In Q4 2026, the U.S. AI computing power sector remains the undisputed star of global capital markets, but the debate over “whether this is the start of a new industrial revolution or a replay of the 1999 internet bubble” keeps heating up as financing structures shift. From the tightly interlocked mega-orders between OpenAI and chip and cloud vendors to data center construction increasingly reliant on bonds and private credit, market focus is moving from “how strong is demand” to “who is footing the bill and how long it can last.” This article lays out the three focal points of the current bubble debate and offers practical observation frameworks and coping strategies for everyday investors.
Focal Point 1: The Circular Deal Mystery—Are Investment-for-Order Deals Sustainable?
So-called “circular deals” refer to upstream and downstream players in the AI value chain trading investment for orders and orders for investment, forming a closed-loop cash flow. The most representative examples are the series of collaborations between OpenAI and NVIDIA, AMD, and Oracle: NVIDIA announced plans to invest capital in OpenAI in stages while locking in large-scale chip supply; AMD offered warrants in exchange for multi-year computing orders; and Oracle, as disclosed, signed cloud service contracts with OpenAI on the scale of hundreds of billions of dollars. Upstream players fund downstream customers, who then use that money to buy upstream products—a structure that has drawn considerable skepticism.
Supporters argue this resembles the long-term order arrangements between aircraft makers and airlines, a normal business practice during a capacity ramp-up; skeptics counter that if downstream customers fail to deliver end-market revenue, the revenue recognized by upstream chipmakers will lack real cash flow support, and once financing conditions tighten, the loop could reverse quickly. The key to evaluating this lies not in the deal format itself, but in whether the computing capacity, once online, can generate enough inference usage and enterprise paid demand.
Focal Point 2: From Cash Flow to Leverage, Financing Methods Are Changing
Early AI data center construction was funded mainly by tech giants' own cash flows—Microsoft, Google, Amazon, and Meta together spent hundreds of billions of dollars a year in capex, mostly covered by operating cash flow, which was the bulls' strongest reassurance. Notably, however, since the second half of 2025 two landmark changes have emerged in the financing structure: first, companies like Meta launched large-scale bond issuances, setting issuance records in the corporate bond market at the time; second, off-balance-sheet tools such as private credit and special-purpose SPVs have penetrated data center projects far more deeply, with some projects monetizing future rents in advance through asset securitization.
Leverage itself is not a monster—every infrastructure wave in history, from railroads to power grids to telecom, was inseparable from debt. The real risk signals appear when interest coverage ratios decline, rental yields fall below financing costs, and new supply far exceeds actual demand; only then does leverage amplify downside volatility. Investors should focus on changes in issuers' credit spreads and the pricing of data center asset-securitization products—they are often the first warning indicators to “speak up” in this cycle.
Focal Point 3: The Debate Over GPU Depreciation Periods and Investment Payback Cycles
Another technical debate centers on chip depreciation periods. Some tech companies have extended server depreciation to six years, while a well-known short seller abroad has publicly questioned whether GPUs' actual economic life may be only two to three years, arguing this inflates book profits and plants impairment risks. Objectively speaking, accounting depreciation and economic life are indeed two different things: flagship chips may lose their edge in frontier training markets after two or three years, yet can still serve in inference, edge computing, and industry scenarios. However, if an architectural leap causes old cards to be retired en masse ahead of schedule, the profits and valuations of related companies will come under pressure—a line item to watch closely every earnings season.
The Bulls' Narrative: Demand-Side Evidence Is No Castle in the Air
The bubble thesis is not without opponents; the evidence chain supporting computing demand currently includes at least three strands. First, token volumes and AI-related revenue on leading cloud platforms keep growing fast, with inference workloads taking a rising share, showing computing power is being genuinely consumed. Second, remaining performance obligations at vendors like Oracle have hit record levels—these are contracts signed in black and white, not verbal intentions. Third, power has become the hardest physical constraint: queues for grid connections, lengthening transformer lead times, and renewed interest in nuclear and gas plants all precisely indicate that demand has outstripped real supply capacity.
Historical experience is also instructive: even when genuine infrastructure waves like railroads, power grids, and fiber optics were followed by bubble bursts, their physical assets kept working for decades afterward. The question was never “is there a bubble,” but how well the purchase price matches the quality of the underlying assets.
Three Valuation Frameworks to Help You See the Truth
Framework 1: The Funding Structure of Capex
Split tech giants' capex into “operating cash flow coverage” and “external financing.” The higher the former's share, the more sustainable the boom; if the external financing share rises rapidly and financing costs climb, be on alert—this is often a leading signal of a cyclical peak.
Framework 2: The Revenue Conversion Rate of Computing Investment
Keep tracking the ratio of cloud vendors' AI-related revenue growth to capex growth. When revenue growth persistently lags spending growth and the gap keeps widening, valuation expansion is usually nearing its end; conversely, if inference revenue starts accelerating, the bubble thesis collapses on its own.
Framework 3: Physical Constraints of Power and Supply Chain
Grid access, HBM memory, and advanced packaging capacity are this cycle's most real “demand validators.” If supply bottlenecks persist, demand is genuinely strong; if bottlenecks ease quickly and rental prices fall markedly, beware of valuation compression triggered by oversupply.
Four Practical Tips for Investors
- Prefer cash flow-generating names: Chip leaders and cloud giants are not cheap, but their cash-generating capacity and risk resilience far exceed those of second-tier AI computing concept stocks that rely on external financing, making them better able to withstand drawdowns.
- Beware purely financing-driven companies: For computing leasing and IDC firms whose revenue depends heavily on a few major clients and which run on high leverage, be sure to control position sizes and closely track their debt maturity profiles and renewal rates.
- Use diversification to hedge uncertainty: Individual-stock black swans are hard to predict; consider participating via AI computing-themed ETFs covering chips, power, network equipment, cooling, and other segments—among them, “picks-and-shovels” segments such as power and cooling are relatively less hit by the bubble controversy.
- Set rebalancing discipline: Take partial profits in stages when sentiment turns euphoric, keep AI computing exposure within your own risk tolerance, and firmly avoid chasing highs with leverage.
Conclusion: Bubble or Not, Cash Flow Will Have the Final Say
Looking back at the 2000 internet bubble—the bubble did burst, yet Amazon, Google, and their peers went on to change the world; the survivors' common trait was holding real cash flow before the winter arrived. The U.S. AI computing power sector today sits in a mixed state of “real demand plus speculative capital.” For everyday investors, rather than predicting when the bubble will burst, keep a close eye on three verifiable indicators—financing structure, revenue conversion, and power constraints—and use discipline and diversification to ride through the cycle, letting data rather than emotion make the call.
