Jump Trading: Quantitative Trading at GPU Scale and Where WorkingAgents Fits

What Jump Trading Does

Jump Trading is a privately held quantitative trading firm headquartered in Chicago with offices in New York, London, Shanghai, Singapore, Sydney, and Amsterdam. Founded in 1999, Jump operates across global markets – equities, fixed income, commodities, foreign exchange, and digital assets – using proprietary research, algorithms, and infrastructure to provide liquidity and capture trading opportunities at scale.

Jump does not manage external capital. It trades its own money, which means every system, every model, and every infrastructure decision is driven by a single question: does this improve trading performance?

Products and Services

Proprietary Trading Jump is a market maker and liquidity provider across global exchanges. It does not have “customers” in the traditional sense – its counterparties are the exchanges and market participants it trades against. Revenue comes from capturing spreads, exploiting pricing inefficiencies, and providing liquidity where markets need it. Jump trades equities, futures, options, fixed income, commodities, and foreign exchange across dozens of exchanges worldwide.

Jump Liquidity A dedicated institutional unit providing liquidity solutions to exchanges, broker-dealers, and institutional counterparties. Jump Liquidity focuses on consistent, programmatic market making at scale – the kind of always-on liquidity that exchanges depend on for tight spreads and deep order books.

Research and Engineering This is the core. Jump’s research teams build quantitative models, machine learning systems, and statistical frameworks that drive trading decisions. The engineering teams build the infrastructure that runs those models – from low-latency networking (microwave and millimeter-wave towers for inter-exchange communication) to custom FPGA and GPU computing clusters. Jump has been building on NVIDIA GPUs and CUDA for over 15 years, starting with the Fermi architecture and scaling through every generation since.

Jump AI (Machine Learning Research) A central R&D group within Jump focused on deep learning, foundation models, and LLM-based agents for quantitative research. The team builds custom foundation models trained on proprietary financial data, develops LLM agents for research automation, and pushes the boundaries of AI applied to market microstructure, signal discovery, and execution optimization. Jump AI partners with academic institutions including UCL for PhD research programs.

Jump Crypto Jump’s digital assets division. Jump Crypto was one of the most active participants in DeFi and blockchain infrastructure, building core protocol technology including:

Firedancer – a high-performance Solana validator client written from scratch in C, designed to dramatically increase Solana’s throughput and resilience. Firedancer is now maintained by the Solana ecosystem as critical infrastructure.
Wormhole – a cross-chain messaging protocol enabling token and data transfers between blockchains. Jump built the initial implementation. Wormhole has since been spun into an independent entity.
Pyth Network – a decentralized oracle providing real-time market data to DeFi protocols. Jump is a founding publisher, feeding institutional-grade price data on-chain.

Jump Crypto also engaged in market making across centralized and decentralized exchanges, providing liquidity for digital asset markets.

Jump Capital Jump’s venture arm with over 173 portfolio companies spanning fintech, AI/ML, blockchain, data infrastructure, and enterprise software. Notable investments include BitGo, Mina Protocol, TRM Labs, and various AI startups. Jump Capital gives the firm exposure to emerging technologies and partnerships across the ecosystem.

Prediction Markets Jump has expanded into prediction market making, providing liquidity on platforms like Kalshi. This applies Jump’s core market-making expertise to a new asset class – event contracts on elections, economic indicators, weather, and other real-world outcomes.

The Infrastructure

Jump’s infrastructure is purpose-built for speed and scale:

GPU computing: 15+ years on NVIDIA GPUs, from Fermi to Hopper. Large-scale CUDA deployments for quantitative research, deep learning training, and real-time inference.
Low-latency networking: Microwave and millimeter-wave communication towers connecting major exchange data centers. Custom network hardware for nanosecond-level optimization.
FPGA acceleration: Custom field-programmable gate arrays for trading logic that needs to execute faster than software can manage.
Global data centers: Colocated servers at major exchanges worldwide, with custom cooling, power, and networking infrastructure.

The Synergy Map

Jump Trading operates at the intersection of quantitative finance, GPU-scale AI research, and global trading infrastructure. The governance challenges are unique – and growing as AI agents move from research tools to production trading systems.

1. Governance for AI-Driven Trading Research

Jump’s research teams build and deploy models that directly influence trading decisions worth billions. As LLM agents and foundation models move from research assistance to automated research workflows, governance becomes critical.

WorkingAgents provides:

Per-agent permissions – a research agent that can query market data and run backtests but cannot submit orders. A signal discovery agent that can read proprietary datasets but cannot export data externally. An execution research agent that can analyze fills but cannot modify live trading parameters.
Audit trails – every model query, every data access, every research workflow logged with researcher identity, agent identity, and full context. When a regulator asks “how did this trading strategy get developed?”, the answer is in the audit trail.
Three-checkpoint guardrails – pre-execution validation on research queries (preventing accidental access to restricted datasets), real-time monitoring of agent behavior (catching runaway research processes), and post-execution review of outputs (ensuring proprietary signals don’t leak).

2. Agent Isolation Across Trading Desks

Jump trades across multiple asset classes, geographies, and strategies. Each trading desk operates with its own models, data, and risk parameters. Cross-contamination between desks – even accidental – creates regulatory risk and competitive disadvantage.

WorkingAgents’ Virtual MCP Servers enforce hard boundaries:

The equities desk’s AI agents see equities tools, data, and models
The crypto desk’s agents operate within digital asset systems
The fixed income desk’s agents access bond market infrastructure
Research agents can read across desks (with approval) but cannot write to production systems

This is not access control at the application level – it is infrastructure-level isolation. Each desk gets its own Virtual MCP Server with its own permission set. No agent can cross a boundary that hasn’t been explicitly configured.

3. Digital Asset Operations Governance

Jump Crypto’s infrastructure – Firedancer, Wormhole, Pyth – operates at the intersection of traditional finance and decentralized systems. AI agents interacting with on-chain systems need governance that spans both worlds.

WorkingAgents provides:

Transaction-level permissions – agents that can read on-chain data via Pyth but cannot initiate transfers. Agents that can prepare cross-chain messages via Wormhole but require human approval for execution.
Circuit breakers – automatic halts when agent behavior in DeFi deviates from expected patterns. An agent providing liquidity on a DEX that suddenly attempts to drain a pool gets stopped.
Cross-system audit trails – a single trail connecting on-chain actions to the agent that initiated them, the researcher who authorized them, and the risk parameters in effect.

4. GPU Infrastructure and Model Governance

Jump has been running NVIDIA GPU clusters for 15 years. As these clusters increasingly run AI models that influence or execute trades, governance of the model lifecycle becomes essential.

WorkingAgents’ MCP Gateway can govern:

Model access – which researchers can deploy which models to which GPU clusters
Inference governance – every inference call logged with input, output, model version, and latency
Resource boundaries – agents cannot consume GPU resources beyond their allocation, preventing one research project from starving others

5. Regulatory Compliance at Scale

Jump operates under SEC, CFTC, FCA, MAS, ASIC, and dozens of other regulatory frameworks. Market making activities are subject to market manipulation rules, best execution requirements, and increasingly, AI-specific oversight. The EU AI Act classifies AI-assisted financial trading as high-risk.

WorkingAgents is designed for SOC 2, HIPAA, GDPR, and FedRAMP compliance:

Immutable audit logs satisfying regulatory examination requirements
PII detection preventing agent-mediated data leaks across jurisdictions
Injection detection preventing prompt manipulation that could influence trading decisions
Encryption at rest and in transit for all agent communications

Jump’s scale – trading across every major global exchange – means compliance is not one framework but dozens, simultaneously. A governance platform that enforces consistent policies regardless of jurisdiction is not a feature request – it is operational infrastructure.

6. Jump Capital Portfolio Governance

Jump Capital’s 173+ portfolio companies increasingly deploy AI agents. Each portfolio company has its own governance needs, but Jump Capital benefits from a consistent governance framework across its investments.

WorkingAgents’ one-instance-per-customer model means each portfolio company gets its own governance instance. Jump Capital maintains oversight through a federated view – aggregate metrics on agent usage, compliance posture, and risk across the portfolio, without accessing any individual company’s data.

Value Proposition

For Jump Trading

WorkingAgents gives Jump a standardized governance layer for its expanding AI agent ecosystem – from research LLM agents to trading desk automation to crypto infrastructure operations. Instead of building custom governance per use case (research vs. trading vs. crypto vs. venture), Jump gets a platform that enforces permissions, guardrails, and audit trails consistently across every agent, every desk, and every jurisdiction.

The specific unlock: production AI agents in trading workflows. The gap between “research prototype” and “production trading system” is governance. Jump’s researchers build powerful AI systems. WorkingAgents makes them auditable, permissioned, and compliant enough to run in production.

For WorkingAgents

Jump Trading is a validation deployment in the most demanding computing environment in finance. If WorkingAgents can govern AI agents at a firm that has been running GPU infrastructure for 15 years, trades billions daily across global exchanges, and builds protocol-level blockchain infrastructure, it validates the platform for every other quantitative trading firm, hedge fund, and proprietary trading operation.

Jump’s NVIDIA relationship – 15 years of GPU computing, likely the longest in finance – adds credibility in the NVIDIA ecosystem that no other financial partnership could provide.

Together

Jump’s AI research builds the trading models. WorkingAgents governs the agents that deploy them. Jump’s GPU infrastructure provides the compute. WorkingAgents’ MCP Gateway provides the control plane. The result is quantitative trading infrastructure where AI agents are powerful, auditable, and constrained – exactly what regulators, risk managers, and compliance teams need to see before AI moves from research tools to production trading systems.

GTC Approach

Jump Trading has been building on NVIDIA GPUs since the Fermi generation – 15 years of deep CUDA expertise in finance. They are one of NVIDIA’s most significant financial services customers. Their presence at GTC 2026 (March 16-19, San Jose) is likely, given their GPU computing heritage and expanding AI research operations.

The Conversation Starter

“You have been running GPU clusters for 15 years – longer than almost anyone in finance. Now your AI research team is building LLM agents and foundation models on that infrastructure. When those agents move from research notebooks to production trading workflows, how are you governing what they can access and what they can do?”

This targets the exact transition Jump is navigating: from GPU-accelerated quantitative research to AI agents that participate in trading decisions. The governance gap between research and production is where WorkingAgents lives.

Key Talking Points

“We govern AI agents at the infrastructure level, not the application level.” Jump’s team thinks in infrastructure. They built microwave towers, custom FPGAs, and GPU clusters. WorkingAgents speaks the same language – governance as infrastructure, not as a compliance checkbox.
“Per-agent identity with capability-based permissions. Your research agent gets a keycard for data access. Your execution agent gets a different keycard for order management. No agent crosses a boundary you didn’t configure.” Maps to how trading desks already think about access – but extends it to AI agents that make autonomous decisions.
“Every agent action logged with full context – who triggered it, what it accessed, what it produced, what guardrails fired.” For a firm under SEC, CFTC, and FCA oversight, audit trails are not optional. This is the difference between “we use AI” and “we can prove to regulators exactly how our AI operates.”
“Your crypto infrastructure – Firedancer, Wormhole, Pyth – creates agent workflows that span traditional and decentralized systems. Our governance spans both.” Shows understanding of Jump’s unique position straddling TradFi and DeFi.
“Jump Capital’s portfolio companies will need governance too. Our model is one instance per customer.” Plants the distribution seed – 173 portfolio companies adopting AI agents is 173 potential WorkingAgents deployments.

What to Ask Them

How is the AI research team handling the transition from research prototypes to production trading agents?
What governance framework exists for agents that can influence or execute trades?
Are they evaluating MCP or other standardized protocols for agent-to-system communication?
How do they enforce data boundaries between trading desks when AI agents can query across systems?
What does Jump Capital’s AI governance recommendation look like for portfolio companies?

What to Offer

A proof of concept on one Jump AI research workflow: take a research agent that queries market data, runs analysis, and produces trading signals. Run it through WorkingAgents’ governance layer. Demonstrate per-agent permissions (the agent can read data but not submit orders), audit trails (every query logged with researcher identity), and guardrails (proprietary signals flagged before they leave the research environment). Show that governance adds milliseconds, not minutes – critical for a firm where latency is measured in microseconds.

Sources: