While Anthropic, OpenAI, and the media talk of an “AI Doomsday,” serious problems have emerged for the AI economy. Revenues aren’t growing and the biggest startups are delaying their IPOs, thus the tech sector needs a distraction, which the claims of an “AI Doomsday” provide.

The reality is that the recent declines in revenues for hyperscalers and the lack of growth in revenue in Oracle’s latest quarterly earnings (June, July, August) suggest that the pop is close, not to mention Larry Ellison’s on-again-off-again plan to sell Oracle stock.

Why am I expecting a pop? Because without growing revenues, the hyperscalers will not obtain the profits needed to justify their huge capital expenditures on data centers, thus will likely cut back on those capital expenditures. That would spell the end of growth for the semiconductor suppliers, whose profits are largely financing the AI boom by subsidizing the big losses by the AI labs.

We are in the greatest boom of all time

Let’s back up. We are in the midst of the AI boom, the greatest boom of all time. Data centers are being constructed at the fastest pace of infrastructure building in the history of the U.S., and this construction has caused the cash flows for the hyperscalers to go negative; Nvidia is the biggest beneficiary of this construction, and share prices are at record highs, again with Nvidia being the highest, followed by the hyperscalers. Furthermore, electricity and memory-chip prices are rising and Americans are angry. What could go wrong?

Recent data suggests that the prices of tokens, which are basic units of the data used to process AI models, have fallen since mid-May and surprisingly, token revenues for the hyperscalers have fallen since the beginning of August (see Figure 1). Apparently, paid usage of AI models has not risen sufficiently to cover the falling token prices and suddenly the uneasy balance between the AI Labs, hyperscalers, and the semiconductor suppliers is in danger of being overturned.

Until recently, the huge losses of the AI Labs were subsidized by the hyperscalers, and more recently by Nvidia and other semiconductor suppliers. This subsidization, including lots of free use, fuelled the demand for the AI models, thus the very rapid construction of new data centers that require rapidly increasing revenues to cover data center expenses. The falling token revenues, however, now make it harder for the hyperscalers to pay for those new data centers as they continue to come online.

Oracle’s quarterly earnings announcement was a shocker

Oracle’s recent quarterly earnings announcement showed that their revenues also didn’t grow much in the quarter covering June, July and August (see Figure 2), completely consistent with the declining revenues for all the hyperscalers mentioned above.

Oracle’s revenues grew from $19.2 billion in the previous quarter to $19.3 billion in the most recent quarter. It is highly likely that Oracle’s revenues did not grow in August while growing some in the previous two months.

A trend of flat or declining revenues is very bad for Oracle and the other hyperscalers for a lot of reasons. Here’s one: Because the hyperscalers have committed to many of these data centers over the next few years, they must find revenues to cover 100% of the payments from day one of their operation.

Each time a data center comes online, the hyperscalers are on the hook to find revenues to cover the costs. Yet, the revenues needed to cover the operation of data centers are not only no longer growing rapidly, they declined in August. Suddenly, the whole AI economy is thrown into question.

This includes Oracle, whose share price has fallen the most and who pays the highest interest rates because it has the most debt. As shown in Figure 2, not only were revenues largely flat between the last two quarters, their capital expenditures increased by more than 50% in the most recent quarter and have more than tripled in the last year.

The other hyperscalers have made the same big financial commitments to build data centers, thus are in a similar situation as Oracle. I expect their revenues to have declined quarter to quarter when they report earnings in late October.

Why did the token prices fall?

Delving deeper, why did the token prices fall? “Price cuts, user migration to low-cost open-source models, intensifying competition, and local inference are structurally eroding the per-token pricing model.”

What’s important to realize is that token processors have not really gotten better at processing tokens, rather it is that fewer tokens are needed because low-cost open-source models, particularly from China, are being more widely used.

So, the token price and revenue declines are really about which of the companies in the AI value-chain will get weaker or stronger. Up until now, the AI Labs were dependent on the largesse of Nvidia and also the hyperscalers, but now the hyperscalers look like they are the weakest link in the chain. Their revenues will be insufficient to cover their expenses so someone must save them, most notably Nvidia and the other semiconductor suppliers, which are the only profitable companies left in the value chain.

But how much can Nvidia do? And what will investors do to their stock and the stocks of the hyperscalers when they do this math? They have been assuming that revenues for data centers would rapidly increase to cover capital expenditures, but the latest data (and talk of a pause in AI development) will make it harder for them to believe that.