Following the US midterm elections, investors have expressed concerns that tighter regulation could slow the buildout of the US’s AI infrastructure. However, the outlook for AI may depend less on regulation than on the widening gap between demand for compute and the industry’s ability to supply it. In a market where physical capacity is scarce and slow to adjust, existing infrastructure becomes more valuable, and policy decisions may affect where investment is made in the future.
The recent weakness in AI-related equities reflects concerns that tighter regulation, permit constraints and local political opposition could slow data centre construction, each influencing the pace, cost and location of investments differently. However, while equity markets have become focused on the legislative risk on the supply side, the companies building AI infrastructure are no longer investing against hypothetical demand but are responding to demand that already exists.
The question for investors is therefore not whether AI demand exists but whether supply can catch up. Cloud providers report strong demand, and business adoption is broadening. As firms move from experimentation to implementation, token consumption, a useful proxy for AI activity and demand for compute, has risen rapidly over the past two years. ‘Compute’ refers to the processing power required to train and run AI models.
Companies building AI infrastructure… are responding to demand that already exists
Tokens are basic units of text that an AI model reads or generates. Demand is increasingly driven by ‘agentic AI’ models that can reason, plan, and execute tasks with limited human intervention. Compared with a conventional chatbot query, these processes consume far more computing resources through hundreds of AI-generated steps and consume as much as 1,000 times more tokens, depending on the task and model. AI adoption is therefore expanding in terms of users and more compute per task. At the same time, ‘Recursive Self Improvement’ (RSI), in which AI models help to develop successive generations, is accelerating the pace of innovation and adding further pressure on demand for computing resources.
Developments surrounding Nvidia, the world’s biggest semiconductor supplier, illustrate this distinction. The departure of the firm’s chief executive has introduced uncertainty around succession and the company’s ability to identify or influence future AI trends. However, while Nvidia plays a pivotal role in the AI ecosystem, investor concerns around CEO succession are largely company specific, and do not undermine the strength of underlying AI demand.
Physical infrastructure constraints
Until now, much of the AI investment debate has been dominated by semiconductors, specifically advanced graphics processing units (GPUs) and high bandwidth memory (HBM). But their shortage is part of a much larger challenge. In other words, what increasingly matters is the AI industry’s ability to orchestrate its four critical ingredients: data, skilled labour, compute, and electrical power, into a single ecosystem. Every AI model, no matter how sophisticated, depends on energy, cooling, and networking infrastructure and AI remains a digital technology with deeply industrial requirements.
Much of the ecosystem remains effectively sold out
Data centres for example require more than land and electricity. They also depend on specialised networking equipment, electricians, engineers, project managers and construction workers. Such bottlenecks cannot be solved by software, and many were created before AI became a focus for investors. In 2023, when interest rates were higher than today, many companies delayed plans to expand their capacity. When AI demand surged after that, they scrambled to respond. However, building a new semiconductor manufacturing facility, data centre or expanding memory production takes years to complete, test and power up.
This explains why much of the ecosystem remains effectively sold out. Producers of HBM chips report that their order books are full, equipment makers say they are heavily booked up and hyperscalers are competing for capacity. The gap between demand and supply remains substantial and provides a buffer against short-term fluctuations in sentiment or policy. In any high-demand, supply-constrained market, existing and soon-to-be-available infrastructure becomes more valuable, further improving the economics of investments.
The next phase of the US-China AI race
Physical constraints not only shape the economics of AI’s infrastructure, but also where that infrastructure is ultimately built. If US state-level regulations were to slow AI’s buildout, in response to concerns around resources and local impacts on health or water use, demand for AI services would not stop. It would simply increase the chances that training, inference and deployment move to a different market. That raises the question, which economies have the infrastructure to deploy AI at scale?
This distinction is important in the context of the strategic competition between China and the US. While this is often presented as a ‘race’ and assumes that ‘winning’ means developing the most powerful AI models, economic value is created by rolling out the models for widespread use, not just by innovation alone.
The domestic US policy debate will eventually, we believe, reflect a bipartisan political determination to remain ahead in technological competition with China. This strategic risk for the US can be summed up as the danger of becoming a ‘token importer’ rather than a ‘token producer’. Just as economies have competed for manufacturing leadership, they may start to compete for leadership in the production and delivery of AI services.
Strategic risk for the US can be summed up as the danger of becoming a ‘token importer’ rather than a ‘token producer’
While China’s access to the most advanced semiconductors may be further restricted, the country is competitive in terms of engineering talent, and benefits from quantities of data generated by one of the world’s largest and most digitised consumer economies. China has also spent years expanding its domestic electricity capacity across multiple technologies, creating a power system with far more generating capacity than the US.
Uncertainties over Nvidia’s leadership succession also raise questions for the AI race. If Chinese advances in AI depend on access to US-designed semiconductors through unofficial channels, or depend on ‘distillation,’ a technique that allows simpler AI models to learn from more advanced ones, then tighter US export controls could widen the technological gap between the two countries. If, instead, recent Chinese advances reflect domestic innovation, additional US restrictions may slow rather than stop development.
Diverging paths
China has already demonstrated an alternative approach. To date, China’s limited access to the most sophisticated semiconductors has meant that rather than relying on ever-larger or frontier-technology models, developers have often focused on efficiency, optimisation and rapid deployment. As a result, China is not leading development into the most powerful models but may be able to embed AI throughout its economy quickly, widely and cost-effectively. DeepSeek, a Chinese AI developer, and Huawei, the country’s largest telecommunications business, are for example working on an ecosystem combining domestic capabilities rather than relying on Nvidia hardware. It is an open question whether China can catch-up using such techniques, and whether it can perfect its chip circuit printing technologies. For investors, Nvidia’s leadership change is shifting AI thinking to how innovation evolves when access to technology, capital or computing power is constrained.
Early November’s AI weakness reflected the unwinding of a crowded trade and concerns about near-term policy developments, rather than a fundamental deterioration in AI demand. Business adoption continues to rise, token consumption is still surging and capital expenditure on an upward trajectory, even if growth rates moderate. Meanwhile, companies across the AI supply chain remain capacity constrained.
AI is a digital revolution with physical foundations
Within our equity allocations, we favour AI-exposed markets including the US, Japan, South Korea and Taiwan. IT remains a preferred sector, as do utilities, where companies are benefiting from investments in power generation, grids and electrification, and financials, which is also positioned to benefit from AI adoption. Within the tech sector, we see AI value increasingly accruing to owners of scarce infrastructure and supply-chain bottlenecks. We therefore favour exposure to hyperscalers with resilient business models, and semiconductor and component firms that own constrained inputs. We are highly selective in software and hardware stocks.
AI is a digital revolution with physical foundations. Data centres, transmission networks and power generation, semiconductor fabrication plants and specialised labour are all in limited supply. In recent weeks, financial markets have underlined how quickly sentiment can move, but infrastructure evolves much more slowly. For investors, that distinction remains central to capitalising on long-term AI opportunities.
CIO Office Viewpoint
AI’s physical constraints still matter