What Happens When AI Eliminates the Bottleneck That Protects Your Margins?

AI business model disruption may not begin when a competitor uses artificial intelligence to operate your existing model more efficiently. It may begin when AI helps eliminate the bottleneck that made your model profitable in the first place. The broader operating consequence is captured in Bottleneck Migration: when AI weakens one constraint, scarcity—and therefore value—moves somewhere else.

That possibility sits inside a genuine industrial paradox. AI is increasing demand for electricity, advanced chips, cooling systems, grid connections, construction capacity, capital and specialist talent. The International Energy Agency projects that global data-centre electricity consumption could more than double to around 945 terawatt-hours by 2030, with accelerated servers driven primarily by AI accounting for almost half of the increase. The scarcity is not theoretical. It is appearing in power-density constraints, transformer lead times, interconnection queues, semiconductor supply and the financing required to build infrastructure at unprecedented scale.

It is therefore reasonable to question any easy promise that AI will create abundance. The first visible effect is often the opposite: concentrated demand enters an inherited system that was not designed for it. Yet the same technology is also being applied to plasma control, materials discovery, grid optimization, energy storage and computing architecture. AI may intensify scarcity in the first order while accelerating the research, design and coordination needed to make some of those scarcities less binding in the second.

A profitable bottleneck can make a temporary business model look permanent.

Scarcity rent is rational, but scoping the AI business model disruption around it is dangerous.

Economists use the term scarcity rent for value captured because access to a constrained resource, capability or channel is limited. There is nothing inherently illegitimate about earning it. A chip manufacturer that can supply scarce accelerators, a cloud provider that can offer capacity immediately, or a utility able to connect a new data centre has every right to charge what the market will bear. Scarcity signals where capacity is valuable and directs investment toward expanding it. Heck, Tesla quickly rose to become a trillion dollar company after offering a product that all other car manufacturers had neglected for decades, by introducing the novelty of the scarcely available “compelling” electric vehicle. Manufacturers with much larger delivery capacity picked up the pace, but it is undeniable that Tesla’s rise was partially due to a lack of alternatives.

The strategic danger appears when an organization confuses the price of access to a bottleneck with the permanent value of the company. A constraint may protect margins for years. It may fund research, infrastructure and exceptional returns. But if the constraint weakens, the premium attached to controlling it can collapse much faster than the organization can redesign itself.

This is why the most important question is not whether current AI infrastructure companies deserve their profits. They do not need to apologize for meeting urgent demand. The question is whether their capabilities remain distinctive when compute becomes cheaper, energy becomes more distributed, thermal constraints change, or customers no longer need to queue for the same scarce resource. The strongest businesses will not merely defend the bottleneck. They will use the capital, knowledge, customer relationships and infrastructure accumulated during scarcity to shape what follows.

Companies preserve solutions long after the problem begins to decay

Organizations often describe transformation as a contest between an incumbent solution and a better version of that solution. That framing is comforting because it leaves the original problem intact. A bank can improve the branch. A retailer can improve the store. A media company can improve the physical catalogue. A manufacturer can improve the efficiency of the established product. The organization remains recognizable; only its performance changes.

Some transformations are more destructive. They do not improve the incumbent answer. They make the original question less important. At the beginning of the twentieth century, fast-growing cities faced a literal horseshit problem. An urban service company could have become exceptionally efficient at collecting manure, optimizing routes and processing waste, and still watched its market contract as motor vehicles changed the transport system producing the problem. The solution did not fail. The problem decayed.

This pattern is easy to recognize in retrospect and difficult to act on in the present. Once a solution works, companies organize around it. Assets are purchased, specialists are hired, incentives are attached to revenue, governance is designed around delivery, and the solution becomes part of corporate identity. The organization becomes increasingly capable of answering a question that the market may be preparing to stop asking.

Kodak did not simply fail to notice digital photography

The familiar Kodak story is often reduced to an incumbent that invented a digital camera and then ignored it. The historical record is more instructive. Kodak introduced a professional digital camera system in 1991, launched a consumer digital camera in 1995 and later introduced the EasyShare system. Its own corporate history documents those developments. By 2008, Kodak’s portfolio included digital cameras, online imaging, retail kiosks, printers, workflow software and digital production systems alongside film, paper, chemicals and photofinishing, as described in its SEC filings.

Kodak saw digital technology. It possessed imaging science, brand recognition, patents, distribution and decades of customer understanding. The deeper failure was not awareness but economic transfer: reassembling those capabilities around a world in which image capture no longer depended on film consumption and chemical processing. Canon and Nikon demonstrate that camera companies could survive the transition. The outcome was not technologically predetermined. The strategic challenge was deciding which assets belonged to photography as a customer outcome and which belonged only to the film-based scarcity system.

Blockbuster solved a bundle of problems that streaming dissolved

Blockbuster offers a different and richer example. Its stores did more than transfer a cassette or disc for a fee. They gave customers access to recent films, confidence that popular titles might be available, a place to browse, recommendations, reservations, snacks and a ritual around choosing the evening’s entertainment. Those were real customer problems, and Blockbuster built an extensive operating model to solve them.

Streaming did not build a friendlier rental shop. It progressively made the rental shop an indirect answer to the customer’s actual goal: watching something desirable now. Netflix’s 2012 annual report described subscribers instantly watching an unlimited catalogue across televisions, computers and mobile devices while its DVD service still existed in parallel. The transition absorbed or invalidated much of the physical problem bundle: travel, inventory, reservations, returns and per-rental transactions, rather than merely performing each element better. Blockbuster’s history had multiple causes, but the strategic pattern is clear: a company can optimize a successful solution while the market removes the need for the solution’s defining machinery.

This distinction matters for the shape of the AI-era firm. A company should not ask only how AI can make the current workflow faster. It should ask whether the underlying constraint still deserves to organize the value chain. Transformation does not always improve the incumbent solution. Sometimes it dissolves the problem the solution was designed to solve.

The lighting lesson: stop cooling the wrong architecture

In 2008, during my first internship in Germany, I worked at a studio-lighting company just as LED technology was becoming commercially relevant. LEDs were still expensive. Getting closer to the technology coincided with moving into a new apartment in Munich, and it influenced my decision to invest in LED lighting there.

What stayed with me was the underlying inefficiency of the incumbent design. A conventional incandescent lamp releases roughly 90% of its energy as heat, according to the US Department of Energy. The immediate engineering response could have been better ventilation, stronger air conditioning, more heat-resistant fixtures and studios designed to tolerate the waste. LED technology asked a more consequential question: why should producing light require generating so much unwanted heat in the first place?

The transition did not merely improve the old solution. It changed the architecture producing the desired outcome. Equivalent illumination could be created with a fraction of the electricity, and much less energy had to be managed as waste heat. The lighting industry moved from treating heat as an inevitable cost of illumination toward a system in which far more of the input served the intended purpose.

Are we treating heat as an unavoidable cost of computation because we are still optimizing the wrong architecture?

That question now hangs over AI infrastructure. Data-centre operators must solve today’s thermal problems, and better liquid cooling, immersion systems, data-centers in outer space, heat reuse and facility design are valuable. But those responses should not freeze the current relationship between useful computation, electricity consumption and waste heat into a permanent assumption. Photonic interconnects, workload-specific accelerators, alternative semiconductor materials and lower-power architectures may change the thermal profile of computation itself. None is guaranteed to replace conventional computing or eliminate cooling. The point is narrower: today’s heat profile is a design outcome, not a law of nature.

Companies that earn scarcity rent from present architectures should therefore invest in both horizons. One horizon expands supply and manages the immediate bottleneck. The other asks what happens when the bottleneck is redesigned. This is the same operating-model distinction developed in Redesigning Work With AI Starts When I Stop Trying to Use AI: inserting intelligence into the inherited system is less consequential than reconsidering the system around a changed constraint.

Three industrial pathways from scarcity toward abundance

The case for abundance should not rest on a catalogue of speculative breakthroughs. It should rest on mechanisms through which AI can lower the cost of search, simulation, control, experimentation and coordination in technically difficult domains. Three pathways are sufficiently concrete to examine without pretending their outcomes or timelines are known.

1. Fusion, distributed energy and the changing role of the grid

Fusion remains an engineering and commercialization challenge of extraordinary complexity. AI does not repeal plasma physics, manufacture reactor materials or guarantee an economic power plant. It can, however, improve specific parts of the development system. Researchers supported by the US Department of Energy have integrated machine learning into real-time tokamak control, adjusting magnetic-field perturbations in response to changing plasma conditions. A separate deep-reinforcement-learning system used sensor inputs and simulations to adjust confinement fields and avoid tearing instabilities. The demonstrated contribution is better adaptive control and access to experimental states, not a promised date for commercial fusion.

The credible abundance thesis is therefore cumulative. AI may help researchers explore more configurations, identify promising materials, improve diagnostics, preserve learning between experiments and control systems in real time. Each gain can shorten part of the search process without resolving every remaining dependency. If fusion eventually becomes commercially viable, its effects would extend beyond powering AI data-centers. It could alter the scarcity assumptions embedded in energy-intensive manufacturing, desalination, transport fuels and industrial heat.

Fusion is not the only route to a less constrained energy system. Solar generation, batteries, electric vehicles, controllable loads and microgrids are already changing the relationship between producers and consumers. The National Renewable Energy Laboratory’s OptGrid work coordinates distributed energy resources at the grid edge so that devices can balance supply and demand, support reliability and improve resilience. The future utility may create less value by selling every unit of centrally generated electricity and more by orchestrating millions of producers, storage assets and flexible loads.

Our grandchildren may find it strange that most households were treated only as electricity consumers, rather than as participants in a distributed energy system. That does not mean the grid disappears. A more distributed system may make grid coordination, balancing, cybersecurity and resilience more important. The business model changes because the utility’s durable value moves from exclusive supply toward orchestration, trust and system stability.

2. Computing architectures that generate less waste

The present response to AI’s power density is appropriately practical: secure more electricity, build facilities, improve cooling and place infrastructure where grids can support it. Yet the long-term economic opportunity may be architectural. Accelerated computing can become more specialized. Memory movement can be reduced. Interconnects can become more efficient. Photonics may carry or process some information with lower energy requirements. Software can route workloads according to latency, carbon intensity, heat and available capacity.

The lighting analogy matters because it prevents a category error: better cooling is valuable in the same way better ventilation was valuable in a studio full of incandescent lamps. It solves the immediate operating problem. It does not prove that the underlying conversion architecture should remain unchanged. The companies that lead the next era may be those that can translate today’s semiconductor expertise, fabrication relationships, software ecosystems and customer access into systems that deliver more useful computation per unit of scarce infrastructure.

This also changes how executives should interpret the current chip boom. Scarcity can produce exceptional economics, but extraordinary scarcity rent may conceal the difference between temporary market power and durable platform value. Durable value is more likely to reside in the capability to keep improving performance, energy efficiency, developer access and system integration after supply catches up. The bottleneck may move from chips to power, from power to interconnection, from interconnection to data, and from data to human validation. Strategy must follow the movement rather than defend the last profitable constraint.

3. Batteries, materials discovery and constraints that stop mattering

Battery-electric vehicles are often judged against a century of optimization in combustion technology and fuel infrastructure. Early comparisons make range, charging speed, cost, weight and cold-weather performance look like permanent categorical disadvantages. Some constraints will remain physical. Others may be temporary features of an immature chemistry, manufacturing system or charging network.

The Department of Energy’s work on AI-enabled materials discovery describes closed-loop systems connecting prediction, synthesis, characterization and analysis. The credible promise is not a magical battery generated by a model. It is a shorter and better-informed search across materials, electrolytes, interfaces, manufacturing conditions and failure modes. AI can help decide which experiments deserve physical execution and make the learning from those experiments easier to reuse.

If energy density, charging, lifecycle, safety and grid integration improve together, range anxiety may not be solved by pursuing an absurd number on a dashboard. It may become strategically irrelevant because the combination of vehicle capability, charging availability and energy management no longer interrupts the customer’s intended journey. The market then stops rewarding the company that merely manages today’s anxiety and starts rewarding the company whose assets help make the anxiety obsolete.

Abundance is not automatic

The transition from cognitive leverage to physical abundance passes through organizations, capital markets, infrastructure and public institutions. A model can identify a better material without producing it at scale. A control system can improve a plasma experiment without resolving reactor maintenance. A grid optimizer can coordinate distributed devices only when regulation, standards, incentives and ownership permit them to participate. Technical possibility is not operating capability.

Efficiency can also increase total demand. When a task becomes cheaper, more people perform it and new applications become viable. This rebound effect is visible across computing history and is likely to remain relevant for AI. More efficient inference may reduce energy per task while the number and complexity of tasks grow faster. Abundance in one layer can therefore create scarcity in another.

Capital concentration creates another risk. The companies with the greatest access to compute and infrastructure may use their advantage to outbid others for scarce resources rather than expand broadly available capacity. Research breakthroughs can remain proprietary. Public grids may carry costs created by private demand. Regions with power and water may capture investment while others fall further behind. The fact that AI can accelerate a solution does not determine who owns it or who benefits.

This is why the argument should make both AI absolutists uncomfortable: the reflexive critic is wrong to assume that present resource consumption is the final state of the system, the indiscriminate optimist is wrong to assume that intelligence automatically converts into abundant energy, materials or infrastructure. The net effect is empirical, conditional and shaped by operating-model choices.

The earlier analysis of the AI Resource Paradox established a smaller mechanism: AI can release cognitive capacity by lowering the cost of preserving, structuring and resuming complex work. The industrial question is what that released capacity builds. More analysis and more options are not abundance. They become productive capacity only when organizations can validate, fund, execute and retain the learning.

From defending the bottleneck to transferring value

Executives should not respond by abandoning every profitable scarcity-based business. That would be theatrical rather than strategic. Current demand must still be served. Infrastructure must still be financed. Customers will continue paying for access, reliability and performance while constraints remain real. The more disciplined response is to separate the value captured from the bottleneck from the capabilities that can travel beyond it.

A utility may possess grid expertise, regulatory relationships, trust, field operations and control systems that remain valuable when customers become producers. A semiconductor company may possess architecture, fabrication knowledge, software ecosystems and developer relationships that remain valuable when individual chips become less scarce. An energy-storage company may possess materials data, manufacturing discipline and safety capability that remain valuable when today’s chemistry is replaced. A cloud provider may possess orchestration, security and reliability capabilities that remain valuable when raw compute becomes cheaper.

The transfer is not automatic. Kodak possessed significant digital capability yet struggled to translate enough of it into an equally powerful economic position. Blockbuster possessed brand, locations, customer knowledge and distribution but did not reassemble those assets around immediate access before the physical rental system lost relevance. Assets survive on a balance sheet; capabilities survive only when the organization can apply them to a changed customer outcome.

This review should sit alongside conventional market and technology planning. Most scenario exercises ask what happens if demand rises, a competitor enters or regulation changes. An abundance scenario asks a more uncomfortable question: what happens if the constraint protecting the business weakens by 90%? What would customers still pay for? Which assets become stranded? Which capabilities become more valuable because they can accelerate the new system?

The answer may reveal that a company has been defining itself too narrowly. A utility is not necessarily an electricity seller. It may be a trusted orchestrator of energy flows. A chip company is not necessarily a supplier of scarce silicon. It may be a platform for useful computation. A vehicle manufacturer is not necessarily a producer of machines optimized around fuel scarcity. It may be an integrator of mobility, storage and grid participation. The more transferable definition is usually closer to the customer outcome and further from the incumbent mechanism.

Build for the world after the constraint

AI is already inside a resource paradox. It consumes scarce physical capacity while releasing new forms of cognitive capacity. It drives demand for power and cooling while helping researchers model plasmas, coordinate distributed grids, search materials and rethink computing architectures. Neither side cancels the other. The relevant question is whether institutions can convert the second-order capability into systems that eventually relieve the first-order pressure.

For executives, the harder implication is that current scarcity may be both an opportunity and a warning. It can produce exceptional margins, justify investment and reveal where customers are desperate for capacity. It can also encourage an organization to defend the constraint because the constraint has become economically comfortable. The moment of highest profitability may be the moment when the successor model most needs funding.

Kodak and Blockbuster are useful not because their histories can be reduced to simple morality tales, but because they expose the difference between recognizing a technology and transferring an organization. The lighting transition shows the physical equivalent: sometimes the breakthrough is not better management of the waste created by the old architecture. It is an architecture that produces far less waste in the first place.

AI may not create abundance. It may amplify demand, concentration and environmental pressure faster than industry can respond. But dismissing its resource appetite as proof of permanent scarcity is as intellectually weak as declaring every research demonstration the beginning of a utopia. Serious strategy must hold both possibilities at once, and prepare the company for the possibility that the bottleneck moves.

The defining leadership task is learning to build value for a world in which the constraint that created your company no longer matters.

The immediate question is therefore not simply how much value your company can extract from today’s scarcity. It is whether the capabilities, relationships, data, infrastructure and judgment built during that scarcity can help create the abundance that follows. Companies that answer that question early may not merely survive the disappearance of the bottleneck. They may be the ones that eliminate it.