THE AI CHIP POWER STACK: ARCHITECTURE OF CONTROL
By Erik Kling | AXISYNC Decision Architecture™
There is a question that almost no one in the AI industry is asking correctly.
The question most organizations ask is:
which AI model should we use?
Which cloud provider?
Which GPU vendor?
These are application layer questions. They are not unimportant. But they are the wrong starting point for any organization that wants to understand where AI capability actually comes from — and more critically, where control over that capability actually resides.
The right question is structural: at which layer of the AI power stack does control concentrate, and what are the implications of operating above that layer?
The Stack Most Executives Never See
See Visual: AI Chip Power Stack — Architecture of Control
AI infrastructure is not a single system. It is a layered architecture, and each layer has a different relationship between visibility and control.
At the top sits the application and model layer. ChatGPT, Gemini, Copilot, Claude. The products that most people interact with, and that generate the headlines. This layer is highly visible and, paradoxically, the least structurally powerful. Models are increasingly interchangeable. Application interfaces are easily replicated. The organizations operating only at this layer are, in the language of supply chain theory, price takers rather than price setters.
Beneath that: compute systems — the hyperscale data center infrastructure operated by Microsoft, Google, Amazon, and Meta. This layer owns the operational scale of AI.
Running a frontier model at commercial volume requires infrastructure investment measured in tens of billions of dollars annually. This is where the hyperscalers have built their first layer of structural advantage: capital deployment so large that meaningful competition requires sovereign-level commitment.
Beneath that: AI chip design and architecture. This is where the standards are set. NVIDIA, AMD, Google TPU, Amazon Trainium, Microsoft Maia. The chip architecture layer determines computational efficiency, software ecosystem compatibility, and ultimately what AI workloads are economically viable at scale.
Beneath that: manufacturing. The physical fabrication of chips at advanced process nodes. TSMC dominates this layer with a concentration that has no parallel in any other critical industry. Every chip iteration from NVIDIA depends on TSMC's process technology. TSMC's advanced nodes — 3nm, 5nm — account for 74% of its wafer revenue, and AI accelerator revenue is forecast to grow at a compound annual rate of roughly 50% through 2029.
Beneath that: equipment. The machines that make the machines. ASML's extreme ultraviolet lithography systems sit here — a subject Post 3 addresses in full, because the ASML position is the most structurally significant chokepoint in the entire global AI capability stack.
At the foundation: energy. The ultimate limiter, as Post 1 established. Intelligence is constrained by compute. Compute is constrained by energy.
NVIDIA: The Software Moat Hiding Inside the Hardware Story
NVIDIA is widely understood as a chip company. That understanding misses the most strategically important dimension of its position.
NVIDIA's own filings frame competitive dynamics as platform-versus-platform, not chip-versus-chip. In its FY2026 annual report, NVIDIA frames its competitive differentiation around the integrated combination of hardware, software ecosystem, networking, and developer tooling. Alphastreet
The core of that platform is CUDA — a proprietary software architecture and programming model that NVIDIA has developed over nearly two decades. NVIDIA's CUDA platform is effectively the standard gauge of the AI development world. A competitor might design a more powerful chip, but if it cannot run on the global CUDA network — the millions of lines of code, libraries, and accumulated knowledge developers have already created — its hardware advantage is severely diminished. Aminext
NVIDIA's moat is systems-level: chips, interconnect, software, benchmarks, and partner ecosystem. The effect is a virtuous cycle: best silicon leads to best-benchmarked systems, which leads to more customers, which leads to more feedback and data to tune compilers and libraries. Essentialbizmarketing
This matters enormously for anyone thinking about where leverage concentrates in the AI stack. NVIDIA does not merely sell chips. It sells the standard that the rest of the ecosystem builds around. NVIDIA CEO Jensen Huang estimates between $3 and $4 trillion will be spent on AI infrastructure by the end of the decade, with vendor lock-in through CUDA forcing teams to pay premium prices across the AI supply chain. Built In
That is not a technology observation. It is an architectural one. The organization that sets the standard controls the terms under which everyone else operates.
The Hyperscaler Response: Silicon Sovereignty
The largest technology companies understand this dynamic precisely — which is why they are all building their own chips.
Google spent eight years and $13 billion developing TPU architecture because its core algorithms represent 80% of its compute needs. Amazon's Trainium follows similar logic — Trainium chips cost 50% less than equivalent NVIDIA hardware while delivering comparable performance for large language models. FourWeekMBA
Custom ASICs from Google, Microsoft, Amazon, and Meta are growing at a 44.6% CAGR, targeting the inference workloads that now represent two-thirds of all AI compute. Introl
The strategic intent behind custom silicon is not primarily about cost, though cost matters. It is about controlling the stack. The hyperscaler that controls its own silicon controls its own performance roadmap, its own cost structure, and its own supply chain — three things that no amount of NVIDIA procurement can deliver. Oplexa
The dual-track strategy now visible across all major hyperscalers — deploying both custom ASICs for predictable inference workloads and NVIDIA GPUs for flexible training
— is not a transitional arrangement. It is an architectural choice to maintain optionality across both layers simultaneously. This is exactly what organizations that understand stack dynamics do: they avoid single-layer dependency by building redundant control at the layers that matter most.
NVIDIA has responded with characteristic intelligence. NVIDIA's NVLink Fusion strategy lets hyperscalers plug custom ASICs into NVIDIA's rack architecture — ensuring NVIDIA stays embedded even when it isn't the primary compute chip. The company is repositioning from chip supplier to rack-scale infrastructure provider, extending its footprint down the stack even as the hyperscalers build upward. It is a move that would have made Mumford recognize the pattern immediately: the organism of control adapts to preserve its position, migrating from the contested layer to the layer beneath it. Siliconanalysts
What the Stack Tells You About Strategy
The AI chip power stack is not a technology diagram. It is a map of where leverage concentrates and where dependency is created.
Every organization engaging with AI is operating at some layer of this stack — and accepting the terms set by the layers beneath it. Most organizations are operating at the application layer and accepting the terms of every layer below. That is not inherently wrong. But it should be a conscious decision, made with full awareness of what is being accepted.
The organizations building durable AI capability — whether corporations or nations — are asking a different set of questions. Not which model is best, but which layers they need to influence or control to preserve meaningful strategic optionality. Not which cloud is cheapest today, but what their dependency profile looks like across the full stack over a five-to-ten-year horizon.
Platform dominance is reinforced not only by technology, but by habit and institutional inertia. The decisions made now about which layers to engage at, which vendors to standardize on, and which ecosystems to build within will create the dependency structures that shape strategic choices a decade from now. Medium
The stack is not neutral. Every layer carries embedded choices about who controls what, on whose terms, and at whose price. The question is whether those choices are made consciously — or by default.
The question is no longer whether the stack matters. Washington and Brussels answered that this week. The question is which architecture will govern it — and on whose terms.
The Layer Question
Post 3 takes this analysis to its sharpest point: ASML.
One company. One layer. One chokepoint that the entire global AI capability stack passes through — and a paradox that reveals everything about the gap between holding a critical asset and controlling an architecture.
The pattern is the same.
Only the geography changes.
Architecture determines optionality.
Optionality determines leverage.
Leverage determines control.
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