How JPMorgan’s New AI Lab Aims to Cut Trading Latency

JPMorgan Chase’s new AI Market Lab is building foundation models and agentic systems from scratch, using order books, tick data and alternative data across multiple asset classes, a level of vertical integration that goes well beyond the AI tooling most banks have deployed to date. JPMorgan’s LLM Suite now supports 1,000 business applications used by 200,000 employees, with Global CIO Lori Beer publicly prioritising data security in the bank’s accelerated rollout. What follows is a practical breakdown of how investment banks are structuring that integration, from data infrastructure through to governance.

Data Infrastructure First

Reducing trading latency with AI starts at the data layer. Banks manage petabytes of historical and live market data, order book snapshots, trade executions, news feeds, economic indicators, and the infrastructure must ingest, process and surface that data for AI models in microseconds. A 2026 study on LLM agents in financial trading found that textual data from news articles and financial reports are the most heavily used inputs, reflecting LLMs’ strengths in natural language processing.

Ingesting Real-time Market Data

Banks deploy ultra-low-latency ingestion pipelines that pull direct feeds from exchanges, often bypassing traditional database layers entirely. Kdb+ databases handle time-series data with nanosecond precision; Kafka or equivalent message queues provide throughput and fault tolerance. A single equity desk can ingest millions of messages per second at peak. Proprietary order flow and alternative data such as satellite imagery remain institutional-tier inputs, largely inaccessible to non-bank participants, according to a September 2026 analysis.

Building a Low-Latency Data Lakehouse

Banks pair that ingestion layer with specialised data lakehouses that combine the flexibility of raw data lakes with the schema enforcement of data warehouses. Microsoft Azure hosts JPMorgan’s private cloud AI environment; other banks run comparable on-premise builds. The lakehouse gives data scientists access to historical tick data, order book depth and macroeconomic indicators for feature engineering with minimal latency, constructing a 5-millisecond moving average, for instance, requires processing gigabytes in near real-time. By 2026, LLMs had made natural language querying of financial databases routine, with analysts running queries directly against market databases to screen for conditions like rising short interest ahead of earnings.

Normalisation and Feature Engineering

Raw market data is noisy. Normalisation standardises formats, removes outliers and handles missing values. Feature engineering then extracts predictive signals: volatility metrics, order book imbalances, spread dynamics, sentiment scores from news feeds. A sentiment agent might process thousands of sources in near real-time, feeding contextualised event summaries rather than raw headlines to the model. Fine-tuned LLMs have made large-scale alternative data analysis commercially viable where it previously required specialist teams.

Model Development

JPMorgan’s AI Market Lab is specifically tasked with transformer-based and time-series foundation models for systematic trading, architectures chosen for their ability to capture temporal dependencies across sequential market data. For broader market analysis, ensemble methods such as gradient boosting and random forests remain common for their predictive stability across asset classes.

Training and Backtesting

Model training runs on high-performance computing clusters with NVIDIA GPUs and requires extensive historical market data. A July 2026 arXiv paper comparing LLMs on technical market analysis found that GPT-4 Turbo achieved the highest annualised return and Sharpe ratio among general-purpose models in simulated backtesting, while FinGPT showed competitive risk-adjusted performance attributed to domain-specific fine-tuning.

Continuous Retraining

Models degrade as market conditions shift, the phenomenon known as concept drift. Banks run automated retraining pipelines on daily, hourly or intra-day cycles depending on strategy frequency. Alongside retraining, real-time drift detection has become a standard component of AI governance in trading, replacing earlier static, policy-based oversight.

Real-time Execution

Moving from model output to live order flow requires low-latency APIs and, in many cases, direct memory access to minimise communication overhead between the AI layer and the Order Management System. The OMS handles regulatory checks and exchange routing; the AI layer handles decision augmentation or, within predefined risk parameters, automated execution. Alpaca, an API-first trading platform, offers commission-free API access specifically for connecting custom AI models to live markets.

Smart Order Routing

AI agents analyse real-time microstructure, liquidity, bid-ask spreads, latency across venues, to select the optimal exchange or dark pool for each order. The routing decision can be made within nanoseconds. Best execution, the regulatory benchmark these routers are measured against, is increasingly a quantitative standard that regulators will assess AI against directly, according to an October 2026 report. Latency at the agent layer is now a first-order architectural concern at major funds, with proprietary multi-agent systems handling execution and risk management in parallel.

Adaptive Execution and Monitoring

Large orders are broken into child orders, with the AI adjusting timing and sizing dynamically based on real-time feedback. A sudden volatility spike triggers automatic strategy modification to contain slippage. This continuous feedback loop compounds over time, improving execution quality and reducing transaction costs.

Governance and Compliance

SR 26-2, issued jointly by the Federal Reserve, OCC and FDIC on April 17, 2026, updates model risk management requirements, documentation, independent validation and monitoring across a model’s lifecycle, addressing data quality, accuracy, stability and input sensitivity. But the guidance explicitly excludes generative and agentic AI from its scope. Banks deploying systems like JPMorgan’s foundation models are, for now, building their own governance standards rather than working against a federal AI-specific framework.

Explainability and Audit Trails

Regulators require transparency on AI-driven decisions affecting client outcomes or market integrity. Banks are deploying SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) to surface how models reach their outputs. JPMorgan’s approach to AI security extends to its LLM Suite, which Beer described on October 22, 2026 as a “highly JPMorgan-secure ecosystem where we’re confident with our data and how our data is being leveraged”, language that reflects both the audit requirement and the bank’s data residency posture. Every model input, output and decision point is logged for regulatory review.

The Regulatory Stack

Beyond SR 26-2, banks must comply with the EU’s Digital Operational Resilience Act, which has applied since January 2025, alongside a proliferating set of US state AI laws, 341 were proposed across 45 states in 2026 alone. Compliance teams are running AI-powered regulatory change management tools that monitor thousands of sources and flag relevant updates. JPMorgan’s early block of ChatGPT in 2022 over data loss risks was an early signal of the bank’s risk-first posture on external AI tooling.

The gap is notable: the newest federal guidance on model risk doesn’t cover the newest kind of model banks are deploying, leaving continuous validation of systems like JPMorgan’s foundation models to internal governance rather than a settled regulatory standard.