00:16 Mike Powell (Rapid Addition, moderator)

Is this working? Yes, good. Welcome to the last session. I’m conscious we’re between you and cocktails, but we’ll try to keep it interesting. We’ve had a fantastic session today speaking about South African capital markets in the context of harvesting growth, growing and accessing liquidity, and improving market resilience. Underpinning all of that, ultimately, is technology, so this is the technology session to finish off the day. I’m joined by three esteemed colleagues representing some major infrastructure and technology firms, and we’re going to have a chat about some of the key trends. Before we do that, could you each briefly introduce yourselves, starting with Eleni?

01:06 Eleni Coldrey (Equinix)

Hi everybody. Eleni Coldrey from Equinix. Equinix is the world’s largest data centre operator and home to the largest colocated capital markets ecosystem.

01:21 Gordon McArthur (Beeks Group)

Hi folks. Gordon McArthur, CEO and founder of Beeks. We’re a cloud computing company, responsible for the JSE’s Colo 2.0, and we work with exchanges, banks, the buy side and the sell side on low-latency infrastructure.

01:37 Alex Mirarchi (Amazon Web Services)

Alex Mirarchi from AWS, also a cloud computing company. I lead business development for financial market infrastructure across all asset classes, as well as trading connectivity in the cloud.

01:45 Mike Powell (Rapid Addition, moderator)

Thanks. So, three very experienced industry colleagues who I’m sure have lots to say on technology. I’m hoping to turn this into an open discussion rather than me reading out questions, but to kick off, could each of you give a couple of minutes on what you see as the key technology trends in capital markets, specifically around trading? Eleni, maybe start with you.

02:30 Eleni Coldrey (Equinix)

Sure. We host 17 different metros where exchange infrastructure sits around the world, so we see different trends in different regions. One of the biggest, and there’s been a lot of talk about it today already, is how to find growth: how to scale globally and find where the alpha is. That may be in new territories you need to move into, or in new segments. There’s a lot of talk in Europe about retail and access to the Savings and Investments Union, for example, and we see a lot of activity from emerging markets trying to get access to the major financial centres. On the regulatory front we see a lot about resilience, both from the EU’s Digital Operational Resilience Act and from customers themselves, in terms of creating a stable technology foundation, because this is critical infrastructure. So resilience on the network front and resilience on the technology platforms. And in terms of technology, I think everyone’s waiting to say AI. A lot of thinking around that, a lot of testing, and waiting to see how it really plays out in the market.

03:50 Mike Powell (Rapid Addition, moderator)

Gordon, from your point of view?

03:52 Gordon McArthur (Beeks Group)

Similar. A lot of the conversations we have are around clients trying to manage an ever-growing technology roadmap and footprint. People are talking a lot about emerging markets, Asian exchanges, South American exchanges, so how do we manage that infrastructure footprint, the networking and market data requirements across 40 or 50 colos globally, with an ever-shrinking team of people to support it? A lot around that. AI we’re going to talk about, so we’ll cover that later.

04:30 Alex Mirarchi (Amazon Web Services)

From a cloud adoption perspective, one of the biggest trends we’ve seen over the last decade or so is increasingly latency-sensitive trading applications running on cloud infrastructure. That’s both critical matching engines across market structures and market participants running their execution engines, algos and OEMs in the cloud. We see that on two dimensions. One is public cloud, an AWS region or a Local Zone. The other is hybrid, where we deploy an Outposts rack or a hybrid solution within a customer’s exchange data centre or an Equinix facility. That’s across asset classes. A lot of the world’s crypto exchanges run on AWS, like Binance, Coinbase and Kraken. A lot of fixed income, currency and commodity marketplaces are increasingly deploying in the cloud, which makes sense: they’re quote-driven, not as latency-sensitive as equities, and without fair and equal access requirements. Examples include Trumid, Broadridge’s LTX platform and Abaxx Exchange, a new commodities exchange launched in Singapore. Even equities exchanges: Aquis in the UK migrated its primary and secondary listing exchange to AWS last year. So it’s a global trend. We’re also working with the industry to launch cloud-native colocation environments. About a year and a half ago we partnered with crypto exchanges and high-frequency trading firms trading in the cloud to launch the world’s first colocation environments on AWS, which Coinbase and One Trading are using. The hybrid deployment model really picked up steam about three and a half years ago, when Nasdaq announced a strategic partnership with AWS. One of the primary elements was migrating all of its North American markets to AWS using Outposts. Since then, they’ve gone live with four exchanges and saw about a 10% latency improvement. TMX in Canada also launched an ATS in the US on that solution, and we’re seeing broad interest from other exchange operators, like the JSE, and market participants.

06:43 Mike Powell (Rapid Addition, moderator)

We’ll come back to some of those points later: where cloud, on-premises and colo fit across that spectrum of technology solutions, and you three are well placed to talk about that. One thing I’d like to do in this session is avoid falling into what people are doing in New York, London and Tokyo. South Africa is a very interesting market. It’s sophisticated and mature, with strong regulation and competing liquidity pools, which we talked about a lot today. But ultimately the pool is only so big, so there are cost sensitivities. To be honest, there are cost sensitivities in pretty much every financial market at the moment, particularly some of the emerging regional markets, so we should bear that in mind. Let’s dive straight into the AI discussion, and look at it from an infrastructure perspective, because we’re not quants, we do the plumbing. Market data is absolutely fundamental to efficiently operating capital markets. It’s getting more expensive, but people also want to extract more value and more actionable insight from it, and there are great new tools coming to market to do that. Machine learning has been around for a while; AI has been the big hot topic, or hype topic. What are you seeing in practical AI use cases being deployed in capital markets, ideally in trading, and what are the challenges or hurdles to adoption on the trading and execution side?

08:44 Gordon McArthur (Beeks Group)

I can start. AI is the ultimate hype cycle. To answer what use cases we see using proper predictive AI at the moment: very few. We have lots of conversations with banks, brokers and exchanges who’ve been given budget to do AI, and nobody really knows what that means. We have bizarre conversations with departments that have big budgets and don’t know what they’re doing. There are AI initiatives, and we’re looking at some things. Beeks is in a lucky position because we capture all the data an exchange creates, so we have the data to start running models and tools on top of. There are things like anomaly detection, where something odd is going on in your infrastructure and we can detect it, using more machine learning than AI. But when you start looking into forecasting, and latency prediction across venues so you can tailor your algos for specific performance, that’s difficult and very early stage. It’s like blockchain: people will spend lots of money, there will be lots of failures and some successes. It needs to be dragged back to reality, because a lot of the models we’ve looked at and are building ourselves are difficult. Maybe not so much in emerging markets, but the big US exchanges are pumping out around 100 gigabits per second. The compute and infrastructure you need to analyse that and make meaningful decisions in colo is huge and cost-prohibitive. It’s an expensive hobby at the moment for a lot of people.

10:38 Mike Powell (Rapid Addition, moderator)

That’s an interesting point, and we might come back to data consumption and power requirements. Eleni, maybe you could pick up on this. There is a lot of hype around AI-driven quant funds and so on. First, we should say what we’re talking about. It all sits under AI, but we’ve been doing machine learning for years, and as a speaker on an earlier panel said, we’ve been using algos and data modelling for years. I think we’re really talking about generative AI and how much of a black box that is. Gordon mentioned some of the safer use cases, analysing vast amounts of data for surveillance or system performance, and that seems like a safer entry point for these tools.

11:40 Eleni Coldrey (Equinix)

I was going to pick up on the hype cycle and where we are in it. There’s a clear distinction between training and tuning your models, and putting them into real-life production for inference. HFT and prop trading firms have had research clusters growing for a long time. We’ve definitely seen a trend where customers who might have run those workloads in exchange colocation are now separating them, because of the volume of data and the size of compute the research clusters require. That’s where you really have to understand workload placement, for any analytics platform, whether or not we give it an AI badge. Those consuming the volumes of data you’re talking about are separating those workloads from the low-latency workloads. There are examples in Europe of firms building huge clusters in Norway or Finland because of power costs. The interesting piece is that the network element, because of the volume of data, is starting to impact the business case. Whereas initially training was put wherever you wanted, it is now coming slightly closer to the market so network costs don’t kill the business case. When it comes to inference and putting it live, there’s a whole value chain of players involved. Are we seeing GPUs in exchange colocations? Absolutely: deployed by HFT firms and the prop community, and also by data vendors and risk analytics providers. But more often, what I hear from tier-one sell-side customers is that they could test these technologies in their own labs, yet the pace of change in AI technology is so fast that the tech is obsolete. So they’re telling the OEMs in the AI space, the NVIDIAs of the world, don’t ship me the boxes, I’ll test in a hyperscaler. We’re seeing a lot of testing in hyperscalers, and then when it comes to inference in production, it moves into private infrastructure. That privacy piece is really key: privacy, regulation and compliance, particularly in a space with so much IP. It’s a journey, and where you are as a firm in that cycle dictates where you place it.

14:29 Mike Powell (Rapid Addition, moderator)

Alex, you must be loving it. All this new data being crunched and models being run, and you work for one of the biggest cloud providers.

14:37 Alex Mirarchi (Amazon Web Services)

I guess so. I’m twitching over here because I’ve been programmed to talk about generative AI and how the world’s going to change. Two really important distinctions, and I couldn’t agree more. Traditional AI has been around for well over a decade in capital markets, at every stage of the trade lifecycle. One of the big trends from cloud adoption has been moving data from siloed on-premises databases to more modern, agile storage such as a data lake or data warehouse, so you can consume massive data feeds from multiple exchanges, integrate them with alternative, semi-structured and unstructured data sets, and run a whole range of analyses in real time. Generative AI really turbocharged that. Before OpenAI emerged, hundreds of thousands of firms were using AWS for their machine learning training needs; now I can’t keep track of the rate of increase. The hype cycle really hit in 2023, when everyone had a generative AI strategy, but I had many conversations with customers asking what an LLM is and how it differs from traditional machine learning. In the back half of 2023 and early 2024 we saw an explosion of proofs of concept, and in the last six months more and more have gone into production. At-trade is a little trickier because of latency requirements and the need to balance them against the cost of doing inference at scale. On the pre-trade side, a lot of customers use generative AI to streamline efficiency, extract information and drive product differentiation. Bridgewater, for example, launched a quantitative research assistant that can conduct financial analysis at the level of a first or second-year quant and respond to natural-language and code-based queries. Vanguard launched a series of investment glide paths, dynamically adjusting asset-allocation baskets for US retirement accounts. On premises they could achieve around 20 glide paths; by using the cloud they increased that to 1,200 unique paths, using traditional AI, and they now use generative AI to recommend solutions for salespeople to offer customers. In trading, a lot of innovation is still on the traditional AI side. The shiny object is generative AI, but exchanges are using AI to launch AI-based order types. Nasdaq launched Dynamic M-ELO about a year and a half ago, the industry’s first AI-based order type, trained on AWS. It dynamically adjusts holding periods and the crossing of liquidity, and by launching the AI version of the order type they increased fill rates by 20% and markout performance by 11%, so an increase in market quality. We’re also seeing trade recommendation engines proliferate, especially in fixed income. On the post-trade side, surveillance is a huge one. A lot of exchanges use generative AI to streamline investigative work. Nasdaq, FINRA and others have spoken publicly about how they previously used traditional AI to scale analysts’ work, and are now integrating generative AI to make information extraction and investigation faster. The numbers being quoted are around a 35% productivity improvement. Customers are using generative AI for settlement, extracting confirmations from emails, Bloomberg chats and even recorded chat transcripts. CCPs are using it to predict margin calls and settlement-failure risk and to minimise manual processes. And of course it helps with coding, so if you have to update a mainframe running COBOL or an old version of Java, we’re seeing a lot of uptake. One last point, because it’s important and relates to your question about limits to adoption: customers are launching common generative AI platforms for their entire workforce to bridge the skills gap, so you don’t have to be a quant or machine learning expert. You can be someone non-technical like me and build generative AI applications. BBVA, JPMorgan and others have spoken publicly about building common platforms that appeal to both citizen developers and machine learning specialists to build AI applications at scale.

19:23 Mike Powell (Rapid Addition, moderator)

A lot of exciting things are happening and the industry is moving fast. My concern is what that means for a cost-pressured, mid-tier broker. This is probably the wrong panel to talk about the applications and tools that will make it easier, but one thing this panel can talk about is that with any modelling, whether machine learning, AI or generative AI, it’s the data that makes the model. Our industry is highly fragmented. In a typical trading ecosystem, different systems generate data in different formats, and orchestrating that data is a challenge and can be very costly. Gordon, Beeks is one of the top managed infrastructure providers. Is that the sort of thing you get involved in, and how do you help firms manage large data sets and data infrastructure?

20:28 Gordon McArthur (Beeks Group)

Managing it, and more around moving it. As people trade on more and more venues, moving data around the world, and in and out of hyperscalers, is a very expensive exercise. So more and more people are looking to store in colo and archive into the cloud. The cost of storage reduces quite substantially, but moving, monitoring and managing data is more difficult than storing it. Again, this is very US-centric, but if you’re trying to arbitrage US data volumes anywhere else in the world, moving that data around is hugely expensive, so people are always looking for ways to manage and monitor it more effectively. We get involved, but we’re predominantly a low-latency infrastructure company, not a cloud storage company, so we do a little of it.

21:35 Eleni Coldrey (Equinix)

To add to that, because of data transit costs and the complexity around them, the advice is: move your model to the data. Look at how the large market data vendors architect across the world for ingestion and distribution. On a smaller scale, that’s how everyone needs to be thinking. If you’re smaller, to your point about democratising access to analytics, there are ISVs and service providers helping do this at scale. But effectively, where is the data originating? We’re not just talking about market data. There are all sorts of other signals: internal data you want to integrate, alternative data sources. The excitement around AI and the models you can run now is about where you can get the signals that give you differentiation. So understanding where the data comes from, and architecting your models around the physical origination points, is really key. That’s why we’re having this conversation: we’re the plumbers of the industry, putting all of this together, and the balance between efficiency and performance is what makes or breaks the business case.

23:01 Mike Powell (Rapid Addition, moderator)

That’s interesting. I spent a lot of years at Reuters running their global market data business, and when you talk market data, a lot of people think of buying data from exchanges, Reuters or Bloomberg, or LSEG today. That’s almost the tip of the iceberg. There are huge volumes distributed from global trading venues and data consolidators, but the amount of data generated internally, from analytics, position keeping, risk and so on, is significant. One thing we’re starting to see as a massive opportunity is that a lot of that data is formatted in different ways. You’ve got different protocols: FIX, native, binary, all sorts of formats, and it needs to be enriched or combined with other data for certain uses. A big challenge for a lot of banks is how they manage that information through their trading ecosystem, externally and internally, not just from an infrastructure and plumbing perspective but from a data model perspective. That creates a lot of friction and challenges in the industry.

24:06 Alex Mirarchi (Amazon Web Services)

Really good points, but I have a slightly different take. I’ve heard about the pain points from a data egress perspective. For reference, there’s no charge for landing data within AWS or other cloud providers, to my knowledge. What we see from a workload perspective is that large volumes of data aren’t really shipped from on-premises or an exchange up to the cloud and back again. I don’t know of any customer, and I speak to quite a few, who is landing the OPRA feed in AWS and then shipping it out again. What we do see is customers consuming data into AWS and using it to solve the pain points you mentioned: getting end-to-end visibility of their entire data estate across business lines, geographies and asset classes to power back-testing and research on the quant fund side. Many have spoken publicly about doing this for well over a decade. Bridgewater, AQR, QRT, Acadian and Balyasny do their back-testing, research and strategy development in AWS, and once they’ve identified their signals and algorithms they push them back on premises, into a Beeks or an Equinix. That created a pull effect for data vendors to host their data natively in the cloud, starting with Bloomberg and Refinitiv, now part of LSEG, then FactSet, S&P and Vela making historical data natively available on AWS. A lot of exchanges have made their data available natively in the cloud too, such as LSEG, NYSE, Nasdaq, B3 and JPX; the list is very long. We also saw a lot of banks, to your point about end-to-end visibility, move to real-time risk calculations rather than end-of-day batch, removing bottlenecks in high-performance computing with cost-optimised deployment models. JPMorgan, Morgan Stanley and others have talked about running millions of risk calculations a second, which they couldn’t do on premises, at about half the cost. It’s not just market participants; exchanges do this too, with the goal of a golden source of truth for their entire operations, then data warehouses for specific use cases like surveillance, billing, reporting, risk and analytics, and a clear view of a customer from exchange to post-trade to settlement. JPX has talked about building its J-Lake platform as a common data solution for internal and customer-facing workflows, and Nasdaq has spoken about its single source of truth. FINRA’s market surveillance platform, one of the largest data repositories in the world, runs in AWS purely to ingest multiple data sources at high volume, optimise cost and enable analysts to do more with fewer resources. But again, for latency-optimised deployments, those definitely remain on premises, next to the matching engine.

27:31 Eleni Coldrey (Equinix)

To pick up on some of that: there are different data requirements for every application a firm runs, which speaks to the complexity you mentioned, Mike. Any single firm has multiple front, middle and back-office applications that all need data but have different latency and depth profiles. So how does a firm create a cohesive market data or analytics strategy when it might need data in two or three different clouds, because its vendors have selected AWS or Google, and also on premises and in colo? It’s a complex picture. ISVs have a role in simplifying that, but ultimately what the buy side and sell side need is vendors that can offer the right data wherever they need it, in an easily consumable fashion. The distribution strategy of an ISV or market data vendor is a critical competitive advantage. You have to reach customers where and when they need it, and that’s going to be cloud, multi-cloud, hybrid cloud, on premises and exchange colo. It’s for the vendors to simplify that.

28:54 Mike Powell (Rapid Addition, moderator)

That’s absolutely right. We have an increasingly complex industry. Regulation makes it more complex, and the innovation coming to market, which is great and creates commercial opportunities, layers on more complexity. We can safely say our world’s not getting any simpler any time soon. Thinking about firms that are budget-stretched, certainly resource-stretched, and have more and more on their plate: Gordon, maybe start with you. What’s the state of SaaS and managed services, and how are they helping remove complexity for the average sell-side or buy-side firm?

29:41 Gordon McArthur (Beeks Group)

People are increasingly looking for a one-stop solution for trading infrastructure. There’s network complexity and infrastructure complexity, and people are trading on 40 or 50 venues around the world. Middle and back-office applications seem to have moved into the more generic hyperscaler clouds, and that’s fine. We talked about this last year: the front office and the matching engine remain in colo. I’m sure Alex will disagree with me, but I’m not aware of a single major equities exchange looking at moving its matching engine into the cloud. So this hybrid complexity is going to continue and be pervasive for a long time. There is no one-stop solution where everything moves up into the cloud. The hybrid architecture the world runs in at the moment will continue, and that creates cost. We’re asked more and more for end-to-end managed services across huge colo estates globally to try to keep people’s costs down, and I don’t see that ending.

31:14 Mike Powell (Rapid Addition, moderator)

I think that’s right. I don’t see that spectrum of solutions going away; the balance might change. As a business leader at a financial institution, you have to balance business outcomes, performance in terms of latency and throughput, and budget. Something has to give on that spectrum, so you’ll always have a mix to optimise results. Eleni, I guess you see that from your world.

31:45 Eleni Coldrey (Equinix)

I’d just add that the industry finds the efficiencies, and the industry is clustered in ecosystems where participants across the value chain sit under the same roof so they can exchange data easily and quickly, whether for low-latency reasons or for large exchanges of data in non-latency-sensitive workloads. Yes, it’s hybrid cloud, but it’s also about identifying where your value chain is and how you can move seamlessly between clouds and colo. That’s the role we play: we see ourselves as the edge between capital markets and the cloud, where that interchange can happen seamlessly. If you want the efficiency of the hyperscalers, perfect, those are the workloads. If you reach a certain scale where private infrastructure is more cost-effective, or you have privacy needs, bring it private and use a private cloud vendor. It’s about finding the edge that lets you move applications where you need them and meet your upstream and downstream. The ecosystem has naturally found a home next to each other, and I think that continues. There’s fragmentation in trading to an extent, but the clustering of the industry is stronger than ever.

33:04 Mike Powell (Rapid Addition, moderator)

Interesting. I’m conscious of time, so I want to give the audience an opportunity. Does anyone have any questions? Any energy at this time of day? No. The panellists will be upstairs afterwards, so you can corner them then.

33:40 Alex Mirarchi (Amazon Web Services)

I have one thing to add. I definitely agree with quite a few of the points raised. When you think about market microstructure, it comes down to the market type, but also the clustering, the liquidity centres of gravity that have existed globally for some time. That’s why we’ve invested a lot over the last five years in hybrid solutions, like the Outposts deployment I mentioned, which has been optimised for ultra-low-latency performance. That happened because Nasdaq had migrated its less latency-sensitive workloads, mid and back office and pre-trade, as Gordon mentioned, and said the next thing for us is the matching engine and meeting fair and equal access requirements. Honestly, in a cloud region we can’t do that today. That’s why the markets I listed running in a region either aren’t regulated to that degree, are running micro-auctions, as Aquis does on the equity side, or are quote-driven, so not as latency-sensitive. Achieving the performance institutional traders want for regulated equity and options exchanges is the technical challenge. So instead of forcing something that isn’t a natural fit today, we worked with our customers to bring the cloud to them. That took about two years of engineering. From there we came up with the concept of a private Local Zone, a fleet of services that could be GPUs, where we’re seeing more interest for inference at the edge, or ultra-low-latency offerings. We’re also hearing interest in FPGA-enabled instance types in cloud infrastructure. That’s where we’re investing, because those centres of gravity are where exchanges and market participants are, and we want to work with customers, provide choice and meet their requirements.

35:24 Mike Powell (Rapid Addition, moderator)

We’re just about there. I know you all have lots to say, so keep this to a sentence or two. Going down the line: over the next 12 months, what’s the one technology trend you expect to see more of, or to have a major impact on the market?

35:49 Eleni Coldrey (Equinix)

As a data centre operator, we’re looking at securing power, securing cooling and securing capacity. There’s high growth, and that’s the business we’re in.

36:06 Mike Powell (Rapid Addition, moderator)

It’s a shame we didn’t get into that, because particularly somewhere with the challenges of load shedding, all this new data analytics and AI is hugely consumptive of power. That’s going to be a key issue, here and globally.

36:20 Gordon McArthur (Beeks Group)

For us, selfishly, it’s more about moving the cloud to the exchange rather than the exchange to the cloud. Over the next 12 months we hope to get more exchanges moving our stack into exchange colocation to provide a service.

36:37 Mike Powell (Rapid Addition, moderator)

Excellent. Alex, to wrap.

36:40 Alex Mirarchi (Amazon Web Services)

More innovation on the data side: building data lakes and data warehouses to power agile analytics and AI applications, traditional and generative. You’ll see more partnership announcements with market infrastructure providers using the cloud to innovate. And that initial trend I mentioned: more low-latency trading using cloud infrastructure, either at an exchange, as Gordon mentioned, or in-region, to meet performance requirements. There were a lot of commas in that one long sentence. I cheated a bit.

37:10 Mike Powell (Rapid Addition, moderator)

Brilliant. Thanks for your insight; very experienced industry practitioners in the technology space. Thank you for your time, and thanks everyone for listening.