Always on Markets, Tokenisation, Deterministic Models & the Race for Sovereign Compute: AI in Trading – 14th September 2025
From secondary markets always-on, AI-driven era: OpenAI’s MCP plug-and-play promises; Nasdaq, Robinhood and BlackRock renewed push for tokenised shares and 24/7; Mira Murati’s deterministic AI improves reproducibility (just not yet accuracy); and the UK’s $2.5 bn OpenAI–Nvidia data-centre, FCA AI framework and ASML’s Mistral stake highlight the global race for sovereign compute – making AI plus secure data governance the next trading edge – here are the five things to watch this week for AI in Trading.
1. MCP and the New Wave of Agents
OpenAI has rolled out full Model Context Protocol (MCP) support in ChatGPT’s developer mode, promising a universal “plug-and-play” layer for agents to connect to data sources like Gmail, Notion or Slack without bespoke integrations (https://platform.openai.com/docs/guides/developer-mode) . As Stuart Winter-Tear notes: “It’s powerful. It’s also messy.” (https://www.linkedin.com/posts/stuart-winter-tear_mcp-just-landed-in-chatgpt-be-careful-activity-7372889066100498432-a4JX?utm_source=share&utm_medium=member_desktop&rcm=ACoAAARqE5IBeeTUq6qzU4Eqccq_UO6-OMeR6Ao) Researchers have already demonstrated prompt-injection attacks, malicious servers and data-exfiltration exploits – one demo used a simple calendar invite to hijack ChatGPT and leak emails. Text prompts are no substitute for enforceable policy and governance, and stability gaps remain.
Why this matters for Trading:
The potential of MCP servers removes one of the biggest operational barriers to AI-driven trading: connecting models to diverse market-data and execution systems. With a single protocol you can hook into market feeds, OMS and EMSs, risk engines and compliance archives without building dozens of custom APIs. That means lower latency, faster onboarding of new data vendors and easier permissioning and monitoring in a regulated environment. But given the perceived risks – this is still WIP. Expect to hear much more on this going forward though – one to watch.
2. All-In Summit 2025: Tokenised Markets Take Shape & the new role of Exchanges in Private Markets
At the All-In Summit (7–9 Sept 2025), Nasdaq CEO Adena Friedman outlined plans to make Nasdaq a full “market-architecture” provider – covering tokenisation of equities, 24/7 trading, and streamlined post-trade processing to reduce cross-border frictions (shortform summary). She described SEC discussions to align traditional markets with blockchain infrastructure and highlighted Nasdaq’s expanding private-markets platform, similar to PISCES – although the UK version has design choices tailored for the regulatory environment: intermittent trading windows, stricter eligibility, controlled price discovery (https://www.fca.org.uk/markets/pisces-private-intermittent-securities-capital-exchange-system). Separately, Robinhood CEO Vlad Tenev announced a pilot to let retail investors trade tokenised stakes in private companies such as SpaceX or OpenAI – digital coins each backed by a slice of a private stake.
Why this matters for Trading:
If more and more private stakes can be freely tokenised and traded, why would a company need to go public? Secondary-market liquidity may need to adapt to increasingly bridge public and private markets, demanding new technology and connectivity between these pools of capital and creating fresh opportunities for traders – and more demand for technology.
3. BlackRock’s Plan to Tokenise All ETFs (https://www.bloomberg.com/news/articles/2025-09-11/blackrock-seeks-to-tokenize-etfs-after-bitcoin-fund-breakthrough)
- BlackRock plans to tokenise all ETF shares as blockchain tokens trading 24/7, across Tokyo, London or New York.
- No minimum investment or market-hours constraint—buy $50 of the S&P 500 at 2 a.m. Sunday if you want to.
- ETF tokens can be used as DeFi collateral—backing loans, earning yield or participating in smart contracts.
Why this matters for Trading:
What happens to traditional brokers when ETFs trade around the clock on-chain? We are likely to see a massive upgrade in pre-trade analytics and execution, and trading desks both buy and sell-side will need to lean even more on AI to decide when, where and how to trade in a market that never closes.
4. Addressing AI Randomness: Reaching Replicability – just not yet Accuracy
Mira Murati’s Thinking Machines Lab has traced inconsistent LLM outputs to GPU-kernel randomness, not “creativity.” Their paper Defeating Nondeterminism in LLM Inference introduces batch-invariant kernels for core operations such as matmul and attention – ensuring same input, same output every time.
Why this matters for Trading:
Trading models will now be able to ensure:
- Stable back-tests & model comparisons—true signal separated from noise.
- Audit-ready compliance—regulators can verify model decisions don’t drift with server load.
- Operational efficiency—deterministic outputs enable caching and lower compute cost.
But, as Stephen Klein of Curiouser.AI rightly warns, replicability is not accuracy: a wrong forecast repeated perfectly is still wrong. Determinism provides the plumbing for trustworthy testing, just not a guarantee of correct predictions.
5. Power Shifts in AI Infrastructure & Regulation: Silicon Valley Faces a Global Fight-Back
The UK is expected to unveil a $2.5 billion OpenAI–Nvidia data-centre investment during President Trump’s forthcoming state visit (FT). The flagship site at Nscale’s Loughton campus will launch with 50 MW of power (expandable to 90 MW) and 10,000 Nvidia Blackwell GPUs by end-2026, with design capacity for roughly 45,000 GPUs, mirroring the scale of OpenAI’s earlier “Stargate Norway” project.
At the same time the FCA has published its AI approach and launched AI Live Testing (FS25/5), focusing on governance, transparency and accountability and giving firms a supervised environment to trial AI models. In Europe, ASML has taken a €1.3 billion stake in Mistral AI to deepen the region’s AI hardware–model integration (Reuters).
Why this matters for Trading:
Greater regulatory clarity and large-scale sovereign compute investment reduce both infrastructure and compliance risk for AI-driven trading and lay the groundwork for more resilient, locally governed secondary markets. With Europe and the UK rapidly expanding their AI capacity, genuine global competition for capital and liquidity is emerging – but the trading technology stack reboot is still WIP.
Perplexity’s plan to integrate WhatsApp into its financial-services offering has hit a serious headwind after WhatsApp’s former security chief, Attaullah Baig, filed a lawsuit against Meta alleging the company let around 1,500 engineers access sensitive user data with no audit trail, breaching a 2020 FTC privacy settlement. Baig says he raised the issue with the SEC and OSHA and was then fired in retaliation. The case raises serious questions about WhatsApp’s data security and Meta’s compliance with U.S. privacy agreements – particularly relevant when sensitive trading data may be exchanged over messaging platforms.
London Stock Exchange Group (LSEG) has embedded its REDI execution management system (EMS) into LSEG Workspace to unify front-office workflows – execution, content and analytics – in a bid to promote “smarter trading.” REDI EMS complements the existing LSEG TORA order and execution management system (OEMS) and portfolio management system (PMS) spanning equities, fixed income, FX, derivatives and digital assets.
The future of trading will hinge not just on AI but on secure, portable and fully connected data. As regulators tighten scrutiny of messaging platforms and as exchanges such as LSEG integrate more execution and analytics tools, firms must ensure front-to-back data governance – including messaging – not just AI strategies are robust.
As always, thank you for reading, and let me know what you found most useful, what you disagreed with, and what you would like to see more of next time.
Best wishes
Rebecca
