China’s AI race took another leap this week: Beijing-based Moonshot AI unveiled Kimi K2 Thinking, an open-source reasoning model outperforming GPT-5 and Claude Sonnet 4.5 across multiple benchmarks – including a 44.9% score on Humanity’s Last Exam – and at a training cost of just USD 4.6 million (https://venturebeat.com/ai/moonshots-kimi-k2-thinking-emerges-as-leading-open-source-ai-outperforming). When state-of-the-art intelligence becomes open-source and affordable, the same capability that powers research labs can rapidly enter trading and the wider market ecosystem. This week’s developments from the rise of retail AI agents to the classification of ‘trading AI’ as high-risk illustrate how AI diffusion continues to disrupt the traditional structure of markets – and how the industry will need to adapt and respond as a result.

1. Retail and Wealth Go Direct: AI Agents at Scale

At the research layer, Menos AI, a San Jose-based startup, is using LLMs to assess the accuracy and reasoning of analyst calls through a “Voice Scoring” system now being piloted by global-macro hedge funds (www.menosai.com). Arta has expanded its Arta AI wealth-management platform globally, adding Bank of Singapore, Hong Leong Bank and Ethivo Asset Management – each adviser now has a generative-AI ‘Sidekick’ connecting market data, CIO research and multi-asset portfolios to automate analytics, risk and client communication (https://www.prnewswire.com/news/arta-finance/). Similarly, BigData.com’s Research Agent uses AI to map trade-related risks across supply chains, revenues and IP exposure in minutes, cutting days of thematic analysis into a single workflow – read more here https://bigdata.com/enterprise.

Why this matters for trading: while these tools focus on wealth and advisory channels, they illustrate how retail and adviser-led flows are becoming faster, more data-driven and reactive. This is likely to further tighten feedback loops, fragment liquidity and raise intraday volatility. As AI is embedded further in workflows – from evaluation to execution, market-makers and algos will need to adjust pricing and risk models to reflect this AI-led trading behaviour – even if some institutions still hold the view that AI should not be directly involved in trading.

2. Exchanges Going Direct

Eurex has confirmed that its new Sponsored Access model, effective November 2025, will allow buy-side firms and retail intermediaries to connect directly to its order books through clearing sponsors (https://www.eurex.com/ex-en/find/news-center/news/Eurex-lowers-market-access-barriers-with-new-Sponsored-Access-model-4756214) significantly reducing latency and trading costs while expanding near-institutional access to exchange liquidity. In parallel, LSEG’s strategic partnership with Nasdaq will integrate Nasdaq eVestment™ private markets datasets – including Market Lens insights, hedge fund data, and LP intelligence – into LSEG’s Workspace and Datafeeds, broadening transparency and decision-making capabilities across private investments (https://www.lseg.com/en/media-centre/press-releases/2025/lseg-announces-strategic-partnership-with-nasdaq).

Why this matters for trading: Exchanges are continuing to shift traditional matching venues into AI-enabled execution platforms, offering participants deeper liquidity insights but also increasing structural complexity and fragmentation. The ability to build secure, data-grounded workflows using LSEG data across tools like Copilot and Teams demonstrates how trading and workflow design are converging. As smart agents begin managing collateral, routing and analytics autonomously, firms will need to invest in robust AI infrastructure and data governance to stay competitive in an increasingly fragmented, algorithmically adaptive market landscape.

3. EU AI Act: Trading AI is ‘High-Risk’

There is ongoing debate that the EU AI Act could classify AI systems used in trading – including execution algorithms and decision-support models – as “high-risk” (https://www.goodwinlaw.com/en/insights/publications/2024/08/alerts-practices-pif-key-points-for-financial-services-businesses).  This stems from the Act’s decision to delegate supervision of financial-sector AI to existing regulators (the EBA, ESMA, and EIOPA) and from Recital 158, which calls for consistent and equal treatment across the financial sector. This would subject algo trading and market-making systems to requirements for explainability, validation, and governance – introducing new obligations around transparency, testing, and auditability. In parallel, Singapore’s Cyber Security Agency (CSA) has released a draft Addendum on Securing Agentic AI, introducing stronger standards for AI autonomy, human oversight, and data integrity. Aimed at AI systems capable of acting independently, the Addendum highlights the importance for trading and investment firms to strengthen controls ensuring that autonomous decisions remain safe, explainable, and aligned with organisational intent. The public consultation on the draft is open until 31 December 2025 (https://www.csa.gov.sg/news-events/press-releases/csa-releases-an-addendum-to-support-system-owners-in-securing-agentic-ai-system/).

What this means for trading: For trading operations, these developments would signal a further shift in regulatory and operational expectations given the potential for systemic risk, necessitating greater compliance controls, explainability mechanisms, and ongoing model validation as well as continuous monitoring and reporting frameworks to maintain compliance in an increasingly AI-regulated market landscape.

4. Rethinking Model Risk for Generative AI

In a similar vein, new research, Move Fast Without Breaking the Bank (Wicker, Szpruch & Mørk, 2025 –  ), outlines how financial institutions can adapt model risk frameworks like SR 11-7 and SS 1/23 for generative AI. The paper proposes a three-pillar Model Risk Management framework focused on governance and tiering, design standards, and testing and monitoring. This includes mapping AI use, managing dependencies like APIs, ensuring transparency and robust data governance, quantifying uncertainty, and continuously testing for failures and emergent behaviours.

What this means for trading: Effective risk management will require vendor transparency, continuous monitoring, and human fallback mechanisms, as traditional model risk frameworks are no longer sufficient – cue more on Trading Venue Perimeter regulation?  The ability to balance innovation speed with robust oversight will determine how safely and effectively financial institutions can deploy large-scale GenAI systems in production.

5. Back to the SLM
With the growing challenges of LLMs, Smaller language models (3–7B parameters) are being flagged again as the solution – but this time paired with knowledge graphs to outperform larger models at lower cost. Graphs externalise structured facts, reducing errors and latency while keeping data private. This modular architecture makes AI suitable for real-time, explainable trading and risk analysis, supporting secure deployment on-premises or at the edge. Read more here – https://www.linkedin.com/in/anthony-alcaraz-b80763155/.

As AI narrows information asymmetries, lowers analysis costs, and democratizes access through open-source models like Kimi K2, intelligence itself is becoming a commodity rather than a competitive edge. The real advantage will come from how effectively firms integrate and apply AI within their operations and decision-making. The next phase of market evolution will hinge on balancing open innovation with regulatory compliance and liquidity resilience, ensuring that AI serves to enhance market integrity and stability rather than undermine them.

Thank you for reading. As always, I’d love to hear your thoughts – what you agree with, where you disagree, and what you’d like to see more of next time.

Best wishes,

Rebecca