The Two Construction Economies: AI Capital Spending, Interest Rates, and the Limits of Resilience

The Two Construction Economies: AI Capital Spending, Interest Rates, and the Limits of Resilience

Private construction has split into two sharply different economies. Using December 2023 as the chart’s baseline, data-center construction spending rises persistently and reaches roughly $4 billion above that level by 2026. Construction outside data centers first advances during 2024, then reverses hard and finishes near $10 billion below the baseline. The resulting gap is not a minor sector rotation. It is a visible expression of how unusually concentrated investment demand has become, how differently projects respond to the cost of capital, and how a strategic technology build-out can coexist with broad cyclical weakness.

The tempting interpretation is that data centers are “immune” to higher interest rates. That is too strong. Rates still change discount factors, financing costs, hurdle rates, land values, and the timing of marginal projects. What the chart supports is a relative claim: data-center construction has so far been much more resilient than the rest of private construction. Explaining that resilience requires separating demand strength from rate insensitivity, nominal spending from real capacity, and projects already committed from projects still competing for capital.

Reading the Divergence Correctly

The chart is an index-like comparison of changes from a common date, not a complete accounting of construction levels. A gain of about $4 billion for data centers and a decline of about $10 billion elsewhere describe changes relative to December 2023; they do not mean data centers are larger than all other private construction. The economically important fact is the slope and persistence of the divergence. One category continues to add spending while a vastly broader category loses momentum.

The “all other” line aggregates heterogeneous markets: offices, warehouses, retail, factories outside the relevant classification, lodging, health facilities, and other commercial projects. Weakness in an aggregate can reflect several simultaneous adjustments rather than one universal recession. Still, a decline of this magnitude signals that ordinary projects are facing a tougher clearing environment. Expected rents, occupancy, credit availability, construction costs, and required returns no longer support as much activity as they did at the baseline.

Nor does the chart identify causality. Higher rates are a plausible contributor because construction is duration-heavy and commonly debt-financed, but the same period can contain tighter bank standards, post-pandemic overbuilding in selected property types, weaker rent expectations, labor and equipment bottlenecks, and shifts in tax incentives. Data centers, meanwhile, benefit from a distinct demand shock: accelerated spending on artificial-intelligence training and inference, cloud services, storage, networking, and redundant computing infrastructure.

The proper question is therefore not “Do rates matter?” but “Why has the positive demand and commitment shock in data centers dominated the negative financing shock more effectively than elsewhere?” That formulation respects what the evidence shows while avoiding a causal conclusion the chart cannot establish.

Construction Is a Long-Duration Asset

A construction project converts a large upfront payment into a stream of uncertain future cash flows. In a stripped-down model, value is the present value of expected net operating income minus development cost: V equals the sum of expected cash flows discounted at a risk-adjusted rate, less initial investment. When the discount rate rises, distant cash flows lose more present value. Because permitting, building, leasing, and stabilization take years, development behaves like a long-duration asset.

Suppose a project costs $1 billion today and is expected to generate $90 million annually in stabilized cash flow. At a 7% capitalization or discount benchmark, the implied value of a durable stream is far higher than at 10%. The exact valuation depends on growth, depreciation, terminal value, taxes, and risk, but the direction is unambiguous: a three-percentage-point increase can erase a large fraction of development equity. Projects near the margin are cancelled, resized, or delayed.

Financing compounds this effect. A developer that borrows 60% of cost sees annual interest expense rise materially when loan pricing moves from 4% to 7%, even before considering fees or hedges. Lenders may simultaneously lower loan-to-cost ratios, demand more preleasing, and apply harsher stress assumptions. The sponsor then needs more equity just as the equity hurdle rate is rising. This joint tightening is more powerful than the policy rate considered in isolation.

Asset prices transmit the signal back to construction. Higher long-term yields raise required capitalization rates for stabilized property. If market value falls below replacement cost, new building stops even when existing space remains occupied. James Tobin’s q framework captures the intuition: investment is attractive when the market value of installed capital exceeds its replacement cost. For much conventional commercial property, higher yields and weaker expected rents have pushed q downward.

Why Ordinary Private Construction Has Weakened

The broad decline is consistent with several mutually reinforcing channels. Office demand remains impaired by hybrid work, making lenders reluctant to fund additional space and forcing owners to focus on refinancing existing buildings. Warehouse construction can cool after an exceptional e-commerce expansion. Retail and lodging respond to consumer demand but also to local supply, wages, and property valuations. Multifaceted weakness is exactly what an aggregate “all other” series can conceal.

Bank balance sheets matter because many smaller and mid-sized developments rely on relationship lending rather than bond markets. When deposit costs rise and commercial-real-estate exposures attract scrutiny, banks ration credit through covenants and lower advance rates. Bernanke and Gertler’s financial-accelerator logic applies: declining collateral values weaken borrower net worth, which widens external-finance premiums, depresses investment, and can further reduce collateral values.

Construction costs also remain relevant. Even if some material inflation moderates, skilled labor, electrical gear, transformers, insurance, and compliance can remain expensive. A project can face both a higher numerator and a lower denominator: development cost rises while the discounted value of future rent falls. That combination destroys feasibility faster than either shock alone.

There is also a pipeline effect. Spending records work currently put in place, not merely fresh decisions. The 2024 rise outside data centers may have reflected projects financed or started when rates were lower and demand assumptions were stronger. As that backlog completed, fewer replacements entered the pipeline. The subsequent fall can therefore be the delayed manifestation of monetary tightening rather than an instantaneous response.

The Demand Shock Behind Data Centers

Data centers occupy a different point in the investment cycle. Demand for compute is not simply a normal expansion of office space; it is tied to a platform transition. Training frontier models requires clusters of advanced accelerators, high-speed networking, cooling, backup power, and specialized buildings. Inference then turns model use into an ongoing capacity requirement. Cloud migration, video, enterprise software, cybersecurity, and data localization add non-AI demand.

The willingness to spend depends on expected strategic value as well as near-term rent. Large technology firms may treat capacity as an option on future products, market share, and lower unit computing costs. Under real-options theory, uncertainty can sometimes delay irreversible investment, but it can also increase the value of securing scarce capacity when being late risks losing a winner-take-most market. The competitive cost of underbuilding may exceed the financing cost of overbuilding.

This is especially true when the investor has a strong balance sheet and high operating cash flow. A hyperscaler funding construction internally is less exposed to bank lending standards than a leveraged property developer. Its relevant comparison may be the return on a vertically integrated computing platform, not the stand-alone rental yield on a shell building. Higher rates still raise the corporate hurdle rate, but the project’s expected strategic surplus may remain comfortably above it.

Long-term customer agreements and committed procurement can further stabilize the pipeline. Capacity reservations, cloud contracts, equipment orders, utility interconnection deposits, and land purchases create sunk costs and contractual obligations. Once a campus reaches an advanced stage, stopping can be more expensive than continuing. Observed spending therefore contains both current optimism and the momentum of decisions made quarters earlier.

“Resilient to Rates” Is Not “Immune to Rates”

Investment occurs when expected return exceeds the weighted average cost of capital by enough to compensate for execution risk. A powerful demand shock raises expected cash flows; higher rates raise the cost of capital. Data-center spending can rise when the first movement dominates the second. That outcome does not imply the second movement is zero.

Consider two hypothetical projects. A conventional building expected to earn an 8% unlevered return becomes unattractive when its required return rises from 7% to 9%. A data-center campus expected to earn 16% may remain attractive after the same two-point increase. Both are rate-sensitive, but only the marginal conventional project crosses the cancellation threshold. The chart is consistent with this threshold model.

Rates also operate indirectly through end customers. A well-funded platform company may borrow little, yet its customers, venture ecosystem, and equity valuation respond to discount rates. If expensive capital reduces software formation or enterprise technology budgets, future compute demand softens. Rate sensitivity can be delayed and distributed across the network rather than visible in the developer’s loan coupon.

The distinction matters for forecasting. Calling the sector immune encourages investors to extrapolate spending without asking what return the next unit of capacity will earn. A better claim is that current projects possess unusually large demand cushions, strategic value, and financing capacity. Those cushions can absorb substantial tightening, but they are finite.

Nominal Spending Is Not Physical Capacity

The chart measures dollars, so an increase can reflect higher prices as well as more buildings. Data centers are intensive in electrical systems, cooling, generators, specialized labor, and increasingly dense computing environments. If the cost per megawatt rises, nominal construction spending can increase even when delivered megawatts grow more slowly. This is especially important during supply bottlenecks.

Analysts should separate four quantities: dollars spent on structures, megawatts of powered capacity, rack density, and effective computing output. A newer facility may support more computation per square foot, while a power-constrained shell may add little usable capacity. Semiconductor efficiency and software optimization can lift output without equivalent construction. Conversely, redundant power and cooling can raise construction cost without directly increasing sellable compute.

Inflation adjustment alone is insufficient because the input mix differs from generic construction. Relevant deflators should reflect switchgear, transformers, cooling systems, engineering, and local labor. The nominal $4 billion rise since December 2023 is still economically meaningful—it represents demand for resources and capital—but it should not be interpreted mechanically as a proportional rise in real capacity.

The same caution applies to the roughly $10 billion decline elsewhere. Falling nominal spending during a period of elevated construction costs can imply an even larger contraction in physical activity. Thus the real divergence could be larger or smaller than the nominal chart depending on sector-specific inflation. Unit-based measures are essential for resolving the ambiguity.

Power Is the Binding Factor

The decisive scarce input is increasingly electricity rather than land or financial capital. Large campuses require reliable power, grid interconnections, substations, transmission upgrades, and often backup generation. Interconnection queues can stretch for years. A project with financing and customers may still fail to proceed because incremental power is unavailable at the needed time and place.

This changes the economics of location. Traditional data-center hubs benefit from fiber networks, skilled operators, and established ecosystems, but congestion raises power prices and delays. New clusters may form near generation, transmission capacity, or supportive regulators. Land that is unremarkable for ordinary commercial use can become valuable if it has a credible power path.

Power constraints also make commitments lumpy. Utilities and developers must invest ahead of realized demand, while hyperscalers reserve capacity to preserve options. Announced requirements may exceed eventual use because each participant seeks insurance against scarcity. That behavior can sustain near-term construction yet create a later digestion period if reservations are duplicated or utilization disappoints.

The social cost extends beyond the developer’s bill. Grid upgrades, emissions, water use, local rate design, and reliability create political questions about who pays. If regulators require data centers to fund dedicated infrastructure or impose moratoria, project economics change. The relevant cost of capital includes institutional permission and infrastructure scarcity, not merely Treasury yields plus a credit spread.

Concentration Creates Both Productivity and Fragility

The investment surge can raise aggregate productivity if abundant compute improves research, software, logistics, health care, and business processes. In growth-accounting terms, it expands a form of productive capital and may raise total factor productivity when firms reorganize around it. A construction mix tilted toward high-return infrastructure is not necessarily unhealthy simply because other categories shrink.

Yet concentration creates macro fragility. If a small group of firms accounts for a large share of incremental investment, aggregate capital expenditure becomes sensitive to their earnings, governance, and expectations. A modest revision to the expected profitability of AI could affect equipment orders, utilities, construction employment, and local tax bases. Correlated decision-making turns a sector story into a macro factor.

The accelerator principle is relevant: investment responds to changes in expected demand, not only its level. Compute demand can remain high while construction growth slows if firms decide that existing and already-ordered capacity is sufficient. Markets often confuse a high level of spending with a permanently high growth rate. The second derivative eventually matters.

There is a familiar historical pattern. Railways, electricity, telecommunications, and fiber networks produced genuine long-run productivity gains while also experiencing overinvestment and financial losses for some owners. Social usefulness does not guarantee attractive private returns at every entry price. Infrastructure can be transformative and still be built ahead of monetizable demand.

A Capital-Cycle Interpretation

Edward Chancellor and Marathon-style capital-cycle analysis begins with a simple proposition: high expected returns attract capital, and the supply response eventually compresses those returns. Data centers currently display the early-to-middle stage—scarce capacity, strong demand, large commitments, and rising investment. The key analytical task is to identify when supply growth begins to outpace profitable demand.

Unlike a simple commodity plant, data-center supply is differentiated by location, latency, power reliability, chip availability, and customer relationships. Scarcity can persist even while aggregate square footage rises. But differentiation does not repeal competition. If many sponsors underwrite the same optimistic utilization, lease rates and returns can fall while power and equipment costs remain high.

The capital cycle also separates beneficiaries. Chip designers, equipment vendors, utilities, engineering firms, owners of powered land, developers, cloud platforms, and application companies occupy different bargaining positions. A surge in construction spending does not imply equal economics across the chain. The scarce bottleneck captures rent; commoditized capacity bears the residual risk.

Investors should therefore ask where scarcity migrates. At one stage it may be accelerators, then transformers, then interconnection rights, then specialized labor, and eventually customers able to monetize inference. Equity value accrues to persistent scarcity, not simply to the largest capital budget.

Financing Structure Explains Part of the Gap

Conventional construction often sits in project-level entities with substantial floating-rate debt. Refinancing risk is immediate and transparent. Data-center investment may sit on investment-grade corporate balance sheets, be funded from internal cash, or use long-dated contracts to obtain infrastructure capital. This produces different exposure to the same yield curve.

Weighted average cost of capital still applies. Internal cash is not free: shareholders could receive it or deploy it elsewhere. But firms with low leverage and high margins are not forced to cancel merely because bank credit tightens. Their adjustment occurs through capital allocation—buybacks, acquisitions, research, and construction compete—rather than through a lender refusing to renew a loan.

Private credit and infrastructure funds can also finance campuses against contracted cash flows. Long-duration investors may value inflation linkage and predictable usage, while securitization or joint ventures recycle capital. Such structures broaden the buyer base, though they can hide leverage and create maturity mismatches if underwriting assumes perpetual hyperscale demand.

The gap in the chart is thus partly a gap between balance sheets. Monetary policy is transmitted unevenly: strongly capitalized strategic investors can continue, while bank-dependent marginal borrowers retreat. Aggregate tightening reallocates investment toward actors with cash, collateral, and pricing power.

Macro Implications of the Two Economies

The divergence weakens the usefulness of headline construction totals. Strong data-center spending can mask recessionary conditions across a wider set of properties, just as a booming energy sector can obscure weakness elsewhere. Policymakers and investors need diffusion measures: how many categories, regions, firms, and projects are expanding?

Employment effects may also differ. Data-center construction creates substantial engineering and building work, but completed facilities can employ relatively few permanent workers compared with offices, retail centers, hotels, or factories. A dollar of capital expenditure is not a fixed quantity of labor income. The geographic concentration of campuses further limits broad spillovers.

Monetary policy interpretation becomes harder when strategic investment resists tightening. Central banks raise rates to cool aggregate demand, but sectors with exceptional expected returns can continue absorbing labor, equipment, and power. Other interest-sensitive sectors then bear more of the adjustment. The result is not policy failure; it is heterogeneous transmission.

Fiscal and industrial policy add another layer. Tax credits, accelerated depreciation, semiconductor policy, energy incentives, and public infrastructure can lower effective project costs. When private strategic demand and public support point in the same direction, construction may remain robust despite restrictive monetary conditions. Analysts should avoid attributing the entire divergence to either rates or technology alone.

What Would Falsify the Resilience Thesis?

Several signals would indicate that the current strength is beginning to succumb to financial or economic limits. First, project cancellations and delayed starts would rise even when announced demand remains high. Second, lease pricing or cloud-compute prices would weaken relative to power and equipment costs. Third, hyperscalers would reduce capital-expenditure guidance or shift spending from buildings toward more efficient utilization of existing assets.

Fourth, interconnection reservations could be surrendered, revealing that option demand had exaggerated genuine requirements. Fifth, financing spreads for specialized developers could widen while loan-to-cost terms tighten. Sixth, utilization rates might lag physical delivery, producing excess capacity despite continued headline spending. These are more informative than any single month of construction data.

A rate-driven test would look for differential responses around changes in real yields and credit conditions, controlling for compute demand and construction costs. Econometrically, one would use project-level starts, financing type, sponsor balance-sheet strength, location, and precommitment status. The hypothesis is not zero sensitivity; it is lower near-term elasticity for projects with strong sponsors and contracted demand.

The chart should therefore motivate research rather than close the case. Time-series co-movement cannot isolate policy effects, and the post-2023 sample is short. Natural experiments involving regional power access, lender exposure, or staggered financing resets could provide stronger identification.

A Practical Dashboard for Investors

Track spending, but pair it with physical and economic indicators. Megawatts under construction, energized capacity, interconnection-queue progress, transformer delivery times, vacancy, preleasing, and utilization show whether dollars become productive assets. Power prices and grid-upgrade obligations reveal the true marginal cost.

Follow the full corporate cash-flow loop. Compare capital expenditure with operating cash flow, free cash flow after investment, depreciation, and incremental revenue attributable to compute. Rising capex is attractive when it creates durable cash flows; it is dangerous when depreciation and financing costs arrive before monetization.

Watch market prices for forward information: data-center real-estate yields, infrastructure-credit spreads, utility valuations, equipment backlogs, and implied volatility. But avoid circularity. Equity enthusiasm can lower perceived capital costs and encourage spending, which validates revenue forecasts temporarily. A reversal can operate in the opposite direction.

Finally, monitor the weak side of the two economies. Bank delinquencies, refinancing gaps, commercial-property transactions, architectural billings, permits, and contractor employment help determine whether the $10 billion decline is stabilizing or deepening. The macro outcome depends on the net effect, not on the most exciting category.

Scenario Analysis: From Soft Landing to Capacity Glut

A useful base case combines continued but slower data-center growth with stabilization elsewhere. Projects already contracted move forward, power bottlenecks limit the pace of delivery, and falling inflation permits some easing in financing conditions. Ordinary construction stops deteriorating as loan coupons reset lower and property prices discover a clearing level. In this path the gap remains wide, but it no longer expands at the recent rate. Data-center suppliers enjoy healthy backlogs while returns gradually normalize. The macro effect is constructive: capital formation remains positive without producing an unchecked scramble for inputs.

The upside case requires compute demand to monetize faster than capacity arrives. New applications raise willingness to pay, inference becomes embedded in daily workflows, and efficiency gains expand usage through a Jevons-paradox effect: cheaper computation generates more total computation rather than less. Strong cash conversion validates current investment, while grid additions and better cooling relieve bottlenecks. Conventional construction also recovers as rates decline. In that world, today’s divergence was an early signal of a broad productivity boom rather than merely a narrow pocket of strength.

The downside investment case is a capacity glut. Sponsors extrapolate extraordinary demand, duplicate reservations, and build against customers whose needs overlap. Model efficiency improves faster than usage expands, or firms discover that many applications cannot support premium pricing. Energized supply then arrives just as desired capacity is revised down. Lease rates weaken, utilization disappoints, and specialized assets trade below replacement cost. Construction spending can remain high for a time because projects are difficult to stop, then fall abruptly once the committed pipeline is delivered.

A separate downside macro case combines sticky rates with broader property stress. Data centers continue building, but their dollar contribution is too small to offset refinancing losses, bank caution, and declining activity elsewhere. The visible champion distracts from deteriorating diffusion. Credit losses tighten standards further and local governments face weaker assessments from conventional commercial property. In this scenario, aggregate construction and employment weaken even though technology capital expenditure remains impressive.

Probabilities should update with evidence rather than narrative. Strong bookings without cash realization favor caution; rising utilization and unit economics support the upside. Faster power delivery raises near-term growth but also increases eventual glut risk. Lower rates help both economies, although they may revive conventional construction more strongly because its projects sit closer to the hurdle-rate margin. Scenario analysis makes clear why a single chart cannot justify either technological euphoria or blanket pessimism.

Portfolio and Policy Conclusions

For portfolios, the chart argues for discrimination rather than a blanket construction view. Beneficiaries with control over scarce power, network connectivity, mission-critical equipment, or profitable compute demand may sustain returns. Highly leveraged owners of undifferentiated capacity face a different distribution. Valuation must incorporate both the secular demand story and the capital-cycle supply response.

Cross-asset implications are similarly conditional. Persistent strategic construction can support industrial demand, selected utilities, electrical equipment, and local labor even while high rates pressure property credit and regional banks. It can also keep investment and power demand firmer than conventional cyclical models predict, influencing inflation composition and the equilibrium path of long yields.

Policy should focus on whether investment creates social value without shifting hidden costs to ratepayers or taxpayers. Transparent interconnection rules, appropriate cost allocation, faster permitting, grid investment, and competition can improve outcomes. Subsidies should be evaluated against additional capacity actually induced, not spending that would have occurred anyway.

The central conclusion is disciplined. Since December 2023, data-center construction has moved roughly $4 billion above its baseline while other private construction, after rising in 2024, has fallen to about $10 billion below. This is compelling evidence of relative resilience and concentrated strategic demand. It is not proof that interest rates have ceased to matter.

Rates are one force in an equilibrium shaped by expected cash flows, financing structure, sunk commitments, power scarcity, technology competition, and public policy. Data centers have withstood tightening because their demand shock, sponsor quality, and option value have so far outweighed the higher cost of capital. The rest of private construction shows what happens when that cushion is absent.

The phrase “two economies” is therefore more than rhetoric. It describes an allocation mechanism: scarce capital and resources are flowing toward projects with the strongest perceived strategic returns, while conventional projects fail higher hurdle rates. Whether this becomes a productivity-enhancing reallocation or an overbuilding cycle will depend on utilization, monetization, and the durability of scarcity. The chart captures the divergence’s beginning; returns will be determined by how it ends. The most useful discipline is to keep both sides of the evidence in view. Strong construction confirms that sponsors possess conviction, financing capacity, and access to constrained inputs; it does not by itself confirm that future customers will pay enough to earn the assumed return. Broad weakness confirms that higher financing costs and lower asset values are binding elsewhere; it does not imply every delayed project is economically wasteful. Capital markets will eventually reconcile those claims through prices, cancellations, utilization, and cash flow. Until then, relative resilience should be treated as a measurable fact and permanent immunity as an untested hypothesis.