You're Building on Infrastructure You Don't Control — and 72% of AI Leaders Know It
Your company runs on a model someone else trained, on hardware someone else built, in a data center someone else operates, connected by cables someone else laid, powered by energy someone else controls, using minerals someone else extracted.
That is the actual infrastructure stack your AI product runs on. The model is layer seven. Your team owns layer seven. Layers one through six belong to other people.
Most AI architecture conversations start and end at layer seven. They debate which foundation model to use, whether to fine-tune or RAG, how to handle context windows, which orchestration framework to chain. These are real engineering decisions. They are also decisions made inside a box that nobody drew on the whiteboard.
Info-Tech Research Group's AI Trends 2026 report, drawn from over 700 global IT leaders, found that 72% cite data sovereignty and regulatory compliance as their top AI-related challenge — up from 49% the prior year. The same report found that only 19% have fully implemented AI governance frameworks. That is not a coincidence. The awareness is there. The gap is structural.
The Six Layers Nobody Draws
Start at the bottom and work up.
Minerals. GPU manufacturing depends on cobalt, tantalum, and rare earth elements. The Democratic Republic of Congo supplies roughly 70% of the world's cobalt. China processes about 60% of rare earths globally. Neither of these concentrations is a trade secret — they are geological facts that no policy can undo at meaningful speed.
Energy. Training a frontier model consumes electricity at the scale of a small town. Running inference at production volume is not much better. The International Energy Agency projected in early 2026 that data center electricity consumption would double by 2030. The constraint is not whether power exists — it is where it exists and what it costs. Your inference bill is partially an energy bill, and energy markets are volatile.
Compute hardware. NVIDIA controls approximately 80% of the AI accelerator market. TSMC manufactures the leading-edge chips that make this hardware possible and operates primarily in a geography that is the subject of ongoing geopolitical tension. This is not commentary on NVIDIA or TSMC — both are exceptional companies. It is a map of concentration that exists regardless of your feelings about it.
Data centers and cloud networks. AWS, Microsoft Azure, and Google Cloud collectively handle the majority of cloud AI workloads worldwide. Cloudflare, Equinix, and a small number of submarine cable operators own the network paths between them. When Cloudflare had a major routing incident in 2023, it knocked down significant portions of the internet's visible surface — not because of any fault with the application layer, but because of a dependency four layers below it.
Model weights. OpenAI, Anthropic, Google DeepMind, and Meta have trained the foundation models that most production AI systems either use directly or derive from. Even "open" weights like Llama were trained by one organisation and could have licensing terms changed, export restrictions imposed, or safety modifications applied that break existing integrations. This is not a hypothetical: OpenAI has changed its terms of service multiple times since 2020, each time with real consequences for production deployments.
APIs and inference services. The layer your team actually touches. The one you designed the product around. The one that can change pricing, rate limits, or deprecation schedules with 90 days notice.
Why Full Sovereignty Is Not an Engineering Option
The Brookings Institution published a detailed analysis in February 2026 titled "Is AI Sovereignty Possible? Balancing Autonomy and Interdependence." The conclusion is precise and worth quoting directly: full-stack AI sovereignty is "structurally infeasible" for almost any country, let alone any company, because AI is "a transnational stack with concentrated choke points across minerals, energy, compute hardware, networks, digital infrastructure, and other elements."
Notice that phrasing. Structurally infeasible. Not expensive. Not inadvisable. Not politically complicated. Structurally infeasible — meaning the constraint is physical and logistical before it is financial or regulatory.
McKinsey's 2026 analysis of sovereign AI ecosystems reaches a compatible conclusion from the enterprise direction: most organisations that list sovereign AI on their roadmaps for 2026 "have a detailed strategy, action plan, budgets, and workload tiering" — but the migrations themselves take three to four years, and the bottleneck is "not primarily technology limitations but the organisational work required to move regulated workloads."
This is the gap that explains NTT DATA's numbers. Their 2026 Global AI Report surveyed nearly 5,000 senior executives across more than 30 markets. Ninety-five percent said private and sovereign AI are important. Only 29% are actively prioritising sovereign AI implementation in the near term. Nearly 60% of AI leaders identified cross-border data restrictions as a major challenge. And only 38% report high confidence in their cloud security posture — the foundation everything else rests on.
That spread between 95% and 29% looks like apathy from the outside. It is not. It is a rational response to understanding what implementation actually requires.
The Honest Architecture Question
The question that matters is not "how do we achieve full sovereignty?" It is: "which specific dependencies carry which specific risks, and what is the minimum viable mitigation for each?"
These are different questions. The first leads to paralysis or expensive theatre. The second leads to actual engineering decisions.
A Zapier survey published in early 2026 found that 81% of enterprise leaders are concerned about AI vendor dependency, and 47% report that at least one key business function would stop working if their primary AI vendor experienced significant downtime or a major policy change. Only 6% say they could switch AI vendors without material disruption. Forty-seven percent and six percent in the same answer. The exposure is understood. The preparation is not.
Brookings calls the alternative to full sovereignty "managed interdependence" — an approach that maps dependencies by layer, prioritises feasible interventions, diversifies suppliers and partners, and embeds interoperability and portability through technical standards and procurement decisions. It is the engineering equivalent of fault tolerance: you do not eliminate failure modes, you isolate them so they do not cascade.
In practice, this looks like a set of specific decisions. Using multiple inference providers with automated failover rather than hardcoding to a single API. Maintaining contracts with two cloud providers even if one is primary. Ensuring that fine-tuned weights are stored in infrastructure you control rather than within a provider's managed service. Tracking which of your business functions are single-threaded through a single external dependency and treating that as a risk item, not a vendor relationship item.
None of this is glamorous. None of it ships a feature. It also does not require you to build your own GPU fab.
What the 72% Number Actually Means
There is a version of the data sovereignty conversation that is almost entirely about regulation — GDPR, EU AI Act, DPDP in India, state-level US laws. That conversation is real and matters for compliance. It is not the same conversation.
The structural dependency problem exists regardless of your regulatory environment. A company operating entirely within one jurisdiction, subject to none of the cross-border data restrictions that 60% of AI leaders cite as a challenge, still runs on compute it does not own, through networks it does not control, using models it did not train. The regulatory layer adds urgency. The structural layer was already there.
The gap between 72% citing data sovereignty as a top challenge and 19% having implemented governance frameworks is not explained by ignorance. The NTT DATA report is explicit that most organisations understand the problem. What they lack is a governance framework that treats infrastructure dependency as a first-class engineering concern rather than a compliance checkbox.
The architecture documents in most organisations have a section on model selection, a section on data pipelines, a section on security. They rarely have a section called "dependency map" that traces every external control point from the API layer down to the physical substrate. That document does not exist in most shops. Building it is not the same as achieving sovereignty. But it is the prerequisite for any honest conversation about risk.
There is a version of this that organisations are already navigating poorly with AI governance at the governance layer — but governance frameworks that only address the model layer are missing the deeper problem. The stack goes further down than the terms of service.
The Honest Close
The question is not whether you control your AI infrastructure. You do not. The question is whether you know precisely which parts you do not control, what would happen if each of those parts changed terms, and whether you have designed for that failure mode or just hoped it would not come up.
Hope is not an architecture decision.
What specific layer of your stack would you be most exposed by if the vendor changed pricing by 40% next quarter?
Sources: NTT DATA 2026 Global AI Report · Info-Tech Research Group AI Trends 2026 · Brookings — Is AI Sovereignty Possible? · McKinsey — Sovereign AI Ecosystems · OpenMetal — Why Enterprise AI Is Hitting an Infrastructure Wall in 2026 · Photo by Brett Sayles via Pexels