The firms furthest along with cloud have stopped arguing about whether market data belongs there. What occupies them now are the questions that outlasted the migration: where the on-premises boundary finally holds, what turns out to cost money, and whether entitlement frameworks built for headcount can survive machines that consume data at machine speed. Getting to the cloud resolved none of those issues.
That was the shape of the discussion at a recent A-Team Group webinar entitled ‘Optimising Cloud, Marketplaces & Managed Data Services‘, which brought together Christina Hois, Managing Director of Capital Markets Data and Analytics at BMO Bank of Montreal; John Heisler, Industry Principal for Financial Services AI Solutions at Snowflake; and Jakub Pater, Director of Real-Time Managed Distribution Service, and Adrian Murray, Head of Product for Pricing and Reference Services, both of London Stock Exchange Group.
An early audience poll set the tone. Asked to describe their firm’s current approach to market data sourcing and distribution, respondents split fairly evenly between predominantly cloud-native or managed-service models and fully cloud-native delivery. Nobody described themselves as predominantly on-premises – a result that would have looked very different a few years ago. With the destination largely agreed, the discussion focused on what happens after arrival.
Where the line settles
The boundary between cloud and on-premises has not dissolved so much as hardened around a single variable: latency.
Hois, the practitioner on the panel, was direct about where BMO draws that line. Heavy analytics, reporting and data storage – anything where instant processing is not required – runs in the cloud. Time-sensitive workloads do not. “In general terms, we’re keeping our time-sensitive workloads on-prem, so anything to do with trading systems or critical market data,” she said. The split maps onto the lifecycle of a trade: pre-trade and settlement-side workloads move; the execution-adjacent tier stays put. The line is not fixed – BMO’s roadmap has advanced materially since the start of the year – but the critical market data tier remains on-premises by design rather than by lag.
Pater, whose remit at LSEG is real-time streaming data, described the same boundary from the supply side. LSEG now offers a lightweight, cloud-native feed that has been widely adopted. The next frontier, he said, is the full tick feed – far heavier, and until recently assumed to be unsuitable for cloud delivery. Clients are beginning to explore whether a cloud-based application can work efficiently with full tick data, and whether ingestion of that feed is stable in a cloud environment. It is proof-of-concept work, not production. The line is moving through the front office by one validated workload at a time, rather than through a decision to migrate everything.
The scale pressure underneath that experiment is considerable. Pater said LSEG’s global network peaked at around 10 million messages per second five years ago, typically at the US market open; it now runs above 22 million updates per second at peak, with an internal projection of 50 million by the end of the decade. Geopolitically driven spikes have already overwhelmed some client applications, and firms have taken filtered, lighter versions of the service to cope. The volume curve is bending upward at the point when firms are trying to decide what their infrastructure should look like.
Cost
Cost predictability is routinely sold as a benefit of managed services and cloud-native delivery. The audience disagreed. Asked to name the single biggest barrier to expanding their use of cloud, managed services or data marketplaces, the largest group chose cost and unpredictable pricing, with governance, entitlements and regulation close behind. Nobody selected performance, latency, vendor lock-in or integration complexity – a result Murray attributed to the composition of an audience already substantially cloud-resident.
Murray unpicked the return-on-investment framing itself. Cost is not usually the dominant ROI measure, he argued; scalability, security and resilience tend to matter more than immediate cost reduction. But firms have historically underestimated the total. The simple view – compute, storage, licence costs – leaves out data-transfer and egress charges, the cost of replicating across regions or providers in a multi-cloud setup, the asymmetrical compute demands of AI workloads, and the engineering effort of building and integrating the cloud in the first place. Until a firm fully migrates, it dual-runs across on-premises and cloud infrastructure, paying twice for a period.
For Hois, the least-visible cost is the custom development work that goes understated – “it’s not baked in anywhere and it’s not thought of the way it should be thought of,” she said. The hidden variable is integration: in a hybrid configuration held together by design, certain asset data stays on-premises for latency reasons, other data sits in the cloud, and the work of combining the two is rarely costed upfront. Security and reference-data masters make it harder still, with more sourcing options, more integration decisions, and entitlements that become genuinely complex once data placement is treated as a strategic question rather than a technical one.
Governance, done well, is not about boiling the ocean, Hois argued; it is about understanding which data confers competitive advantage and placing and governing it accordingly. Data has become a line of business in its own right – something her team treats as an asset with a direct link to a business objective or a revenue stream, or it does not get worked on at all.
Pater set the operational counterweight against the idea of the cloud as an escape. The shift from on-premises removes the hardware refresh cycle – the capital expenditure and project mobilisation that comes round every three to five years – and moves firms from a capex to an opex model. But the operational burden does not vanish. “You don’t keep it in your basement anymore,” as he put it, yet firms still need skilled internal people to run and monitor the platform, together with a robust FinOps function to manage the consumption costs that replace the old capital cycle. He added that the difficulty of recruiting subject-matter experts to run complex market data platforms has itself become one of the stronger drivers of managed-service adoption – firms are outsourcing the day-to-day not only to save cost but because the expertise is hard to hire.
Entitlements under AI
Within an institution’s own walls, entitlements are largely a solved problem: roles, lease-based access, and visibility into any discrepancy between policy and practice. The data marketplace model – a single copy of data shared one-to-many from provider to consumer – has itself become the accepted standard for data distribution and consumption, displacing the SFTP files and the early-morning calls when a delivery failed before the market opened, which Heisler described as a level of loss the industry no longer tolerates.
What changes with AI is the arithmetic of exposure. When the consumers of entitled data are people, the growth in the number of potentially mis-entitled users is analog: a firm can only hire so many people, and headcount moves by a handful over a quarter. An AI consumer does not obey that constraint. “I could spin 50 agents up in a matter of 30 seconds,” said Heisler, and the risk exposure broadens accordingly, against contracts written between first and third parties on the assumption of human-scale consumption.
Demonstrating to a data provider that consumption of its data inside an institution remains within contract therefore becomes more urgent. Good governance, Murray said, rests on a firm’s metadata strategy, a central data catalogue with lineage and rights tagging, and telemetry that builds an audit trail of who accessed what. That machinery makes applying entitlements manageable within a firm; the exposure, both agreed, is at the boundary, where data traverses multiple cloud platforms and end-to-end traceability is easily lost. Heisler placed a current innovation specifically: using an AI layer to enforce the differences between one contract and another for the same data product – enforcement that keeps pace with consumption rather than trailing it.
Hois’s one governance discipline, stressed above the rest, was data quality caught as close to source as possible. The further data travels from source, the more errors it accumulates, and a quality problem should never be allowed to roll past end of day.
From chatbot to workflow
For Heisler, the value of AI is not in the chatbot but in reworking existing workflows so that AI sits where human judgement adds something and mechanical steps stay mechanical. Asking a chatbot the same question every day is a step backwards – that belongs in a stored procedure or a visualisation. His example of the real thing is buy-side research: a single workflow pulling first- and third-party data, structured and unstructured, that runs the deterministic parts as SQL and reserves AI for the parts that need judgement.
The prerequisite firms underestimate is the semantic layer that allows data to be accessible to an intelligence layer. When a human says ‘trade,’ they know which tables and columns hold the relevant information; an AI does not, and a semantic strategy supplies both that context and consistency of consumption across systems. At one end, conversational interfaces over first-party structured data are well understood; at the other, genuine reasoning across combined first- and third-party data is already running in production. Between the two sits the distance between an early chatbot success and a full agentic workflow, most of it semantic, and, Heisler suggested, the place where the timeline is being compressed fastest by AI coding assistants and similar tooling.
Hois closed the loop with her data-quality argument: using AI to maintain the very data quality that makes other AI applications work, building a compounding knowledge base rather than a system that answers the same question differently each time. BMO is building an AI digital assistant that already sits on her org chart, performing tasks at VP and Director level, with the expectation that its knowledge should compound rather than reset.
What changes, and what doesn’t
Asked what the picture would look like in two years, the panel split between what accelerates and what stays put. Hois expected a wholesale redefinition of how firms protect their most valuable assets. Pater expected cloud and managed-service adoption to accelerate, freeing clients to move from maintaining infrastructure toward value-generating work. Murray forecast a rearchitecture wave: firms that began by lifting and shifting data into the cloud will start rebuilding to take genuine advantage of what AI and cloud can offer, rather than replicating on-premises designs in a new location. Heisler put it as a nesting doll – firms will move on to higher-order problems, with today’s hard problem encapsulated and taken for granted, while data continues to sit at the centre of whatever comes next.
The firms furthest down this road have stopped arguing about the destination. The cost, the latency line and the entitlements pressure that remain are real, but they are the problems of arrival rather than of transit – and the firms confronting them now are the ones that will define what comes after.
Subscribe to our newsletter
