This week’s IMF Spring Meetings underscored what Rodrigo Buenaventura described as a “turning point” for global markets. AI, tokenisation, stablecoins, and quantum computing are converging to reshape market structure into a deeply interconnected system. But at the same time, the Financial Times highlights a growing policy gap – with frontier AI capabilities accelerating faster than regulatory frameworks and U.S. resistance to federal AI regulation, responsibility is shifting toward private sector coordination – Read here – making global standard-setting bodies such as IOSCO increasingly essential. Against this backdrop, AI in trading is moving beyond execution optimisation towards full front-to-back system integration – forcing a European regulatory rethink of governance, control, and market stability. Here’s what I learnt this week on AI in Trading:

1. AFM: AI as the Future Core Operating Layer of Markets

The Authority for the Financial Markets positions AI as the new operating layer of capital markets – improving efficiency and price formation, but introducing systemic risks through data dependency, model opacity, and tightly coupled autonomous systems. They argue that future market integrity will hinge on model design, data governance, and firm accountability – not the pre trade risk controls currently in place for most firms. – Read more here

Why this matters for trading: With the EU AI Act requiring continual monitoring of high-risk systems from August 2026, control is shifting from monitoring of individual trades to supervising systems.  The AFM paper acknowledges that the market is likely to evolve into a mixed eco-system where well-governed AI models will interact with less trusted opaque ones, meaning market stability will depend on real-time oversight of model interaction, data integrity, and feedback loops in the ecosystem – not just pre-trade controls.  The paper also provides an indication of how regulatory oversight is likely to adapt – extending MAR controls, revisiting Trading Venue Perimeter, Agent identification and containment, along with adapting for self-learning algorithms – noting yet again that “accountability does not change with automation”.

2. Runtime Governance: From Controls to Continuous Enforcement

As discussed in the FIX AI and Algo working groups, traditional pre-deployment controls (e.g. Model Risk Management) are increasingly viewed as insufficient for AI-driven workflows, driving a shift to continuous, runtime governance. A new paper by Lukasz Szpruch, Agus Sudjianto, Tanveer Bhatti and Gary Ang – “Scalable Runtime Governance for Agentic AI in Financial Services” introduces a runtime governance framework for agentic AI. The model moves from static controls to policy embedded directly into execution, structured across four tiers – from assistive tools to fully autonomous systems with irreversible impact. – Read more here

Why this matters for trading: This provides a blueprint for aligning AI governance with market structure. Combined with evolving FIX tags and newly proposed identifiers like the vLEI, it could provide future classification, monitoring, and constraint of agent-driven workflows in real time – shifting oversight from observation to active monitoring and containment of agentic activity.

3. The Unbundling of the Terminal: AI-Native Execution Stacks

After a week of yet more AI execution intelligence announcements – this time from TS Imagine (read here) and Glimpse AI (read more here) there was also another challenger to the BBG terminal from Fincept Open Source (read more here). Along with announcements from Microsoft 365, this emphasises the continued gradual shift away from the “trading terminal” model to agentic, always-on AI layers embedded into everyday workflows  – providing access to real-time financial data, analytics, and research enabling firms able to build custom AI agents without a dedicated terminal environment. – Read more here

Why this matters for Trading: the transition to systematic, agent-driven trading is accelerating the democratisation of capabilities. Differentiation is no longer about who has access to the best tools, but who can most effectively integrate proprietary data, govern autonomous agents, and translate insight into execution across increasingly fragmented and automated markets. Fintechs are no longer competing to replace the Bloomberg terminal. They are competing to become the intelligence modules inside an AI-native workflow layer.

4. Time as Critical Infrastructure in Agentic Markets

As highlighted during World Quantum Day, even frontier technologies remain dependent on precise, trusted time; without it, accuracy, security, and coherence break down. This post from Sarah Young is a great reminder on the risk of “unpriced systemic risk” – read more here.

Why this matters for trading: In increasingly distributed, agent-driven systems, timing becomes a systemic dependency. GNSS alone is insufficient – resilient, auditable synchronisation is required to prevent drift, ensure determinism, and maintain market integrity under stress. Timing is no longer a utility; it is core infrastructure. – Read more on the paper here

5. The Physical Constraint: Data Centres, Power, and AI Scaling

As highlighted by Fieke Jansen, the Dutch grid operator TenneT has temporarily refused to connect several newly planned data centers to the electricity grid due to the risk of overload. This has triggered legal action from an Australian developer, which could see TenneT face penalties of €500,000 per day starting June 1 as TenneT is state-owned and funded through electricity bills, the financial risk could ultimately fall on the public – read more here. Also highlighted is Amsterdam’s request to data centre operators, including Equinix, to reserve a portion of their capacity for European governments and businesses as part of a broader push for digital sovereignty – Read here. Equinix responded promoting its global, customer-neutral business model – highlighting the rising structural tension between public policy goals and how hyperscale and colocation providers operate. – Read more here

In the US the backlash against AI-driven data center expansion has led to Maine becoming the first US state to ban data center construction – read more here. There are now communities blocking over $156bn in data center rollouts according to Jigar Shah – read more here. He argues that the solution is not to halt AI build-out, but to redesign it: require data centers to adopt flexible, interruptible power usage, shift utility planning from overbuilding to optimization, and leverage distributed energy (batteries, virtual power plants) to reduce peak demand. Without reform, costs are socialized while gains remain private, and as a result public resistance will continue to intensify.

Why this matters for trading: AI performance is no longer just a function of model quality – it is constrained by compute, power, and architecture. Even as greater technological innovation becomes available, as costs scale, inefficiencies such as “tokenmaxxing” and poorly designed agent systems (“AI heat death”) become material risks. The competitive edge shifts towards access, efficient architecture, resource optimisation, and resilient infrastructure design.

With most buy-side firms still dependent on broker algorithms, what’s interesting is that trading technology is shifting from traditional vendor offerings to agentic, always-on AI layers embedded into everyday workflows but still with the same regulatory obligations. The AFM also refer to the possibility of regulating possible unauthorised trading environments before AI agents are deployed at scale. Along with this regulatory oversight, there are also practical implications of the new technological constraints as Quantum evolves – and more basic access to energy and water to keep the technology running. Those who can provide access to ever smarter execution models while monitoring complex autonomous systems at scale will be those who drive future execution and shape liquidity. As markets become more automated, systemic and interconnected globally, the absence of common global standards will amplify the risks of fragmentation, opacity, and systemic instability – yet even more reason for greater industry collaboration.

Thanks again for reading – more to follow next week. As always let me know your feedback/comments

Many thanks

Rebecca