[Apologies for the longer-than-usual post—fair bit to unpack here.]
AI doomerism is top of mind these days, but we have a different and more prosaic question: when will AI firms start making money?
The answer is a key input into a case we made last year in the NYT: that AI is in a financial bubble. We still think so, and, if anything, our concerns have deepened.
It’s not just that the gap between the industry’s spending and investors’ returns on that spending looks characteristic of a financial bubble, as we initially argued. It’s that the gap has widened, such that the revenue flows needed to close it look implausibly large. If our assessment is correct, investor patience is likely to run out, and, depending on the pace at which they rush for the exit, the bubble will either pop or start to deflate.
In the time since we pointed to historically high valuations amidst low revenues, the stock market has increased by 15 percent, while equity valuations of AI firms are up by more than 25 percent. In other words, market investors are mostly waving off bubble concerns, of which we are far from the sole purveyors. They apparently believe—or they think enough other investors believe—that in the relatively near future, currently unprofitable companies will begin raking in profits at historically unprecedented rates.
But we see a problem with that assumption. Using a standard financial valuation model—see the methodology section below for details—we estimate that since 2024, just six firms (Google, Amazon, Microsoft, Meta, Oracle, and SpaceX) who are heavily invested in AI (so-called “hyperscalers”) have spent about $1.2 trillion on AI investments while only making $277 billion in revenues—a gap, so far, of about a trillion dollars. Revenues would need to be between $2.4 trillion (if they simply broke even on the investments) and $3.8 trillion (if they needed to match the return they receive on their existing businesses) over the next six years. This is equivalent to their current incremental AI revenues increasing between 13-45 times in just six years (the range depends on revenue flow assumptions).
To make up for their relatively slow start, the hyperscalers need to quickly become hyper-profitable. Just in the next year, they’d need to post revenues between $520-850 billion, roughly three to four-and-a-half times what we estimate they made this year. If we strip out circular revenues—those from a firm in which the lab has an ownership stake—the future revenues must be even higher.1
Part of what makes the revenue burdens so large in each year is because a tech-driven timing glitch means expenses are spread over a relatively short period of time. Investments like computer chips have a relatively short shelf-life; tech firms generally believe the chips that do computations for AI will generate revenue for 5-6 years. After that, the firm writes the value of the investments down to zero. This constraint makes it imperative that such investments quickly start generating revenue, which in practice requires two things: AI needs to be widely and rapidly adopted and it needs to be valuable enough to the firms using the technology that they’ll spend increasingly large sums on it.
But, as the revenue numbers noted above show, that’s not what we’re seeing. Simply put, AI firms are losing a race against time. Investments are far outpacing profits, and while we may be wrong, we’re hard pressed to see how the latter can catch up to the former. In fact, we fear the companies are about to fall even further behind; industry consensus estimates and plans from the hyperscalers themselves state that this is only the beginning of their AI capex spending spree. Adding forecasted expenditures out to 2030 (see figure below) results in these six firms spending $5.7 trillion in the five years from 2026-2030, requiring $13.1-$18.7 trillion to be earned in revenue in the next ten years.
Just how unrealistic would a turn to profitability of that magnitude be? The total amount of incremental AI revenue—from their AI business, not, in the case of Amazon, e.g., from their shipping you cat food—needed to just break even in the next 10 years is roughly equal to all the revenue these firms earned in the past 10 years. Put differently, businesses like Amazon and Google—which took 30 years to get their current size--would need to create another Amazon-and-Google sized AI business, on top of their existing ones, in the next 10 years.
Part of the reason for this gap between spending and profits is one that has been observed in essentially every technological revolution, many of which, we hasten to add, have made huge and indelible contributions to growth and living standards. Diffusion of new technology not only requires firms to adopt it, but also to figure out how to best integrate the technology into existing workflows, a process that can take considerable time. For example, in 1900, only five percent of mechanical power was electrified. It would be another 30 years before 80 percent adoption was achieved.
Firms don’t all move together in adoption, and even early adopters need time to figure out how to best integrate the technology into their workflow. Electrifying the factory floor was only the first step in gaining increased productivity from electricity; the real benefits came after employees understood how to best arrange the factory floor to benefit from a superior power source. Similarly, the productivity and income gains from new technologies like railroads or the internet were realized years after their initial invention.
If we’re right, and AI profitability continues to be lapped by spending, investors will gradually realize that promised future profits are taking too long to arrive, resulting in lower share prices and diminished AI investment. That’s a problem for anyone with a stock portfolio, as well as the broader economy, as AI has been supporting economic growth through both investment and wealth effects (the positive correlation relationship between higher wealth and higher consumer spending).
For these reasons, we would like to be wrong about our bubble call, which, of course, may be the case. But the numbers we’re seeing suggest the initial gaps between spending and profits that got us thinking about bubbles in the first place have only widened. What happens next is a matter of time, and that may be running out for the AI industry.
Methodology
We use a simple discounted cash flow (“DCF”) model to calculate required revenues in each year. The model takes as given how much AI capex has or will be spent in a given year and then solves for the discounted revenue required to 1) break even, discounting at the firm’s weighted average cost of capital (“WACC”) with generous assumptions on already earned AI revenue (the “low” scenario), 2) generate a hurdle rate that is relatively standard among firms, often 3-4 percentage points above the firm’s WACC with slightly less generous incremental AI revenues (the “baseline” scenario), and 3) generate an NPV equal to what their existing businesses earn above their cost of capital and attributing zero AI revenue that the firm hasn’t explicitly disclosed (the “high” scenario). We discuss each of the model’s elements and inputs in turn.
AI Capex
We estimate the capex that firms are spending specifically on AI by using company-provided figures when given. When firms instead only report capex across all business segments, we estimate AI capex by estimating from company documents what dollar amount of the capex was for non-AI purposes (e.g., Amazon buying a fulfillment warehouse) from 2024-2026. After 2026, we assume non-AI capex grows by seven percent. The residual consensus estimate after seven percent growth is then attributed to AI capex.
How much of the capex is equipment vs. buildings has been disclosed by the firms at different points in time. For example, Microsoft has indicated 2/3rds of AI capex is equipment, while 1/3rd is on datacenters, while Amazon has indicated it’s closer to 50/50, but these figures change over time as the price of chips and buildings change. Given this, we assume AI capex of 60 percent on equipment and 40 percent on buildings for each firm.
Depreciation and Taxes
Equipment and buildings have different useful lives and as a result are depreciated over different time horizons. We assign the useful life of the equipment based on company documents, which largely give a value of 5-6 years (some critics argue this is too high, which would support our “time-glitch” point above), while it is standard to depreciate buildings over 39 years. For taxes, we use the firm’s effective tax rate, per Bloomberg. We incorporate tax savings from the depreciation as a benefit the firm receives in each year.
AI Revenues
For 2024-2026, we estimate firms’ incremental AI revenue, i.e., the revenue they earned from AI that was either new AI business segments or revenue growth due to AI in traditional business segments. Two facts complicate the task.
First, the firms themselves rarely quote figures for direct AI revenue. Indeed, a Goldman Sachs June 10 analysis found that “corporate commentary during Q1 earnings season painted a similar picture to economy-wide surveys that suggest enterprise adoption is nascent. Roughly 54% of companies discussed AI in the context of productivity on their earnings call. However, just 11% of companies quantified the AI productivity gains on a specific use case and only 2% of companies quantified the impact of AI productivity on earnings (vs. 10% and 1%, respectively, last quarter). There was little differentiation in company margins and share price reactions between companies that discussed using AI and those that did not.” When firms do give explicit quantifications of AI revenue (or AI impact on earnings), we use those revenue figures directly.
The second challenge is because separating out AI’s dollar contribution to existing business segments is itself not easily measurable, even by the firms themselves. That is, while Google may be deploying AI across a variety of processes to help sell ads, it is difficult to disentangle how much of the ad revenue would have occurred without the AI. Put differently, it is difficult for Google—and everyone else—to measure AI’s marginal revenue productivity of capital when deployed in its traditional business segments.
To address this, we attribute revenue most generously in the “low” scenario, e.g., aggressively assuming 50 percent of revenue growth since 2023 in one of Meta’s advertising segments is driven by AI, slightly less generously in the “baseline” scenario by e.g., stripping out advertising revenue gains, and in the “high” scenario assume no incremental revenue is generated by AI unless it is explicitly disclosed by the firm.
Discount rates
For the “low” scenario, we calculate a conservative WACC by 1) assuming a 10-year Treasury rate of five percent, 2) an equity risk premium of five percent, and 3) taking each company’s CAPM beta and weighted average cost of debt (after tax) from Bloomberg. In the baseline and “high” scenarios, we assume the same WACC as in (1) but require NPV equal to capex plus 10 percent (in the baseline scenario) or the cost of capital plus the historic return earned above that (in the “high” scenario).
If the firms don’t hit these revenue targets next year, the investments might still be profitable, but the burden of future revenues becomes successively larger. For example, if incremental AI revenues “only” increased by 66 percent next year (which would be on trend with the most generous AI revenue attribution assumptions) to $313 billion, they would then need to make $612 billion-$1.2 trillion in 2028 for the investments to be on track to be profitable.