Goldman Sachs Says the AI Trade Is Entering a Consolidation Phase

Goldman Sachs says the artificial intelligence trade has entered a consolidation phase after a sharp reversal exposed crowded positions, leveraged funds, and fragile market plumbing. The distinction matters. Consolidation suggests that investors are resetting expectations after an extended rally, rather than abandoning the earnings and infrastructure demand supporting AI companies.

A short item distributed through rsshub 36kr summarized the bank’s view that strong earnings and reduced investor positioning should ease extreme volatility. The underlying argument is more nuanced than a simple call to buy the dip. Goldman’s strategists see improving conditions, but they also acknowledge that leverage can turn an ordinary decline into forced selling.

That creates the central tension for investors. AI demand and corporate spending remain substantial, while the vehicles used to trade that demand have become increasingly aggressive. The resulting market can look calm at the index level even as individual semiconductor, memory, networking, and data-center stocks swing violently.

The recent decline therefore tests two competing explanations. One says the AI investment case has weakened because spending is too high and valuations ran too far. The other says a fundamentally intact theme became overcrowded, then suffered a positioning reset that cleared some excess without ending the broader trend.

Goldman currently leans toward the second interpretation. Its evidence includes continued earnings growth, lower hedge-fund exposure, and sustained demand for computing infrastructure. However, the next earnings cycle must show that enormous capital budgets are producing revenue and profit, not merely more capacity.

The Sell-Off Changed Positioning, Not the Entire AI Thesis

Goldman’s core judgment is that the recent decline removed excess positioning faster than it damaged the fundamental AI investment case.

Technology stocks entered the reversal with investors heavily exposed to many of the same themes. Those positions included semiconductor producers, memory suppliers, optical-networking companies, data-center equipment makers, and the hyperscalers financing new infrastructure.

Crowding does not automatically mean that an investment thesis is wrong. It means many participants own similar assets for similar reasons. When volatility rises, those investors can respond together, creating a wave of selling that overwhelms company-specific information.

Goldman’s prime-brokerage data offers a window into that process. Vincent Lin, co-head of Prime Insights and Analytics at Goldman Sachs, said hedge funds’ gross exposure had reached five-year highs in early June. Net exposure had also risen to a four-year high.

After the retreat, both measures fell to roughly the 60th to 65th percentile of their three-year ranges. That left positioning near the middle of its recent distribution, rather than at an extreme.

Lin described the episode as a healthy reset, while stressing that positioning was not completely washed out. Hedge funds had taken risk off the table, but they had not collectively rejected AI infrastructure.

That distinction helps explain why the rsshub 36kr summary described a possible consolidation instead of a lasting collapse. A consolidation occurs when prices digest earlier gains and investors reassess expectations. It can include sharp declines, failed rebounds, and wide differences between individual stocks.

The decline also followed a familiar pattern in momentum-driven markets. A successful theme attracts more capital, rising prices support greater risk-taking, and new investment products make exposure easier to obtain. When the direction changes, those same forces operate in reverse.

A striking example came from Situational Awareness, the AI-focused hedge fund founded by former OpenAI researcher Leopold Aschenbrenner. Axios reported that the fund sold its public-equities portfolio to Citadel during the rout. The fund had recently reported about $20 billion in assets under management.

That sale does not establish what happened across the entire market. It does show how quickly an AI-focused portfolio can become part of the event it was designed to trade. A concentrated investor facing losses or redemptions may need liquidity regardless of its long-term beliefs.

The reset was also unusually broad within technology. Goldman characterized the selling as the greatest cumulative reduction in technology exposure within the history of one of its datasets. Yet the bank’s conversations with hedge funds still indicated fundamental interest in AI companies.

This is why the recent action cannot be interpreted through prices alone. Falling prices might reflect lower earnings expectations, weaker demand, excessive valuations, forced deleveraging, or several factors at once. Goldman’s positioning data gives the deleveraging explanation greater weight, but it does not eliminate the others.

The resulting market is more selective. Investors can no longer treat every supplier connected to AI infrastructure as an interchangeable beneficiary. They must distinguish companies with visible orders, durable margins, and pricing strength from those supported mainly by thematic demand for their shares.

Strong Earnings Give Goldman’s Consolidation Case Its Foundation

The bullish side of Goldman’s argument depends on profits carrying more of the market than expanding valuation multiples.

In May, Goldman’s Shawn Tuteja said S&P 500 earnings per share were tracking 17% above the prior year. That marked a sixth consecutive quarter of double-digit earnings growth. Expectations for the following 12 months had risen to 13%.

At the same time, the market’s forward valuation multiple had compressed from 22 times earnings to 21 times. A 21-times multiple is not historically inexpensive, but the combination suggested that earnings were doing more work than simple investor enthusiasm.

The AI portion of the market contributed heavily to those results. Hyperscalers were continuing to buy chips, networking equipment, power systems, cooling technology, and construction capacity. Their spending supported revenue across a supply chain extending well beyond model developers.

Tuteja said hyperscalers had committed to $755 billion of capital expenditure during 2026, representing 38% annual growth. Goldman presented those figures as evidence that demand remained active across AI infrastructure.

Capital expenditure, or capex, is money used to acquire long-lived assets such as servers, data centers, and electrical equipment. It supports suppliers immediately, but it creates a harder question for the companies funding it. Those buyers eventually need returns that justify the spending.

The market has already moved through several layers of this buildout. The first major beneficiaries were advanced semiconductor companies. Attention then expanded toward cloud platforms, data centers, power suppliers, memory producers, optical networking, and specialized cooling systems.

Goldman said the optical-networking theme had gained more than 100% during 2026 by mid-May. Liquid-cooling shares, another equipment category highlighted by the bank, had risen about 30%. Such gains demonstrate demand, but they also increase the risk that prices anticipate years of flawless execution.

The case for durable spending now increasingly rests on actual AI use. Goldman’s investment-banking researchers said enterprise adoption was broadening beyond chatbots and coding tools. Agentic systems, which can execute multistep work rather than only produce an answer, were increasing computing requirements.

One company interviewed by Goldman said its top 5% of enterprise customers consumed three times as many tokens as the median customer. Tokens are the small text units that models process and generate. Rising token use generally requires more inference capacity, which is the computing used when deployed models respond to users.

Goldman argues that constrained supply can stretch infrastructure spending across several years. Its enterprise adoption research points to sustained demand as companies move from experiments toward operational workflows.

Still, supplier revenue and buyer returns are not the same thing. A chip manufacturer can benefit from large orders even if a cloud provider later struggles to monetize the resulting capacity. The present earnings strength validates infrastructure demand, but it does not settle the economics of the entire system.

This is the most important limitation in the rsshub 36kr account. Strong current profits can support a consolidation thesis, but markets price future cash flows. Investors will keep asking whether AI revenue can rise quickly enough to cover depreciation, energy, financing, and development costs.

The next stage should therefore produce a wider performance gap. Companies selling scarce components may retain pricing power. Businesses relying on generalized enthusiasm will face greater scrutiny. Hyperscalers must show that new services, advertising improvements, cloud workloads, and internal efficiency gains are turning capex into measurable returns.

Why Leveraged ETFs Can Turn a Pullback Into a Rout

The recent volatility was not only a verdict on AI companies; it was also a consequence of how investors packaged and financed their exposure.

Leveraged exchange-traded funds promise a multiple of an asset’s daily return. A two-times semiconductor ETF, for example, seeks to deliver twice the underlying index’s daily move. The fund must rebalance frequently to maintain that target.

This creates a mechanical feedback loop. When the underlying assets rise, the fund often needs to buy additional exposure. When they fall, it needs to sell. Traders describe this behavior as short gamma because the rebalancing tends to amplify the prevailing direction.

Tuteja warned about this structure before the latest sell-off. He explained that a stock expected to fall 3% after negative news might instead decline 10% when leveraged products and other forced sellers reduce exposure together. The same mechanism can magnify gains during a rally.

The risk grows when leveraged funds hold volatile securities with limited trading depth. A small product creates little market impact. A large product forced to transact near the close can move prices, particularly when several funds follow similar mandates.

Goldman’s Brian Garrett later said one in five ETFs included some leveraged or inverse component. Among ETFs issued during 2026, the share was one in three. That growth made the market more aggressive and increased the potential footprint of automatic rebalancing.

The problem was especially visible outside headline US indexes. Garrett observed that implied volatility in South Korea’s KOSPI had risen above levels seen during the global financial crisis. The index had become heavily influenced by a small number of volatile stocks tied to the AI supply chain.

Concentration can conceal this stress. An index combines many securities, so gains in one group can offset losses elsewhere. Low index volatility can coexist with extreme movement in individual shares, creating what Garrett called a difference between the stock market and a market of stocks.

In that environment, an investor watching only the S&P 500 or Nasdaq may underestimate the pressure beneath the surface. Semiconductor, memory, and equipment shares can experience large daily moves while diversified indexes appear comparatively stable.

The ETF volatility discussion also clarifies why cleaner positioning can reduce future turbulence. When hedge funds lower exposure and leveraged products lose assets, fewer participants must sell during the next decline.

However, cleaner does not mean clean. Lin said post-sell-off hedge-fund exposure remained around the middle of its recent range. Investors also reduced their macro hedges while selling AI holdings, a process he described as de-grossing.

A macro hedge is a position designed to offset broad market risk, such as an index put or short futures position. If investors remove both long positions and protective hedges, their balance sheets shrink. That lowers immediate exposure but can leave the remaining market sensitive to a renewed rise in correlations.

Correlation measures how closely securities move together. Low correlation benefits stock pickers because company-specific developments matter more. Rising correlation indicates that investors are again treating many securities as expressions of one shared risk.

This mechanism supports Goldman’s expectation that volatility should ease after the reset. Less crowded positioning reduces the fuel available for forced selling. Yet the same market structure remains in place, including leveraged ETFs, options activity, concentrated indexes, and highly connected AI supply chains.

The consolidation thesis therefore does not promise calm trading. It suggests that the probability of repeated, self-reinforcing declines has fallen because investors have already removed substantial risk. A fresh negative catalyst could still reactivate the machinery.

The Real Opponent Is Earnings Delivery Versus Positioning Excess

The decisive contest is not AI bulls against AI skeptics; it is operating performance against the excess expectations embedded in crowded trades.

Goldman’s constructive view works only if earnings keep validating the infrastructure cycle. Lower positioning can stabilize prices, but it cannot create revenue, improve margins, or generate returns on invested capital.

The AI trade includes businesses with very different economics. Nvidia and other chip suppliers sell constrained computing equipment. Cloud providers finance data centers. Utilities and industrial companies provide power and cooling. Software developers seek revenue from models and applications.

A single label can obscure these differences. Some infrastructure suppliers recognize revenue as equipment ships. Hyperscalers bear years of depreciation and operating expenses. Application companies must persuade customers to pay enough for AI services to cover inference costs.

This separation becomes more important during consolidation. In a broad rally, investors often reward exposure to a popular theme. In a slower market, quarterly results and guidance create sharper distinctions between companies.

Goldman’s May assessment offered a strong starting point. Earnings growth remained high, and AI-related demand appeared resilient despite inflation and weaker economic growth. Tuteja even described AI as comparatively resistant to pressures affecting consumer and cyclical companies.

That defensive interpretation has limits. AI spending is financed by companies with strong balance sheets, but it remains discretionary capital allocation. Executives can delay projects if expected returns fall, financing becomes more expensive, or electricity constraints impede deployment.

Interest rates add another layer. Higher long-term bond yields increase the discount rate applied to future earnings. This effect can weigh most heavily on companies whose valuations depend on profits expected many years from now.

Tuteja highlighted the 30-year Treasury yield crossing 5% in May as an important psychological threshold. The precise level changes with the market, but the underlying relationship remains relevant. Expensive capital places greater pressure on companies to demonstrate near-term returns.

The skeptical case also focuses on concentration. AI-related companies account for a large share of major US indexes, which means disappointing results from a few businesses can affect passive investors far beyond dedicated technology portfolios.

Goldman has previously warned that narrow leadership and rapid momentum gains often precede higher volatility and weaker returns. That does not identify a market peak, but it challenges the assumption that strong index performance proves broad economic health.

The rsshub 36kr summary emphasizes improving prospects after the shakeout. A more complete reading is conditional. Reduced hedge-fund and ETF exposure removes one source of instability, while earnings must still overcome demanding valuations and immense capital requirements.

There is also a verification gap around the latest strategy note. The public newsflash provides a brief paraphrase, while Goldman’s public market discussions supply related data and reasoning. Without the complete client note, readers should avoid treating every detail as a firm forecast for all AI stocks.

Goldman itself distinguishes institutional views from individual market commentary. Its public transcripts state that speakers’ opinions can change and do not necessarily represent the bank as a whole. They are also explicitly presented as information, not investment advice.

That caution matters because different Goldman teams examine different parts of the market. Prime-brokerage specialists focus on hedge-fund positioning. Volatility traders examine options and ETF flows. Equity strategists assess earnings, valuations, and portfolio construction.

Their views can still form a coherent picture. Positioning had become extreme, market structures magnified reversals, and earnings remained comparatively supportive. After deleveraging, the balance between fundamentals and technical pressure became less unfavorable.

Yet a coherent explanation is not a guarantee. A consolidation can resolve through renewed gains, a prolonged sideways market, or another decline. The outcome depends on whether profits catch up with expectations before a new macroeconomic or company-specific shock arrives.

Readers should therefore resist two easy narratives. The sell-off did not prove that AI investment had failed. Strong infrastructure spending does not prove that every AI-linked valuation is justified.

What the Goldman Sachs AI Outlook Still Cannot Settle

Three unresolved issues can weaken the consolidation call: capex monetization, persistent leverage, and rising correlations across AI assets.

The first issue is monetization. Hyperscalers must show how infrastructure spending translates into cloud revenue, advertising gains, subscription growth, or lower internal costs. Broad claims about future productivity will not provide enough evidence.

Investors should compare capital expenditure with incremental revenue and operating income. They should also watch depreciation, because servers and networking equipment create expenses over their useful lives. Rapid hardware replacement can make this burden more significant.

The accounting timing can temporarily flatter the story. Cash leaves when companies build data centers, while depreciation spreads through future income statements. Revenue may arrive gradually or remain difficult to attribute directly to AI.

The second issue is leverage. Hedge funds reduced exposure, but Goldman said positioning was not washed out. Leveraged ETFs and options activity also remain structural features of the market, rather than temporary participants.

A calmer period can attract risk back quickly. If volatility declines and prices recover, funds may rebuild positions, retail traders may return to call options, and leveraged products may gather new assets. That process can recreate the same vulnerability under different prices.

The third issue is correlation. Lin identified rising stock correlation as a key signal to monitor after the de-risking. If correlations remain low, earnings and company-specific execution can drive returns. If they rise sharply, macro pressure may again overwhelm fundamental distinctions.

The recent liquidation of a high-profile AI portfolio demonstrates this risk. Axios reported that Situational Awareness sold its public-equities holdings to Citadel amid heavy losses in AI stocks. Its experience does not predict another liquidation, but it shows how concentrated exposure can transmit stress.

The AI fund sale also complicates the idea that reduced positioning is automatically bullish. Some selling improves future market balance. Other selling can signal that losses, financing pressure, or redemptions remain unresolved.

Another uncertainty concerns the composition of demand. Training frontier models requires enormous clusters, while inference spreads computing needs across deployed products. A shift toward inference can broaden demand, but it can also favor different hardware and software suppliers.

Efficiency improvements create an additional tradeoff. Better chips, optimized models, and cheaper inference can lower the cost of each task. Lower costs may expand usage, but they can also reduce revenue expectations for suppliers if demand does not grow quickly enough.

Power availability can restrict both sides. Data centers need electricity, transmission capacity, cooling, and permits. Strong chip orders do not guarantee that completed systems can be energized on schedule.

Geopolitical policy remains another variable. Export controls, tariffs, and US-China technology restrictions can alter demand for advanced processors and manufacturing equipment. They can also accelerate alternative supply chains that compete for capital.

None of these uncertainties invalidates Goldman’s conclusion that conditions have improved. They explain why improved conditions should be described as a better setup, not a resolved investment case.

The best evidence for the consolidation thesis will come from narrower price reactions. If companies with strong orders and improving margins outperform weaker peers, the market is returning to fundamental discrimination. If the entire group continues moving together, positioning still dominates.

Three Signals to Watch During the Next Three Months

The next phase will be decided by hyperscaler returns, hedge-fund exposure, and the correlation between individual AI stocks.

The first signal is capital-expenditure conversion. Upcoming hyperscaler results should disclose spending plans and provide evidence that AI capacity is producing revenue or measurable efficiency gains.

Investors should focus on changes in cloud growth, AI service demand, operating margins, and management guidance. Rising capex alongside improving revenue would strengthen Goldman’s consolidation thesis. Rising capex without clearer returns would weaken it.

The distinction is important because infrastructure suppliers can post excellent results before their largest customers establish attractive economics. The market needs evidence from both sides of the transaction.

Goldman’s own traders identified hyperscaler earnings as their central near-term test. Lin said he was watching spending trajectories and progress converting capex into incremental revenue and profit. That is a more demanding standard than simply maintaining investment.

The second signal is whether hedge funds rebuild exposure. The post-sell-off range of roughly the 60th to 65th percentile indicates room for buying, but it also means the market is not starting from an exceptionally defensive position.

A gradual increase supported by earnings would reinforce the idea that the reset cleared weak leverage. A rapid return toward previous highs would recreate crowding before the fundamental evidence had time to improve.

ETF assets and options positioning deserve attention alongside hedge funds. Continued growth in leveraged and inverse products would increase the amount of mechanical trading around volatile securities.

The third signal is stock correlation. Low correlation would indicate that investors are separating winners from weaker businesses. It would also give active managers more scope to trade company-specific fundamentals.

A sharp rise in correlation would suggest that the AI complex is again behaving as one leveraged macro position. In that environment, even strong companies can decline because investors sell liquid holdings to reduce total risk.

These signals should be read together. Better hyperscaler economics can justify renewed exposure. Moderate exposure can support orderly price discovery. Low correlation can allow earnings quality to determine which stocks recover.

The opposite combination would be hazardous. Weak monetization, rapidly returning leverage, and rising correlation would undermine the claim that the recent rout was merely a healthy reset.

The rsshub 36kr newsflash captured the optimistic direction of Goldman’s view, but the bank’s public analysis supplies the necessary conditions. Strong earnings and cleaner positioning improve the setup. They do not remove valuation, leverage, or execution risk.

For developers, enterprise buyers, and AI product users, this market debate has practical consequences. A sustained infrastructure cycle can lower inference costs and expand available capacity. A spending slowdown can delay data centers, constrain access, and push providers to prioritize profitable workloads.

Enterprise teams should therefore watch product economics alongside financial markets. Growth in real deployments offers firmer evidence than stock momentum. Goldman’s thesis becomes more convincing when expanding token use produces recurring revenue and measurable business outcomes.

The AI trade does not need another uninterrupted rally to remain intact. It needs a quieter market in which earnings can catch up with expectations. Over the next three months, watch whether capital spending produces returns, whether leverage rebuilds responsibly, and whether individual stocks begin trading on their own results.