Criteria for Assessing the AI Bubble

When assessing an AI bubble, the first layer to examine is stock prices—but one must not stop there. If a company’s stock price surges tenfold while its revenue remains stuck at the conceptual stage and customer orders have not translated into cash flow, that is clearly a red flag. However, if rising stock prices are accompanied by concrete orders, booked capacity, and long-term purchase commitments from customers, high market valuations may not immediately signal a bubble. The real danger arises when the market prices the entire supply chain based on the most optimistic assumptions: chipmakers, optical modules, servers, power equipment, data centers, cloud platforms, and AI applications—all assumed to face infinite demand, stable pricing, and declining capital costs. That’s where bubbles are most likely to form.

More important than stock prices is order quality. AI-related orders can be categorized into three types: (1) hard orders that are already signed, delivered, and revenue-recognizable; (2) capacity reserved by customers or framework agreements; and (3) mere expressions of intent, memoranda of understanding, or pipeline figures cited in investor presentations. These carry vastly different weight in valuation. If the market treats letters of intent as revenue, framework agreements as profit, and future capacity as cash flow, valuations quickly become distorted. This is especially critical for AI data centers and GPU cluster investments, which involve large upfront payments, long construction cycles, and rapid technological obsolescence. Orders must be scrutinized for who the customer is, contract duration, presence of cancellation clauses, price certainty, and the counterparty’s ability to sustain payments.

A second, more practical indicator is electricity consumption. Unlike many past tech narratives, AI is not purely asset-light. Training and inference for large models require GPUs; GPUs require data centers; and data centers require power, cooling, land, fiber optics, and transformers. When an AI company claims it is rapidly scaling compute capacity, this must ultimately show up in electricity demand. If stock prices and order narratives surge dramatically while power grid connections, data center construction, rack deployments, and actual utilization lag behind, it means the narrative is running ahead of the real economy. Conversely, if power utilities, grid operators, transformers, power generators, and data center REITs all simultaneously experience demand pressure, it indicates that AI investment has at least entered the phase of real capital expenditure.

However, more electricity consumption isn’t inherently better. Increased power usage only confirms that someone is building—it doesn’t prove the investment will be profitable. This leads to the third and most critical criterion: payback period. The core issue with AI investment isn’t whether demand exists, but how quickly each dollar of capital expenditure can be recovered. Suppose a GPU cluster is extremely costly and, after three to five years, faces performance obsolescence, declining energy efficiency, and falling rental rates—then the company must generate sufficient revenue within a very short window. If the depreciation schedule spans five years but the true economic life is only three, reported profits will be overstated. If customer contracts last just one year while the company supports its valuation with long-term debt and extended depreciation periods, the risk is even greater.

This is precisely where AI valuations are most prone to misjudgment. The market often evaluates AI companies using revenue multiples or EBITDA multiples—but AI infrastructure is not traditional software. True software companies have low marginal costs, so most incremental revenue flows directly to gross profit. In contrast, AI compute platforms must continually purchase chips, build data centers, pay electricity bills, and refresh hardware. If revenue growth each year requires ever-larger capital expenditures, investors should look beyond top-line growth and instead focus on free cash flow, depreciation, lease liabilities, and reinvestment needs. An AI company may appear nearly profitable on the income statement while consistently burning cash on the cash flow statement.

From this perspective, there should be a clear framework for assessing whether the AI bubble exists: stock prices act as a thermometer, orders validate demand, power consumption reflects physical constraints, and payback period delivers the ultimate verdict. Soaring stock prices merely indicate heated market sentiment; rising orders confirm real demand; increasing power usage shows capital expenditures are being deployed; but only a reasonable payback period indicates that this investment cycle has the potential to generate sustainable returns. If all four factors align, high AI valuations can be understood as forward-looking pricing typical of an early-stage industrial revolution. However, if only stock prices rise—while orders lack substance, power consumption hasn’t materialized, and payback periods remain unclear—it’s far more indicative of a bubble.

$SpaceX (SPCX.US)$ $Tesla (TSLA.US)$ $CHINA MOBILE (00941.HK)$ $21Vianet (VNET.US)$ $Advanced Micro Devices (AMD.US)$ $Zhongji Innolight (300308.SZ)$ $Apple (AAPL.US)$

126K Views