AI in trading means data and agents – but if agents are still a decade away, data means datacentres, and datacentres mean dependence on third parties – a near-complete reversal of the traditional trading floor. The next decade of trading won’t be defined by new exchanges or faster fibre, but by data, compute, and the architecture of intelligence. As AI moves from research into live market infrastructure, trading is becoming inseparable from external systems – cloud providers, datacentres, and energy grids. Even if we are still waiting for the agents, the dependency chain is market now. The question is no longer whether firms should rely on external infrastructure, but how these dependencies are incorporated, governed, audited, and made resilient while trading at speed.

1. Latest from Karpathy – No Autonomous Trading Desks Next Year: Agents Need a Decade for the March of the Nines

Andrej Karpathy’s latest remarks offered a dose of realism for those expecting fully autonomous AI systems tomorrow (https://www.youtube.com/watch?v=lXUZvyajciY). His headline point – this isn’t the year of agents, it’s the decade; don’t get ahead of the timelines.

He described current AI agents as brittle – lacking memory, feedback, and persistence arguing that building the connective tissue of cognition will take years of systems engineering. These “disembodied minds” imitate rather than experience, learning patterns from human culture rather than from the world itself. In time, they may evolve into self-improving systems capable of reflection, self-critique, and synthetic “dreaming” cycles to avoid collapse. But the reality is each extra nine of reliability – from 99 to 99.999% – will demand exponentially greater effort. Turning demos into deployable autonomy is a decade-long grind.

Why this matters for secondary markets:
While there will be no truly autonomous trading desks in 2026 – what matters now is building the architecture, the connective tissue between data, models, and execution (as discussed in  (https://www.mindfulmarkets.ai/something-for-the-weekend-agents-infrastructure-the-intelligence-layer/). AI can already enhance memory, feedback, and persistence across the trade lifecycle through retrieval layers and feedback loops that improve routing and pricing over time. But for now – and for potentially the next decade, the human remains in the loop – we are just continually using AI as the junior trader – or as Karpathy describes it, the intern.

2. AWS Cloud Shock – but Markets Kept Moving

The AWS outage that hit last Monday halted services at retail banks, brokers like Robinhood, data providers including LSEG and BMLL, and even UK government portals such as HMRC and Universal Credit. Yet, as highlighted by Nick Dunbar and the team at GlobalTrading, “the dog that didn’t bark was telling”: despite their deep reliance on AWS, major exchanges and trading venues stayed online ((https://www.globaltrading.net/aws-outage-disrupts-trading-data-and-infrastructure-ecosystem/). The market didn’t halt – this time.

Why it matters for secondary markets: the episode was a stress test for modern market infrastructure – and a reminder that resilience is now a competitive asset. Venue continuity proved the value of dedicated colocation, cross-cloud redundancy, and clearly defined failover paths for market-critical systems. But it also exposed how opaque the cloud has become: a macro-scale black box with enormous systemic importance. Firms need contractual auditability – traceable data lineage, preserved model artefacts, and pre-agreed exit plans – backed by robust third-party incident playbooks that cover client communications, best-execution reporting, and post-event algo recertification. In a market increasingly built on shared infrastructure, resilience is the new alpha.

3. Power Is the New Market Structure: From Belgium’s Grid Triage to the Data-Centre Arms Race in Space

What happened: Belgium’s grid operator Elia has proposed a dedicated energy category for data centres following a ninefold surge in connection requests driven by AI infrastructure demand (https://www.reuters.com/business/energy/belgium-mulls-energy-limits-power-hungry-data-centres-ai-demand-surges-2025-10-22/). Across Europe, power and permitting have become the new bottlenecks in the AI build-out, as capex pipelines expand and grid capacity tightens.

Why this matters for secondary markets: As automation spreads across asset classes, the competitive frontier for electronic liquidity providers (ELPs) is shifting from code speed to compute power. In an AI-driven trading environment, performance is determined not just by how fast packets travel, but by where models live – and whether the datacentre has the power, cooling, and GPU density to sustain real-time inference under load. The ability to refresh prices, recalibrate risk, and publish firm quotes now depends on megawatts as much as microseconds.

Firms with scalable, venue-adjacent compute will reprice faster, maintain tighter spreads, and win more SOR flow; those constrained by grid limits or shared infrastructure risk stale quotes and declining hit-rates. Compute has become part of market structure itself – forcing firms to choose between prioritising intelligence (smarter models) or speed (closer compute).

The bigger story is energy policy is now market-infrastructure policy. Grid allocation and utility regulation may determine which regions can host high-compute trading clusters. And as firms like Nvidia’s Starcloud explore orbital data centres (https://blogs.nvidia.com/blog/starcloud/) the race for AI capacity could soon extend beyond the planet – a literal new frontier for trading infrastructure (Read more on this in AI for Planet tomorrow).

4. Risk Controls for Smarter Algos: From “Checks” to Continuous Assurance

What happened: as AI-driven models gain greater autonomy and reasoning capability, their interconnectedness across assets, venues, and now compound algo strategies introduces new modes of failure that traditional static testing can’t anticipate.

Why this matters for secondary markets: Risk control is shifting from one-off validation to continuous assurance. The old model of static UAT is giving way to live certification, combining scenario-based guardrails, real-time drift monitoring, and pre-trade sanity layers such as price bands, venue caps, and participation clamps across multi-asset algorithmic chains. Regulators are watching closely: under DORA and emerging FCA market resiliency frameworks, firms will need to demonstrate not only that algorithms behave as intended, but that they can withstand unexpected market conditions.

This means moving from testing what should happen to probing what could happen. Expect the rise of challenge datasets, shadow-mode testing in production, and enhanced traceability through standards like FIX protocol’s Unique AlgoCertIDs (FIX Trading Community), which will allow firms to identify, audit, and attest to specific algorithm instances in real time. Continuous testing, not static sign-off, is fast becoming the new benchmark for market integrity.

5. Oversight is going live: sandboxes, smart data and continuous supervision

New announcements from the UK’s AI Growth Lab (https://www.gov.uk/government/news/new-blueprint-for-ai-regulation-could-speed-up-planning-approvals-slash-nhs-waiting-times-and-drive-growth-and-public-trust), FCA’s Supercharged Sandbox/Smart Data Accelerator (https://www.fca.org.uk/firms/innovation/ai-lab#section-supercharged-sandbox), EU “Apply AI” strategy (https://digital-strategy.ec.europa.eu/en/policies/apply-ai) as well as Australia’s new AI policy templates (https://www.industry.gov.au/publications/guidance-for-ai-adoption/ai-policy-guide-and-template) continue to point to test-and-learn supervision – less static rules, more live validation of models, datasets and flows.

Why this matters for secondary markets: if AI models are dependent on good data, then we can expect increased regulatory demand for dataset provenance, how models are built, and how their decisions can be explained and verified. Tools that can turn unstructured inputs, like trading voice data, into usable information are opening new possibilities for automation, but managing and using that data responsibly remains a challenge. Clear standards are needed to ensure data quality and consistency are in place to still enable innovation. Global initiatives such as the Financial Accounting Standards Board (FASB) are proposing  updates to the 2026 Data Quality Committee Rules Taxonomy (DQCRT) (https://www.xbrl.org/news/fasb-proposes-updates-to-2026-data-quality-rules-taxonomy/) with comments due by November 10^th 2026.

The fabric of secondary markets is rapidly evolving into an ecosystem of interlocking dependencies – between human and machine, code and power, cloud and regulation. If true AI autonomy remains years away, keeping human oversight central as AI will increasingly serve as the junior trader on the desk.  Cloud resilience has become market stability itself, while compute power and energy access will increasingly define trading performance. Risk control is shifting to continuous assurance through live testing and traceable algorithm certification, and regulation is moving toward real-time supervision built on data provenance and explainability. Together, these trends signal a shift from autonomous trading systems to accountable intelligence infrastructure – where resilience, compute, and data quality have become the new sources of alpha.

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