Oil prices stand above US$100/barrel, the Strait of Hormuz has not yet returned to normal opening, inflation and interest rate pressures have re-emerged, and expectations of a Federal Reserve interest rate cut have become more fragile. According to the traditional macro framework, this is not the most comfortable environment for highly valued technology stocks. However, U.S. stocks hit new highs, and the AI ​​chain continues to be chased by funds.

Song Xuetao, a macro analyst at Guojin Securities, pointed out in a research report on May 25: "The current AI market is in a stage of rational enthusiasm. The bubble has emerged but is not out of control." The key to this sentence is not the "bubble" but the "rational" enthusiasm: Agentic AI has moved from an auxiliary tool to an autonomous execution tool, allowing the market to see more clearly for the first time the closed business loop of AI from "burning money" to "making money".

On the rational side, the proliferation of Agent applications has brought rapid growth in Token consumption, reasoning computing power requirements, and ARR of leading manufacturers; on the fanatical side, the valuation has already eaten up the growth expectations for 2027-2028 in advance. As of May 20, the forward P/E ratio of the seven largest companies in the U.S. stock market is about 35 times, and the remaining 493 companies in the S&P 500 are about 25 times. What this premium implies is not the logic of ordinary growth stocks, but that the penetration rate of AI must reach 5 to 8 times that of past technological revolutions.

But what really determines whether the AI bull market can continue is not the performance of a single quarter, nor a certain popular application, but three variables: in the short term, liquidity shocks, especially the unwinding of oil prices, inflation, interest rates and Japanese yen carry trades; in the medium term, industry realization, whether the AI penetration rate can match the current valuation; in the long term, harder constraints such as energy, power grids, employment, social resistance and hardware technology mutations.

Agent changes from "co-pilot" to "main driver", and the market begins to reward capital expenditure

In the past round of AI transactions, the market was most worried about the giants spending money too quickly: huge investments in data centers, GPUs, and cloud infrastructure, but the revenue recovery path was not clear enough. The change in Agentic AI is that it is no longer just a Copilot-style auxiliary tool, but has evolved into an Autopilot-style autonomous execution tool.

This has two consequences.

First, Token consumption is accelerating again. The first round of demand after the emergence of GPT came from the improvement of model capabilities, and the second round of demand after the launch of Agent came from the explosion of inference computing power. Autonomous execution of tasks means longer contexts, more complex steps, and more frequent model calls. Inference is no longer a leftover after training, but has become the main battlefield that continues to consume computing power.

Second, revenue expectations were revised upwards. After the proliferation of representative Agent applications such as Openclaw and Claude Cowork, the annual recurring revenue of model manufacturers has grown rapidly. The mid-year estimates cited in the material show that Anthropic's full-year ARR expectations have been raised from $9 billion at the beginning of the year to $44 billion, doubling on average every six weeks. If the trend continues, next year's ARR is expected to exceed $300 billion.

This explains why the market no longer simply punishes Capex. As long as revenue grows fast enough, capital expenditures turn from a burden to a moat. As a result, hardware chains such as Nvidia, Broadcom, and optical modules and storage have regained support.

With oil prices above US$100, why can AI assets continue to rise?

This round of AI assets rising against oil prices is not because macro risks have disappeared, but because several forces have temporarily overwhelmed the risks.

The first is the spread of demand in the industrial chain. The inference stage not only requires GPU, but also CPU, optical modules, and storage are included in the high-prosperity logic. 800G/1.6T optical modules are in short supply, and demand for high-end storage is rising. Light Counting predicts that 800G transceiver shipments will more than double in 2026, 1.6T port shipments will grow from a small base in 2025 to tens of millions, and 1.6T chipset sales will exceed $2 billion in 2026 and maintain high growth rates over the next three years.

Secondly, the performance of technology giants is too strong. The S&P 500 EPS growth rate in the first quarter was approximately 27.1%, a new high since the fourth quarter of 2021. Three companies, Meta, Alphabet and Amazon, contributed 70% of the index’s profit increase. As long as these heavyweight companies continue to make money, the suppression of the index by the oil price shock will be pushed back.

The third is that U.S. growth is increasingly dependent on AI infrastructure. In the past few quarters, AI infrastructure investment has contributed to more than half of U.S. GDP growth. Total non-agricultural and retail sales data are acceptable. Although the employment structure has differentiated, it will be difficult for the market to immediately switch to stagflation trade before the total volume weakens significantly.

There is also a more direct factor: large technology companies are not as sensitive to oil prices as aviation, express delivery, railways, chemicals, automobiles, tourism and other industries. They are more afraid of electricity prices than oil prices. When the traditional real economy is squeezed by oil prices, it is easier for funds to group into AI assets, combining "risk-off" transactions and growth transactions.

Valuation has already eaten up the good days in 2027-2028

The danger of the AI market is not that there is no industrial support, but that the market is pricing too quickly.

The seven largest companies in the U.S. stock market have a forward P/E ratio of 35 times, and the remaining 493 companies in the S&P 500 have a P/E ratio of 25 times. Behind this valuation difference, there is a very smooth future: in the next 3 to 5 years, AI infrastructure will continue to expand, and demand for computing power, cloud, data centers, and semiconductors will remain high; AI will continue to penetrate advertising, search, cloud services, office software, code generation, financial risk control, customer service, investment research, content and other scenarios; revenue contribution and efficiency improvements will be realized at the same time.

But technological revolutions are rarely so smooth. It took about 40 years for electricity to go from its invention to large-scale implementation on the assembly line, and about 25 years for computers. The current diffusion rate of AI being priced by the market requires it to be 5 to 8 times faster than these general-purpose technologies.

It's not impossible, but the margin for error is thin. As long as the commercialization of AI applications is slower than capital expenditures, inference demand cannot catch up with training demand, or depreciation and power costs start to eat into profit margins, valuations will react first. The correct direction of the industry does not mean that the stock price can advance indefinitely.

The biggest short-term risk: interest rates run faster than ARR

The real short-term pressure comes from liquidity.

If the Strait of Hormuz remains open for a long time and oil prices remain above $100 or even continue to rise, inflation will spread from energy prices to the service industry, transportation and raw materials. U.S. PPI has risen to 9.8% year-on-year in April, the highest level since October 2022. Once inflation solidifies, the Fed's policy path will be forced to rewrite.

The swap market has priced in the Federal Reserve raising interest rates 0.8 times this year, and the European Central Bank and the Bank of England even raising interest rates more than 2 times. At the same time, doubts about the policy independence brought about by the change of the Federal Reserve and increasing disagreements within the FOMC are also weakening the market's confidence in future easing.

Japan is also a gray rhinoceros. Japan has long been a financing pool for global leveraged trading, but the depreciation of the yen and inflationary pressures have forced the Bank of Japan to issue a tightening signal, and the 30-year JGB yield has risen to above 4%. If Japan's financing costs continue to rise, triggering the unwinding of global carry trades, it will be difficult for highly valued AI assets to survive alone.

There was already a preview on May 15: the 10-year U.S. bond yield exceeded 4.5%, the 30-year bond yield exceeded 5%, highly crowded momentum trading cooled down, the Philadelphia Semiconductor Index fell by about 4% in a single day, and the Nasdaq fell by about 1.5%. This is not evidence of a trend reversal, but it illustrates that crowded trading is extremely sensitive to interest rates.

The most critical short-term comparison is simple: whether ARR (annual recurring revenue) can be revised faster than interest rates. If not, funds may first be reduced to the hardware segment with higher certainty; if liquidity continues to deteriorate and AI revenue expectations cannot continue to be revised upward, valuation pressure will be significantly amplified.

More difficult issues in the medium and long term: organization, power, employment and hardware routes

The mid-term test is the industry’s fulfillment. General technological revolutions usually do not rise in a straight line, but "first accelerate, then decelerate, then accelerate again." First there is the wave of capital, then there is organizational integration, and finally there is the release of productivity. The Internet also experienced investment booms, capital expenditure expansion and asset bubbles in its early days. Real productivity improvements only gradually emerged over many years.

The difficulty with AI pricing now is that it almost requires rapid adaptation of corporate organizational structures, rapid retraining of workers, rapid adoption of business models, and no strong conflicts at the social level. This speed has not been common in human history.

Long-term constraints are tougher.

The first is energy and infrastructure. AI data centers require a large amount of electricity and cooling water. Grid expansion, transformers, and energy storage are not variables in the PPT, but real bottlenecks. If AI infrastructure continues to push up electricity costs across society, regulatory and social backlash will heat up.

The second is employment and consumption. AI can improve corporate efficiency in the short term and reduce the demand for engineers, customer service and other positions; but if technological unemployment outpaces the creation of new jobs, residents' consumption power will be weakened. Improvements in B-side efficiency will ultimately depend on C-side purchasing power. If the non-AI sector falls into recession, it will be difficult for AI to stand out in the long term.

The third is social acceptance. At the beginning of the year, there was a rush to install Openclaw among all Chinese people, but the American people's resistance to data centers pushing up electricity prices and technological unemployment is rising. This affects the speed of AI penetration.

The fourth is the mutation of hardware technology. If there is an engineering breakthrough similar to the "DeepSeek moment" and computing power, storage, and transmission efficiency are greatly improved, then the most scarce hardware links today may suddenly become surplus. The logic of high prosperity in the hardware chain is not irreversible.

The long-term prospects of the AI industry remain optimistic. If we do not consider the social contradictions caused by technological unemployment and the restructuring of production relations, AI does have the opportunity to increase total factor productivity and help the economy get rid of stagflation pressure. Even if the financial market deleverages midway, the remaining data centers, low-cost technologies and proven application scenarios may become the basis for the next round of industrial expansion.

But stock pricing is not the industry vision itself. What needs to be verified most in this round of AI bull market is whether the ARR, ROI and technology penetration speed that the market is currently betting on can continue to be realized in an environment where oil prices, inflation, interest rates and social constraints have all hardened. The correct direction can only explain why there is a bull market; the speed of realization determines whether the bubble will get out of control.