History does not repeat, but it does rhyme. For investors watching the $4 trillion AI infrastructure buildout, the question is whether the rhyme is about to turn ominous.
The global market for AI-related hardware and services is projected to reach $68 billion by 2033, up from $39 billion in 2025, according to industry estimates. But a research analysis by Huachang Securities strategist Zhang Yu, published July 29, draws on the 2000 dot-com bubble to argue that the current cycle may be approaching a peak — and that the warning signals are already visible in company financial statements.
"Free cash flow and EBITDA were the two most reliable leading indicators of the 2000 peak, with their inflection points preceding the stock top by about one quarter," Zhang wrote. His team examined 18 financial metrics across four categories of dot-com-era companies — network operators, equipment makers, component suppliers, and IT infrastructure firms — and found that only five or six had consistent predictive power.
The framework matters because the current AI investment cycle shares structural similarities with the late-1990s telecom buildout. Tech giants including Microsoft, Amazon, Google, and Meta are projected to spend a combined $240 billion on capital expenditures in 2026, much of it directed at AI data centers and GPU clusters. The question, Zhang argues, is whether those investments will generate returns before the financial strain becomes visible.
The Productivity Paradox
Academic research on general-purpose technologies suggests a consistent pattern: productivity gains lag investment by 10 to 40 years. The steam engine, electricity, and information technology all followed this trajectory, though the lag has shortened with each successive wave. For AI, the consensus among economists surveyed by Zhang's team points to 2030 as the inflection point when productivity acceleration should begin in earnest.
The mechanism is straightforward. General-purpose technologies require complementary investments in organizational structure, workforce training, and business model redesign before their benefits materialize. The OECD, in a study of 23 countries, found that only the top 5% of firms consistently capture efficiency gains from new technologies, while most small and medium enterprises struggle to adapt.
"Technology diffusion is a slow process," Zhang wrote. "The benefits of AI will not be evenly distributed, and the market may be pricing in an adoption curve that history suggests is unrealistic."
What the Financial Signals Show Now
For the current cohort of US tech stocks, the picture is mixed. First-quarter 2026 earnings showed all core financial indicators in "healthy range," according to Zhang's analysis, with no clear inflection point yet visible. But the second-quarter reporting season, now underway, will be critical.
The dot-com playbook suggests a specific sequence of deterioration: operating cash flow growth slows first, followed by declining coverage of capital expenditures by internal cash generation, then revenue deceleration, and finally an absolute decline in free cash flow and EBITDA — the point at which stock prices historically peaked.
F5 Inc., a networking and security company whose hardware sits inside most major AI data centers, reported fiscal third-quarter results on July 27 that showed no such deterioration. Revenue rose 11% to $865 million, beating the $834.6 million consensus, while adjusted earnings of $4.73 a share topped estimates by 18%. The company raised its full-year revenue growth outlook to 9% to 10% from 7% to 8%, citing AI inference workloads as a structural demand driver.
But the broader picture is less clear. Swedish education technology company Albert, which is pivoting to AI-powered math learning, reported second-quarter revenue of SEK 35 million, down 4% from a year earlier, with annual recurring revenue falling 10% to SEK 125.6 million. The company's EBITDA from continuing operations improved to minus SEK 1.6 million from minus SEK 6.6 million — a sign of cost discipline, but also of a business still in transition.
Why Credit Markets Won't Warn You
One common investor instinct is to watch bond markets for signs of distress. Zhang's analysis of the 2000 bubble suggests this is a mistake. Using WorldCom as a case study, he showed that credit rating agencies did not issue negative outlooks until eight to nine months after the Nasdaq peaked, by which point WorldCom's stock had already fallen roughly 80%.
"Fixed-income investors price for solvency; equity investors price for growth expectations," Zhang wrote. "Credit indicators are lagging, not leading, signals for equity tops."
Current conditions reinforce the point. CDS spreads for major US technology companies remain near historic lows, bond subscription multiples exceed three times — well above the roughly 2.5 times for US Treasury auctions — and credit ratings are at the highest levels. None of this provides advance warning of an equity peak, Zhang argued.
What Comes Next
The second-quarter earnings season, which runs through mid-August, will provide the first real test of Zhang's framework. If free cash flow margins at major cloud providers and AI hardware suppliers begin to compress, or if operating cash flow growth decelerates while capital expenditure continues to rise, the historical pattern would suggest a market top within one to two quarters.
For now, the data supports both narratives. AI infrastructure spending shows no sign of slowing, and companies like F5 are reporting accelerating demand. But the productivity gains that would justify the investment remain theoretical, and the historical record suggests they may take years to materialize — longer, perhaps, than the patience of equity markets.
This article is for informational purposes only and does not constitute investment advice.
