NextFin News - The AI trade is facing a harder question than whether companies still want artificial intelligence: can financing, permitting and earnings support the scale of investment investors have begun to expect? Goldman Sachs Asia-Pacific equity strategist Timothy Moe highlighted pressure points around Nvidia, US politics and opposition to data centers in a Nov. 20 interview, while Goldman’s own research shows that AI investment remains exceptionally large. The market debate is shifting from whether AI demand exists to who funds the buildout, who captures the returns and how quickly physical capacity can be delivered.
That distinction matters because the AI trade can weaken without AI adoption collapsing. Spending may continue to rise while the cost of capital rises faster, approvals take longer or customers demand clearer evidence of returns. Moe’s focus points to a market entering that second phase. The key risk is not that artificial intelligence stops mattering; it is that the investment chain becomes less synchronized than the valuation narrative assumes.
The Demand Story Is Still Large, but the Bar Has Moved
Goldman Sachs Research put the 2026 consensus estimate for capital spending by large hyperscaler AI companies at $527 billion, up from $465 billion at the start of the third-quarter earnings season. The upward revision matters because it shows that spending expectations continued to rise even after investors had absorbed years of aggressive data-center investment.
Goldman also estimated that AI capex recently equaled 0.8% of gross domestic product. That is large enough to affect equipment suppliers, utilities, construction and semiconductor demand, but it remains below the 1.5% of GDP or more reached during earlier technology booms. On Goldman’s comparison, hyperscaler capex would need to reach $700 billion in 2026 to match the peak intensity of the late-1990s telecommunications investment cycle.
Those figures support a durable demand thesis without proving that every AI-linked equity can preserve its valuation. Spending is an input, not a return. The same dollar of capex can support several suppliers during construction, while the eventual economic profit may concentrate in a much smaller group of platforms and infrastructure owners. That is why concern about Nvidia and other AI-related stocks can intensify even while aggregate demand remains high: the debate has shifted from volume to distribution.
The pace of spending shows why the bar is rising. Hyperscalers spent $106 billion on capital expenditures in the third quarter of 2026, including AI and non-AI projects, up 75% from a year earlier. Analysts expected growth to slow to 49% in the fourth quarter and 25% by the end of 2026. A move from 75% growth to 25% growth is not a collapse in absolute investment, but it changes valuation mathematics. High-multiple companies need growth that arrives early enough and lasts long enough to justify the premium.
Goldman’s broader estimate adds geographic complexity. Its research said a commonly cited $794 billion forecast for hyperscaler spending could understate total global AI investment by about $200 billion while overstating US AI investment by roughly the same amount. The opportunity may therefore be broader than a narrow group of US technology companies, but the capital cycle may also be more exposed to different regulators, utilities and industrial bottlenecks.
The first tension is clear: the buildout can be economically real and still produce an unstable equity trade. The question is who pays, who earns and who gets permission to build.
Financing Is the Transmission Mechanism
The direct effect of AI uncertainty is usually described as weaker sentiment toward technology stocks. The deeper mechanism runs through financing. Data centers, power capacity and network equipment require large upfront outlays, while the revenue benefits arrive over time and are distributed unevenly among cloud providers, model developers, chip companies and customers.
When financing is cheap and investors accept long payback periods, that mismatch is manageable. When bond issuance rises and credit investors demand more compensation, each project must clear a higher hurdle. Goldman commentary on AI uncertainty said credit issuance in 2026 was about three times the 2025 level in the cited discussion, with supply increasingly outweighing demand and credit spreads widening. The point is not that AI demand has failed. It is that the capital structure supporting the demand is becoming more contested.
The first-order effect is higher funding costs and a lower present value for long-duration growth. The second-order effect is more consequential: selective financing favors companies with strong balance sheets, secured capacity and clearer cash flows, while marginal or highly leveraged projects are delayed. Capital can migrate toward the largest platforms and most defensible infrastructure even as the wider AI narrative loses momentum.
That creates a cross-market feedback loop. Wider credit spreads increase the cost of data-center construction; slower construction delays equipment orders; delayed orders pressure supplier earnings estimates; weaker earnings confidence raises the equity risk premium; and a higher equity risk premium makes new financing more expensive. The loop can weaken AI-related equities without a fundamental reversal in long-term adoption.
It also changes the meaning of a capex beat. Early in the cycle, higher spending is read as proof of demand. Later, it can be read as evidence that platforms must spend more to defend their position. The same dollar can signal opportunity or defensive escalation depending on the return on invested capital that investors expect.
That is why the financing debate is more important than a simple headline about “AI uncertainty.” If the market concludes that customers are demanding capacity but the providers cannot earn an adequate return on it, higher spending becomes less bullish. Investors then distinguish between companies that own scarce capacity, companies that finance the buildout and companies that depend on the next order.
The market’s expectation gap is therefore measurable through the relationship between capex and monetization. Spending growth of 75% in the third quarter can support a premium when revenue and margins validate it. The expected slowdown to 25% by the end of 2026 requires a different proof: returns must improve even as the physical growth rate moderates. Without that proof, the market can lower multiples before it lowers demand estimates.
Permitting and Politics Could Turn a Cycle Into a Bottleneck
The immediate market pressure is cyclical, but the permitting problem could become structural. Higher volatility, tighter credit and changing expectations can mean-revert as earnings catch up. A shortage of power interconnections, local opposition to data centers or a slower approval regime does not automatically self-correct.
The interview’s focus on US political developments and opposition to data centers matters because AI infrastructure is physical. Model training and inference require land, electricity, cooling systems, transmission capacity and fiber connections. Software demand can scale rapidly; a data center cannot. The investment cycle therefore depends on local authorities, utilities and communities whose incentives differ from those of technology investors.
That creates a bottleneck between announced capex and productive capacity. A company can raise its spending plan, but it cannot instantly turn that plan into megawatts, permits and operating facilities. If approvals slow, the market may see a temporary gap between demand forecasts and revenue realization. Suppliers exposed to the buildout could experience order-timing volatility even if the eventual need for capacity remains intact.
The strongest bullish counter-thesis is that scarcity protects the trade. If power and advanced chips remain constrained, companies that secure capacity can preserve pricing power. Goldman’s comparison with earlier technology booms reinforces the argument that current AI capex, at 0.8% of GDP, remains below the 1.5%-plus peaks reached in earlier cycles. From this perspective, investors may be overreacting to a normal pause before a much larger infrastructure phase.
That counter-thesis attacks the cautious view at its foundation, and it is credible. Demand can remain stronger than supply for years. Large platforms may continue investing because the cost of falling behind is greater than the cost of excess capacity. A temporary valuation reset could then create room for the next expansion rather than mark the end of the cycle.
But scarcity does not answer the allocation question. It can benefit incumbent platforms while hurting marginal projects. It can move value from chip vendors toward utilities, contractors and owners of permitted power capacity. The existence of demand does not guarantee that every AI-linked asset receives the same share of the economics.
The structural test is observable. If approval times, grid-connection queues and power prices continue to rise while hyperscaler capex remains elevated, the bottleneck is no longer only a sentiment event; it is a supply-side constraint that can reshape returns across the chain. If approvals normalize and new capacity comes online without a material rise in power costs, the structural thesis weakens and the problem remains mainly cyclical.
What the Nvidia Concern Says About Market Leadership
Concern about Nvidia is a proxy for a broader question: can the market’s leading AI suppliers maintain earnings growth as customers become more selective? Nvidia sits near the center of the buildout because accelerator demand links hyperscaler budgets to semiconductor revenue. That position also makes the company a focal point for any debate over inventory, customer concentration, competition or the pace of capital deployment.
The first-order reading is that doubt about Nvidia threatens the entire AI trade. The second-order reading is more nuanced. If the leader’s earnings remain strong but its valuation multiple contracts, capital may rotate toward less crowded beneficiaries rather than leave the AI complex entirely. If Nvidia’s results weaken because hyperscalers cut orders, the effect would travel backward through memory, networking, equipment and data-center construction.
This is why a single theme can show dispersion. Hardware suppliers may benefit from near-term orders, while software and enterprise applications wait for customers to prove a return on investment. Large platforms may keep spending to protect strategic position, while smaller firms face a higher cost of capital. The trade becomes less about “AI” and more about duration, balance-sheet strength and bargaining power.
Goldman’s capex data provides a valuation discipline. Spending growth is forecast to slow from 75% year over year in the third quarter to 25% by the end of 2026. If revenue and profit growth do not replace that deceleration, investors can demand a lower multiple even if spending remains historically high. The risk is not necessarily that AI investment stops; it is that the marginal dollar produces less incremental earnings than the market assumed.
The most important falsifying signal for this cautious interpretation would be renewed acceleration in hyperscaler capex growth alongside stable or rising returns on invested capital. If quarterly capex growth remains near the third-quarter 75% pace rather than moving toward the expected 25% year-end rate, and if management teams show measurable monetization, the argument that the trade is entering a financing-led digestion phase would be wrong.
Outlook: Three Paths for the AI Trade
Over the short term, the AI trade will remain sensitive to headlines about Nvidia, data-center approvals, credit spreads and political developments. That is a liquidity and sentiment horizon. Equity prices can react before end demand changes because they discount the timing and financing of future cash flows.
Over the medium term, earnings delivery will decide whether the trade broadens or narrows. The base case is digestion rather than collapse: hyperscalers keep spending, but growth slows toward the 25% year-end rate cited in Goldman’s research; financing becomes more selective; and returns concentrate in companies with balance-sheet capacity, secured power and visible customer demand. Under that scenario, AI remains a real investment cycle but leadership becomes more uneven.
The upside case requires two developments together. Spending would need to remain near or above the $527 billion consensus estimate, while enterprise customers show enough productivity or revenue gains to validate the investment. The market could then look beyond the financing debate and treat the cycle as an early infrastructure phase rather than a late-stage bubble.
The downside case is not simply that AI demand disappears. It is that capex stays high while monetization disappoints, credit issuance expands faster than demand and permitting delays prevent new capacity from becoming productive. Under that path, long-duration growth assets could be marked down even as physical AI investment continues. Suppliers with concentrated exposure to a few platforms would be most vulnerable to order timing and bargaining pressure.
Over the long term, the structural question is whether AI infrastructure becomes a durable layer of the economy or an investment wave whose returns are captured mainly by a narrow group of platforms. Goldman’s estimate that global AI investment may exceed the commonly cited hyperscaler figure suggests that the economic footprint is broad. Broad investment, however, can also mean broad competition, and broad competition can compress returns.
The decisive indicators are not limited to chip demand. Investors will need to track hyperscaler capex against revenue growth, credit issuance and spreads, data-center approval and grid-connection timelines, power costs and evidence of enterprise monetization. A demand story becomes an earnings story only when capacity turns into cash flow.
Moe’s warning is best understood as a warning about the quality of growth, not the existence of AI. The near-term pressure is cyclical at the valuation and financing level; infrastructure constraints could become structural if they delay supply for long enough to change who captures the economics. That combination means the next phase will require proof of returns rather than another promise of spending.
The AI trade is not breaking because demand vanished; it is being repriced because the cost, permission and ownership of demand now matter as much as its existence.
Explore more exclusive insights at nextfin.ai.
