Goldman Sachs on the AI boom: Strong earnings will outweigh valuation concerns before the investment cycle peaks, but volatility will rise further

Author: Zhuifeng Trading Desk, Wall Street CN The AI rally is not a simple replay of the 1999–2000 bubble. Goldman Sachs believes the more critical question now is that while earnings and capex are still being revised upward, market prices have already priced in a great deal of optimistic expectations, and investors' sensitivity to narrative shifts is rising. According to the Zhuifeng trading desk, Goldman Sachs judged in a June 22 research note that the AI investment boom may continue, and near-term market expectations for its scale may even need to be revised higher. However, the note also pointed out that a large amount of value has already been priced in ahead of time, making the market more vulnerable to any news that challenges the optimistic AI narrative. The primary risk for AI trades is no longer just a "valuation bubble." Forward P/E ratios have not clearly spun out of control, because earnings expectations have been revised upward in tandem. What really needs to be tested is whether the current strong earnings can be sustained after the capex cycle peaks. For investors, strong earnings may continue to overwhelm valuation concerns before the peak of the AI investment cycle appears. But as incremental market cap becomes increasingly dependent on optimistic assumptions, stock volatility could rise further, and the value of downside protection is also increasing.

AI is not 1999, but the market has already run ahead of the macro picture

Goldman's core judgment is that today's AI cycle is not like 1999–2000, which was built on a combination of extreme valuation expansion, macro overheating, and financing imbalances. Currently, fundamentals have not deteriorated significantly and are even still strengthening. AI-related corporate earnings are strong, capex plans continue to be revised upward, and the market therefore has reason to keep buying related assets. Compared with the late 1990s, forward valuations have not seen the same degree of runaway expansion. But this does not mean risk is low. The market cap growth of AI-related companies has already significantly exceeded what benchmark macro income calculations would justify. To explain current prices, one must assume that AI winners can capture above-normal productivity dividends over the long term. In other words, the core bet of the current market is not that "valuations can expand indefinitely," but that "super-high earnings can be sustained."

What truly resembles the 90s is investment intensity; other bubble signals have not yet appeared in sync

The late-stage 1990s tech bubble had four typical signals: investment staying at abnormally high levels, declining macro profit margins, rapidly rising corporate financing needs and leverage, and a widening current account deficit. Currently, the only signal that has clearly emerged is the first one—accelerating AI capex. The research note states that tech investment as a share of GDP has already broken through the highs of the 1990s, and the pace of increase is even faster. Hyperscaler expectations for 2026 capex have risen nearly 80% compared to six months ago. Following the current trajectory, AI-related investment could approach, or even exceed, the peak of the 1990s tech investment boom in the coming years. However, this capex cycle still differs from that era. First, its duration has not yet reached the length seen in the late 1990s. Second, its breadth is not as wide. 1990s tech investment was more like an economy-wide expansion, whereas today's AI capex is more concentrated in hyperscale cloud providers, semiconductors, and related infrastructure chains. The most critical contrast at the macro level lies in profits. In the late 1990s, corporate profit margins peaked and began to decline after 1997, as rising wages and unit labor costs eroded profits. The current situation is different: corporate profits as a share of GDP remain near highs, and productivity growth has not been fully offset by a similar acceleration in wages as it was back then. Corporate financing is also not following the same path. Hyperscaler free cash flow has clearly declined, and the ratio of capex to operating cash flow has risen significantly. But looking at the entire corporate sector, the gap between savings and investment has not deteriorated notably, because profit growth has largely offset the rising investment rate. External imbalances are also different. In the late 1990s, the US current account deficit widened; currently, the current account deficit is actually narrowing. At least from the perspective of macro imbalances, the current AI cycle has not yet exhibited the typical cracks seen at the tail end of that bubble.

$27 trillion in market cap added, exceeding the benchmark macro ledger

Changes at the market level are more aggressive. Since late November 2022, the value added by AI-related companies is approximately $27 trillion, up from about $19 trillion as of November 2025. Meanwhile, traditional valuation metrics for the US stock market remain at historical highs; the Shiller cyclically adjusted P/E ratio was only higher in late 1999 and 2000. However, there is one key difference between this rally and 1999: earnings expectations are also being revised upward rapidly. Because EPS expectations are rising, forward P/E ratios have not moved up in tandem this year even as stock prices continue to climb. Recent gains have been driven more by earnings than by pure valuation expansion. The problem is that the macro ledger does not offer support of the same scale. Benchmark calculations suggest the present value of new capital income for the US economy from AI productivity gains is about $9 trillion. Even using a more conservative market scope, looking only at "purer AI" companies, the related value added is about $14 trillion; if 25% of the gains from other AI-related companies are added, the scale is about $17 trillion—still above the benchmark calculation.

To justify current prices, one must bet that winners capture more profit over the long run

Current market prices are not entirely inexplicable, but they require more optimistic assumptions. These assumptions include: faster AI adoption, higher AI-driven productivity gains, capital capturing a larger share of economic returns, or US companies capturing more global AI revenue. One optimistic path outlined in the research note is that US companies capture 50% of global related revenue, the capital income share is significantly above the economic average, AI adoption is faster, and discount rates are lower. Only when multiple conditions hold simultaneously does the potential value more easily cover the current market cap increase. The most compelling optimistic narrative is that AI-related companies can capture a higher share of the productivity dividend over the long term. So far, this narrative is indeed supported by earnings. Semiconductors, cloud providers, and infrastructure beneficiaries have strong profits and high margins—it is precisely these earnings that support the market. But this is also the point of fragility. Early in a productivity acceleration, profit shares typically rise; over a longer horizon, competition, investment expansion, and new waves of innovation can erode excess returns. AI industry concentration is high, and technological characteristics may favor capital owners, but there is still no answer as to how long the moats of current winners can be maintained.

The biggest risk is shifting from a "valuation bubble" to an "earnings bubble"

The AI investment boom itself is generating substantial profits. Companies that sell chips, sell computing power, and build data centers directly benefit from rising capex. As long as the investment peak has not been approached, upward earnings revisions may continue to outweigh valuation concerns. But if the market simply extrapolates the next two to three years of strong earnings far into the future, risk will rise. Capex cannot grow at the current intensity forever. Once the investment cycle peaks, the earnings trajectory for the companies currently benefiting most directly may become harder to judge. This is also why "forward P/E is not expensive" does not necessarily equal cheap. Cyclical and commodity companies often do not look expensive at the peak of a cycle because the earnings denominator is too high. Whether the AI infrastructure chain will face a similar problem depends on how long the investment intensity can last, how quickly AI returns materialize, and whether technological innovation reduces reliance on high-intensity capex.

AI may be masking weakness in the non-AI economy

Compared to the 1990s, there is another important difference in the current macro backdrop. In the late 1990s, US domestic demand was very strong, with real domestic demand growing at an annualized rate of nearly 6% in the final two years, and consumption, residential investment, and non-tech investment were all robust. Capital inflows driven by the Asian and emerging market crises, a strong dollar, and global commodity price deflation actually masked domestic overheating in the US, allowing the cycle to extend further. The situation is now reversed. The US economy outside of AI is not that strong. Non-tech investment is weak, consumption growth is far below that of the late 1990s, and real disposable income has grown at an annualized rate of about 1% over the past two years, compared to 5%–6% in the late 1990s. This means the AI boom may not be adding fuel to an already overheating economy, but rather offsetting weakness in areas outside of AI. As a result, the extreme bubble and typical imbalances preceding the 2001 recession seen in 1999–2000 may be less likely to appear; but if the AI narrative stumbles, the non-AI part may not provide sufficient support.

Volatility shifts gears, portfolios need more downside protection

Market structure has already changed. Credit spreads remain tight, a path different from the gradually rising credit stress of 1998–2000. But equity volatility is beginning to rise more noticeably. In recent months, single-stock implied volatility has increased, and US single-stock options skew has shifted lower, with demand for calls rising relative to puts. At the same time, implied correlation has fallen to very low levels, suppressing index volatility, but longer-term index volatility is also slowly creeping up. Gains have also become more concentrated. Broad index performance is still more moderate than in the late 1990s, but the semiconductor index's gains over the past few years have approached the late-stage performance of the Nasdaq back then. In April and May, the Nasdaq, South Korea, Taiwan, the SOX semiconductor index, and a basket of non-profitable tech stocks all posted consecutive two-month gains at multi-year highs. As long as the peak of the investment cycle has not appeared, strong earnings may continue to dominate the market. But as prices become increasingly dependent on optimistic assumptions, the value of protecting the downside rises. The path forward looks more like staying in the trade while using put protection, or replacing some spot exposure with call options to control drawdowns. On the rates side, there is also a reverse risk: if the vulnerability of the non-AI economy is exposed after the AI investment peak passes, the probability of a significant decline in interest rates at that time could be higher than usual.