The Exchange Is the Last Institution That Thinks It's Immune

By Guy Melamed, Chairman of Mela Partners

There is a sentence worth sitting with before we begin. In every major structural disruption in financial markets, the institution closest to the center adapted last. Not because it lacked intelligence or resources. Because it had the most to lose from acknowledging the shift early, and the most institutional momentum pushing it toward the assumption that its position was permanent.

The telegraph made regional price arbitrage obsolete in a decade. Electronic trading dismantled the specialist system in a generation. High-frequency trading reordered the economics of liquidity provision before most exchange executives had finished reading the first academic papers on the subject. Dark pools drew off order flow for years before regulatory frameworks caught up. And each time, the exchange survived. It adapted, modernized, and continued. A sophisticated reader will recognize this pattern immediately and draw the obvious conclusion: exchanges have seen this movie before. They always find a way.

That objection deserves a direct answer, because this time the mechanism of disruption is structurally different from anything that preceded it. Electronic trading automated the execution of human decisions. High-frequency trading accelerated the execution of human strategies. Autonomous AI systems do not accelerate or automate human decisions. They replace the human in the decision loop entirely, continuously, across every layer of the market simultaneously. That is not a faster version of what came before. It is a different category of change. And the exchanges that treat it as the former will discover, at cost, that it was the latter.

Consider what happened in a single 48-hour window this week, and read it not as a set of product announcements but as a map of where the market is going.

Interactive Brokers formally opened its client accounts to Claude, Anthropic's flagship AI model. The system does not answer questions or generate reports. It drafts trades, proposes execution strategies, and generates trade instructions tied directly to portfolio analysis. The firm's own language was precise: "informed by agentic technology, controlled by the client." The agent is inside the account. The human increasingly serves as the approval layer rather than the primary source of analysis. Simultaneously, Rockefeller Capital Management, founded in 1882 and serving some of the most sophisticated institutional and family clients in the world, announced it is building a new AI-enabled wealth management platform in collaboration with Anthropic. This is not a firm experimenting at the margins. It is a firm that has concluded the future of investment intelligence runs through an AI orchestration layer, and has committed its architecture accordingly. Magnetar Capital, the $18 billion alternative investment fund, announced an AI-powered fund in which hundreds of machine agents conduct the research and generate investment ideas, while humans retain final trading decisions. And Mastercard unveiled Agent Pay for Machines, a settlement protocol designed for AI agents to transact with one another autonomously, without a human authorizing each payment.

Taken together, these are not signals of a coming shift. They are evidence of a shift already underway. The question is what it means for each of the three layers on which the exchange's economic model depends.

The Data Problem

The traditional exchange generates value at three distinct points: listing fees, transaction fees, and data licensing. The data layer is the quietest of the three, but for many of the world's leading exchanges it has become the most strategically significant revenue stream.

The pricing model, built on per-seat and per-terminal access, reflects a world in which the consumer of market data is a human analyst or portfolio manager, operating within the bandwidth of human attention and processing speed. Intelligent agents break that premise at the foundation. Where a senior analyst team might meaningfully process data across dozens of securities per day, an AI-driven market participant simultaneously monitors thousands of instruments, cross-references alternative data sources, models scenario probabilities, and generates actionable signals, continuously, around the clock. The consumption is not ten times greater. It is categorically different in kind.

When machines become the primary consumers of exchange data, the per-seat licensing model does not just underperform. It becomes the wrong unit of measurement entirely. Anthropic has already demonstrated what happens when an AI layer sits above legacy data infrastructure. In May 2026, it released ten specialized financial agent templates integrated with FactSet, S&P Capital IQ, and Moody's, positioning Claude as an orchestration layer above existing financial data providers. Bloomberg's terminal has been the defining moat of financial data for three decades, built on a proprietary interface that professionals must learn, pay for, and depend on daily. As AI agents access the same underlying data through APIs and deliver it through continuously-acting interfaces, the terminal risks becoming one interface among many rather than the exclusive point of access. The exchange data business faces an identical structural exposure, and at an earlier stage of the cycle, which means the window to act is still open.

The Infrastructure Question

For two decades, exchange leadership has pointed to the matching engine as the defensible core of the business. Speed in nanoseconds. Regulatory primacy. Depth of liquidity. These advantages are genuine. But they rest on an assumption that has not yet been stress-tested under the conditions now emerging: that routing decisions are made by human portfolio managers, or by algorithms executing fixed human mandates.

This is where the distinction from previous waves of automation becomes critical. Algorithmic trading optimized the execution of human decisions. It answered the question of how to get a trade done efficiently once a human had decided what to do. Machine agents operate at a fundamentally different level. They continuously redefine the mandate itself, simultaneously evaluating strategy, positioning, risk exposure, and cross-asset opportunity in real time, without waiting for a human review cycle. The question is no longer how to execute an order. It is what to do across an entire portfolio, right now, and where to do it. Routing in that context becomes a live optimization problem, not a fixed instruction. The NYSE partnership with NVIDIA, HPE, and Redpanda to rebuild market infrastructure on AI-optimized silicon reflects this understanding directly: that the matching engine of the future must be designed for a market where many of the participants making decisions are not human. The system already processes more than 1.1 trillion messages per day. The challenge is no longer throughput. It is architecture. Jeffrey Sprecher of ICE made the same point with characteristic bluntness at the Bernstein Strategic Decisions Conference in May. Describing blockchain settlement tests underway internally at NYSE, SEC approval being sought for 24/7 tokenized equity trading through a sister ATS, and a private credit data partnership with Apollo already in motion, he said: "I feel like we've been caught up a little bit in the SaaS apocalypse," arguing that the market is penalizing ICE for uncertainty around AI and data monetization at precisely the moment the company is building the infrastructure to capitalize on both. LSEG reached the same conclusion in its own infrastructure review: infrastructure is no longer adjacent to trading strategy. It is trading strategy.

The Governance Opportunity

This is the section that separates a strategic discussion from market commentary, and it is where exchanges have both the most to gain and the least time to act.

Autonomous systems operating in live markets generate regulatory questions that existing frameworks were not designed to answer. When Claude drafts a trade at Interactive Brokers and a client approves it, the fiduciary question is genuinely unresolved. Who is responsible for that decision? The client? The broker who deployed the system? The model that generated the recommendation? Current securities law assigns fiduciary responsibility to the human advisor or institutional manager. It has no established framework for a world in which the recommendation originates from an autonomous system and the human role is reduced to confirmation.

The explainability problem is equally unresolved. When a fund running hundreds of AI research agents takes a significant position based on signals generated across a network of models, the audit trail is not a memo or a recorded conversation. It is a sequence of model outputs, weighted inputs, and probabilistic assessments that no single human can fully reconstruct or defend before a regulator. When Mastercard's Agent Pay for Machines settles a transaction between two autonomous agents, the question of regulatory accountability is genuinely open. Neither agent is a legal person. The firms that deployed them may not have full visibility into the chain of decisions that produced the settlement.

These are not edge cases. They are the normal operating conditions of the market being built right now by the most sophisticated participants in the world. The SEC, the FCA, ESMA, and their counterparts across Asia are actively working through these frameworks. They will produce regulation. The question is whether exchanges are in those conversations shaping the outcome, or subject to the result afterward.

Historically, regulators have turned to exchanges whenever new market structures required governance architecture. It happened with electronic trading. It happened with derivatives clearing after 2008. That role is available again. The exchanges that engage proactively on fiduciary accountability for AI-driven systems, on explainability standards, on disclosure requirements for autonomous execution, and on the regulatory architecture of agent-to-agent transactions, will not just shape the rules their own market operates under. They will establish themselves as the legitimate governance authority of AI-native capital markets. That is a structural position worth far more than any single product or partnership.

The Three Challenges, and the Window

The exact timeline remains uncertain, but the transition is likely to unfold in phases over the next three to five years, and it will not affect all exchanges equally. Those with established data businesses built on terminal-style licensing face the most acute near-term exposure. Those with cloud partnerships, API-first architectures, and active regulatory relationships are better positioned, not because they are safe, but because they have the infrastructure to move faster.

Every exchange faces three challenges in this transition: the economics of its data business, the readiness of its execution infrastructure, and its governance positioning relative to the regulatory architecture being built right now. These are not independent problems. They are the same underlying shift viewed from three angles, and the institution that addresses them as a unified strategic question will be better positioned than the one managing them in silos.

Interactive Brokers has made its bet. Rockefeller Capital Management has made its bet. Magnetar has made its bet. Mastercard has made its bet. The gap between where your most sophisticated participants are heading and the assumptions embedded in your current infrastructure strategy belongs on every exchange board agenda before the next quarter closes.

In every previous disruption, exchanges adapted after the infrastructure changed around them. Agentic markets may be the first transition where waiting for proof means surrendering the ability to shape the outcome. The only question that matters now is whether the exchange is at the table where the terms of AI-native capital markets are being written, or reading about it afterward.