As AI capital expenditures surge and valuations become increasingly expensive, Goldman Sachs cautions the market: the real risk often emerges when growth rates begin to slow.

In a recent strategy report titled "The Widening and Narrowing of AI Trades," Goldman Sachs Global Investment Research noted significant volatility in AI-related trading, driven by two opposing forces: continued "better-than-expected" capital expenditure from tech giants on one side, and rapidly rising investor concerns about "AI disrupting traditional industry profit pools" on the other.

Buoyed by strong capital expenditure guidance, memory chip stocks have surged an average of 55% year-to-date. Conversely, software stocks have plummeted 24%, weighed down by fears of AI disruption. The same AI theme is manifesting in nearly opposite trends across different segments of the market.

Goldman Sachs categorizes this volatile AI trading into four distinct stages, with their respective stock price trajectories now diverging sharply.

The first stage involves compute leaders, such as NVIDIA, which are facing scrutiny over "over-earning," recently exhibiting a disconnect where profit estimates are revised sharply higher while the stock price remains stagnant.

The second stage encompasses AI infrastructure, including memory, equipment, and servers. Driven by robust capital expenditure guidance from tech giants, these sectors have experienced a sustained rally, with memory stocks up 55% this year.

The third stage includes AI-enabled applications, such as software services. These have faced panic selling due to extreme market fears that AI will disrupt their traditional business models, leading to a 24% plunge in software stocks year-to-date.

The fourth stage involves AI-driven productivity enhancements in non-tech industries. With actual financial returns still unclear, stock prices in this segment have largely moved sideways recently.

Faced with this extreme divergence, the report indicates that both the currently soaring "infrastructure winners" and the plunging "application losers" harbor their own underlying risks.

**Peaking Capex Growth Brings "Multiple Compression" Risk for AI Infrastructure**

The market must first digest another potential upward revision to capital expenditure expectations.

According to consensus estimates compiled by Goldman Sachs, AI capital expenditure by tech giants is projected to reach $667 billion by 2026. This figure is $127 billion higher than estimates at the start of the Q4 earnings season and represents a 62% year-over-year growth rate.

The flip side of these substantial capex increases is the squeeze on free cash flow. The report emphasizes that "super cloud vendors' capital expenditures are approaching levels that would exceed 90% of their operating cash flow this year, a ratio even higher than during the dot-com bubble." More specifically, Goldman Sachs estimates that 2026 capex could consume 92% of these giants' operating cash flow.

To fill this massive funding gap, the giants are forced to significantly cut shareholder returns. Aggregate stock buybacks for these companies were reduced by 15% in 2025; the proportion of cash flow used for buybacks plummeted from 43% in early 2023 to just 16% currently. Simultaneously, companies like Oracle and Google have begun frequenting the bond market for funding.

Goldman Sachs expects the absolute level of capital expenditure could still be revised higher within the year. Since the fiscal years for Oracle and Microsoft end in May/June, the upcoming Q2 earnings season might act as a catalyst for another round of upward revisions to spending expectations.

However, Goldman Sachs warns that the core risk lies not in the "absolute value" but in the "growth rate." "We expect consensus estimates for super cloud vendor capex still have modest upside, but we still project capex growth will peak later this year."

This anticipated slowdown in growth could become the "Achilles' heel" for AI infrastructure stocks.

**H2 Risks for "AI Infrastructure": Slowing Spend and the "Over-earning" Trap**

Goldman Sachs stresses, "Once capital expenditure growth slows, the revenue growth and valuations of some AI infrastructure stocks will appear extremely vulnerable."

The logic is direct: orders, revenue, and profits in the infrastructure supply chain are highly sensitive to the pace of capex growth. When the market narrative shifts from "accelerating every quarter" to "still growing but no longer accelerating," the most fragile part of the valuation is often the "growth premium."

Goldman Sachs states plainly that many AI infrastructure-related sectors have seen significant multiple expansion in recent years. Historical experience shows that investors typically assign lower valuation multiples to companies experiencing slowing growth.

This is the core meaning of the so-called "multiple compression" mentioned in the report's theme: even if profits continue to grow, once the market begins to worry about the sustainability of that growth, multiple contraction can offset the support for stock prices provided by rising earnings.

Among the subsectors listed in the report, manufacturing equipment, servers & networking, foundry & IDM, and power & utilities generally trade above their five-year average valuation multiples.

Goldman Sachs believes the current "latest bottleneck" within infrastructure is concentrated in the memory segment.

The report notes that major memory stocks have risen an average of about 145% since early Q4 2025 and are up about 55% year-to-date. Goldman attributes most of this rally to profit improvements driven by strong demand and price increases.

They also point out that memory stocks trade at an average forward P/E of about 12x, which is below both the broader market and their own five-year averages, suggesting they are not superficially "expensive."

But Goldman Sachs immediately offers a warning using NVIDIA as an example: when the market starts to worry that a company is in an "over-earning" state, the stock price may stop tracking upward revisions to earnings.

From late 2022 to mid-last year, NVIDIA's stock price and earnings grew in sync by 12x, while its valuation multiple remained largely stable. Recently, however, the dynamic has changed.

Goldman points out: "Over the past five months, despite a 37% surge in NVIDIA's forward earnings estimates, its stock price has been essentially flat."

Goldman characterizes this phenomenon as the market psychology of "over-earning": when a company performs exceptionally strongly at a cyclical peak, it can easily invite concerns about intensifying competition and demand sustainability, ultimately manifesting as "profits remain strong, but the valuation contracts."

For trading, this implies that even if infrastructure companies continue to deliver strong short-term results, investors will become more discerning about the "second derivative of growth" and "whether multiples can expand further."

**Near-Term Tech Giant Divergence: Focus Shifts from Capex to "Return"**

Goldman Sachs judges that divergence in performance among tech giants will persist in the near term.

This is because, as quarterly capex growth stabilizes broadly in the first half of 2026, market attention will pivot to the question of "whether AI investments are generating any return."

The report provides a stark comparison: tech giants' free cash flow yield is about 1%, near historical lows, while the rest of the S&P 500 offers about 4%.

As free cash flow weakens and conversion rates decline, capital will naturally seek alternative options. Goldman states plainly that "investors are increasingly allocating capital elsewhere."

**AI Application Layer: A "Fine Line" Separating Winners from Losers**

If the tension in the infrastructure layer is "how much faster can capex go?", the tension in the application layer is "who gets disrupted, and who captures the new revenue?"

Goldman Sachs judges that the diffusion of AI trading to the application layer is a natural path of technological development: once the infrastructure is built, value creation shifts from "selling shovels" to "transforming business models," recouping earlier investments by reshaping profit pools.

However, this makes equity outcomes more "idiosyncratic." Goldman emphasizes that future analysis will rely more heavily on company-specific judgments, such as competitive positioning, barriers to entry, and pricing power.

A key sentence in the report highlights the core uncertainty at the application layer: "Amid an still uncertain ultimate competitive landscape, the line is fine between a company being perceived as an AI revenue 'winner' and one facing 'disruption' concerns."

A direct consequence is that investors are currently not assigning high valuations to the potential for AI-driven revenue growth for many public companies.

Goldman says, "Contrary to our expectations, investors are pricing in little upside from AI-related revenue for public companies; instead, AI applications from private companies are receiving the most attention."

The report cites progress from several private companies: Anthropic launching its Claude Cowork tool; Insurify introducing a comparison application within ChatGPT; Altruist developing tools for personalized tax strategies for wealth management clients.

Such cases reinforce a concern in public markets: even if AI generates new demand, the resulting revenue may not necessarily flow to public companies.

**Why "Losers" Face a Tough Near-Term: Disruption Fears Hard to Disprove Short-Term**

The other side of the application layer is the damaging impact of disruption narratives on valuations.

Goldman Sachs notes that market focus in recent weeks has centered on "AI disruption risk."

The report states that software stocks have fallen about 23% over the past six weeks, and "despite short-term earnings resilience, investors are increasingly questioning the sector's long-term growth prospects."

Goldman offers a very clear judgment here: "Concerns about AI disruption are difficult to disprove in the short term."

They further note: for companies already labeled by the market as "potentially disrupted by AI," stock price stabilization depends first on earnings stabilization; but "this uncertainty around disruption is unlikely to be resolved near-term."

Goldman specifies the conditions for "application layer losers" to recover: "Investors will likely require either multiple quarters of evidence demonstrating business resilience or a more significant valuation discount relative to the market before re-engaging on a large scale."

This is the current dilemma for sectors like software: short-term earnings might be acceptable, but the market is trading based on fears of a "long-term reallocation of profit pools."

**Quantifying Disruption Risk: Two Metrics from Goldman Sachs**

On assessing "who is more vulnerable to disruption," Goldman offers two vectors (while noting other dimensions like regulatory barriers and market power exist).

First, workforce exposure to AI automation. Goldman notes recent increased concern about white-collar job displacement. They collaborated with economists to estimate the proportion of companies' wage bills exposed to AI automation, combined with observing the ratio of "labor costs/revenue."

Goldman cautions this metric is a "double-edged sword": AI can boost efficiency but also replace jobs. In terms of market performance, over the past six months, sectors with lower exposure have been rewarded, while those with higher exposure have been penalized.

Second, tangible asset intensity. Goldman measures this using "(assets - cash - intangible assets) / revenue" and constructs industry-neutral, equal-weight baskets. They observe that companies with "heavier" assets have recently significantly outperformed those with "lighter" assets, to a degree "beyond what macro conditions typically explain." Similarly, goods-producing companies have outperformed service-oriented companies.

For investors, these clues suggest not that "heavier assets are better," but that the market is using them as "proxies for moats/barriers to entry" to counter application-layer uncertainty.

**Three Catalysts: Goldman Pins the "Inflection Point" on Second Half 2026**

Goldman Sachs believes three catalysts are needed for tech giants to regain market leadership.

Their baseline judgment is that these catalysts are "more likely to materialize in the second half of 2026."

First, AI revenue must accelerate. Market reactions during earnings season have proven that whenever revenue growth exceeds expectations, investors regain confidence in the return on AI investments.

Second, visibility on a bottoming in free cash flow resulting from slowing capex growth. Goldman believes that once signs of a cash flow bottom appear, the market may revert to pricing these companies on earnings rather than cash flow, reducing valuation volatility. They explain, "Slowing capex growth will give investors hope for a rebound in free cash flow. This will prompt a repricing based on profitability." Currently, the giants' 24x forward P/E sits only at the 14th percentile over the past decade, making valuations highly attractive.

Finally, the fading of macroeconomic tailwinds. Goldman Sachs economists expect the US economic cyclical acceleration to peak around mid-year and moderate in the second half. As macroeconomic红利 recedes, capital will inevitably flow back towards these tech giants with high long-term certainty.