The Bottleneck We Are Removing

I have been trying, with mixed success, to get better at golf.

Anyone who has tried to improve at golf knows the frustration. The answer is not simply to swing harder. In fact, swinging harder usually makes things worse. Real improvement comes from identifying the bottleneck — grip, balance, tempo, rotation or sequencing — and removing it.

The same principle applies to businesses, markets and economies. Progress rarely comes from adding more effort to the same system. It comes when a true bottleneck is removed.

History doesn’t repeat, but it does rhyme

The Industrial Revolution, in its most important sense, was not simply a story about machines replacing people. Beginning in the late 1700s and accelerating through the 1800s, machines began to substitute for one specific kind of human input: physical strength. The steam engine did not make a worker more skilled; it removed strength from the productivity equation. Coal mines, mills and factories could produce far more output with less human and animal effort.

The transition was not painless. Jobs were displaced, wages were disrupted, and communities had to adapt. But the long-term result was not permanent mass unemployment. It was the most dramatic improvement in human living standards ever recorded. Entire categories of work — railroad engineering, electrical installation, industrial chemistry, photography and accounting at scale — became major employers in the new economy.

We believe we are now at the front edge of an equivalent moment in history. The bottleneck being removed this time is not physical strength. It is skilled knowledge work.

The cognitive bottleneck

Today, we do not lack physical capacity to manufacture goods. Global factories often run below maximum throughput. We do not lack raw materials in any permanent sense. What we lack is enough trained, capable knowledge workers to do the analytical, creative and judgment-intensive work that separates productive economies from stagnant ones.

The evidence is visible across the labour market. Skilled software engineers, doctors, lawyers, scientific researchers, pharmaceutical chemists and experienced financial analysts command premium wages because demand for their work generally exceeds supply.

This is the crucial insight: the true ceiling on much of modern economic activity has been the supply of skilled knowledge workers, not merely the supply of capital, raw materials or physical capacity.

In 2025 and 2026, that constraint began to loosen in a commercially visible way. 

The first proof point: software engineering

The clearest signal is software engineering, where AI tools have been deployed earliest and most aggressively.

Software was the first frontier because the work happens in symbols, on a screen, with measurable outputs. AI coding assistants are changing how software is written, reviewed, tested and deployed. Leading frontier model developer Anthropic has seen revenues grow from $1B to $19B in 15 months¹. 

That is not the demand curve of a curiosity. It is the demand curve of a tool users cannot imagine giving up.

Software engineers have not suddenly become obsolete. Instead, the amount of software being built appears to be expanding. Companies that previously could not justify a feature, workflow or internal tool can now build it.

This pattern of improved productivity, demand expansion and work reorganization is consistent with prior technological waves. When the steam hammer arrived, the blacksmith did not simply disappear; the metallurgist emerged. When textile machinery scaled production, the textile engineer was created. When a true bottleneck is removed, the economy often produces much more in ways that were previously uneconomic or unimaginable.

Over the next five years, we believe this pattern will appear across skilled cognitive work: radiology, legal research, financial analysis, engineering, accounting, drug discovery and customer support. Some tasks will be automated, but the larger effect may be that each trained professional supervises more output and covers more ground.

The bottleneck moves from “do we have enough people?” to “what should we do with this new capacity?”

The physical infrastructure of cognitive abundance

This is not a marginal capital cycle. It is a generational infrastructure build comparable in scale to the railroads, the electrical grid or the interstate highway system — compressed into a much shorter period.

Artificial intelligence may feel like a software story, but at scale it is also a physical story. The cognitive revolution requires data centers, electricity, networking systems, semiconductors, water, cooling systems and metals.

According to BofA Global Research’s recent infrastructure analysis, global data center capacity is expected to double by 2030, requiring approximately $7 trillion of cumulative capital investment. By 2030, data centers are expected to consume more electricity than Japan, drive more than 20% of demand growth in advanced economies, and account for an estimated 3% of global electricity consumption, up from approximately 1.5% in 2025.²

Capital expenditure guidance from U.S. cloud providers — Amazon, Microsoft, Google and Oracle — points to roughly $585 billion in 2026 and $700 billion in 2027.3 These cloud providers — known in the industry as “hyperscalers” — are reinvesting an extraordinary share of operating cash flow into capital expenditure, reminiscent of earlier infrastructure booms. U.S. data center power demand alone is expected to rise from roughly 32 gigawatts in 2025 to approximately 95 gigawatts in 2030, requiring an estimated $600 billion-plus of grid investment.² 

The bottlenecks within the bottleneck

Capital is readily available, but several physical inputs to AI infrastructure are themselves bottlenecked. That is what makes the opportunity particularly interesting.

BYOP – Bring Your Own Power

An AI server rack consumes the equivalent power of 65 households when running at full capacity. To get ahead of a crippling surge in electricity demand, the U.S. policy direction has shifted toward requiring hyperscalers and AI infrastructure builders to “build, bring or buy” their own incremental power, rather than relying entirely on existing electrical grids and passing the cost to ratepayers.

Natural gas turbines, transformers, switchgear, grid-scale electrical systems and power-management equipment have become critical constraints. Turbine and transformer lead times have stretched to several years, prices remain well above pre-pandemic levels, and backlogs are measured in years, not quarters.⁴

Water and Cooling

Water may be the most overlooked input in the AI supply chain. Today, a single 100-word AI prompt consumes roughly half a liter of water. As we extrapolate data center growth beyond 2030, water demand is projected to double. This is not sustainable over the long term.2

As rack densities rise, the challenge moves from delivering power to removing heat. That shift increases demand for liquid cooling, chillers, heat exchangers, pumps, sensors, controls, water treatment and site-level water planning.²

Critical Metals

Data centers and associated electrical infrastructure require copper, aluminum, steel and other industrial metals. Cooling and backup power systems are particularly metals-intensive. AI data centers are expected to consume a growing share of global copper demand by 2030, adding to demand from electric vehicles, electrification and grid expansion.2

Metals may represent a relatively small share of total data center capital expenditure, but they can increasingly determine project timelines.

The lesson is straightforward: every link in the AI delivery chain is constrained by something physical. The economic value is likely to concentrate in whoever owns, manufactures, designs or controls the constrained input.

Implications for our investment framework

Markets have begun repricing for this reality, but in our view the repricing is incomplete and may take five to ten years to fully play out.

Through prior eras of major infrastructure transition, the largest and most durable returns often came not from the visible novelty itself, but from the businesses that supplied the inputs required to scale that novelty. The fortunes of the Industrial Revolution were built not only by inventing better engines, but by mining coal, milling steel, generating power and financing the networks that connected them.

The parallel today is compelling. The most durable franchises of the AI era may not be limited to the companies building the models. They may also include the companies that supply the physical inputs required to make those models useful at scale.

These are not all “AI companies” in the narrow sense. Many are old-economy businesses with new-economy demand curves. They may appear less glamorous than the models themselves, but they often possess hard physical assets, long-duration contracts, installed bases, regulatory complexity, technical expertise and capital-intensive moats (durable competitive advantages) — characteristics that can support durable compounding.

That is the structural framework underlying our portfolio positioning.

We are deliberately emphasizing businesses that own or supply irreplaceable pieces of the physical backbone of AI. We prefer bottleneck suppliers where capacity is difficult to replicate quickly: long-cycle manufacturing, specialized engineering, regulatory complexity, installed-base service revenue, scarce physical assets and mission-critical products with high switching costs.

We are also being more selective with businesses whose moat is primarily information-based and therefore more vulnerable to commoditization by the AI tools transforming the economy.

Valuation discipline remains essential. A great theme can still become a poor investment if purchased at the wrong price. Our focus is not simply on finding companies with AI exposure, but on identifying businesses where demand durability, moat quality, balance sheet strength and expected return justify the valuation.

In simple terms, we want exposure to the companies helping AI become real in the physical world.

Risks worth taking seriously

It would be intellectually dishonest to present this thesis as a one-way bet.

The timing and path of returns are uncertain. AI may unfold faster than prior technological revolutions, but the path will not be linear. Periodic pauses in capital spending, valuation pullbacks, swings in investor sentiment and overcapacity worries are inevitable. Capital deployed at peak euphoria can take years to recover, even when the thesis is right.

The distribution of value — who actually keeps the gains — is also uncertain. Productivity gains may accrue to consumers, workers, infrastructure suppliers, model providers, hyperscalers or some combination of all of them. If AI-augmented workers retain most of the value, or competition drives model economics toward commoditization, some equity returns may disappoint relative to the apparent size of the opportunity.

There are also physical, political and environmental risks. Power generation, transmission lines, water usage, permitting and local community resistance will all matter. The AI build-out may be digital in purpose, but it is physical in execution — and physical infrastructure is rarely built in a straight line.

We hold these risks alongside our base case rather than allowing them to dominate it. The discipline is to size positions so we can survive cyclical declines and remain invested through them.

We also want to be plain about one short-term reality. AI-related companies have been one of the main reasons the market has done well over the past two years. To be clear, these are largely diversified businesses with established product lines and customer bases — AI-related spending has accelerated their growth, not defined it. Any group that has led the market for that long will, at some point, lag behind other parts of the market for a stretch — when investors get nervous, when money moves into other sectors, or when a quarter of data center spending looks uneven. Over the longer term, however, we continue to believe this is where the strongest returns will come from. We have sized our positions so we can stay invested through the inevitable short-term swings rather than chase any single quarter or year.

Closing thoughts

The cognitive bottleneck is beginning to loosen. The infrastructure required to remove it is being built at unprecedented scale. The companies supplying that infrastructure can compound through this transition — provided we remain disciplined on valuation, balance sheets and position sizing.

We will be watching three signposts:

Hyperscaler capital expenditure guidance into 2027 and 2028.

Measurable diffusion of AI into non-software white-collar work such as radiology, legal research, financial analysis, drug discovery, engineering and accounting.

Relative earnings revisions of physical infrastructure providers versus pure software application businesses.

The visible novelty is the AI model. The lasting wealth may lie in the infrastructure that makes models useful at scale.

As always, we remain focused on separating durable structural change from short-term market enthusiasm. If you have any questions about your portfolio or how we are positioning through this environment, please do not hesitate to reach out to the team.

Best regards,

Marc Correia

References

Anthropic annualized revenue trajectory; data sourced from JPMorgan, China Artificial Intelligence: Addressing Ten Key Questions for Investors, March 27, 2026.

BofA Global Research, Transition Investing — Data centers: power-hungry, water-thirsty, metals-needy, $5tn opportunity, April 21, 2026. Data center, water, electrical and metals statistics in this commentary are primarily sourced from this report unless otherwise noted.

Goldman Sachs, AI Hyperscaler Reinvestment, 2026; cross-referenced with Deutsche Bank, Public Cloud & AI Update: Reflecting on C1Q Results, May 4, 2026, 

Eaton Corporation & GE Vernova Q1 2026 earnings releases and earnings calls, May 2026; data center backlog and order growth disclosed by the companies. 

Disclaimer

RBC Dominion Securities Inc.* and Royal Bank of Canada are separate corporate entities which are affiliated. *Member-Canadian Investor Protection Fund. RBC Dominion Securities Inc. is a member company of RBC Wealth Management, a business segment of Royal Bank of Canada. â / ™ Trademark(s) of Royal Bank of Canada.  Used under licence.  © RBC Dominion Securities Inc. (2026). All rights reserved.

This information is not investment advice and should be used only in conjunction with a discussion with your RBC Dominion Securities Inc. Investment Advisor.  This will ensure that your own circumstances have been considered properly and that any action is taken based upon the latest available information. The strategies and advice in this report are provided for general guidance.  Readers should consult their own Investment Advisor when planning to implement a strategy. Interest rates, market conditions, special offers, tax rulings, and other investment factors are subject to change. The information contained herein has been obtained from sources believed to be reliable at the time obtained but neither RBC Dominion Securities Inc. nor its employees, agents, or information suppliers can guarantee its accuracy or completeness.  This report is not and under no circumstances is to be construed as an offer to sell or the solicitation of an offer to buy any securities.  This report is furnished on the basis and understanding that neither RBC Dominion Securities Inc. nor its employees, agents, or information suppliers is to be under any responsibility or liability whatsoever in respect thereof.   The inventories of RBC Dominion Securities Inc. may from time to time include securities mentioned herein.

This commentary is based on information that is believed to be accurate at the time of writing and is subject to change.  All opinions and estimates contained in this report constitute RBC Dominion Securities Inc.'s judgment as of the date of this report, are subject to change without notice and are provided in good faith but without legal responsibility. Interest rates, market conditions and other investment factors are subject to change. Past performance may not be repeated. The information provided is intended only to illustrate certain historical returns and is not intended to reflect future values or returns.