Agentic AI: Where are the opportunities?

Key Points

• The defining feature of the agentic AI era is the surge in token consumption, where cloud service providers stand to benefit.
• Functions such as planning, memory management, tool execution, and multi-agent orchestration rely heavily on CPUs and memory rather than GPUs alone.
• The advancement of AI has led to an unprecedented surge in electricity demand. Autonomous agents, which run continuous reasoning loops or long-duration tasks (such as coding and browsing), can consume significantly more energy than simple one-off queries.
• Over the past three years, progress in AI has been measured by raw intelligence — larger models, sharper responses, and more fluent text. That phase now appears to be maturing. The next frontier is agency: AI systems that not only respond, but execute.

Structural Shift Or A Hype Cycle

Agentic AI systems such as OpenClaw have seen rapid adoption, driven by their ability to perform real-world actions rather than simply generate text. These systems can autonomously execute terminal commands, manage local files, complete web forms, and control browsers within a secure sandbox environment.

As of 22 April 2026, OpenClaw has processed approximately 17.2 trillion tokens this month alone, significantly surpassing the usage levels observed in any standalone LLMs.

In our view, current advancements in AI are clearly supporting the transition towards agentic systems. However, the technology remains early-stage and operationally inconsistent. Key challenges include inconsistent task execution and high costs arising from substantially increased token consumption. That said, We expect these constraints to ease over time through continued improvements in AI capabilities, as well as more advanced chips that reduce the cost per token.

On the positive side, near-term token consumption is likely to be supported by Chinese models. As users can choose which models to run their agents on, this trend has created a structural cost advantage tailwind for lower-cost Chinese offerings. On average, Chinese models are priced at roughly one-sixth the cost per token of US alternatives from companies such as OpenAI and Anthropic. This cost advantage is partly driven by China’s relatively inexpensive and abundant electricity for data centres, which helps offset the limitations of less powerful and less energy-efficient domestic chips.

Combined with ongoing efficiency gains in software and algorithms — driven in part by the need to innovate under US technology restrictions — this creates a compelling value proposition. Indeed, daily token consumption in China has surged to approximately 140 trillion as of March, up from 100 trillion at the end of 2025, according to government estimates.

Figure 2: Chinese AI models are cheaper than western rivals per token

Source: Artificial Analysis, Data as of 13 April 2026.

Agentic AI to Provide a Further Boost to AI-Related Sectors

We believe the industry is transitioning into the era of agentic AI. While adoption remains at an early stage, the long-term outlook for this theme is highly compelling and continues to create a broad set of investment opportunities. Below are some of the opportunities for investors to capture within this agentic AI era.

Investment Opportunities
Cloud Computing

The defining feature of the agentic AI era is the surge in token consumption. According to OpenRouter, token usage has spiked sharply since early 2026, when AI agents began to gain widespread attention.

Cloud computing remains the core enabler of the AI agent era, primarily by providing the elasticity and vast computational power required for autonomous reasoning. Unlike traditional single-turn inference models, AI agents operate in iterative “loops” — repeatedly reasoning, using tools, and self-correcting — which significantly increases token consumption and GPU utilisation.

For cloud providers and hyperscalers, this represents a shift from sporadic usage to continuous, high-intensity compute demand, thereby justifying the substantial AI-related capital expenditure directed towards accelerated infrastructure. By hosting these agents, cloud platforms are evolving from simple storage providers into “AI operating systems”, where compute and logic are tightly integrated.

In early April 2026, Anthropic restricted third-party agent harnesses, specifically targeting OpenClaw — the open-source AI agent framework that had rapidly gained traction among developers. Notably, Anthropic argued that OpenClaw users were effectively paying around USD 200 in subscription fees for usage that would otherwise amount to USD 1,000–5,000 in API value.

The scale of the issue was significant. More than 135,000 OpenClaw instances were estimated to be running at the time of the announcement, and industry analysts highlighted a price gap of over five times between flat subscription costs and equivalent API-based usage for heavy agentic workloads. In short, agentic AI is driving a structural boom in token consumption.

From an investment perspective, we believe hyperscalers stand to benefit the most from this surge in computing demand, particularly Alphabet Inc., Microsoft, and Amazon.

Featured products: Unit Trust: 1) Fidelity Global Technology A-ACC-USD 

Read More: Digital Economy (Internet) update: Sailing into headwinds, but staying the course; Maintain 3.5 Star

Figure 3: API Token-Consumption growth

Source: Bloomberg Intelligence. 

Semiconductors

While GPU demand remains strong, each model call increasingly requires greater coordination, memory, and system-level compute. AI workloads are increasingly constrained not by compute, but by system architecture efficiency (memory bandwidth, data movement, interconnect latency, and overall system coordination). Data centres are no longer centred around a single chip, but rather fully integrated system architectures, where each layer is designed to eliminate bottlenecks between compute, memory, and data flow.

Traditional generative AI — such as early versions of ChatGPT, Claude, and Gemini — operates reactively: a user inputs a query and receives a response. These are what researchers describe as passive tools: they lack persistent memory of prior interactions, cannot take real-world actions, and are unable to delegate tasks to other systems.

Agentic AI removes these limitations. Such systems can retain context across sessions, plan multi-step strategies, utilise tools (including browsers, code editors, APIs, and databases), and coordinate with other agents to parallelise complex workflows.

Crucially, functions such as planning, memory management, tool execution, and multi-agent orchestration rely heavily on CPUs and memory rather than GPUs alone. As a result, earnings estimates for major players in the memory chip sector have undergone successive upgrades. Companies such as SK Hynix and Samsung Electronics are expected to deliver earnings growth of close to 400% for FY2026, while Micron’s earnings growth is projected to exceed 600% over the same period.

Featured products: Global X Asia Semiconductor ETF

Read More: Global X Asia Semiconductor ETF (HK:3119) and the Asian Semiconductor Supply Chain Hegemony

Figure 4: Earnings growth for SK memory chip companies

Power generation

The advancement of AI has led to an unprecedented surge in electricity demand. Autonomous agents, which run continuous reasoning loops or long-duration tasks (such as coding and browsing), can consume significantly more energy than simple one-off queries. A recent study found that agentic AI can use between 62 and 136 times more energy per query than conventional single-inference AI models.

Electricity demand from data centres rose by 17% in 2025, with AI-focused data centres growing even faster — far outpacing global electricity demand growth of around 3%. While power consumption per AI task is declining rapidly, with efficiency improving at an unprecedented rate, total electricity consumption from data centres is still expected to double by 2030, with AI-focused data centres’ power usage projected to triple over the same period.

At the same time, AI deployment is increasingly constrained by a range of physical bottlenecks, limiting the pace at which data centres can expand in the near term. Supply chains for critical energy infrastructure — such as gas turbines and transformers — as well as advanced chips and IT components, have tightened over the past year. GE Vernova’s order backlog for combined-cycle gas turbine (CCGT) equipment is fully booked for the next four years. Meanwhile, the growing pipeline of data centre projects is placing increasing strain on planning and regulatory systems, delaying grid connections and other necessary approvals.

To address these energy challenges, the technology sector is adopting new approaches.

1)      Increasing use of renewable energy. Corporate power purchase agreements for renewables signed in 2025 accounted for 40% of total.

2)      Nuclear. The pipeline of conditional offtake agreements between data centre operators and small modular reactor (SMR) nuclear projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts today.

3)      Onsite gas-based power generation. Gas turbines that directly bypass the need to go through traditional grids which is subjected to slow connections.

All these has resulted in an unprecedented demand for grid connections. Companies like Eaton, ABB, Schneider Electric provide the provide the critical electrification required to integrate, stabilize, and distribute that power to sensitive AI servers.

ETF feature: First Trust NASDAQ® Clean Edge® Smart Grid Infrastructure Index Fund 

Read more: Data centre infrastructure outlook: AI is crunching on power demand!

Table 2: Accuracy, latency, and GPU energy consumption when servicing a single agent request on HotpotQA. 

Source: Jiin Kim, Byeongjun Shin, Jinha Chung, Minsoo Rhu. The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective. 2506.04301v2 [cs.LG].  07 January 2026. Numbers in parentheses indicate the relative increase over ShareGPT (the conventional single-turn inference). LATS and Reflection both represent AI agents.

Key Takeaways

The emergence of Agentic AI—autonomous systems capable of reasoning and executing complex workflows—represents a structural inflection point from passive assistance to active productivity. This transition provides a powerful structural tailwind for the orchestration and infrastructure sectors, reinforcing our conviction stance.

 Declaration:

For specific disclosure, at the time of publication of this report, IFPL (via its connected and associated entities) and the analyst who produced this report hold a NIL position in the abovementioned securities.

This research report was prepared with the assistance of artificial intelligence (AI) tools. iFAST Financial Pte Ltd does not rely exclusively on AI for content generation; the content of this report – including all investment theses, ratings, price targets and conclusions – has been independently reviewed and verified by the research analyst(s) to ensure accuracy and professional integrity.