NextFin News - Artificial intelligence adoption is accelerating into 2026, but the industry's growth is running into a hard physical ceiling: there is not enough computing capacity to meet demand, and the constraints are likely to persist for years. Michelle Weaver, a US thematic research strategist at Morgan Stanley, laid out the warning plainly this week as the S&P 500 climbed 0.30% and the Nasdaq jumped 0.59% on a tame inflation print. Compute has become a scarce resource, she said, and the bottlenecks holding back supply are structural enough that they will not clear quickly.
The constraint is not just about chips. Power, labor, and political approval for new data centers are all tightening at once, creating a multi-year squeeze on the infrastructure that the AI boom depends on. For investors, the implication is a rotation in what "owning AI" means — from the companies designing the fastest models to the companies controlling the electrons, the land, and the permits that let those models run at all.
The Bottleneck: Undersupplied Compute in an Overheated Buildout
Weaver's central claim is deceptively simple: demand for compute is systematically higher than supply, and the gap is not closing. "We are very much undersupplied. We are seeing compute being a constrained resource," she said. The scarcity is compounded by what she described as three distinct choke points for data-center buildout — power, people, and politics — each of which operates on a different timeline and none of which can be solved by spending alone.
"The power bottlenecks, political bottlenecks, labor bottlenecks – that will keep a check on supply for the next few years."
That framing matters because it reframes the AI investment story. Through 2024 and into 2025, the market rewarded the enablers of AI — semiconductor designers, data-center builders, and power companies. The question now is whether the constraint shifts from "who can build the most capacity" to "who already has capacity," a change that would reprice winners and losers across the entire technology complex.
The scale of the shortfall is large enough to affect corporate planning. Morgan Stanley's power analysis projects a US power shortfall through 2028 of 44 gigawatts before accounting for innovative "time-to-power" solutions that bypass the standard grid-interconnection process. Weaver has separately put the constraint in similar terms, telling a Morgan Stanley podcast that the thematic research team is estimating a nearly 40-gigawatt shortfall in power needed for data centers — "multiple New Yorks worth of power," in her words. Even after assigning probabilities to the workarounds, the bank's base case still leaves a 1 GW to 11 GW supply deficit through 2028.
To put that in perspective: data centers are expected to require roughly 68 GW of additional electricity between 2026 and 2028, while projects already under construction account for about 15 GW and another 15 GW is covered through available or contracted grid capacity. The arithmetic does not close without delays, cancellations, or higher costs. Morgan Stanley's analysts, led by Stephen Byrd, have called AI computing demand "the most important technological shift in modern history," and warned that "AI infrastructure stocks are at the center" of a transition whose "non-linear rate of AI improvement" is creating broader asset-valuation impacts.
The time-to-power solutions that could narrow the gap are specific and finite: natural-gas turbine transactions that might add 15–20 GW, fuel-cell deployments adding 5–8 GW, and nuclear-powered data-center deals adding 5–15 GW. Each is real, each is capital-intensive, and each still requires permits, interconnection agreements, and local acceptance. The residual deficit is not a modeling artifact — it is the part of the shortage that cannot be engineered away.
The capital math behind the buildout is equally unforgiving. Morgan Stanley has framed AI infrastructure financing as a multi-trillion-dollar question, and the reason is arithmetic, not rhetoric. A single gigawatt of data-center capacity can require several billion dollars of investment once land, buildings, transformers, switchgear, and backup generation are counted. Closing even the lower end of the 1 GW to 11 GW deficit therefore implies tens of billions of dollars of incremental spending that must be committed years before the revenue it supports is visible. In a higher-rate environment, that is capital that competes with buybacks, dividends, and non-AI projects inside the same corporate budget. The bottleneck is not only physical — it is financial, and it is governed by hurdle rates.
Why the Constraint Is Structural, Not Cyclical
The critical analytical question is whether this is a cyclical shortage that will mean-revert as capital floods in, or a structural regime shift that reorders the industry. The evidence points to structural — and that distinction determines whether the scarcity premium is a trading opportunity or a permanent repricing.
A cyclical shortage has a clean self-correcting mechanism: high prices attract capital, capital builds capacity, capacity restores balance. That loop works when the binding constraint is a single input with a liquid supply chain. The AI-compute bottleneck fails that test on three counts.
First, the power constraint is a permitting and grid-connection problem, not just a generation problem. A natural-gas peaker plant or a new transmission line can take a decade to site, permit, and interconnect. Local moratoriums on data centers are already in force across more than 150 US jurisdictions spanning dozens of states, and community opposition has turned new construction into a political contest rather than an engineering one. Capital cannot accelerate a political process by writing a check.
Second, the labor constraint is demographic and specialized. Data centers need electrical engineers, high-voltage technicians, and construction crews with experience in mission-critical facilities — workers who cannot be trained at the pace the buildout requires. This is not a wage problem that higher pay solves quickly; it is a pipeline problem measured in years.
Third, demand is non-linear while supply is linear. Morgan Stanley has described the rate of AI improvement as "non-linear," meaning compute demand can double faster than any physical infrastructure pipeline can respond. When demand grows exponentially and supply grows linearly, the gap widens even as absolute capacity rises. Scarcity is a feature of the geometry, not a temporary imbalance.
For comparison, the 2020-2022 semiconductor shortage was cyclical: it was a single-supply-chain disruption with a clear price signal, and it resolved as fabs came online and demand normalized. The current compute bottleneck is different because the constraint has migrated downstream from chips to the physical plant that houses them — and the plant is governed by permits, grids, and communities, not by foundry capex cycles. In the earlier episode, the fix was money. In this one, the fix is time, and time cannot be bought.
This is the crux of the structural call: the constraint has moved from a layer of the stack that responds to capital expenditure to a layer that responds to civic process. That is a regime change, not a cycle. Mean reversion requires a mechanism that pulls the system back to equilibrium, and no such mechanism exists when the binding inputs are permits, skilled workers, and grid interconnection slots. The shortage persists not because investors are unwilling to pay, but because the things that must be built cannot be built fast enough at any price.
The Second-Order Effect: Scarcity Migrates Down the Stack
The first-order effect of a compute shortage is obvious: hyperscalers and AI developers cannot run every workload they want, and those with reserved capacity gain pricing power. The second-order effect is subtler and more important for investors — the bottleneck changes which layer of the stack captures the economic rent.
In a world of abundant compute, value accrues to the best models and the widest distribution. In a world of scarce compute, value accrues to whoever controls the scarce input. That shifts the center of gravity from software margins to infrastructure ownership. A company with a signed power-purchase agreement and an energized site becomes more valuable than a company with a better model but no place to run it.
This dynamic also creates a cross-asset transmission channel. Power scarcity flows into wholesale electricity prices, which flow into data-center operating costs, which flow into the unit economics of AI services. If the marginal cost of a token is set by the marginal cost of a megawatt-hour, then AI profitability becomes partially a function of the power market — linking technology valuations to natural-gas prices, nuclear capacity, and grid congestion in a way that did not exist five years ago. A technology investor now has exposure to PJM congestion charges whether they want it or not.
The transmission chain runs further. Higher power costs raise the hurdle rate for new AI services, which slows the rollout of marginal applications, which reduces the near-term revenue growth of the software companies that the market has been rewarding. The bottleneck therefore acts as a throttle on the very adoption curve that equity valuations are discounting. This is the third-order expectation gap: the market is pricing AI as if the constraint were a temporary speed bump, when the mechanism suggests it is a persistent governor on growth.
The expectation gap is where the risk sits. The market has largely priced AI as a software-and-chip story. If the binding constraint is power and permitting, then the consensus winners may be over-owned and the consensus laggards — utilities, independent power producers, electrical-equipment makers, and grid-services companies — may be under-owned. Morgan Stanley itself has noted that "AI infrastructure stocks are at the center" of the transition, but the market has not fully re-rated the implication that infrastructure now includes the grid, not just the server rack.
There is also a positioning asymmetry worth noting. The AI-enabler trade has been one of the most crowded positions in the market for two years, while the power-infrastructure trade, despite strong performance, remains more dispersed across sectors — regulated utilities, independent generators, turbine manufacturers, fuel-cell developers, and electrical equipment. Crowding matters because a scarcity thesis that is correct but already fully owned offers little upside; a correct thesis that is only partially owned offers both repricing and multiple expansion. The compute bottleneck points toward the latter.
The Counter-Thesis: Innovation Could Shrink the Constraint
The strongest case against the scarcity thesis is that technology will out-innovate the bottleneck. Model efficiency has improved dramatically — smaller models trained on less compute have exceeded expectations in early 2026, and inference optimization, sparsity, and specialized silicon all reduce the compute required per unit of output. If efficiency gains outpace demand growth, the shortage evaporates without a single new power plant.
Weaver herself has cautioned that investors may be surprised by "just how powerful those models are" even without massive compute volumes, which cuts both ways: powerful small models reduce demand pressure, but they also accelerate adoption and expand the total addressable market for AI, potentially adding more demand than they subtract. The counter-thesis also has a policy dimension. If the shortage becomes politically salient, governments can fast-track permitting, designate data centers as critical infrastructure, and subsidize transmission — compressing the timeline that the private market cannot. This is the "time-to-power solutions" lever Morgan Stanley models, and it could reduce the shortfall materially if deployed aggressively.
But the counter-thesis requires two things to go right simultaneously: efficiency must improve fast enough to offset non-linear adoption, and policy must move faster than local opposition. History suggests policy accelerates only after a shortage becomes visible in prices or service delays — meaning the relief arrives after the scarcity premium has already been earned by incumbents. Efficiency gains, meanwhile, have a poor track record of reducing total resource consumption: they lower the cost per unit of output, which expands the set of economically viable applications, which increases total demand. This is the Jevons paradox in modern dress, and it has held in energy, computing, and transportation for two centuries.
The falsifying signal is specific: if US data-center power additions exceed 30 GW per year by 2027 while AI workload growth decelerates below 40% annually, the structural-scarcity thesis is wrong and the constraint is cyclical after all. Until then, the default assumption should be that scarcity persists.
What to Watch: Beneficiaries, the Exposed, and the Signals
Short term (sentiment and liquidity): Expect volatility in AI-infrastructure names as the market digests the scarcity message. Companies perceived as capacity-constrained may be marked down; companies with visible capacity may be marked up. This is a positioning trade, not a fundamentals trade. The August 12 session, in which the S&P 500 rose 0.30% and the Nasdaq gained 0.59% on a benign inflation report, is a reminder that macro headlines still dominate intraday price action — the bottleneck is a multi-year thesis, not a daily catalyst.
Medium term (fundamentals): The beneficiaries are owners of energized data-center sites, independent power producers with contracted capacity, electrical-equipment suppliers, and grid-services firms. The exposed are AI developers and hyperscalers without secured power, whose capex plans may face delays or cost overruns, and whose margins may compress if wholesale power prices rise faster than they can pass costs through. Within the semiconductor complex, the asymmetry is subtler: chip designers with guaranteed access to advanced packaging and co-located power gain a relative advantage over fabless peers competing for the same constrained infrastructure.
Long term (structural): If the scarcity thesis holds, the AI industry bifurcates into a capacity-haves and capacity-have-nots dynamic. Incumbents with power and permits gain a durable moat; late entrants face a higher barrier than at any point since the buildout began. This is not a bubble-pop scenario — it is a consolidation scenario. The companies that survive the bottleneck are not necessarily the ones with the best technology, but the ones with the most reliable access to the inputs that technology requires.
Base case: compute remains undersupplied through 2028, with the 1 GW to 11 GW residual deficit forcing prioritization of workloads and premium pricing for reserved capacity. Upside case: aggressive policy fast-tracking and breakthrough efficiency gains close the gap by 2027, restoring a more balanced market. Downside case: the deficit widens beyond 11 GW as adoption accelerates, triggering service rationing and a sharper repricing of capacity-constrained names.
The signals to watch are concrete: quarterly data-center power-connection queues by region, wholesale electricity prices in key hubs like Northern Virginia and Texas, hyperscaler capex guidance versus realized capacity additions, and any federal permitting-reform legislation. Each of these is observable and dated.
The central judgment: the AI boom is not running out of ideas; it is running out of physics. And physics does not respond to earnings calls.
Data as of market close August 12, 2026. All figures attributed to Morgan Stanley research, company disclosures, or public market data.
Explore more exclusive insights at nextfin.ai.
