ... of the students in my class this week and you're doing the case study today, lo and behold, a lot of that was done and put together by Fable as kind of a...

Actually, it was really, really interesting, right? It's a case study based on an existing Ares deal that needs plenty of information to put a model together, and then a whole PowerPoint presentation for the actual case study. And there was a lot of back and forth to get it where I needed to get it. But you know what? For the amount of, I'm not going to say amount of effort, but compared to when I was doing some of this work with other models before, the results I was getting just straight out of the Fable application in terms of download and kind of review were actually pretty good. Much better than even a couple of months ago.

So I think maybe starting the process off in the desktop app and just saying, "Okay, instruct the subagents accordingly, get me the first draft," and then continuing to work in the Excel add-in interface.

Maybe that's the current best workflow.

Oh, Graham, I think we have another episode awaiting us then.

Yeah.

We haven't prepared the exact for the Excel add-in.

But yeah, this was definitely a little bit clunky.

And in terms of the tasks, so basically Fable is amazing for planning and high-level structure and even the fact that it was the one that was orchestrating, working out which tasks should be done by which subagent.

And if we go down through the tasks, basically the data mining, Fable has recommended that Sonnet do that piece of work, and it gave me the actual prompts that I needed to use within Sonnet.

So I just toggled- Okay ... the actual model within the window.

And then finding the consensus estimates, again, that's just data mining.

And then for Opus, that kind of comes right at the end for more complex reasoning.

Basically, it says to use Opus for just basically a check after the model's been built, stress test the logic, make sure it's kind of all cohesive, and that- Yeah. Okay ... it kind of sense checks this at the end.

So I think it's not a bad solution. I mean, the one thing I have read is that if you're juggling between the different models within the Excel add-in, it actually does use quite a few tokens just handing over tasks between the different models.

So you can- Yeah ... offset the efficiencies, I think, if you've got too much juggling between models. But I did ask Fable to optimize the efficiency, so I'm assuming that this is kind of the best use of those subagents and minimizing handovers.

Until you realize you've burned through all your tokens just switching and had no idea.

Well, honestly, I haven't hit my limits for this week.

Okay.

And I've definitely been using Fable quite prolifically.

Okay. And in fact, I even did a test run just using Fable for the same merger model, and the outputs weren't hugely different.

So that gives me a little bit of confidence.

Okay.

And let's dive into the actual model itself.

Yeah, let's see what we got.

Okay. So we have our merger model, our short form model, starting off with the market and deal assumptions. So we have the offer details here.

So we've got cash consideration of $88.82 per share and also an exchange ratio of one for one.

So effectively, it's a cash and share mix in the consideration. Okay? And that gives a total value of $398.41.

So I think what does make it interesting, I think in this deal is that you have got that nice mix of consideration, which is quite common when you've got very large deals and it is a mega deal. You do need a bit of stock in there.

So yeah, so that's the kind of the initial setup.

And then Claude has grabbed the standalone company data.

So we've got the debt and cash, and some of the earnings numbers, both EBITDA on an LTM basis and then EPS forecasts. I think they've done a good job in terms of grabbing the adjusted numbers.

Those are the ones that the analysts tend to focus on because of cleaned, adjusted EPS figures. So and that's for 2026 and 2027.

And this is still a live deal, so it hasn't closed yet.

So- Right ... probably the focus will be on the 2027 numbers. Okay.

And then we've got some of the key assumptions that are going to be used in the numbers. So we've got the cash used from the balance sheet, the tax rate, the interest rates, and then that important number, that's the one we all like to focus on, is the synergies. $2.75 billion of synergies there.

So, a good decent bit of strategic overlap, hopefully for this transaction.

Go on.

And what's the net income? We got what, $8 billion for Union Pacific, $3.2 for Norfolk Southern.

So the synergy number is material for sure, but it's not crazy.

I would use that word in the sense that, was it Paramount's...

Not Paramount Sky Dance, but Paramount Warner Brothers where combined net income was $100 and synergies were $400. Something like that.

Something sort of- Yeah. No, you're absolutely right, Graham.

I mean, actually, to be fair, the numbers, it's quite punchy when you compare that to the LTM EBITDA of Norfolk Southern.

It's the- Mm ... $5.5 billion of EBITDA versus three tax synergies of 2.75.

So it's still quite a punchy number.

But- Yeah. And I went on the... They have a whole website to basically sell the benefits of the merger, and they say there's going to be no layoffs and more union jobs created and all this stuff. I don't know.

What are the costs you really cut out of combining just these two separate rail networks? I mean, I don't know. Or is it literally just we own all the rail networks and we're going to put prices up and it's the revenue synergies? So I'm sure there's some overheads in there, but as you point out, a really important point, which for me, unionization of a workforce is a big challenge when it comes to cost savings because it does create huge obstacles to headcount reduction.

Yeah.

But yeah. So yeah, maybe it's a lot of overheads, but yeah, I'm not an expert on railroads, unfortunately. So right. So let me move down.

We've got a few more adjustments, a little bit more advanced actually, Graham.

This is something that you spotted when you first shared the results, which is some slightly more advanced items in the calculations. They've got some adjustments for step-ups where you revalue the assets of the target in the calculations of earnings per share, and that's a little bit...

That's not always one that we would do in our first cut on the numbers, is it? That one in particular is really used to figure out what your increase in depreciation amortization is going to be, and if you make the argument, well, that's a non-cash expense, so who really cares? I mean, you care in the sense you get a tax shield on it, so that's something.

Yeah.

But a lot of times, and we were talking a lot of times when we look at this on a cash basis, you just don't even think about that kind of adjustment.

Absolutely. Cash EPS all the way. So, it's been a little bit more sophisticated than we probably need when it comes to that.

We then have our sources and uses of funds. So that's really standard clear layout.

So they've done a good job there. Most importantly, the sources equal the uses, which is always what you need for a deal to be able to go ahead.

And- Indeed.

Yeah. Is there anything particular that you want to highlight from that, or should we just leave it there? Actually, I was just... I just think I'm going to correct myself on the whole cash tax savings from the depreciation step-up because I'm assuming this is a...

We all know this is a stock deal, not an asset deal.

So the step-up doesn't actually get you a cash tax savings, so.

Oh, yeah. Okay, great. And then this is where it starts to get interesting, I guess, is where we've got the accretion dilution analysis.

So effectively, we're combining the net income of the two companies.

We're pulling through some adjustments to things like all the extra interest on any debt to use to finance the deal, interest lost on cash used to finance the deal, and then once you've got your pro forma net income, we then calculate the pro forma diluted shares. That takes into account all the new shares issued as part of the equity component into consideration, and that allows you to calculate pro forma EPS. And you can compare that to the standalone EPS of Union Pacific, and that gives you a nice bit of EPS accretion.

In the second year, so that's 2027, on a pro forma basis.

So we've got double digits EPS accretion.

And that presumably is just the synergy ramp from 50 to 100%, right? Yeah. So that's what I think is quite nice about this particular model is that you can flex this. So you can say, well, actually by 2027, maybe it's only going to be 75% synergy run rates achieved, and it just allows you to flex the outputs based on your sensitizing that number.

Yeah. I mean, we're looking at a reasonably well put together...

I mean, it's a short form model. I mean, to call it a model even is a little bit of an overstatement perhaps, but it's well constructed.

Yeah, absolutely. And then if we keep going down, we have the other critical metric that I would always have alongside EPS accretion and dilution, and that's pro forma leverage. Because as hopefully analysts are aware, that EPS accretion is very easy to achieve if you leverage up because debt's cheap as a source of financing.

Yep.

So there's a risk that you just achieve EPS accretion by using too much leverage.

So having an idea of what any debt financing does to your leverage numbers is really helpful. So here they've calculated the pro forma debt, pro forma net debt, combined LTM EBITDA, including synergies, and that allows us to calculate pro forma net debt to EBITDA, including synergies.

I'll just highlight the number at the bottom there, 2.9 times.

So it just gives us a bit of reassurance that the debt used to finance the deal doesn't result in them being a huge risk of credit rating downgrade or anything like that.

Yeah. And the one thing that'd be interesting to see is just what the standalone leverage positions were before for this deal, because I know a bunch of new debt is getting issued to make this deal happen.

Yeah, that's a really good point, because I don't think there is anything...

No, it hasn't given us anything on a standalone basis. So no, you're right.

It's a bit of a shame that you can't compare that set up in the leverage multiples.

Yeah. And do we just combine EBITDA? I don't know.

I can't remember if up above in the model we had standalone EBITDA, and you can just calculate it quickly yourself.

Ah.

But it would be helpful if Claude had thrown that in.

Yeah, it would be. In fact, what we could do is actually update the prompt.

Should we do that? Oh, yeah.

Okay.

See, I'm so polite, I say please. Please add standalone.

There we go.

And then one thing I noticed while it's tinkering on that, one thing I noticed is the sensitivity analysis down at the bottom is still following those same kind of not data tables, but all one giant formula. It's really hard to audit.

And obviously we finally got the answer why that is the case, because I think it's something like the Excel Claude API does not allow data table creation.

So we've got to run with these kind of crazy tables for the time being.

Yeah.

Or I think last time when I asked it to do a data table, it just literally gave me the setup and even said what prompt, not what prompt, what cell to link to what input, that kind of thing. So it's not that either.

Yeah, absolutely. So because it's running the query at the moment, it's kind of jumping around a bit. But yeah, so what you were referring to there, Graham, the crazy formulas. Yeah, we definitely still have those.

But we do have a decent data table at least, which allows us to see how the EPS accretion varies in terms of financing mix and the level of synergies which can be achieved. So, yeah, we have the outputs, but we just try to ignore the actual formulas which have been used to derive those.

Yeah.

Right. How is our query getting on? It's still running in the background.

Oh, here we go. Let's see.

Oh, got some numbers.

Okay, so it's another full turn of leverage. It's not nothing.

Yeah.

In the context of a corporate credit like this.

Because in the world that I used to work in, looking at something that's 3.4 times leverage, you say, "Oh, that's like senior bank debt, that's nothing." Yeah.

But for a corporate credit like this, it's a little bit more material.

It is. There's obviously a little bit of industry flex, and the one thing I do know about railroads is they're very asset intensive, and you can generally stretch the leverage metrics- Yeah ... where you've got all that track, rolling stock, whatever.

So yeah, we probably would expect the numbers to be higher than for lots of other kind of industries like consumer businesses.

That's totally fair. Yeah.

Yeah. But yeah, as you say, another turn, 1.2 times excluding synergies or 8.8 times extra when you include the synergies. So it is a big jump up in the leverage for them.

Yeah.

So yeah. But it's useful to have that information side by side.

So what do we think about the deal? Is it weird if I say I don't care? I just- No. I think I was quite excited by the Claude elements and the running, using the orchestration of, using Fable. The deal itself, oh, I don't know.

I didn't get excited by this one, Graham.

It's just a bit meh.

Yes. It is a bit meh. The numbers though are not bad.

You've got EPS accretion, you've got leverage, which doesn't look too bad, very scary, so.

Yeah. Well, when you ever see a public merger announcement that's not accretive to EPS, right? True.

It's just like insert synergy, goal seek synergies to make this greater than zero.

True. I think I have actually come across a deal, it was a defensive move by a company where the analyst running the numbers was like, "This isn't EPS accretive." But actually, funnily enough, the actual press release didn't comment on EPS accretion. It talks about all the strategic benefits and things like that.

Got you. Okay.

But yeah. It ticks the boxes on this deal, but yeah, there's not a lot of exciting stuff happening from my perspective.

No. We'll still take trains. They're still going to suck here.

That's it.

Yeah. That's it. But we've learned a new way to work with Claude, which I'm really excited by.

100%.

Great. Well, I hope everyone enjoyed listening to this week's episode of "What's the Big Deal?" I hope you all learned a little bit about how to optimize your use of Claude. And yeah, definitely we'll keep revisiting this topic as we get more and more skills with our use of AI. Tune in for next week's episode.