AI Strategy - INFRASTRUCTURE IS THE NEW CONSTRAINT
- Erik Kling
- Jun 2
- 6 min read
By Erik Kling | AXISYNC Decision Architecture™
For two decades, technology strategy was primarily a software conversation. Who had the best algorithm, the most elegant code, the fastest iteration cycle. The dominant narrative was that physics was a commodity — that energy, hardware, and connectivity were inputs you purchased, not variables you needed to architect around.
That era is ending.
The constraint has shifted from the virtual to the physical, from code to concrete, copper, silicon, and electricity. And most organizations — and most nations — are not yet thinking at the layer where the real decisions are now being made.
The Five Physical Constraints of the AI Era
The next decade of AI development will be shaped not by which company ships the best model, but by five physical constraints that cannot be resolved by writing better software:
Energy.
Compute.
Semiconductors.
Connectivity.
Governance.
Each deserves to be understood not as a background condition but as a strategic variable — one that will determine who gains leverage, who loses optionality, and who ends up dependent on someone else's infrastructure decisions.
Energy: The Binding Constraint
The scale of what is arriving is difficult to fully absorb.
The IEA now projects global data center electricity consumption will reach approximately 1,100 TWh in 2026 — equivalent to Japan's entire national energy consumption. Goldman Sachs Research forecasts a 160–165% increase in power demand from data centers by 2030. Deloitte estimates US AI data center power demand could grow more than thirtyfold by 2035.
The nuclear dimension makes the argument concrete in a way no projection can. Microsoft signed a 20-year power purchase agreement to restart Three Mile Island Unit 1 — a shuttered nuclear facility in Pennsylvania — bringing it back online in late 2024 to deliver 835 MW of carbon-free baseload power to their regional operations. Google and Amazon have followed with their own nuclear agreements.
When hyperscalers restart nuclear power plants to operate their compute infrastructure, energy has moved from background assumption to primary strategic variable.
Everything else in this section is supporting evidence.
That single fact is the argument.
Semiconductors: The Chokepoint Beneath the Chip
Most discussions about AI begin at the model layer. But control does not reside at the model layer.
The true constraints are not in chip design. They are in High Bandwidth Memory and advanced packaging — the less visible layers of the stack where physical limits bite hardest. Even when leading-edge logic capacity exists, packaging throughput and HBM availability cap how many AI accelerators can actually ship. The backend of semiconductor manufacturing has become a first-order growth limiter for the entire AI industry.
This is the pattern Mumford identified in 1934: the constraint is never where visibility is highest. It is in the infrastructure layers that most people do not think to examine.
Governance: The Invisible Architecture
Governance may be the most underestimated of the five constraints, precisely because it operates beneath the level of product and market discussion.
Since 2022, the United States has implemented unprecedented restrictions on chip exports and semiconductor manufacturing equipment — expanded through the Biden administration's AI Diffusion Framework in January 2025, and actively renegotiated since. These controls have imposed license requirements on shipments of advanced AI chips, high-bandwidth memory, and semiconductor manufacturing equipment to over 100 countries, effectively creating a tiered global map of who can access which layers of AI infrastructure.
Governance decisions made in 2024 and 2025 will shape the infrastructure landscape for a decade. That is not a policy observation. It is an architectural one.
What Mumford Understood in 1934
Lewis Mumford published Technics and Civilization in 1934. His central argument was not about machines. It was about the moral, economic, and political choices embedded in the design of technical systems — choices that become invisible once infrastructure is built, and that constrain all future options in ways that later actors cannot easily reverse.
Mumford's most provocative claim was that the key invention of the modern industrial age was not the steam engine. It was the mechanical clock. Not because the clock was more powerful, but because it restructured time itself — and in restructuring time, it restructured what was possible, what was efficient, what was profitable, and ultimately what was human.
The visible technology was subordinate to the invisible architecture.
The same logic applies to AI infrastructure today. The models are the visible technology.
The power grid agreements, the HBM supply chains, the chip packaging throughput, the export control regimes — these are the invisible architecture.
And it is the invisible architecture that will determine the range of choices available to every organization and nation that wants to participate in the next technological epoch.
Heidegger pressed the same question from a different angle. In The Question Concerning Technology, he argued that the essence of modern technology is not any particular machine but what he called Enframing — the tendency of technological systems to reveal the world as standing-reserve, as material to be ordered, optimized, and consumed. The danger was not the technology itself but the way Enframing progressively narrows the space within which humans can exercise genuine choice.
Applied to the present moment: organizations and nations that allow AI infrastructure decisions to be made by others — that treat infrastructure as a commodity to be purchased rather than an architecture to be designed — are ceding their capacity for independent action. Not dramatically, not visibly, but structurally. One dependency at a time.
Most leaders view AI from the application layer downward. Architecture requires reading the system in the opposite direction: from energy upward. Intelligence is not the beginning of the chain. It is the consequence of it.
The Stack That Most Executives Have Never Seen
At the top: models and applications. This is where AI becomes visible. The interface layer that captures attention and generates the headlines.
Beneath that: compute systems. The GPU clusters and server farms that run the models at scale.
Beneath that: chip architecture. The design decisions that determine computational efficiency, power consumption, and capability.
Beneath that: manufacturing. The foundries — primarily TSMC in Taiwan — where chips are physically fabricated.
Beneath that: equipment. The machines that manufacture the chips. Specifically, ASML's extreme ultraviolet lithography systems, without which advanced chip fabrication at leading nodes is impossible. One Dutch company sits at a chokepoint of the entire global AI capability stack.
Beneath that: energy. The electricity, the grid infrastructure, the generation agreements that determine whether any of the above can actually run.
As you move down this stack, visibility decreases. Control increases.
The most important strategic question is not: which model is best?
It is: which layer of this stack do you control, which do you depend on, and what are the terms of that dependency?
The Decision Organizations Are Currently Making — Without Knowing It
Most organizations engaging with AI are making infrastructure decisions by default rather than by design.
They are selecting cloud providers based on familiarity. Choosing chip platforms based on availability. Accepting energy and connectivity constraints as background conditions. Treating governance and export control exposure as legal problems rather than strategic variables.
These are infrastructure decisions. And infrastructure decisions have the peculiar property that their consequences become visible only after the optionality they foreclosed is no longer recoverable.
The organizations that understand what is happening at the infrastructure layer — that can read the stack below the model layer, assess their dependencies, and design their architecture accordingly — will have a durable strategic advantage over those that remain fixed at the application layer.
This is not an argument against using AI. It is an argument for thinking architecturally about AI — which requires going significantly deeper than most current conversations go.
The constraint has shifted. The decisions being made now, in supply chains and energy agreements and chip fabrication alliances and governance frameworks, will shape the range of choices available in 2030 and beyond.
Those decisions are happening whether or not you are at the table.
Architecture determines optionality.
Optionality determines leverage.
Leverage determines control.
This is not a theoretical proposition. It is a description of what is already happening at the infrastructure layer of the global AI economy — visible to those who know where to look, invisible to those who remain fixed at the application layer.
The question for every organization engaged with AI is a simple one: at which layer of the stack are you operating, and which layers are operating on you?
Erik Kling is the author of AXISYNC Decision Architecture™ and RHODES DOCTRINE™, and the founder of AXISYNC Partners LLC. This piece is part of a series on the structural dynamics shaping the next decade of technology, governance, and civilizational architecture.
Comments