The agentic AI inflection - March Trust Update

Key events

• Markets reacted positively to the ceasefire between the US and Iran and hope of a lasting deal after renewed negotiations

• The arrival of true agentic AI marks a critical inflection point: the shift from AI assistance to execution, doing ‘real’ work

• Enterprise adoption of agentic AI has accelerated faster than expected and AI revenues are growing at unprecedented rates

Market review

Global equity markets were weak in March as the escalating conflict in the Middle East weighed heavily on risk assets. Global markets¹declined 5.5%, the US²3.2% and Europe³returned -8.0% as energy-related concerns negatively impacted sentiment (all returns in sterling terms).

The US and Israeli military campaign against Iran, which began on 28 February, intensified throughout the month. Iran retaliated with sustained missile and drone attacks against Israel, US military assets and Gulf state infrastructure. Iran also effectively closed the Strait of Hormuz – a critical shipping channel through which around 20% of the world's seaborne oil trade and a comparable share of global liquefied natural gas (LNG) supply passes – with flows through the passage down 95% from normal levels. This constituted the largest energy supply disruption since the 1970s.

The Brent crude oil price posted its largest quarterly jump since 1990, rising 94%, prompting an emergency release of 400 million barrels of oil by the International Energy Agency, the global body that coordinates energy policy among major economies. This price spike and the potential for shortages led to concerns over global supplies of LNG and other materials crucial for advanced manufacturing, including semiconductors. While clearly a potential headwind, most companies hold four to six months' worth of inventory of these critical inputs and there has been minimal disruption in the near term.

The conflict injected significant uncertainty around inflation, growth and the path of monetary policy. Most risk assets struggled as geopolitical turmoil overshadowed otherwise constructive fundamental trends. The prospect of higher inflation and potential rate hikes instead of cuts saw global sovereign bonds fall and US 10-year Treasury yields spike to 4.5% at their peak.

The rotation away from mega-cap technology into cyclical and value-oriented sectors that characterised early 2026 paused somewhat as the Middle East crisis created indiscriminate selling pressure across markets. After a historic run, gold ETF outflows totalled $12.2bn, the single largest monthly outflow in five years, and the VIX curve inverted and went into backwardation, meaning short-term protection became more expensive than longer-term protection, which typically happens when the market is undergoing extreme turbulence.

February's headline and Core Consumer Price Index – the official measure of inflation – rose 0.3% and 0.2% month on month, largely in line with consensus. Focus has quickly shifted to March's data where all eyes will be on the impact of higher energy prices. While acknowledging that risk, the health of the underlying US economy remains strong, supported by robust corporate earnings growth.

The final day of the month saw a strong rally as the S&P 500 rose 2.7%, posting its best day since May 2025 on comments by US and Iranian officials that both sides were willing to negotiate and bring an end to the conflict. Thankfully, since then, further progress has been made, with President Trump backing down on his threat to wipe out key civilian infrastructure in Iran. For now, a fragile ceasefire remains in place while both sides continue to negotiate terms for a lasting deal. Markets reacted positively to this news, along with US insistence that a deal must include a full reopening of the Strait of Hormuz, but the weeks ahead are likely to remain volatile while we wait to see if a lasting agreement can be reached.

Technology review

At NVIDIA's GTC conference – its annual showcase for AI and computing technology – Chief Executive Officer Jensen Huang confirmed that the Rubin Graphics Processing Unit (GPU), a specialist chip designed to process large amounts of data simultaneously, essential for AI, remains on track for a second half ramp this year and unveiled the next generation Rubin Ultra GPU architecture. NVIDIA also disclosed visibility to more than $1trn in cumulative Blackwell and Rubin revenue from the start of 2025 through to the end of 2027. In addition, the company launched its new compute racks built on Groq’s technology, following a licensing agreement signed in late 2025.

Memory producer Micron Technology (Micron) delivered outstanding results and guidance, with revenue and earnings per share materially ahead of expectations, driven by exceptionally strong pricing in DRAM and NAND – two types of computer memory used extensively in AI systems.

Several off-quarter results released during the month reinforced the strength of the AI infrastructure theme, with demand for computing power, memory and networking continuing to outstrip supply.

In semiconductors, Broadcom delivered strong results, with revenue up 30% year on year. Guidance was materially ahead of expectations, driven by robust AI chip demand of more than 140% year on year, and management projects AI semiconductor revenue – covering custom AI chips and networking – will exceed $100bn in the financial year 2027, representing 10 gigawatts (GW) of computing power. Networking continues to increase in the revenue mix, supporting structurally higher margins. Broadcom also confirmed OpenAI as its sixth custom AI chip customer.

Memory producer Micron Technology (Micron) delivered outstanding results and guidance, with revenue and earnings per share materially ahead of expectations, driven by exceptionally strong pricing in DRAM and NAND – two types of computer memory used extensively in AI systems. AI-driven demand continues to outpace supply, with limited near-term capacity additions supporting further pricing and margin expansion. Despite increased capital expenditure (capex) in 2026, supply growth remains constrained. The company also announced its first five-year strategic customer agreement, signalling a shift toward more durable, secured memory demand. That said, both Micron and Samsung Electronics were impacted by concerns that memory pricing momentum may be moderating. Memory demand for AI is likely to prove durable, but after a strong run and against a more challenging macro backdrop, a period of consolidation seems likely.

Semiconductor materials supplier Elite Material delivered results that were broadly in line with expectations. Demand is accelerating, however, driven by AI-related orders from Amazon, Google and NVIDIA, as well as low earth orbit satellite demand from SpaceX. The company announced a 10% price increase for high-end copper-clad laminate – a specialist material used in the circuit boards that form the backbone of AI servers – and a significant capacity expansion of 60% over 18 months, supporting stronger growth and margin expansion ahead.

Taiwanese thermal solutions specialist Asia Vital Components delivered solid results and guidance, underpinned by robust AI data centre demand. The need to manage the intense heat generated by next-generation AI hardware is driving broader adoption of higher-value liquid cooling solutions across NVIDIA and an expanding range of custom AI chip suppliers, supporting a richer product mix and meaningful margin expansion. Delta Electronics, also exposed to data centre infrastructure, had a strong month supported by GTC and growing interest in high-voltage direct current data centre architectures – a more efficient way of delivering power to AI computing facilities.

In contrast, data centre fibre optic suppliers Coherent and Corning both pulled back after strong performance in the run up to the Optical Fiber Communication Conference. They were also impacted by concerns about macroeconomic headwinds caused by the conflict in Iran.

In software, MongoDB, a developer-focused cloud database platform, reported solid results, but next-quarter guidance came in slightly below expectations and full-year guidance implies a second-half slowdown. Management highlighted a conservative outlook given limited visibility inherent in its consumption model – indicating revenue is tied directly to how much customers use the platform – making it harder to forecast.

Outlook

The near-term outlook is unusually bifurcated. The geopolitical backdrop, centred on the Strait of Hormuz closure and the Iran conflict, represents the most significant external risk to global markets in several years.

The arrival of true agentic AI around the turn of the year marks a critical inflection point: the shift from AI assistance to execution, doing ‘real’ work.

The Trust makes use of Nasdaq put options – that increase in value when the Nasdaq falls; used in the interests of efficient portfolio management to soften the portfolio's sensitivity to market movements in the event of a sharp market drawdown. The put options were helpful towards the end of the month as the market fell and we rolled our exposure to lock in some profits but still maintained some protection in case of a worsening Middle Eastern conflict and inflationary shock. This should provide a meaningful offset to the downside if there is a further market drawdown in the near term, but it is important to note the position is not sufficient to protect absolute returns.

The arrival of true agentic AI around the turn of the year marks a critical inflection point: the shift from AI assistance to execution, doing ‘real’ work. The key underlying driver is improving large language model quality and capabilities. As intelligence increases, accuracy improves and errors fall, enterprise deployment becomes more practical and the economically addressable range of real-world use cases expands.

Three things have converged to enable this.

First, model capabilities have crossed a threshold, especially around coding. Most of the code at leading AI laboratories is now written by coding agents, AI systems that can write and test software autonomously. METR's 'time horizon' metric tracks the length of task that frontier agents can complete autonomously with at least 50% reliability. ChatGPT in 2022 could handle sub-one-minute tasks; GPT-5.2 in late 2025 managed tasks exceeding six hours; Opus 4.6 in 2026 has reached a point estimate of 14.5 hours. Historically, this metric doubled every seven months; over 2024-25 the doubling rate accelerated to roughly every four months, outpacing even the most optimistic forecasts.

Second, the tooling layer – the software infrastructure that allows AI to interact with real-world applications AI that can independently plan and execute complex, multi-step tasks – has matured for different use cases. Claude Code and OpenAI's Codex are agentic solutions for developers, but Anthropic's Cowork has brought agentic execution to non-technical knowledge workers by synthesising research, managing files and completing multi-step workflows on the desktop. Domain-specific plugins and custom skills have further extended utility.

OpenClaw – an open-source personal AI agent whose underlying code is freely available for anyone to use and build upon – went viral in January 2026 and became the fastest-growing open-source project in history. It offers a local personal assistant that connects AI models to a user's local file system, browser and messaging apps with persistent memory, allowing agents to work autonomously across sessions. NVIDIA CEO Jensen Huang described it as "the operating system for personal AI" and launched NVIDIA's own NemoClaw wrapper, which offers enterprise governance and security guardrails on top.

Third, the way agents interact with tools has fundamentally changed. Previously, each time an agent needed to use a tool – for example, to query a database, check a file or call an application programming interface – it required a separate round-trip through the model, with each step's output feeding into the next step's input. Errors compounded and the amount of information an AI model can hold in its working memory at any one time (the ‘context window’) filled up with intermediate results. New approaches, such as Anthropic's Programmatic Tool Calling, allow agents to batch multiple tool calls in code executed within an isolated digital space where actions can be taken safely without affecting other systems, filtering results before they reach the model. This shifts orchestration from natural-language reasoning to structured, deterministic code, and improving reliability on complex, multi-step tasks. OpenClaw's skills architecture works on a similar principle: agents follow bundled instructions and scripts rather than reasoning through each step from scratch.

The impact is staggering. Enterprise AI adoption has lagged consumer use for three years because early models were too unreliable to perform valuable work or justify the change management required. A year ago, we expected the enterprise and agentic inflection to arrive in 2027. The pace of model improvement and the innovations in tooling described above mean it is happening now and has materially accelerated since the launch of the latest generation of models including Anthropic's Claude 4.6 Opus and Google's Gemini 3.0 in late 2025.

Most technological cycles evolve incrementally, but AI is not following that pattern and our aim is to help investors capture the AI opportunity and navigate the dramatic change we believe lies ahead.

We have long stated we feel this cycle is not a normal technology cycle and it will be even more impactful than even previous general-purpose technology cycles. Instead, we are in the early phases of a rare period of discontinuous technological progress, with change happening in sudden leaps rather than gradual steps, which will drive significant productivity benefits for companies, provide a deflationary growth tailwind for global growth and likely reshape almost every sector.

Anthropic's own revenue trajectory is strong evidence that we have finally hit the inflection point and that the next 12-24 months will see unprecedented change. Anthropic has grown from $1bn in annualised revenue in late 2024 to $9bn by the end of 2025, before accelerating to $14bn by mid-February 2026, $19bn by early March and a staggering $30bn one month later. The revenue curve looks exactly like a moment of discontinuity that nobody – including us – had predicted six months ago. Demand is now inflecting sharply, driven by AI model performance and agents that can sustain longer chains of reliable work. It is too early to say for sure, but if we are correct, and the global economy holds together and the Middle East crisis passes, then this could be a strong period for equity investors, at least those on the right side of the AI trade.

The speed of change is also inflecting higher: during the month, a leaked blog post indicated that Anthropic's newest Mythos model represented "a step change" in capabilities, which Anthropic subsequently confirmed. The new model reportedly gets "dramatically higher" scores on coding, reasoning and cybersecurity, which caused cybersecurity stocks to sell off (the Trust has very limited exposure to cybersecurity). The release has apparently been delayed while the company works with key ecosystem partners, including cloud providers and large cybersecurity vendors to put in place the fixes required due to the security flaws and encryption breakthroughs it has identified already.

Our team is already using agents in the research process, helping to accelerate screening, research and analysis and seeing very strong results. The agentic shift is transformative for AI infrastructure too because agents working autonomously in parallel in the background can consume 1,000x to 1,000,000x more computing tokens – the units of data that AI models process – than standard queries, driving new demand for high-end AI infrastructure. Alphabet has talked about needing to double AI serving capacity every six months to meet demand. This reinforces our longstanding view that inference demand – the computing power required when AI models are actively being used – will become an increasingly important driver of compute demand and hardware investment. Inference capacity is directly tied to AI revenue generation and return on investment.

The data points from GTC 2026 were also supportive: $1trn in NVIDIA platform orders through 2027, Vera Rubin delivering 10x performance-per-watt improvements, enterprise agentic AI reaching an inflection point and AI laboratory revenues growing at unprecedented rates. Supply chain checks from our investment team of 12 amid a recent extensive travel programme in the US and Asia indicate that AI demand and supply chain resilience remain robust. AI data points continue to underpin our constructive view, with Mythos reportedly a 2x improvement over previous versions.

Meanwhile, OpenAI closed a $122bn funding round, valuing the company at $852bn including the latest round of funding, and Coreweave closed its first investment-grade financing round secured by GPUs – notably the first investment grade-rated financing secured by high-performance computing infrastructure and an associated customer contract. Investment-grade financing refers to borrowing that has been assessed by credit rating agencies as carrying a low risk of default.

We believe the market is underestimating the exponential growth in AI demand that lies ahead. Most technological cycles evolve incrementally, but AI is not following that pattern and our aim is to help investors capture the AI opportunity and navigate the dramatic change we believe lies ahead. The challenge for investors will be to balance the macro uncertainty and risks against a rare moment of discontinuity where AI capability compounds and inflects rather than advances linearly.

1. As per the MSCI All Country World Net Total Return Index

2. As per the S&P 500 Index

3. As per the DJ Euro Stoxx 600 Index