One of the World’s Largest Hedge Funds on Its 86x Growth in Token Spending | Edited Transcript

How Man Group uses AI for investment research, systematic strategy development, data integration, token budgeting, governance, and knowledge sharing.

Chapter Timestamps

00:00 Direct investment predictions are not AI’s primary use

05:00 Man Group builds technology across its investment stack

07:30 AI agents expand the research a portfolio manager can cover

10:00 A podcast revealed constraints beyond scarce GPUs

12:30 Agents now generate and test systematic trading models

15:00 Safe organizational adoption is the main bottleneck

17:30 The data architecture combines three distinct layers

20:00 Metadata and shared semantics matter more than fine-tuning

22:30 Decentralized budgets and education govern model spending

27:30 Longer autonomous tasks are driving greater token use

30:00 Trading hypotheses require a written economic rationale

33:04 AI lowers barriers to less-structured markets and instruments

35:27 New hires must learn to supervise end-to-end automation

37:48 AI democratizes skills while moving work toward planning

40:28 AI alpha depends on an integrated market ecosystem

43:11 Embedded engineers reduce the need for outside restructuring

45:34 Shared workflows capture expertise without exposing every signal

48:16 Token growth still lacks a defined financial payoff

Made with: The Transcript Desk Chrome Extension

Full video:

Man Group CTO Gary Collier and head of data and AI Tushara Fernando explain how the asset manager applies generative AI across discretionary research, systematic strategy development, coding, and operations. They describe an agent-based research system that has produced 15 to 20 human-approved trading models, alongside an 86-fold increase in token consumption since January.

Transcript

00:00-02:30

Joe Weisenthal: Bloomberg Audio Studios: podcasts, radio, news. Hello, and welcome to another episode of the Odd Lots podcast. I’m Joe Weisenthal.

Tracy Alloway: And I’m Tracy Alloway. Joe, I’m very interested in AI. No, really, I am.

Joe Weisenthal: That’s a surprise.

I’m very interested in the investment context—specifically, the actual implementation. How do investors use it? Obviously, that’s an incredibly important question for reasons that need no explanation. But I also think it raises very interesting puzzles about what the technology is used for.

I remember when ChatGPT came out and people were asking it, “What stock should I buy?” We also did that prediction-markets episode recently, and one thing you definitely can’t get much value from is asking, “Which contract should I buy?” or “What’s inflation going to be, so I can trade this contract?” That doesn’t mean there aren’t interesting uses. It just seems that the crudest version of “using AI for investing” is obviously a dead end.

Tracy Alloway: Here’s the question I have. We’ve been through technological revolutions in investing before. Notably, we had robo-advisors—remember those? We had systematic investing and high-frequency trading.

My big question is: How much of the current use of AI is basically an iteration or improvement on those kinds of machine-learning dynamics, versus something more substantial or revolutionary? Is it a minor evolution, a moderate evolution, or something big?

Joe Weisenthal: This came up a little bit in our conversation with Ian from Hudson River Trading. I’m glad you brought up the quant example because, in some sense, big data has been part of quant investing from the very beginning. How do you establish that value stocks outperform expensive stocks? You need a lot of data, computers, and the ability to run the math.

02:30-05:00

Joe Weisenthal: But then there’s this idea that, yes, maybe cheap stocks outperform expensive ones, and momentum stocks outperform stocks with poor momentum—things that seem to be true, but we don’t really know why. There’s a lot of disagreement about the source of these quant effects.

That makes me think about AI because we can get outputs from AI models that are obviously remarkable. They can recognize patterns, but they can’t really explain how they arrived at those patterns. I’m very interested in where this leads us in investing and whether we get ideas and strategies that work but that we can’t articulate in plain English.

Tracy Alloway: I’m glad you brought up data as well, because one thing you hear at basically every finance conference nowadays is the importance of data when it comes to running LLMs: making sure your data is processed and clean, that you have a lot of it, and that, hopefully, you have proprietary data. The people we’re going to talk to have a lot of data.

Joe Weisenthal: Totally. There was one more thing—I don’t know if you caught it last week. We’re recording this on July 7, and while you were on vacation, a very interesting paper came out from Bridgewater about its use of proprietary data to fine-tune an open-source version of Qwen for the specific purpose of identifying which financial information is newsworthy and which isn’t.

They found that the combination of open-source models and proprietary data produced better results than the leading U.S. frontier models. That is very interesting to me. These are the types of things I’m curious about.

Tracy Alloway: Where does the secret sauce—or the alpha—actually come from?

Joe Weisenthal: Especially when everyone is going to have access to really intelligent models.

Anyway, enough talk from us. I’m really excited because we have the perfect guests to discuss the implementation of AI within investing and asset management. We’re speaking with Gary Collier, Man Group’s CTO, and Tushara Fernando, the firm’s head of data and AI.

Gary and Tushara, thank you both so much for coming on Odd Lots. Let’s start with Gary, and then perhaps both of you can go one after the other. Could you give us a quick description of your roles at Man Group and what you specifically do?

Gary Collier: I can start with that. It’s perhaps useful to give a bit of context about the firm. I describe Man Group as a full-spectrum active asset manager.

05:00-07:30

Gary Collier: By “full spectrum,” I mean that we cover alternatives and long-only strategies; public and private markets; and systematic and fundamental discretionary investing. We also have a solutions business that can bring all of that together into customized client mandates.

I would describe the CTO role as full-spectrum too. Across all of the above, we’re very opinionated about what every part of the technology stack looks like, from the choice of servers and networking devices all the way through to end-user-facing software.

Similarly, from left to right, we build an awful lot of our own technology—from custom data feeds through research frameworks and trading systems, as well as our middle- and back-office operating platform. So it’s very much a full-spectrum role.

Tushara Fernando: On my side, I look after data and AI. That covers everything from market data and alternative data to all of the context and knowledge that we plug into our AI models.

From an AI perspective, it’s about how we provide these great new capabilities to our quants, researchers, and fundamental portfolio managers.

Tracy Alloway: Tushara, I wanted to ask you about this because I think your title used to be head of data and machine learning, and now it’s head of data and AI. How do you think about the difference between those two terms—machine learning versus AI?

Tushara Fernando: That’s a really good question, and I think it links to one of your earlier points about whether we’re using traditional machine-learning techniques or generative AI.

Historically, machine learning in a quant firm was really about using more traditional techniques—looking at things like linear regressions and neural networks to try to predict some future outcome. With the onset of generative AI, it’s now more than that. It’s about enabling people to create things, as well as using those traditional techniques.

That’s why we changed the title of the role: It now encompasses both the generative-AI aspects and the more traditional machine-learning methods.

Tracy Alloway: That makes sense. If I were sitting in a Man Group office today—and Gary laid it out perfectly well; you have a lot of different roles, from discretionary, traditional portfolio managers to systematic quants—and I were shadowing one of your traders or PMs, what would they be doing on their screen today that differs from what they were doing in, say, 2022?

07:30-10:00

Tracy Alloway: What are they doing differently with generative AI?

Gary Collier: If you walked across the floor, you would see different windows and different tools in use now, regardless of the role—whether it’s trading, quant research, or discretionary investing.

I was chatting with a couple of discretionary analysts last week, and they commented that everywhere you look across the discretionary floor, everybody has an AI-focused window on their screen. It’s affecting, improving, and augmenting pretty much every role we have in the firm—in different ways depending on the role, of course. But there’s no role that’s unaffected by AI.

Joe Weisenthal: What are they doing? Let’s say I’m a traditional long-only stock picker—the classic role we often imagine at an asset manager. I now have access to some of the world’s most advanced models. What am I doing with them?

Tushara Fernando: Think about how a PM like that would traditionally work. They want to look across a broad set of names and access as much data as possible for those names. They want to look at earnings reports, broker research, alternative data, and podcasts.

But there’s only so much time in the day. There are only a few hours, and there are 50 names. How are they going to cover them all? How do they get to the really important insight?

AI has allowed us to access all those different types and modalities of data—podcasts, alternative data, and broker research reports—and synthesize them into what is actually meaningful. A PM can obtain that insight asynchronously. They can go away for a coffee or come in the next morning, and an agent will have distilled a change that occurred on the internet and presented the insight that is meaningful to their portfolio and investment thesis.

10:00-12:30

Tushara Fernando: One recent example involved a PM covering the AI trade. One of the important questions was: Where is the bottleneck? It’s hard to know exactly where it is.

There was a podcast with the head of engineering at a large hyperscaler. He said it was increasingly important to have more and more GPUs to train models, but that data-center capacity was becoming scarcer. It was increasingly difficult to find data centers large enough to train these models.

You could take a couple of things from that. One is that GPUs are scarce. The other is that we’re likely to need better networking between these large data centers in the future.

That information came from a head of engineering speaking on a podcast. This isn’t someone a PM would usually interact with. He doesn’t attend the investor conferences they go to, and he’s not somebody they often have access to. But through AI, the PM was able to have an agent transcribe the podcast and synthesize the information, giving them better insight into that investment idea.

Tracy Alloway: I feel like podcasts are an important source of alpha, Joe. Everyone should be listening to podcasts.

Joe Weisenthal: The non-joke version of that is something people say. We had Alex Namas on the podcast, and the question was: What will be scarce after AGI? People who can get good guests on their podcasts—people who can reach those guests.

Tracy Alloway: Okay, so we’re clearly talking our own book.

Setting that aside, it sounds like most of what you’re doing is augmenting research. You’re allowing investors and managers to be more efficient in their research processes.

There’s a lot of talk nowadays about agentic workflows: the idea that, instead of having humans drive every step of a particular trade or project, you have a system that can conduct the research, generate ideas, test hypotheses, and even execute on them. Is that something you’re working toward, or is it still too far away to contemplate?

12:30-15:00

Gary Collier: No, that’s not too far away at all. In fact, it’s something we’ve been working on for quite some time.

If we shift the focus to the quant and systematic parts of the business, think about it this way: Technology plus quant techniques gives you the ability to build systematic strategies. So what happens if we add AI to that mix? We gain the ability to systematize the way we build systematic strategies in the first place, effectively providing a big force multiplier—a big leverage multiplier—to our quants and researchers.

For well over a year—actually, about a year and a half—we’ve been building a system that can handle all the different parts of the quant-research process. That includes idea formation: looking at academic papers and carefully labeled datasets; reasoning about the content of those papers and datasets; and asking whether the economic hypotheses within them could be real, or at least worth testing.

Other agents then write the code to encapsulate those ideas, obtain the right market data, and run the appropriate backtests. Further agents evaluate the output of the preceding agents.

This was one of the early ideas that we thought could provide great bang for the buck, and we’ve been working on it for some time. To make this tangible—and demonstrate that it isn’t merely something happening in a lab without practical consequences—at the last count, I think 15 or 20 models had gone all the way through the process.

They began as models ideated by AI, progressed through signal construction and validation, and were then reviewed and validated by a human investment committee. They were deemed fit and proper to trade our clients’ assets. So this is very real. It isn’t make-believe at this point.

Joe Weisenthal: You mentioned hoovering up information that might appear on a podcast, where someone identifies a particular bottleneck in their business. We are on a podcast, so what is your bottleneck?

What would you like to have more of right now to improve your process? Is it compute? Is it data? If you could snap your fingers and have more of something, what would it be?

15:00-17:30

Gary Collier: You can always use more compute, and you can always use more data. But I think the bigger problem is the world of opportunity supported by the AI capabilities we’re seeing—and, of course, the new capabilities being delivered all the time. That envelope is expanding so quickly that keeping up with it from a human perspective can be quite hard.

There’s no shortage of good ideas. Filtering ideas is important, but while we’re used to making lots of changes at Man Group, the level of organizational change required to put all of this into effect is, I think, the bottleneck.

We’re a regulated business, and we have a fiduciary duty. We need to ensure that the things we’re doing and deploying are done with the minimum amount of risk. There’s a lot of work in this field for which the industry doesn’t yet have firm answers.

Evaluating and testing the results of AI processes is good and proper, but that’s a rapidly expanding field. We also have to think about how we run agents across the business in a safe and controlled fashion.

I’m sure we’ve all seen—and smiled at—the stories about AI deleting someone’s inbox, or deleting all their photos when they can’t get them back. We can’t afford to have that type of thing happen at an enterprise level. Making sure that we can move quickly but safely is the bottleneck, if you like.

Joe Weisenthal: That makes sense. Let’s get into a couple of specifics, and I want to return to how you move forward safely.

Both of you have now mentioned filtering, and that is precisely what I discussed in the introduction. There was a paper from Bridgewater about fine-tuning a version of Qwen on its own data to improve filtering, so PMs could use their time more efficiently and focus on the signal.

How do you build that? Is it an off-the-shelf system? What technology and models go into the ingestion pipeline? What does it consist of?

17:30-20:00

Tushara Fernando: For us, it really depends on the type of data. There are broadly three types of data that we look at.

The first is market data, which is often highly structured: tick data, order books, and similar information. We ingest every tick from most exchanges—almost a terabyte of tick data per day.

Then we have alternative and unstructured data, which is much messier and more malformed. There, it’s really about how we structure the data, tag it, and connect it. What is the knowledge layer on top of it? How do we think about the connections between tickers, sectors, and companies? How does AI interrogate that data in a uniform way? What common language connects all those datasets?

The last piece is quite new: our institutional knowledge and context. This is becoming a new layer in our data architecture. How do we tell AI about our processes? How do we make it speak Man Group? How do we tell it the right way to run a backtest?

Those are the three areas we focus on.

Joe Weisenthal: Just to push further on this point: All of that makes a ton of sense to me, especially the latter two categories, which are big problems. We know that generative AI solves a lot of the unstructured-data problem. You can extract a lot of signal from a mass of text dumped into a file.

But what are you building? How does it work? Are off-the-shelf frontier models best for the job, or do you fine-tune open-source models in-house? Today, what is the best technology stack for that part of the process?

Tushara Fernando: Historically, we’ve looked at fine-tuning for a couple of use cases. But where we’re seeing the most bang for our buck at the moment is in properly tagging and structuring the data—in other words, preprocessing it.

We’ve found that if you take a dataset such as credit-card data, AI can look at it and see the columns and rows, but it doesn’t really understand the nuances of what it represents.

20:00-22:30

Tushara Fernando: We therefore have to invest in adding extra color and metadata to that information. We want descriptors that explain the nuances. We might say, “When you look at this dataset, each row means that a person went into a shop and bought something.” You tell the AI that in plain English and provide those descriptors for all your datasets.

The second piece is connecting many datasets. They use very different language and semantics, so we’ve had to invest in a shared language—a semantic layer and a common way of doing things. We tag all the columns with this unified language so the AI can connect different datasets.

That’s very important in idea generation. The AI understands that a field in one dataset is linked to a field in another, and it can quickly navigate among tickers, sectors, and macroeconomic insights.

Tracy Alloway: What do you think is more important: having the latest frontier model or having a beautiful set of structured, tagged, and labeled data?

Tushara Fernando: It depends on the task you’re trying to perform. If you’re looking at a coding task, you really want the latest frontier models. But for a quant-research task, the underlying data is what fuels alpha research. Simply trying to use a frontier model will get you nowhere.

Joe Weisenthal: This is something that comes up in many of our questions about AI deployment, and perhaps both of you could address it.

You have all these different teams, and I assume everybody—including people who don’t know much about AI—intuitively thinks, “I want the strongest version of the model. I want Opus 4.x and whatever else is best.”

For deep computational tasks, engineering problems, or coding problems, those probably are the best. But many people probably don’t need anything like that at all.

How do you think about internally providing tokens and managing consumption—avoiding wasted money from people using the most advanced models, while also giving them enough room to explore and discover the maximum potential value they can obtain from AI?

22:30-25:00

Gary Collier: The economics question is an interesting and important one.

Toward the end of last year, when we knew growth was likely to accelerate rapidly, we modeled what we thought the company would begin to look like in terms of different classifications of users and use cases. We came up with budgets, and this year we federated those budgets out to all the business units within the firm.

I’m a strong believer in pushing decision-making down to the lowest possible level at which it makes sense. That allows people to be agile and make the economics work for them and their departments.

Budgets are fungible, of course. If a department wants to move money from some other area of spending and say, “We think we should buy more tokens,” it is free to do so.

The AI platform we’ve built offers a very rich set of models. It isn’t 100% complete, but it includes all the main frontier models and a number of different open-source and open-weight models.

Joe Weisenthal: Do you have a model that routes queries to the optimal, most cost-efficient model? I know there’s a lot of interest in classifiers that route queries, and the big AI labs have them themselves. Is that something you’ve built in-house? How do you solve that problem?

Tushara Fernando: We could do that, but we’ve chosen not to. We want people to understand the dynamics of how best to use AI and which models to use.

Instead, we’ve leaned on education. We have a very rich dataset showing how people use AI, so we have great insight into it. Some of the things we observed were quite basic mistakes.

Often, people would perform multiple tasks within the same context window. They would try to determine the best way to write an investment thesis, then ask where to go for lunch, and then ask what the weather would be tomorrow.

25:00-27:30

Tushara Fernando: That’s a well-known problem if you’re in the weeds of AI, but we have 1,700 or 1,800 people at Man Group, and their technical understanding of how these things work at a fundamental level varies.

We’ve therefore been focusing heavily on education. We’re very transparent about the budgets people have and how those budgets are being spent, and we talk to them about the different classes of models.

Through that, we’ve seen better results. We’ve also seen people find quite creative ways to reduce token spending and contribute those solutions back to the platform.

Tracy Alloway: Say more about that.

Tushara Fernando: For example, suppose you’re using a coding agent and want to perform some basic Git commands to interact with the version-control system. Often, the coding agent will send the command and then process the entire result returned by that command as tokens.

Instead, there are simple tools and basic steps you can use. Rather than calling an LLM to perform inference and make tool calls, the system can intercept that tool call and handle it outside the agentic loop.

Tracy Alloway: If I were looking at a chart of your overall token consumption, what would I see? I assume it’s upward-sloping despite some of these efficiency efforts, but how steep is the slope at the moment?

Tushara Fernando: Since January, I think token consumption has increased 86-fold.

Tracy Alloway: Wow.

Tushara Fernando: It’s really quite incredible. We weren’t expecting usage to grow the way it has, and the growth has been across the board. It isn’t just in technology and technology-adjacent departments.

We’ve seen people in finance, operations, and the people team using agentic coding workflows. As a technologist, it’s super exciting to give this new capability and power to people who haven’t been able to use it before.

Joe Weisenthal: This gets to a question I’ve been wondering about. It definitely feels like December and January were a pivotal period.

27:30-30:00

Joe Weisenthal: From your perspective, how much of that sharp inflection was driven by the capabilities of the models themselves—the progression from one frontier-model release to the next—and how much came from the emergence of high-quality harnesses such as Claude Code or Cowork?

Those tools allow someone to do things with AI that go beyond asking questions and extend to manipulating real work. Which would you describe as the key driver of that huge acceleration: the model or the harness?

Gary Collier: I think the two are coupled.

Tushara Fernando: One interesting benchmark to consider is METR. It looks at tasks that humans would take different amounts of time to complete, ranging from a couple of minutes to many hours.

What we’re seeing is that, roughly every seven months, the length of a task that an agentic workflow can go away and complete doubles. You can now ask an agent to perform a task that would take a human 16 hours. That changes how you think about teams and how you interact with these agents.

You move from being in the loop—asking an agent to solve a problem such as writing a unit test—to assigning it a problem such as building an entire feature for an application, or even an entire application itself.

I think it’s this scalability that has allowed larger tasks to be completed, resulting in greater token usage.

Tracy Alloway: I want to return to the oversight question. We all know that finance is a highly regulated environment, and you touched on this earlier. You’re still having humans approve many of these model outputs, so there is oversight.

I assume that when a human approves a new model or an output, they have to understand what is actually coming out of it. There has to be some explainability.

If I’m a PM—or perhaps a quant—sitting in front of a risk-management committee or a regulator, what does explainability actually look like? How do I translate the model outputs into something understandable and defensible?

30:00-33:04

Gary Collier: Yes, you’re right. Explainability is super important to us. To be clear, we’re not in the high-frequency-trading business, where people construct huge neural networks, look through multidimensional spaces, and produce outputs that are completely inscrutable. Our trading and holding-period horizons are typically days to weeks or months, so all our trading decisions are ultimately explainable. We never want to be in a position where we don’t know why a trade happened—where the answer is simply, “The AI did it.”

Going back to some of the things we discussed earlier, take the example of AI generating brand-new trading hypotheses based on what it has seen in a dataset and the contents of an academic paper. The system—the agent—goes as far as naming and writing the investment hypothesis.

Tushara Fernando: That’s one of the first things it does, before it moves on to writing code.

Gary Collier: It gives us a real English description of what it thinks the rationale for the trading signal is.

Joe Weisenthal: Obviously, we haven’t even gotten to the future of labor, the humans in the loop, and how many humans we’ll need in the loop in the future. But I’m curious about another way of asking that question: Has AI allowed you to examine markets that you wouldn’t previously have had the bandwidth to cover?

For example, take the Ethiopian stock exchange. A certain amount of human labor would be required just to gain any familiarity with it whatsoever, setting aside everything else. No matter how much potential profit there is in a frontier market, there’s a minimum cost involved. That might take some investment opportunities off the board because the potential profit isn’t large enough to justify the expense of getting up to speed.

This strikes me as something AI could potentially help with by reducing some of those upfront human-labor costs and creating a new opportunity set. Is that something you think about or have seen specifically at Man Group?

Gary Collier: I think it’s probably happening incrementally at the margins, if we add up all the different micro-augmentations we see across the firm.

33:04-35:27

Gary Collier: Here’s a related example: Someone built a relatively simple AI system that takes data from complex-instrument PDFs and specifications and automatically populates our reference-data store. So yes, I think it’s absolutely happening, but it’s the sum of many different parts across the firm.

Tushara Fernando: What we’re also seeing in the systematic space is that you need a few prerequisites to build a systematic strategy. You need connectivity that allows you to trade the instrument, and you need to understand the market price. Those are two fundamental things.

For developed markets, that’s really easy: You look at the order book. But less-developed markets and areas such as crypto or securitized credit are often harder to connect to. They may be traded more verbally, or the contracts may be complicated, making it difficult to understand the price because of unstructured-data nuances in its derivation. We’ve seen that AI allows us to consider accessing those markets systematically earlier than we could before.

Tracy Alloway: Going back to the labor-market side, if I’m a portfolio manager at Man Group and more and more of my job involves using AI for research—or even supervising agents that I may have helped develop—what does that mean for the talent you’re looking for? Are you looking for engineers who can tweak these models? Are you looking for more traditional investors with stronger intuition about how these things might play out, or some combination of those characteristics? What do you look for now?

Gary Collier: I think it’s fair to say—and this is something I’ve been making a strong case for—that everyone we hire into the firm should raise the bar with respect to AI, regardless of the role they’re coming in to do. I think that applies as much in operations as it does in the front office.

35:27-37:48

Joe Weisenthal: What does that mean? If someone says, “I’m capable of raising the bar with respect to AI,” what does that person look like in the recruiting process?

Gary Collier: It means being as familiar with the technology as you could reasonably expect someone to be, given the wealth of information available.

When I’m hiring people, I want bright, motivated people who get things done and are passionate about their subject matter and field of expertise. Nowadays, it’s very hard to fulfill all those criteria and then say, “Well, I don’t know much about AI. I don’t really use it as part of my job.”

Tushara Fernando: The other thing you want is someone who’s more of a long-term thinker—someone who genuinely wants to automate a process from end to end and is happy not to be in the weeds or in the loop.

When you think about technologists, that’s actually quite difficult. People love being in the weeds. They love debugging issues and getting into the nitty-gritty. But fundamentally, you want to level up. You want to be a conductor of these agents. You want to oversee the end-to-end process rather than necessarily being in the weeds—almost like a manager with a lot of technical expertise. Someone who thinks about things in broader, strategic terms is much more valuable than they used to be.

Tracy Alloway: This leads to the other thing I wanted to ask. You could see the arrival of AI tools generating two different outcomes. You might have some people who are extremely good at using AI and become superstars—people who orchestrate many different agents, as you put it. Or you could have a democratizing effect, where a junior employee with less experience can now automate many tasks, learn from AI, and use it for research.

Which are you seeing more of at the moment: the superstar dynamic or the democratization of skill sets throughout Man Group?

37:48-40:28

Gary Collier: I think we’re genuinely seeing both. Given the size of the firm, though, we’re seeing more of the latter. As I said at the beginning of the conversation, almost everybody is using it day to day. Examples of people doing genuinely cutting-edge work are naturally rarer, but there are a fair few of them as well.

Tushara Fernando: Those people often cut across multiple teams, too. That’s difficult because you need to move from spending most of your time executing to spending most of your time planning.

The time required to execute keeps falling. It’s becoming cheaper and cheaper. You can build code and features very quickly, so the focus really needs to be on what we should build, how it all connects, and what process we want to develop across multiple teams. It can be very difficult to take people out of their seats and get them to collaborate and plan a workflow instead of immediately trying to build a proof of concept and execute on the idea.

Joe Weisenthal: Let’s talk more about the 86-fold increase in token spend. If we’d had this conversation in January or February, much of the discussion would have been about what this means for legacy software companies. That was when the big software selloff was particularly intense.

But by July, I feel like the conversation has become: “No, these companies aren’t going after the world of software. They’re going after the world of labor.” You see people discussing the ratio of token spend to employee salaries and saying that the total addressable market is essentially all human labor—even if we may not get there for a while.

Tushara Fernando: I rather hope not.

Joe Weisenthal: Is token spend part of your technology budget, or is it becoming a distinct line item that’s more comparable to labor? When you think about Man Group in 2027 or 2028, do you discuss expected ratios of salary costs to token spend?

Gary Collier: We genuinely haven’t started having that conversation yet. I would guess that, at some point, it will come. The company is set up so that we direct resources to wherever they produce the best economic outcome for us. I’m sure that time will come.

40:28-43:11

Tushara Fernando: One interesting aspect of the token-budgeting process is that we’re increasingly seeing the spending driven not by people, but by agents. Those agents relate to workflows, which raises the question: Who owns the workflows? Is it this department or that department?

That’s a new problem for us—one we haven’t solved yet—but it’s a great problem to have.

Tracy Alloway: Earlier, I brought up the machine-learning parallel. I know you don’t do much high-frequency trading, but we’ve certainly seen a dynamic in HFT where everyone is competing in a race to the bottom. I can’t even remember where we are now—

Joe Weisenthal: In terms of microseconds?

Tracy Alloway: Yes, in terms of the time increments. But it’s that race-to-the-bottom dynamic. Would you expect AI-driven alpha to be competed or arbitraged away relatively quickly as everyone jumps on the same bandwagon? Or are there certain advantages—data and scale, for instance—that you would expect to persist for some time?

Tushara Fernando: I’d say it goes back to your question about where the alpha is. It’s true that it’s now easier for people to onboard datasets, analyze them, and build features. But that doesn’t mean they can trade them.

We’ve been doing this for decades, so we have the capabilities to access these markets. We have relationships with brokers and access to data that isn’t simply available off the shelf. It’s about putting all of those things together: the expertise, market access, all the data we have, the rich market data, and then giving AI access to capabilities such as backtesting and compute. It’s this whole network—this entire ecosystem—that I think drives alpha.

There isn’t one code repository at Man Group that I can point to and say, “That’s where the alpha is.” It’s really a network of different systems interacting with one another. Some features and datasets will definitely become table stakes. They’ll move from being sources of alpha to being risk factors because everybody has them and they move the market. But that isn’t the only way we make money. We’re able to connect different datasets and different ways of doing things, and then actually trade on those signals.

43:11-45:34

Joe Weisenthal: Gary, I want to return to something you said earlier when we were discussing bottlenecks. Of course, everybody wants more data and more compute. Nobody would complain about having those things. But where the rubber meets the road is whether an institution has the capacity—culturally and hierarchically—to get the most out of these tools.

This is clearly a very hot area. Just last week, for example, Microsoft announced something it’s calling the “Frontier Firm,” which is essentially a new subdivision that seems intended to solve this specific problem: going into an organization and identifying its optimal structure.

Arguably, that’s also what a company like Palantir is trying to do with its forward-deployed engineers. We’ve heard about Anthropic sending engineers into Goldman Sachs to help it really leverage—I hate using that word because it’s such a cliché—these tools.

What specifically are you seeing on that front? Do third-party companies come to you and say, “We can work with you to determine the organizational structure of the future for Man Group so that it gets the most out of these tools”?

Gary Collier: I did smile at the multibillion-dollar forward-deployed-engineer division you just mentioned, partly because that’s how we’ve been set up internally for about 15 years.

We’re very big on platforms, and we have several teams—Tushara runs one of them—that build large, cross-cutting elements of platform technology, in his case the data platform. But a large part of the technology team already consists of forward-deployed engineers, and has for a decade and a half. They sit with our quants and our discretionary managers.

One of the things I often say to people I’m interviewing to join the team is, “Let’s walk across the fifth floor here in Riverbank House, and I want you to tell me who the engineers are and who the quants are. I bet you’ll get it wrong.” What you’ll see on their screens is very similar, and that holds just as true today as it did a decade ago.

45:34-48:16

Joe Weisenthal: One last question. I’m curious about something else in the investment context. We know AI works when there’s a large pool of data and it can draw on both unstructured and structured data and communicate across them.

In an entity such as Man Group, are there alignment issues? If I have subject-matter expertise that generates alpha, that might be why I have a seat in the organization or why I’m considered a rainmaker. You hear about this at all kinds of firms, where compensation is linked to someone’s deep expertise in a particular area.

How do you align incentives so that the firm—the Man Group franchise—is actually capturing some of that superstar’s expertise and data? How do you get those people to contribute as much information as possible to a system that requires a lot of data and information?

Tushara Fernando: In some areas, that’s still a work in progress.

Gary Collier: We say “in some areas” because, in others—notably the systematic and quantitative areas of the business—a highly collaborative approach and shared codebases have simply been the way those areas have worked for a long time.

I’m not saying you can walk across the discretionary part of the floor and find all the fundamental investors being quite so free and open. I’ve talked to many of them about their processes. They’re open about 90%, but then there’s the remaining 10%: “This is where my personal value-add lies, and I’m not so comfortable talking about it.”

Even so, there’s a very decent and genuine amount of collaboration there as well. It’s only when you reach the particularly sensitive areas that people may become a little more reluctant to talk.

Tushara Fernando: At a high level, there’s a huge amount of shared workflow. Think about how we backtest or how we read an investment report. These workflows reflect people’s expertise, but they aren’t necessarily the 10% that produces the returns.

Those areas are encapsulated in AI playbooks and AI skills that we put into our knowledge platform, where they can be used across the floor. Some of the particulars around how a strategy works and the investment process are held back in certain cases.

48:16-50:40

Joe Weisenthal: That makes sense. Gary and Tushara, thank you so much for coming on Odd Lots. We really appreciate you taking the time to talk about where you are.

Gary Collier: Thank you.

Tushara Fernando: Our pleasure. Thanks a lot.

Joe Weisenthal: You know what I found really interesting about that conversation? There is still so much to figure out.

Tracy Alloway: Even the basic token budget and where that spending gets allocated.

Joe Weisenthal: An 86-fold increase since January is pretty crazy.

Tracy Alloway: That’s the other thing I was thinking. At some point, an 86-fold increase in token usage needs to show up in another concrete number, whether that’s expense reduction, revenue generation, or income generation. I don’t know how much leeway there is on that.

Joe Weisenthal: Right. At the moment, they’ve had this 86-fold explosion, but they’re still saying, “We want to build an internal router to minimize this.” Even after an 86-fold increase, they’re still at the stage of figuring out which model they want to use and experimenting.

To my mind, that’s actually rather bullish for token spending. It can grow 86-fold and still not reach the point where they say, “We really need to clamp down on this.” Who knows when firms that are just starting out will actually have to impose some token austerity?

Tracy Alloway: I suppose we’ll find out at some point. But the other thing that stood out to me was your question about how you get workers to give up their own secret sauce—the thing that’s effectively responsible for them having a job in the first place. Keystroke surveillance solves all of this, right?

Joe Weisenthal: Yes, although you need them to offer it up. There was a story a while back about Meta starting to train its models on its own internal data, and my reaction was, “They weren’t doing this already?” I was genuinely surprised.

Tracy Alloway: Especially in finance and investment, which are already highly regulated industries and presumably monitoring almost everything anyway.

50:40-52:14

Joe Weisenthal: I had assumed that all these companies building models were already using all the data their own employees generated. But does the rainmaker suddenly start writing everything down with pen and paper so that they aren’t implicitly uploading all their knowledge to the AI? I think that’s a pretty interesting question in itself.

Tracy Alloway: Clearly, there are lots of interesting questions.

Joe Weisenthal: Shall we leave it there?

Tracy Alloway: Let’s leave it there. This has been another episode of the Odd Lots podcast. I’m Tracy Alloway. You can follow me at @tracyalloway.

Joe Weisenthal: And I’m Joe Weisenthal. You can follow me at @TheStalwart. Follow our producers: Carmen Rodriguez at @carmendeiro, Dash Bennett at @dashbot, Caleb Brooks at @calebbrooks, and Kevin Lozano at @kevinlozano.

For more Odd Lots content, go to Bloomberg.com/OddLots for our daily newsletter and all our episodes. You can talk about all these topics 24/7 in our Discord at discord.gg/oddlots.

Tracy Alloway: If you enjoy Odd Lots and like hearing us discuss how finance firms are actually implementing AI, please leave us a positive review on your favorite podcast platform.

And remember: If you’re a Bloomberg subscriber, you can listen to all our episodes completely ad-free. Just find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening.