The biggest investment theme of our generation is entering a new phase. You’ve probably heard the term ‘agentic AI’. This is the idea that you can ask AI to do a task and it will break it down into smaller sub-tasks and work independently, using tools like search engines and databases, to complete the job without your supervision.
That ability to work unsupervised is why agentic AI promises big gains in productivity. There is, however, a problem. It’s one thing for a large language model (LLM) to answer a series of single, human-generated queries in a few seconds or minutes each. It’s quite another to have the model work autonomously for hours at a time, generating its own queries and building its own teams of sub-agents.
AI agents need a lot more compute, a lot more energy, and a lot more infrastructure than the applications most of us have used to date like chatbots. This explains the billions in capital expenditure going into building datacentres, power generation, and other AI-related infrastructure. Over the first nine months of 2025, investment in AI accounted for 39 per cent of US economic growth, analysis from the St Louis Federal Reserve shows.1
But how sustainable is the current rate of capex? Where is that capital flowing? And where are the best investment opportunities likely to appear next?
Is the current rate of AI investment sustainable?
A corollary of the ramp up in AI capex has been falling free cashflow among hyperscalers, which are the big cloud computing providers like Amazon, Microsoft, and Google. This has led to a situation where the current cash earnings structure is probably not sustainable.
“Either capex will need to come down or revenue and absolute free cashflow will need to go up,” says Jonathan Tseng, a Fidelity analyst who covers the semiconductor market. “The underlying question is what spending from customers can be tapped, either enterprise spend or consumer advertising and subscription revenues, and whether LLM-based offerings can access it.”
Tseng sees good reasons to be optimistic that the question is resolved in favour of higher AI revenue as opposed to lower AI capex. He points to the rise of AI ‘harnesses’, which enable LLMs to turn their output into actions, for example by writing and executing code.
Tseng also notes the importance of Anthropic’s decision to develop LLMs’ ability to write code.
“If a model can write code then, by extension, the model can do anything a computer can do,” he says. “Model harnesses like OpenClaw are the ‘missing link’ which turns this idea into reality.”
The implications are profound and underpin the case for hyperscalers’ free cashflow being restored via higher revenues rather than cutting capex. Most white-collar work is mediated via computers. So if an LLM can do anything a computer can do, it brings all that knowledge work into the realm of AI.
“The total addressable market for LLM spending is no longer the IT budget but the broader wage budget of the business world,” says Tseng.
There’s also a strategic bias towards higher capex.
“Big Tech views AI as existential, favouring overinvestment to avoid being left behind,”
says Alex Grant, an analyst and portfolio manager covering communication services. “CEOs explicitly prioritise excess capacity over underinvestment, given the risk of competitive disadvantage. Scarcity of data centre resources creates a game theory dynamic, incentivising firms to secure capacity both for their own growth and to block competitors.”
The appetite to invest at scale may be there, but there are hurdles to overcome that could slow the rate at which agentic AI can be adopted and monetised. Understanding where these bottlenecks are is helpful for judging where capital is likely to flow.
“The market’s behaviour through this cycle has been notably consistent in one respect, in that it disproportionately rewards scarcity,” says capital goods analyst Shreeji Parekh. “Capital has rotated toward every identifiable bottleneck in the datacentre build-out, from accelerator chips and electrical equipment in 2022-23, then air and liquid cooling and the wider server ecosystem in 2024, followed by power generation in 2025.”
Memory chips are the most recent example of scarcity driving price increases. One constraint holding back agentic AI is that compute power has outpaced the speed at which the overall system can get data into and out of chips. The physical pathways between chips, known as interconnects, are typically made of copper and are hitting physical limits in terms of bandwidth, latency, and power efficiency - a problem known as the 'memory wall'.
“The memory wall is a key concern,” says Tseng. “The bottleneck for expensive datacentre compute is increasingly down to connectivity. This will get worse as chips get more powerful. A beefier chip needs even faster interconnects to shovel data in and out of it and ensure it is fully utilised. Hence data connections within the datacentre will inevitably shift from copper towards optical interconnects.”
Where the chips will fall
Looking forward, there are potential ‘picks and shovels’ plays to consider. For decades, computers have improved by making transistors smaller, thereby fitting more of them on each chip and improving computing power. Increasingly, though, the focus is shifting to overall system performance.
“How chips are bonded together is becoming much more important, necessitating more capital-intensive processes such as hybrid bonding and thermocompression bonding,” says Austin Kelly, an analyst covering semiconductor capital equipment. “As a result, the assembly equipment market will go from a low-growth, cyclical market to a cyclical but structurally growing market.”
But Kelly points out that datacentres also need a lot of less sophisticated analogue chips that process signals from the real world, for example to manage power delivery to servers.
“In 2027, I think there will be a realisation that AI demand will drive an exceptionally tight market in analogue chips,” he says. “These dynamics will cause a huge squeeze on capacity, but one that only becomes evident once you have both end markets, non-AI and AI, moving in the same direction. The analogue stocks will rerate once these supply-demand dynamics become obvious.”
Tseng sees similar potential for a demand crunch affecting other types of chips that are less sophisticated than the graphics processing units (GPUs) that LLMs rely on. This is because while an AI agent makes heavy use of GPUs, it also sends instructions to other tools, including long-established software programmes, that use older technology.
“A potential new area of shortage,” he says, “could be server central processing units (CPUs), as increased agentic tool-use creates paradoxical demand for the legacy, non-accelerated compute that these tools primarily run on.”
Elsewhere, our analysts see potential opportunities arising from the construction and operation of the datacentres needed to make agentic AI a reality at scale.
“Hyperscalers are increasingly shifting toward modular and prefabricated builds, shifting value toward companies that can offer full stack swift deployment rather than just supply components or systems,” says capital goods analyst Parekh, who also sees scope for companies that supply heating, ventilation, and cooling (HVAC) and electrical equipment to carry on their strong performance that goes back to 2023.
Selected utilities meanwhile should also do well where they have exposure to datacentre growth and control the whole supply chain of power or water.
“Escalating NIMBYism (Not In My Backyard) across the US will drive datacentre growth to Texas, where there are no zoning laws outside cities and the state government and utility regulators are highly supportive of datacentres,” says utilities analyst Srishti Sinha.
Preparing for disruption
We may not yet live in a world where we can delegate work to an LLM and trust the model to get on with it. Increasingly, though, the barriers to making that a reality are physical rather than technological.
One by one, those barriers are being removed by the brute force of enormous capex. Naturally there is some disquiet about whether the sums involved are sustainable, and what the return on investment will be.
Weighed against that is the enormous prize on offer if AI agents deliver on their promise. In that scenario, compute will remain constrained for years to come as AI takes over more and more of the economy, a favourable set-up for sectors and businesses supplying into that scarcity.
