Every great technology buildout follows the same arc: first the world buys the machinery, then the world learns how to use it. Goldman Sachs equity strategists believe artificial intelligence is approaching exactly that handoff point, and in a new 2027 outlook they are nudging clients to reposition before the crowd does. The opening leg of the AI trade was easy to spot from a distance โ chip designers, data center landlords, power producers, networking vendors, cooling specialists. The next leg is quieter. It belongs to companies that already own unglamorous, tangible things: fuel pumps, gas turbines, delivery vans, warehouse racks, grocery aisles, restaurant kitchens, ski lifts. For these businesses, AI is not the product; it is the wrench that tightens the product. That distinction sounds academic until you follow the money. Infrastructure companies are paid for capacity. Adopters are paid for efficiency, and efficiency flows through an income statement in a way capacity almost never does, because saved cost lands directly in operating margin while added capacity must be filled before it earns a dime. The strategic point is not that AI stops mattering to infrastructure. It is that the marginal dollar of return shifts from the companies selling the picks and shovels to the ones using them to dig faster.
The trigger for the call is a forecast about growth rates, not absolute spending. Goldman expects capital expenditure growth tied to AI to begin decelerating in 2027 โ the first slowdown in the rate of expansion since the current buildout started. Investors routinely misread such warnings as forecasts of collapse. They are not. Decelerating capex growth simply means the second derivative turns negative while total spending keeps climbing to a higher plateau. Markets, however, price second derivatives mercilessly. When a company is valued on the assumption that its addressable market doubles every eighteen months, a shift to mid-teens growth is enough to compress the multiple even if the backlog remains record-setting. That is the mechanical reason rotations happen in technology cycles: the narrative of acceleration is replaced by a narrative of digestion, and capital flees toward whichever set of companies can still show expanding profitability without needing expanding capital budgets. In practical terms, this is a moment to audit your portfolio for how much of its expected return depends on AI infrastructure spending continuing to accelerate, and to ask which holdings would still compound if that spending merely leveled off at a very high number.
There is also a market-microstructure story underneath the fundamental one. Infrastructure winners have spent two years absorbing enormous flows; their valuations now embed a great deal of optimism, and consensus positioning makes them fragile to disappointment. Meanwhile, many of the physical-asset businesses on Goldman’s radar screen remain priced as though the last decade of digital disruption fear will simply repeat itself forever. That creates an asymmetry: limited further multiple expansion required for a good outcome, and a genuine downside cushion from real cash flows generated by hard assets. Rotation phases like this one tend to be driven as much by relative valuation as by absolute fundamentals. When one group trades at a premium built on growth assumptions that are starting to bend, and another trades at a discount built on fear that is starting to look overstated, capital does not need a dramatic catalyst to move โ it only needs permission. A reputable bank publishing a framework that legitimizes the switch can supply exactly that permission. The lesson for individual investors is that the earliest, cleanest gains in a regime change often come not from the new leaders’ brilliance but from the old laggards’ re-rating.
The core of the thesis is what Goldman describes as physically intensive business models โ companies that reach customers through tangible assets rather than through code. The screen looks for firms where hard assets and operating leases represent a large share of the revenue base. At first glance this looks like an anti-technology filter, but it is better understood as an anti-substitution filter. A business built on a fleet, a network, a pipeline, a store footprint, or a distribution route has an economic moat made of atoms: replacement requires enormous capital, permitting, local knowledge, and time. Those same assets, however, are exactly where AI-driven optimization produces outsized operating leverage, because physical operations generate sprawling, messy datasets about routing, maintenance, inventory, and demand that have historically been managed with rules of thumb. Apply better forecasting to a fleet and you cut empty miles. Apply predictive maintenance to a turbine and you cut unplanned downtime. Apply demand modeling to perishable inventory and you cut waste. The asset base is the moat; the data exhaust from that asset base is the opportunity.
Goldman frames part of the appeal as reduced disruption risk, and the logic is worth internalizing. Software can be rewritten by a competitor in a weekend; an insurance underwriting workflow can be partially automated by a well-trained model; a customer service layer can be compressed dramatically. It is far harder to automate the physical delivery of a candy bar to a specific shelf at a specific temperature, or to replicate the snow, lifts, lodges, and local permits behind a ski resort, or to replace the last-mile pump that puts fuel in a tank. AI can attack the administrative skin of these businesses โ scheduling, pricing, paperwork, forecasting, customer contact โ while leaving the physical core intact. That is a crucial asymmetry: disruption risk is bounded because the asset cannot be downloaded, while productivity upside is real because the labor around that asset is expensive and highly exposed to automation. Investors should look for businesses where the labor bill is large relative to revenue but the customer relationship depends on physical presence, since that combination is where efficiency gains turn into margin rather than into price competition.
The bank’s analysts built the list around three overlapping characteristics. First, a high proportion of hard assets and operating leases relative to revenue, which identifies the truly asset-backed operators. Second, presence in industries characterized by physical proximity โ telecommunications services, energy, transportation, retail and food, utilities, and similar sectors where service is delivered in person or through infrastructure. Third, and most interesting, management teams that have explicitly discussed AI improving productivity on earnings calls. That third filter is where judgment matters most. It captures real intent, since executives rarely volunteer operational AI programs unless they are underway, but it also captures performance art, because saying the words is free. The way to separate substance from theater is to listen for specificity: named systems, quantified headcount or hours saved, timelines, capex or opex attached to the initiative, and evidence that the savings show up in reported margins a quarter or two later. Companies that only offer adjectives are telling you they are watching the trend; companies that offer numbers are telling you they are running it.
One detail in the methodology deserves attention because it reveals the mechanism: the final group was filtered to names with at least 30% of wages exposed to AI automation. This is a labor-cost-share heuristic, and it is the engine of the whole idea. If payroll consumes a meaningful slice of revenue, even modest productivity improvements translate into meaningful margin expansion, because the savings do not need to be enormous in absolute terms to matter proportionally. Consider a business where labor is a quarter of revenue and AI-assisted scheduling, dispatch, and forecasting reduce labor hours per unit of output by several percentage points. That improvement compounds across thousands of employees and lands almost entirely in operating income, because no new factory or fleet is required. The nuance is that exposure is not the same as realization. Union agreements, minimum staffing rules, customer service expectations, and the difficulty of redeploying workers all slow the conversion from theoretical automation to actual savings. The best candidates combine high wage exposure with high employee turnover, since voluntary attrition lets a company shrink headcount through hiring restraint instead of layoffs.
Translating the theme into sectors makes it concrete. In telecommunications, AI appears in network optimization, field technician dispatch, proactive fault detection, and churn prediction โ all levers on a cost base dominated by people and physical plant. In energy, the applications run from predictive maintenance on turbines and compressors to seismic interpretation, drilling optimization, and grid balancing, where small improvements in uptime are worth enormous sums. In transportation, the gains come from route optimization, yard management, fuel efficiency, and maintenance scheduling, with margins thin enough that a few percentage points of cost matter enormously. In food and grocery, the money sits in demand forecasting, waste reduction on perishables, and labor scheduling that matches staffing to foot traffic. In retail and restaurants, inventory accuracy, shrink reduction, and dynamic pricing do the work. Notice the pattern: in each case, AI is applied to a cost line attached to a physical operation, not to the creation of a new digital product. Investors screening for these traits should watch operating margin trajectory, revenue per employee, selling and administrative costs as a share of sales, free cash flow conversion, and unit-level metrics like on-time performance, shrink rate, and same-store traffic.
History offers both encouragement and a warning. The 1990s fiber buildout ended with enormous capacity, collapsing prices, and years of pain for the companies that laid the cable โ while the businesses that used cheap bandwidth to build new services prospered later. The cloud capex wave from roughly 2014 onward followed a similar rhythm: infrastructure vendors had spectacular runs, then de-rated as spending growth normalized, while adopters of cloud infrastructure quietly expanded margins for a decade. The shale boom is another useful analog, where service providers captured the early cycle and operators captured the later one. The warning is that capex deceleration can mean two very different things. It can signal digestion, where demand is strong but the buildout has reached sufficiency, which is bullish for adopters. Or it can signal an air pocket, where the economics of the buildout itself have broken down, which is bearish for nearly everyone including adopters, because a broader growth scare would hit cyclical physical businesses hardest. Distinguishing between the two requires watching order books, power contracts, utilization rates, and pricing, not just the capex line.
A clear-eyed view must also catalog the risks. The first is the value trap: cheap-looking asset-heavy businesses can stay cheap for years if returns on capital are structurally poor, and physical intensity often comes with capital intensity, debt, and interest-rate sensitivity. The second is AI-washing, where management teams claim productivity transformation to satisfy an analyst question and never quantify the result. The third is execution risk, because operational AI programs require data hygiene, integration work, process redesign, and change management โ the unglamorous part that consumes budgets without producing headlines. The fourth is labor and regulatory pushback, since automation in visible, unionized, or politically sensitive industries invites scrutiny and can slow savings. The fifth is cyclicality: trucking, energy services, restaurants, and retail all face demand cycles that overwhelm productivity gains in a downturn. And the sixth is that some of this optimism may already be priced in, which is why entry discipline and staggered purchases around earnings dates matter more than conviction alone.
Practical portfolio construction follows from those risks. Rather than buying a basket indiscriminately, build a small, deliberately diversified set of positions across at least three distinct physical sectors so that no single demand cycle dominates returns. Size individual names modestly โ a few percent of an equity allocation โ because the thesis is idiosyncratic and depends on company-specific execution. Stagger entries across two or three quarters so that earnings reports reveal whether margin improvement is genuinely appearing, and treat each report as a checkpoint rather than a formality. Prefer companies that disclose measurable AI outcomes over those that merely reference the technology, and favor management teams with a track record of delivering prior efficiency programs, since operational discipline is transferable. Watch for a clear early tell: a company whose revenue grows modestly while operating margin expands and headcount stays flat is executing; one whose revenue grows modestly while headcount grows just as fast is narrating. Finally, set explicit sell criteria in advance, such as margin targets that fail to materialize within four quarters, so that thesis drift cannot quietly become hope.
Here is the actionable playbook in five steps. First, screen for physical intensity using asset and lease ratios alongside labor cost as a share of revenue, then narrow to sectors where delivery requires proximity. Second, verify the AI claim by reading transcripts for numbers, not adjectives, and by checking whether the efficiency shows up in reported margins within two reporting cycles. Third, size positions conservatively and diversify across sectors to neutralize cyclical risk. Fourth, monitor a fixed dashboard each quarter โ operating margin, revenue per employee, administrative cost ratio, free cash flow conversion, capital spending relative to sales โ and compare each company against its own history rather than against peers. Fifth, rebalance deliberately as the cycle matures, trimming names that have re-rated past their demonstrated improvement. The broader takeaway is that the AI story is graduating from a hardware narrative to an operating-margin narrative, and the winners of that second phase will look nothing like the winners of the first. They will be companies you can drive to, walk into, or plug into โ and the investors who identify them before the label becomes consensus will be the ones who capture the compounding.