The Bottleneck Premium
Why the physical world cannot scale at the speed of intelligence
Research classification: Foundational thesis. Investment recommendation: None. Status: Version 1.0, published October 23, 2026.
The central thesis
Artificial intelligence can multiply the amount of work performed by machines at extraordinary speed.
The physical infrastructure supporting that work cannot expand nearly as quickly.
Digital agents can often be provisioned far faster than power plants, factories and grid infrastructure can be built. Their computation still depends on electricity, cooling, grid connections and industrial equipment that must exist somewhere in the physical world.
This mismatch may become one of the defining economic constraints of the AI era.
The most interesting investment opportunities may emerge several layers removed from the companies building the models.
The bottleneck is not necessarily where the money is being spent. It is where additional spending cannot immediately produce additional supply.
1. Intelligence is becoming abundant. Infrastructure is not.
Our first foundational thesis examined the multiplication of AI agents: a possible transition from occasional chatbot use to persistent personal agents interacting with networks of specialized business agents.
The number of agents alone is not the important economic variable. The important variable is total computation performed: model calls, tokens, active runtime, tool use, reasoning and verification. Ten thousand lightweight agents might consume less computation than a smaller collection of demanding agents.
Nevertheless, the architectural possibility matters. As the marginal cost of machine intelligence falls, new categories of work may become economical. The result could be a substantial expansion in useful computation demanded.
This is a scenario, not a certainty. Improvements in model efficiency, hardware utilization and workload scheduling could offset some of that growth.
But the physical infrastructure question remains:
What happens when demand for computation can expand faster than our ability to supply the electricity required to perform it?
2. The electricity problem is already measurable
The International Energy Agency's 2026 central projection has global data-centre electricity consumption increasing from approximately 485 TWh in 2025 to about 950 TWh in 2030. Electricity consumption from AI-focused data centres is projected to more than triple over that period. These are forecasts, not guaranteed outcomes. IEA, 2026
The IEA also identifies constraints in grid connections, energy equipment, chips, financing, planning and permitting that could prevent more aggressive near-term growth. Electricity-delivery constraints are already a documented cause of delay and could become a major cause of postponement, relocation or cancellation.
A forecast of electricity demand does not mean every proposed data centre will be built.
Electricity must be available at the right location, at the required capacity, with sufficient reliability and at a commercially viable price. An abundance of electricity somewhere else is not necessarily useful.
The relevant commodity is increasingly deliverable power, at a specific location, by a specific date.
3. When the grid cannot arrive, capital looks for another route
Traditionally, large computing facilities depended heavily on utility infrastructure. But new transmission capacity, substations and grid connections can require years of planning, construction and approvals.
For developers attempting to deploy expensive computing equipment quickly, waiting may carry substantial opportunity costs. One response is behind-the-meter generation: producing electricity at or near the computing facility rather than relying entirely on conventional grid delivery.
Natural-gas turbines and reciprocating engines, fuel cells, batteries and hybrid systems are among the possible solutions. These technologies differ considerably in cost, deployment time, emissions, reliability and scalability. We should not assume one will dominate.
What matters is that electricity access has become sufficiently valuable to justify investment in alternative architectures.
Evidence: Enerflex
On October 13, 2026, Enerflex (TSX:EFX) reported a contract to engineer, fabricate and assemble more than one gigawatt of low-emission natural-gas generation equipment for a North American data-centre development. Combined with earlier awards, the company reported more than 1.75 GW of contracted behind-the-meter generation backlog, with deliveries expected through 2028. Enerflex, October 13, 2026
This represents company-reported commercial procurement commitments, not merely management's estimate of a potential market. The release does not disclose contract value, deposits or cash paid.
Gigawatts of contracted equipment are not equivalent to completed generating capacity, recognized revenue or free cash flow. Customer economics, project execution and manufacturing costs still matter.
Enerflex provides evidence that developers are making commercial commitments to secure power outside conventional grid timelines. It does not prove that every company associated with modular generation will benefit.
4. Solving one bottleneck exposes another
Imagine a developer successfully securing dedicated natural-gas generation. The grid-connection problem may be reduced. But the project now requires generation equipment, fuel supply, electrical conversion, protection systems, switchgear, transformers, cables, controls, commissioning and maintenance.
Some of those components have their own manufacturing constraints. The bottleneck moves.
AI workloads → computing capacity → dependable electricity → generation equipment → electrical distribution → qualified components → specialized materials and labour.
This is not a single sequential chain. Several constraints may operate simultaneously, and projects may require different combinations of equipment.
An increase in turbine manufacturing capacity may expose a shortage of electrical integration equipment. An increase in transformer assembly capacity may expose a shortage of qualified components. An increase in equipment availability may expose a shortage of installation crews, permits or energized grid connections.
Capital can relieve individual constraints. It cannot guarantee that every adjacent constraint disappears at the same time.
5. The transformer is a revealing case study
Transformers are essential to much of the electrical infrastructure connecting generation, distribution systems and computing facilities. In 2025, the IEA reported that waiting times for critical grid components, including transformers and cables, had doubled over the preceding three years. IEA, 2025
For manufacturers with established facilities, technical qualifications and customer relationships, this creates the possibility of sustained order growth and stronger utilization.
Evidence: Hammond Power Solutions
Hammond Power Solutions (TSX:HPS.A) reported second-quarter 2026 revenue of C$324.8 million, an increase of 44.7% year over year. Its backlog was 96.9% higher than a year earlier. Management identified data-centre activity as a driver of US sales and large-project orders. Hammond Power Solutions Q2 2026 results
These are company-reported results. Hammond also reported that backlog declined 6.9% sequentially from the first quarter and 3.1% from the fourth quarter of 2025 as shipments exceeded new orders. Margin performance reflected pricing, product mix, operating leverage and tariff effects.
The business may benefit from the structural trend without its shares necessarily offering attractive returns at every valuation. A strong industry thesis and an attractive stock purchase are separate propositions.
6. The surprising bottleneck: materials
In October 2026, Hitachi Energy announced a C$479 million investment in London, Ontario, including Canada's first domestic manufacturing line for transformer insulation pressboard. The company expects the new facility to be operational by 2028. Hitachi Energy, October 2026
The project connects an expanding electrical-infrastructure industry with something far removed from artificial intelligence: wood fibre.
Specialized cellulose materials are used in transformer insulation. Hitachi describes its pressboard as made from pure, unbleached cellulose taken from long-fibred coniferous wood. The material must meet specified mechanical and electrical requirements. Hitachi Energy technical description
Government materials say the London facility is expected to convert wood fibre into up to 20,000 tonnes of electrical insulation material annually. Government of Canada, October 16, 2026
Could parts of Canada's forestry and specialty-pulp industries eventually benefit from increased demand for electrical infrastructure? Possibly. But ordinary lumber, wood chips and commodity pulp cannot be assumed to qualify without the required fibre, processing and electrical properties.
The potential opportunity may therefore reside not in commodity fibre volume, but in specialized processing, manufacturing qualifications and established supply relationships.
The London investment does not establish material incremental demand for any particular Canadian forestry company. It does reveal how far upstream the consequences of electrical-infrastructure expansion can extend.
The same logic applies to electrical steel, copper conductors, specialty cable insulation, industrial enclosures and other components.
7. The smallest supplier may possess the hardest-to-replace capability
This section states a hypothesis to be tested, not a demonstrated conclusion.
Large infrastructure projects can depend on relatively small components supplied by a limited number of qualified manufacturers. A supplier may have spent decades developing production methods, quality controls, certifications and customer relationships. Replacing it could require engineering, testing and qualification work.
That creates the possibility of disproportionate economic value at unexpectedly small points in the supply chain. It also creates an important distinction between scarcity and pricing power.
A component may be essential without being profitable to manufacture. Customers may have alternatives, contractual pricing protections or the ability to redesign equipment.
The investment opportunity exists only when scarcity translates into durable incremental economics for the supplier.
Canadian companies under investigation
Enerflex (TSX:EFX) — Direct evidence of company-reported commercial demand for behind-the-meter natural-gas generation equipment.
Evolution PowerX Corp. (TSX:E; formerly Enterprise Group, Inc.) — Exposure to distributed and modular power generation. Future commercial orders, fleet economics and financing requirements remain essential to the investment case. The corporate name changed on August 4, 2026. Company name-change record
Hammond Power Solutions (TSX:HPS.A) — Reported transformer sales growth and substantial backlog, with valuation and capacity expansion central to the analysis.
Hammond Manufacturing (TSX:HMM.A) — Established electrical and electronic enclosure manufacturing capabilities. No issuer-specific data-centre order or revenue contribution is established in this paper.
Mattr (TSX:MATR) — Through AmerCable and Shawflex, Mattr sells specialized wire-and-cable products into industrial and infrastructure markets. Mattr reported higher North American data-centre wire-and-cable sales volumes in the second quarter of 2026 but did not separately quantify the data-centre contribution. Mattr Q2 2026 results AmerCable markets
These companies are research candidates, not recommendations. Their business models, financial risks, liquidity and valuations differ substantially. No common AI-infrastructure narrative should substitute for company-specific analysis.
8. The paradox: scarcity can destroy demand
There is an opposing interpretation that deserves equal scrutiny.
If electricity becomes too expensive, equipment lead times become too long or construction costs escalate beyond expected returns, developers may delay or cancel projects. The same shortage that supports supplier pricing power can undermine the customer's willingness or ability to spend.
Meanwhile, manufacturers are expanding capacity. If demand grows more slowly than expected, new factories could arrive just as customers reduce orders.
Technology may also change. More efficient inference, better hardware utilization, smaller models, different power-distribution architectures and alternative generation technologies could reduce demand for particular equipment categories.
A company can correctly identify an enormous future market and still invest in the wrong manufacturing capacity. Investors can correctly identify a growing business and still overpay for its shares.
This is why we must distinguish:
- Proposed capacity from funded projects.
- Procurement frameworks from binding purchase orders.
- Orders from delivered equipment.
- Backlog from recognized revenue.
- Revenue growth from sustainable free cash flow.
- Industry scarcity from shareholder returns.
The strongest evidence will come from order conversion, realized pricing, utilization, margins, cash generation and returns on invested capital.
9. What would prove or disprove this thesis?
Sorites Research is intended to preserve falsifiable claims, not rewrite history to accommodate later outcomes.
Baseline and measurement rules
The baseline is all qualifying evidence publicly known through October 23, 2026, including the Enerflex, Hammond Power Solutions, Mattr and Hitachi Energy evidence cited above. Baseline events cannot count as future successes.
The measurement window is October 24, 2026 through December 31, 2028.
Eligible evidence is a regulatory filing, audited or quarterly result, binding-award disclosure, utility or interconnection filing, permitting record, or named-project developer disclosure. Media may corroborate but will not be the sole evidence when a primary record exists.
One project, customer award or factory expansion counts once across revisions. A source must explicitly attribute an order, revenue result, delay or expansion to data-centre or AI electricity demand, or to an associated grid or electrical constraint.
The initial Canadian tracked universe is frozen as EFX, E, HPS.A, HMM.A and MATR. Later additions will be recorded separately and will not be backfilled into the initial basket.
Prospective tests
Prediction 1 — Supplier breadth
By December 31, 2028, at least three distinct equipment categories will produce a new post-baseline issuer-reported binding award or measurable revenue or order increase explicitly tied to data-centre demand.
Prediction 2 — Physical delays
By December 31, 2028, at least five named data-centre projects of 50 MW or greater will experience a delay of at least six months or cancellation, with electricity delivery, grid access or electrical equipment explicitly identified as a cause.
Prediction 3 — Upstream migration
By December 31, 2028, at least two new post-baseline manufacturing investments or acquisitions will explicitly identify shortage or lead-time reduction in a named electrical component or qualified material as a reason.
Prediction 4 — Canadian economic transmission
By December 31, 2028, at least two companies in the frozen Canadian basket will report an explicit data-centre demand linkage in two separate post-publication reporting periods. A pre-publication result or a one-time thematic statement will not qualify.
Measurable weakening conditions
The broader thesis will be considered materially weakened if, by December 31, 2028, either:
- a subsequent IEA central projection for 2030 global data-centre electricity demand falls at least 20% below the approximately 950 TWh baseline used here;
- fewer than two of the four prospective tests are satisfied; or
- reported order or revenue growth across the frozen Canadian basket fails to produce the repeated data-centre linkage required by Prediction 4.
The thesis can still prove directionally correct while individual investments fail because of dilution, excessive capital spending, competition, poor execution or valuation.
10. The investment question
Much of the investment world is concentrating on the companies designing models, manufacturing advanced chips and constructing data centres. Those businesses are important. But they are not the entire system.
Behind every new computing facility is a physical network of energy infrastructure, factories, materials, skilled workers and regulatory approvals.
Some of these industries are mature, overlooked and unglamorous. Some have spent decades operating in markets that offered limited growth. Their circumstances may now be changing.
The opportunity is not to buy every business that can plausibly be connected to AI. It is to identify which existing capabilities are becoming more valuable, which suppliers can capture that value, and which securities have not already priced in the outcome.
The defining constraint on artificial intelligence may not be how quickly we can create intelligence. It may be how quickly we can build the physical world required to sustain it.
Intelligence may scale exponentially. Infrastructure must still be manufactured, transported, installed, connected and operated.
Somewhere in that difference lies the next bottleneck. And possibly the next investment opportunity.
Disclosure and methodology
I am Drew Clement, founder of Sorites Research, which is founder-led. As of October 23, 2026, I hold no direct position in EFX, E, HPS.A, HMM.A or MATR, and I have not held a position in any of these companies to date. If I establish a position in a company covered by this paper, I will update the site's disclosures to identify that change.
I had no company contact connected with this paper. Sorites Research received no payment, sponsorship or other compensation from the companies covered in this paper.
I also founded and operate Archive Signal, a separate business that sources, structures and licenses proprietary datasets. Its buyers include AI developers, creating a commercial interest in AI spending growing. Any specific overlap with a company covered here must be disclosed before publication. This paper uses public sources only. Archive Signal's relationships are not independent validation of this thesis.
AI tools assisted research, source checking and editorial production. Drew Clement remains responsible for the published work; AI assistance is not independent verification.
This is research and opinion, not a recommendation to buy or sell a security or personalized investment advice. Scenarios and forecasts are uncertain. Company-specific financial and valuation analysis is required before any investment decision.
Corrections: [email protected]
Sorites Research | Foundational Thesis SR-002
Published October 23, 2026.