What if the AI infrastructure build-out is NOT overshooting demand — what demand scenarios justify current investment
Is the AI Building Boom Real, or Are We Just Building for Ourselves?
Based on analysis of a 91-node, 305-edge knowledge graph exploring demand scenarios that would justify current AI infrastructure investment.
What Are We Even Talking About?
Right now, the biggest technology companies in the world are spending hundreds of billions of dollars building data centers, buying computer chips, and signing decades-long contracts for electricity. The question this analysis tries to answer is simple: is there enough real demand for AI to justify all of that construction — or are we in a bubble?
Think of it like a city deciding to build a hundred new highways. That only makes sense if millions of new cars are coming. If the cars never show up, the highways sit empty and the money is wasted. The knowledge graph this analysis is based on maps out all the different arguments for why the cars are coming — and a few reasons to worry they might not be.
The Most Surprising Finding: The Most-Connected Idea Has the Lowest Score
Imagine you ran a survey about whether a party would be crowded. The answer “yes, it will be crowded” shows up on nearly every page of the survey — but whenever someone rates how confident they are in that specific answer, they give it a one out of ten.
That is almost exactly what happened here. The concept called the “Inference Jevons Paradox” is connected to 22 other ideas in this graph — more than almost anything else — but it was rated with the lowest possible importance score. The same is true for “AI Power Demand Constraint,” which is connected to 20 things but also scored at the bottom.
Why does this matter? It suggests these two ideas were recognized as central — everything else points to them — but whoever built this map was not confident in what they actually mean. They are the destination, but nobody was sure what to do when you arrived.
What Is the Jevons Paradox, Anyway?
In the 1800s, an economist named William Stanley Jevons noticed something strange: as steam engines got more efficient and used less coal per unit of work, Britain ended up burning more coal overall, not less. Why? Because cheaper, more efficient engines made it profitable to use them everywhere, so total use exploded.
The same logic applies to AI. As AI models get cheaper and faster to run — as the “price per question” drops — people use them far more. A company that could only afford to run 1,000 AI queries per day at the old price might run 100,000 at the new price. Total computing demand goes up, even though each individual task got cheaper.
This graph says that mechanism is one of the most important things to understand about AI infrastructure demand. But it also says the analysts who built this map were uncertain whether it would hold — hence the low confidence score despite the high centrality.
The January 2026 release of a model called DeepSeek — which dramatically cut the cost of AI inference — is the only real-world test case in this entire graph. It is the only completed event, not a projection. And the graph maps it as evidence that cheaper AI does accelerate demand rather than reduce it.
The $2 Trillion Question
The strongest evidence that demand is real is a number: $2 trillion in contracted backlogs — essentially, purchase orders that hyperscalers (Amazon, Google, Microsoft, Meta) have already committed to. This is often cited as proof that the AI build-out is demand-driven rather than speculative.
But the graph identifies a problem with this logic. A portion of those contracts may be circular. AI companies buy computing from cloud providers. Cloud providers invest in AI companies. Those AI companies use that investment money to buy more computing. The contracts look like independent demand, but they might partly be the same money going around in circles.
The graph calls this the “AI Circular Capital Loop,” and it attacks the $2 trillion figure directly — not disputing the number, but questioning whether it proves what it is supposed to prove. The graph also flags 2027 as the year when this becomes testable: if third-party revenue (from companies with no investment relationship with cloud providers) grows alongside capex, the demand is real. If it does not, the circular flow explanation gains credibility.
Five Roads That All Lead to the Same Bridge
The graph encodes a powerful structural argument: AI demand is not coming from one place, it is coming from five independent sources simultaneously.
Think of it like a city where five different industries — manufacturing, tourism, shipping, technology, and healthcare — are all growing at the same time. Even if one industry stumbles, the others keep the city busy.
The five “pillars” in the graph are: agentic AI (AI systems that work autonomously over hours or days), existing contracted backlogs, sovereign AI spending by governments and nations, physical AI like robots, and the shift from traditional software to AI-powered services.
Here is the non-obvious part: all five of those independent demand sources eventually run into the same physical bottleneck — electricity and power infrastructure. The independence of the demand sources does not reduce the constraint; it amplifies it. Five separate highways all converge on one bridge.
The Geopolitical Twist
One of the more counterintuitive findings is that geopolitical tension — particularly the split between US and Chinese AI development — has two opposite effects at the same time.
On one hand, it increases total demand. If the US and China each build their own parallel AI ecosystem rather than sharing one, you need roughly twice the infrastructure to serve the same global market.
On the other hand, it reduces a specific risk. Right now, almost all advanced AI chips are manufactured at a single company in Taiwan (TSMC). A geopolitical conflict that pushes China to build its own chip manufacturing reduces the world’s dependence on that single point of failure.
The graph maps the same geopolitical force as simultaneously expanding total demand and shrinking supply-chain concentration risk. These two effects do not cancel each other out — they just pull in different directions at once.
When Solving One Problem Makes Another Worse
The graph contains several cases where a proposed solution backfires structurally.
The clearest one involves custom chips. Google, Amazon, Microsoft, and Meta are all building their own AI processors — partly to reduce their dependence on NVIDIA. But all of those custom chips are manufactured at the same place: TSMC. The diversification move away from one vendor concentrates supply-chain risk at another. The graph encodes this with a high-confidence edge: custom silicon amplifies TSMC concentration risk.
A similar pattern appears with edge AI — the idea of running AI models on local devices (phones, laptops) rather than central servers. The intuitive expectation is that on-device AI reduces demand for cloud computing. The graph says the opposite: more capable on-device AI creates more use cases, which drives more cloud-level compute demand. The relief valve increases pressure.
The Things That Are Both True at the Same Time
Several tensions in the graph have no resolution — they are genuinely two-sided.
The enterprise adoption gap is one. About 78% of large companies are running AI pilots. That single statistic can be read two completely opposite ways. Optimistic reading: 78% of enterprises are about to convert pilots to full deployment, which will massively increase compute demand. Pessimistic reading: those pilots have already been running for months or years and have not converted, which means the expected demand surge is not materializing. The graph encodes both interpretations with nearly identical confidence weights.
The analogy to the 1990s fiber-optic overbuilding is another unresolved tension. In the late 1990s, hundreds of billions of dollars were spent laying fiber-optic cable in anticipation of internet demand that eventually did arrive — but decades later, after most of the companies building it had gone bankrupt. The graph contains explicit arguments for why AI infrastructure is categorically different (locked-in contracts, physical asset scarcity, multi-decade demand), but the most direct challenge to those arguments is that cheap debt and circular capital flows were also present in the fiber boom. The graph records both the argument and its strongest counterargument without resolving which is correct.
What Would Actually Prove Any of This?
The graph generates several testable predictions — places where real data, over the next few years, will tell us whether the demand is genuine.
The most important one: if the Jevons mechanism is real, then token prices (the cost of running one AI query) should predict compute demand better than enterprise adoption surveys. Every time tokens get 10 times cheaper, compute purchases should go up significantly. That is measurable now.
The second test: nuclear power purchase agreements. Several data center operators have signed 20-year contracts to buy power from nuclear plants. If those commitments are load-bearing — if they represent genuinely locked-in demand that will materialize regardless of near-term AI adoption rates — then nuclear PPA signing volume should be a better predictor of actual data center construction than any survey of enterprise intent.
The third test, explicitly encoded in the graph, is 2027. If enterprise AI workloads as a share of corporate computing spend do not show clear acceleration by the second half of 2027, the main bridge between current investment and future utilization is missing.
Bottom Line: What the Graph’s Structure Actually Shows
The structural analysis of this knowledge graph reveals four non-obvious things.
First, the most-connected concepts are the least-trusted ones. The Jevons mechanism is central to nearly every demand argument in the graph, but the people who built this map assigned it minimal confidence. This is not a contradiction to resolve — it is a signal about where the genuine uncertainty lives.
Second, the primary demand evidence may share a hidden common cause. The $2 trillion contracted backlog and the supply-constraint evidence that validates it are both potentially touched by the same circular capital flow. The graph does not claim they are circular — it claims the question cannot yet be answered.
Third, the bear case and the bull case attack different parts of the argument. The bull case rests on demand independence across five pillars. The bear case does not dispute any of those pillars directly — it attacks the base of the validation chain at the source (the contracted backlog). One is an argument about breadth, the other is an argument about depth.
Fourth, 2027 is the graph’s own resolution date. The map was not built to be permanent — it explicitly encodes a time horizon at which several of its core tensions become empirically distinguishable. The graph is, in a sense, a set of bets with a stated expiration date.