BMLL and Simudyne build Large Market Models for AI trading simulation

BMLL, the Cambridge-founded historical market data provider, and Simudyne, an agent-based simulation platform, have announced a partnership that applies generative AI techniques borrowed from large language model development to the problem of financial market simulation. The collaboration combines BMLL's granular Level 2 and Level 3 order book data with Simudyne's Pulse simulator to pre-train what the two companies are calling "Large Market Models": foundation models of intraday order flow capable of generating synthetic tick data and reacting to a user's live trading activity in real time.

The distinction the partnership draws is between static historical replay and reactive simulation. Conventional back-testing runs a trading algorithm against a fixed tape; the market does not respond. The new platform, the companies say, simulates a market that adapts dynamically as a user's executed orders propagate through it. That distinction matters most to quantitative traders optimising execution algorithms, risk teams stress-testing margin and default-fund models against simulated volatility regimes, and exchanges prototyping changes to matching engine logic or circuit-breaker mechanics before deploying to production.

Generative AI meets market microstructure

The architectural parallel to language AI is deliberate. Simudyne is applying the same pre-training approach that underpins models such as ChatGPT to the sequential, high-frequency structure of order-book data rather than text. The quality of a large language model is constrained by its training corpus; the same principle applies here, which is why BMLL's tick-level historical data is the critical input. BMLL's chief executive Paul Humphrey put it plainly: "High-quality training data is the single most critical ingredient for sophisticated, AI-driven financial modelling."

The partnership is structured through BMLL's Activate Data Credits Programme, an initiative that offers selected partners subsidised access to its Data Lab and Data Feed during the build phase, deferring upfront licence costs until a product reaches commercial adoption. Simudyne participates under that programme, training and hosting its generative foundation models and multi-agent simulation software directly within the BMLL Data Lab environment. The arrangement is a template increasingly visible across the data-infrastructure layer of financial markets: a data custodian offers preferential access in exchange for validation of its data as a training foundation, then monetises through downstream client usage.

Cross-sector read-across: from capital markets to AI infrastructure

The macro significance of this deal sits at the intersection of two converging forces. First, the industrialisation of AI training pipelines is spreading beyond technology companies into domain-specific verticals. Capital markets, with their vast stores of high-resolution, timestamped behavioural data, represent one of the richest substrates available for training generative models. What BMLL and Simudyne are assembling is, in effect, a domain-specific AI training stack with proprietary data at its base, not unlike what pharmaceutical companies are doing with genomics datasets or what defence contractors are doing with sensor telemetry.

Second, the emergence of Large Market Models as a concept signals that the financial industry's AI ambitions are moving from narrow prediction tasks (price forecasting, anomaly detection) toward synthetic environment generation. That shift has significant implications for regulators and exchanges alike. If market participants can train execution strategies in simulators that closely replicate real market microstructure, the boundary between research environment and live market grows thinner. Regulators in the UK and EU, already grappling with AI governance frameworks for financial services, will need to decide how synthetic training environments interact with rules on market manipulation and algorithmic trading conduct.

For capital allocators watching the AI infrastructure trade, the BMLL-Simudyne model points toward a category of companies that are neither AI labs nor traditional data vendors but foundational AI training platforms for regulated industries. Nordic Capital, which acquired BMLL in October 2025, is the first institutional investor to test that thesis at scale in the market data space. Whether other private equity and sovereign capital will follow into similar data-plus-simulation plays across financial services, defence, and life sciences will be a meaningful signal of how the broader convergence infrastructure thesis matures through 2027.