How Much Revenue Does AI Need to Avoid Being a Bubble?

Aiming for a trillion-dollar revenue level by 2029 may become the baseline for AI technology to prove itself (not just Agent ARR, but also contributions from other modalities). The expansion of the financing gap will continue for at least several more quarters, so concerns about free cash flow will not disappear into thin air.

Viewing the AI industry as a whole, it may need revenue nearing the trillion-dollar level around 2030 to cover its accounting costs, while the peak of the external financing gap may appear around 2028, corresponding to a magnitude of approximately $700–800 billion.

Such a high-level break-even point and an external financing gap that is far from peaking mean that AI technology must continuously prove itself through more explicit revenue conversion; sources will not be limited to current Agent-ARR revenue but will also require contributions from other immature models.

The level of the accounting break-even point and the actual revenue growth path will become important references for measuring the health of capital expenditure, which is closely related to the future performance realization of the entire hardware chain.

Under different baseline scenarios, there is huge room for fine-tuning the calculation of the break-even level. The speed of depreciation (different terms for computing power and infrastructure), the interest rate level for cash interest payments (whether credit spreads can remain low), and changes in Capex growth rates under the broadest macroeconomic perspective (whether it peaks in 2028) all have significant impacts on the final result.

In addition, with the continuous construction of data centers and the medium-term alleviation of short-term power supply pressures (grid integration), whether economies of scale will emerge, as well as potential “high-quality computing power peak shaving” or policies similar to “transmitting computing power from west to east,” will all affect the final break-even level.

Our understanding of the essence of the AI era also needs to deepen. For example, computing power is currently a short-term bottleneck for model parameters, but does grid-integrated computing power necessarily train “better” models? Furthermore, how do we define a good model—is it having sufficiently large parameters, or being more cost-effective within a certain capability range? These factors determine whether the grid integration of computing power and electricity will generate economies of scale.

Overall, in this world where AI production factors are highly fluid, identifying the characteristics of any single link is difficult. However, the trillion-dollar revenue magnitude and the hundred-billion-dollar financing gap may gradually become the baseline consensus for measuring the AI bubble in the future stage.

Considering the complex dimensions of the discussion, we adopt common premises for some indicators, mainly focusing on sensitivity analysis regarding the speed of depreciation and CAPEX growth rates. Common indicator assumptions include, but are not limited to, scenario growth rates for Capex, the proportion of capital expenditure for long-life versus short-life equipment, and the ratio of OPEX to existing assets.

It should be noted that for external financing assumptions, we adopt a fixed ratio of External Financing/Capex. On this basis, we discuss the size of the external financing gap when revenue (ARR) equals the accounting break-even level.

1. Faster Equipment Depreciation Requires Faster AI Monetization

In the first part, we focus on the sensitivity analysis of the depreciation period of AI-related assets on the accounting break-even level.

When discussing depreciation, capital expenditure levels (Capex) need to be divided into short-life IT equipment (CPU/GPU/server components) and long-life data center infrastructure. Around the third quarter of 2025, before the spread of the Agent model, there was widespread discussion concerning equipment depreciation, which was essentially concern over the depreciation of rapidly iterating short-life chips.

Taking Google as an example, most internet companies basically classify the depreciation period for AI-related short-life computing equipment at the 5–8 year level, while fixed asset equipment extends to around 20 years.

We assume three scenarios for the depreciation periods of long-life and short-life equipment: Aggressive Scenario (5 years/15 years), Neutral Scenario (8 years/20 years), and Conservative Scenario (10 years/25 years). We observe the ARR revenue level required to achieve accounting break-even for CAPEX under different scenarios, as well as the free cash flow break-even level (same for all scenarios) and the external funding gap (when revenue equals the accounting break-even level).

Under the neutral Capex growth assumption (growth rates of 25%, 15%, 5%, and 0% for the next four years), to achieve accounting break-even for Capex, the neutral depreciation scenario corresponds to AI revenue of $923 billion in 2029 and $1.18 trillion in 2030; the conservative depreciation scenario requires AI revenue to reach $1.06 trillion in 2030, while under the aggressive depreciation scenario, $1.42 trillion needs to be realized in 2030.

From a CAGR perspective, assuming that leading US model providers collectively achieve an ARR level of approximately $200 billion by the end of 2026, the annualized growth rates required for the next 4 years (2026–2030) would be 63.3%, 55.8%, and 51.8%, respectively. It is clearly visible that even with conservative depreciation, the required growth rate demands breakthroughs in both technology and application.

Calculating the external financing gap requires more assumptions, making the results more “static.” Specifically, we assume that for each future year, the AI industry's revenue exactly equals the accounting break-even level shown in the chart above (i.e., covering OPEX, depreciation, and interest expenses). In this scenario, the external financing gap can be simplified as the current year's CAPEX expenditure minus depreciation and other stock-based compensation expenses. Based on this estimation, the external financing gap will peak around 2028, corresponding to a magnitude of approximately $700–800 billion.

Of course, it is clear that our primary premise is that actual revenue levels can keep up with the break-even level. In other words, if revenue falls below the expected realization, the external financing gap will be larger, meaning it may be underestimated compared to actual financing conditions.

For example, ideally, the funding gap in 2026 would be around $600–700 billion. However, what we observe is that bond financing in the first half of 2026 was approximately $250 billion, looking toward $500 billion for the full year. Adding nearly $100 billion in equity financing from Intel and Google, the total approaches $600 billion without considering off-balance-sheet financing and direct credit.

Considering that the flood of investment-grade tech bond issuance has just begun, current overall bond interest rate levels may not reflect potential supply changes, and interest rate levels have also caused some concern in the recent past.

However, it must be pointed out that interest rates themselves have an extremely minimal impact on the self-proof path of the AI bubble. If we focus only on the AI sector (ignoring debt pressure generated by main businesses), a fluctuation of about 2 percentage points in interest rates results in only a difference of tens of billions of dollars annually in interest payment pressure on overall AI debt accumulation.

Interest rates are more of an observation indicator, generally a “result” of changes in industry prosperity. What requires more attention is the cause leading to the change. Therefore, even if the market anticipates tightening by the Fed, “ceteris paribus,” changes in US Treasury yields and credit spreads are insufficient to constitute the fundamental reason for the easing or bursting of the bubble.

2. Sensitivity Level of Capital Expenditure

Next, we discuss the path changes corresponding to differences in capital expenditure under neutral depreciation periods (8 years/20 years) and neutral interest payment levels (7.5%). Of course, it must also be recognized that capital expenditure is a result of prosperity, so whether optimistic, neutral, or pessimistic scenarios, they are merely static descriptions.

For the accounting break-even level, the neutral CAPEX scenario (also using all baseline assumptions) corresponds to $938 billion in 2029 and $1.2 trillion in 2030, while the pessimistic scenario also requires revenue to reach $1.08 trillion in 2030. The optimistic high investment growth scenario seeks a revenue level of approximately $1.4 trillion in 2030.

The path differences for the external financing gap are more significant. Under the pessimistic Capex scenario, the external financing gap may peak this year, while the neutral and optimistic scenarios still need to peak around 2028, with the neutral scenario corresponding to a gap level of approximately $800 billion. It must be emphasized again that the premise of the path assumption is that revenue equals the accounting break-even level. In the high probability event that revenue cannot meet this condition, the scale of the financing gap will be larger, and the peak time will be more delayed.

The above static calculations may change significantly with technological breakthroughs and the emergence of new models.

But overall, aiming for a trillion-dollar revenue level by 2029 may become the baseline for AI technology to prove itself (not just Agent-ARR, but also contributions from other modalities). The reversal of the financing gap will continue for at least several more quarters, and concerns about free cash flow will not disappear. Even adopting pessimistic Capex growth rates and extremely conservative depreciation levels (plus maintaining low interest payment rates), a trillion-dollar leap must be achieved by the end of 2030 at the latest; of course, under more aggressive assumptions, the realization node will move forward to 2028.

Finally, the reality we face is that hundreds of millions of dollars in revenue in 2023 has already transformed into a revenue level exceeding $100 billion in 2026, and the technology itself is evolving at an extremely rapid pace. Therefore, it is normal to feel concerned about the trillion-dollar revenue threshold corresponding to 2029, three years from now, just as it was impossible to predict in 2023, three years ago, that global CAPEX would reach hundreds of billions of dollars and that leading tech companies would burn through all their free cash flow.

Sunk costs are accumulating, but we are still in a stage of fear of missing out: previously worrying about “0 to 1,” now worrying about “1 to 10”; in the face of industrial trends, right or wrong is only visible in hindsight.

Xuetao Macro Notes

Risk Warning and Disclaimer

The market carries risks; invest with caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Responsibility for investments made based on this content lies solely with the investor.