What are the key variables that determine the AI bull market?

Oil prices remain above $100 per barrel, the Strait of Hormuz has not yet fully reopened, inflation and interest rate pressures are resurfacing, and the Fed's rate cut expectations have become more fragile. According to traditional macro frameworks, this is not the most comfortable environment for high-valuation tech stocks. Yet U.S. stocks are hitting new highs, and AI chains continue to be chased by capital.

Guojin Securities macro analyst Song Xuetao pointed out in a research report on May 25: "The current AI rally is in a stage of rational exuberance; bubbles are present but not out of control." The key point of this statement is not "bubble," but "rational" exuberance: Agentic AI moving from an auxiliary tool to an autonomous execution tool, allowing the market to see for the first time the business cycle from "burning money" to "making money."

On the rational side, the diffusion of Agent applications has led to rapid growth in token consumption, inference computing power demand, and the ARR of leading vendors; on the exuberant side, valuations have already priced in growth expectations for 2027–2028. As of May 20, the forward P/E ratios of the seven U.S. tech giants are about 35 times, and the remaining 493 companies in the S&P 500 are about 25 times. This premium implies not ordinary growth stock logic, but an AI penetration speed that must reach 5 to 8 times that of past technological revolutions.

But what truly determines whether the AI bull market can continue is not quarterly earnings alone, nor a single blockbuster application, but three variables: short-term liquidity shocks, especially oil prices, inflation, interest rates, and yen carry trade unwinding; medium-term industry realization, whether AI penetration speed can match current valuations; and long-term constraints such as energy, power grids, employment, social resistance, and hardware technological breakthroughs.

Agent from "co-pilot" to "main driver," markets begin rewarding capital expenditure

In the previous AI trading cycle, the market's biggest concern was that giants were spending too fast: data centers, GPUs, cloud infrastructure investments were huge, but revenue recovery paths were unclear. The change with Agentic AI is that it is no longer just a Copilot-style auxiliary tool but evolving into an Autopilot-style autonomous execution tool.

This has led to two outcomes.

First, token consumption has accelerated again. The first demand wave after GPT emerged was driven by model capability improvements; the second wave after Agent deployment comes from explosive inference compute needs. Autonomous task execution means longer context, more complex steps, more frequent model calls, and inference is no longer just a side task after training but a main battlefield of continuous compute consumption.

Second, revenue expectations have been revised upward. After the proliferation of representative Agent applications like Openclaw and Claude Cowork, model vendors' annual recurring revenue (ARR) has grown rapidly. Mid-year estimates cited in the material show that Anthropic's full-year ARR has been revised from $9 billion at the start of the year to $44 billion, doubling roughly every six weeks. If this trend continues, next year's ARR could exceed $3 trillion.

This explains why the market is no longer simply punishing Capex. As long as revenue growth is fast enough, capital expenditure shifts from a burden to a moat. Nvidia, Broadcom, and hardware chains like optical modules and storage are thus gaining renewed support.

Why can AI assets still rise above $100 oil?

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

First is the diffusion of industry chain demand. Inference stages not only require GPUs but also involve CPU, optical modules, and storage, which are pulled into high prosperity logic. 800G/1.6T optical modules are in tight supply, and high-end storage demand is rising. Light Counting predicts that by 2026, 800G transceiver shipments will more than double, and 1.6T port shipments will grow from a small base in 2025 to tens of millions, with 2026 sales of 1.6T chipsets exceeding $2 billion, maintaining high growth over the next three years.

Second is the strong performance of tech giants. In Q1, the S&P 500 EPS growth was about 27.1%, reaching a new high since Q4 2021, with Meta, Alphabet, and Amazon contributing 70% of the index's profit increase. As long as these heavyweight companies continue to profit, the impact of oil shocks on the index will be delayed.

Third is the increased dependence of U.S. growth on AI infrastructure. Over the past few quarters, AI infrastructure investment has contributed to over half of U.S. GDP growth. Non-farm and retail data are still solid, and although employment structure has become more segmented, as long as the overall economy does not weaken significantly, the market will find it hard to immediately switch to stagflation trading.

Another more direct factor is that large tech companies are less sensitive to oil prices than industries like airlines, express delivery, railways, chemicals, automobiles, and tourism. They are more concerned about electricity prices than oil. When traditional real economy sectors are squeezed by oil prices, capital tends to flock into AI assets, blending "hedge" trades with growth trades.

Valuations have already priced in the good days of 2027–2028

The danger of the AI rally is not the lack of industry support but the market pricing it too quickly.

The forward P/E of the seven U.S. giants at 35 times, and 25 times for the remaining 493 companies in the S&P 500, imply a very smooth future: AI infrastructure will continue expanding over the next 3 to 5 years, with high prosperity in compute, cloud, data centers, and semiconductors; AI will keep penetrating advertising, search, cloud services, office software, code generation, financial risk control, customer service, investment research, and content creation; revenue contribution and efficiency gains will be realized simultaneously.

But technological revolutions are rarely this smooth. Power took about 40 years from invention to large-scale application, computers about 25 years. Now, the market's pricing of AI's diffusion speed demands 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 AI application commercialization lags behind capital expenditure, inference demand cannot keep up with training demand, or depreciation and electricity costs start eroding profit margins, valuations will react first. Having the right industry direction does not mean stock prices can be infinitely advanced.

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

The real short-term pressure comes from liquidity.

If the Strait of Hormuz remains closed long-term, oil prices stay above $100 or continue rising, inflation will spread from energy prices to services, transportation, and raw materials. In April, US PPI year-over-year rose to 9.8%, a new high since October 2022. Once inflation becomes entrenched, the Fed's policy path will have to be rewritten.

The swap market has already priced in about 0.8 rate hikes from the Fed this year, and over two hikes from the ECB and Bank of England. Meanwhile, doubts about the Fed's independence due to leadership changes and increasing FOMC dissent are weakening market confidence in future easing.

Japan is also a "gray rhino." Long a source of leverage trading financing globally, the yen's depreciation and inflation pressures have prompted the Bank of Japan to signal tightening, with 30-year Japanese government bond yields rising above 4%. If Japan's financing costs continue to rise, triggering global carry trade unwinding, high-valuation AI assets will find it hard to remain insulated.

A prelude occurred on May 15: 10-year US Treasury yields broke above 4.5%, 30-year yields above 5%, cooling crowded momentum trades, with the Philadelphia Semiconductor Index dropping about 4% in a day and the Nasdaq down about 1.5%. This is not a trend reversal but shows how sensitive crowded trades are to interest rates.

The most critical short-term comparison is simple: can ARR revisions outpace the rise in interest rates? If not, capital may shift first into more certain hardware segments; if liquidity continues to worsen and AI revenue expectations cannot be revised upward, valuation pressures will intensify.

Longer-term, more complex issues: organization, power, employment, and hardware pathways

The mid-term test is industry realization. General technological revolutions usually do not follow a straight line but "accelerate, decelerate, then accelerate again." There is a wave of capital, followed by organizational adjustments, and finally productivity improvements. The early internet also experienced investment booms, capital expenditure expansion, and asset bubbles, with real productivity gains only gradually emerging over years.

Now, the difficulty in AI pricing lies in its near-requirement for rapid organizational restructuring, worker retraining, quick commercialization of business models, and societal acceptance. Such speed is uncommon in human history.

Longer-term constraints are even more formidable.

First is energy and infrastructure. AI data centers require massive electricity and cooling water; grid expansion, transformers, and energy storage are not just PPT variables but real bottlenecks. If AI infrastructure continues to push up social electricity costs, regulatory and social backlash will intensify.

Second is employment and consumption. AI can boost corporate efficiency in the short term, reducing roles like engineers and customer service; but if technological unemployment outpaces new job creation, household consumption capacity will weaken. Revenue from B2B efficiency gains ultimately depends on B2C purchasing power; if non-AI sectors fall into recession, AI will also struggle to sustain long-term leadership.

Third is social acceptance. Earlier this year, there was a nationwide enthusiasm for deploying Openclaw, but in the U.S., resistance to data centers raising electricity prices and technological unemployment is growing. This will affect AI penetration speed.

Fourth is hardware technological breakthroughs. If there are engineering breakthroughs similar to a "DeepSeek moment," dramatically improving compute, storage, and transmission efficiency, the most scarce hardware segments today could suddenly become oversupplied. The high prosperity logic of the hardware chain is not invulnerable.

The long-term outlook for AI remains optimistic. If we set aside social conflicts from technological unemployment and restructuring, AI indeed has the potential to boost total factor productivity and help the economy escape stagflation. Even if financial markets undergo deleveraging, existing data centers, low-cost technologies, and proven applications could form the foundation for the next wave of industry expansion.

But stock pricing is not the same as industry vision itself. The most critical validation for this AI bull market is whether the market's current bets on ARR, ROI, and technological penetration speed can continue to materialize amid rising oil prices, inflation, interest rates, and social constraints. Having the right direction explains why there is a bull market; the speed of realization determines whether the bubble will spiral out of control.