NextFin News - BlackRock executive Rachel Lord Jewell is signaling that the artificial-intelligence capex cycle is not about to snap back to pre-boom levels. Her view, highlighted in a July 9 market note, is that AI capital spending can stay broadly stable for two to three years even after the surge that has defined the first half of the decade. That is a more important judgment than it first appears: stable capex at a very high level usually means a market is moving from acceleration to infrastructure digestion, not from expansion to contraction.
The distinction matters because AI spending has become a layered ecosystem rather than a single line item. Chip orders, data-center construction, power interconnection, cooling systems and networking equipment are all part of the same spend pool, but they do not move at the same speed. A pause in the growth rate of GPU demand does not necessarily mean the broader buildout is fading. It can mean the opposite: money is shifting down the stack toward the physical constraints that govern how much compute can actually be deployed.
That is why the market reaction to any AI-capex comment now depends less on whether spending is rising and more on what kind of spending is rising. A stable aggregate number can still mask a changing composition, and composition is where the next earnings winners and losers are usually decided. The companies tied to the first wave of AI enthusiasm have already benefited from explosive demand. The next leg may belong to the suppliers that make the facilities, grids and cooling systems usable.
Public estimates underscore the scale of the opportunity and the debate around it. IDC has forecast that global AI infrastructure spending could reach $758 billion by 2029, while broader industry estimates have continued to point to a large and still-expanding market in 2026. Even without treating any single estimate as gospel, the direction is clear: the market is not discussing a small cyclical uptick. It is debating the pace at which an industrial buildout can absorb capital without forcing a reset.
That is the central question for investors. Is the current AI capex phase a cyclical wave that will fade once early deployments normalize, or is it a structural regime in which the buildout remains elevated for years because the underlying compute stack still lacks enough power, connectivity and cooling? Jewell’s framing leans decisively toward the latter.
In that sense, stable AI capex is not a warning sign. It is the footprint of a system that has become too large to move in a straight line.
Stable Spending Is a Different Signal Than Slowing Demand
Jewell’s comment is best understood as a change in the growth rate, not a verdict on demand. That distinction is crucial. In capital cycles, growth can slow even as the absolute amount of spending stays high enough to support entire supplier ecosystems. The market often treats those two facts as if they were the same thing. They are not.
AI is already showing the characteristics of a multistage industrial cycle. The first stage was model training and GPU procurement. The second is the much slower process of building the physical environment around those chips: substations, transmission access, power-generation commitments, server halls, cooling, fiber and networking. If the first stage had been the whole story, spending would likely have peaked earlier and then fallen sharply. Instead, the bottleneck keeps moving deeper into the infrastructure layer.
That shift is why a “stable for two to three years” view can be bullish for some parts of the ecosystem even if it sounds neutral in the abstract. Power equipment makers do not need capex to grow at 50% a year if it is already running at a very high base and stays there. Data-center developers do not need each quarter to top the last one by a wide margin if the underlying utilization rate continues to climb and the backlog remains long. In that environment, stability becomes a form of visibility.
The market’s common mistake is to think in one-dimensional terms: either the AI spend wave is still accelerating or it has already rolled over. But industrial buildouts rarely end that cleanly. They tend to migrate from shortage to queuing, from queuing to capacity rationing, and from capacity rationing to a slower, more durable investment path. That path may still be expensive. It just stops looking like a headline chase.
In the AI case, the mechanism is straightforward. More compute demand requires more servers. More servers require more power density. More power density requires more cooling and more grid access. More grid access requires more permits, interconnections and capital. The chain is not infinitely elastic. When one node tightens, spending shifts to the next node. That is why the market can see “stable capex” and still be underestimating the value of the physical bottlenecks underneath it.
There is a second-order implication here that the market is only partially pricing. If the capital intensity of AI stays elevated for several years, the winners are not necessarily the highest-profile model developers. They are the firms and utilities that convert those models into usable, running capacity. Put differently, the trade is moving from intelligence to infrastructure. That is a different earnings engine.
Why This Looks Structural, Not Merely Cyclical
The structural case is stronger than the cyclical one because the constraints are no longer purely financial. Cyclical capex tends to mean-revert when inventory builds, utilization slips, or demand disappoints. Structural capex does something else: it persists because the system itself has changed, and the old level of infrastructure is no longer sufficient.
AI looks closer to the second category for three reasons. First, the deployment curve is still early relative to the size of the compute stack that enterprises and hyperscalers say they want to build. Second, the limiting factors are physical, not just budgeting choices. Third, the history of prior technology buildouts shows that once infrastructure bottlenecks become the rate limiter, spending can stay high long after the original enthusiasm has cooled.
That does not mean the path is linear. It means the base rate is likely higher than skeptics expect. The capex curve can flatten while remaining elevated. It can even wobble around a plateau for several years. Yet that plateau may still be enough to keep the entire supply chain busy, because the denominator is so large.
This is where the cyclical thesis weakens. A cyclical boom usually comes with three signs: inventories build faster than end demand, pricing power fades, and suppliers begin to talk about normalization. In AI, the talk has moved in the opposite direction. The conversation is now about whether power queues, transformer availability, local grid constraints and thermal design are slowing deployment. That is not the language of an inventory bust. It is the language of a buildout that has run into real-world friction.
Historical comparison supports the point. Large infrastructure waves rarely terminate when the first wave of enthusiasm cools. They tend to stretch into a second and third phase as adjacent bottlenecks reveal themselves. Railroads, electrification and telecom all followed that pattern: the initial spend was obvious, but the supporting network took longer and consumed capital for years after the first excitement faded. AI is behaving more like that than like a one-quarter trade.
That still leaves a serious counter-thesis. If AI monetization fails to keep pace with the capex wave, boards will eventually force discipline. The market is right to worry about that. Capital can outrun revenue. When it does, even a structurally necessary buildout can turn into an expensive overhang.
But the burden of proof for that bearish view is high. To falsify the structural call, investors would need to see a broad downgrade cycle in hyperscaler capex guidance, not just one company trimming plans. They would also need evidence that utilization is lagging badly in newly built capacity and that project pipelines are being delayed rather than merely repriced. Until that appears, the more defensible judgment is that AI spending is normalizing, not reversing.
“The important question is not whether AI investment is still expanding, but whether the ecosystem can absorb the scale of the buildout over the next several years,” Jewell said in remarks highlighted by the July 9 note.
The quote captures the key point. The issue is absorption, not existence. If the ecosystem can keep absorbing capital, the story stays alive. If it cannot, the capex plateau becomes the prelude to a reset.
What The Market Should Watch Next
Short term, the AI trade remains vulnerable to sentiment. Any sign that revenue conversion is slipping, or that a major buyer is becoming cautious, can trigger a sharp re-rating in the most crowded names. That is the cyclical risk. It is real, and it can hit fast.
Medium term, though, the better base case is that AI capex stays high but grows more slowly, with spending migrating from compute hardware toward the infrastructure required to support it. That favors the suppliers with grid access, permitting strength, power equipment exposure, thermal-management expertise and data-center pipelines already in place. It also favors businesses that can show recurring demand from installed AI capacity rather than one-time order spikes.
Long term, the thesis is more clearly structural. If AI remains a major software and productivity platform, the economy will need a much thicker layer of physical infrastructure than the one built for cloud computing alone. That means the capex cycle may settle into a new equilibrium rather than unwind. In that world, the headline risk is not that AI spending disappears. It is that investors keep waiting for the old, explosive growth rate to return even after the market has already moved on to steadier, infrastructure-led expansion.
The upside scenario is simple: AI monetization keeps improving, utilization rises, and the buildout supports a longer period of elevated capex. The downside is equally clear: if revenue growth stalls and capex guidance rolls over across multiple hyperscalers, the market will reprice the whole stack as an overbuilt cycle. The base case sits in between. Spending stays high, growth slows, and the winners shift from the obvious compute names toward the firms that make the facilities work.
The next signals to watch are equally concrete: hyperscaler capex guidance in upcoming earnings updates, backlog and delivery commentary from power and data-center suppliers, and any signs that grid interconnection delays are easing rather than worsening. The clearest falsifier would be a broad and sustained cut in capital-spending plans from the largest AI buyers. If that happens, the structural argument weakens quickly.
For now, the market should treat stable AI capex as a sign of maturation, not exhaustion. The spending boom is no longer just about buying more chips. It is about building the industrial system that lets those chips do useful work.
That is why the real AI trade may be less about acceleration than endurance.
Explore more exclusive insights at nextfin.ai.
