AI's problem: The missing revenues
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Aïda Amer/Axios
Nvidia's AI chips may be flying off the shelves, but they don't seem likely to pay for themselves in the form of higher corporate revenues any time soon.
Why it matters: The U.S. stock market continues to hit new highs, driven in large part by optimism surrounding the coming AI revolution.
- In the best-case scenario from skeptics and cautious optimists, however, the promise of AI will take much longer to materialize than the current investment frenzy suggests.
- In the worst case, it never will.
Between the lines: Either way, billions of dollars in capital is almost certain to be incinerated.
The big picture: Newly published reports from Goldman Sachs, Barclays, and Sequoia Capital have crunched the numbers on how much has been and will be spent on AI-related infrastructure, and how much extra revenue companies will need to make all that spending worth it.
Between the lines: "Overbuilding things the world doesn't have use for, or is not ready for, typically ends badly," Goldman Sachs head of global equity research Jim Covello warns.
- Sequoia's David Cahn similarly cautions against "the delusion that we're all going to get rich quick, because [artificial general intelligence] is coming tomorrow."
Follow the money: Goldman Sachs projects that companies and utilities will spend about $1 trillion on AI capex in the coming years.
- Nvidia data centers will be sold at a pace of about $150 billion per year on an annualized basis by the final quarter of 2024, estimates Cahn.
- Cahn's rule of thumb is that companies in aggregate will need to generate about four times in revenue what they spend on Nvidia data centers in order to cover the cost of energy and their own margins.
Zoom in: The lack of revenue is at the core of the skepticism.
- Cahn notes that OpenAI is still generating the bulk of AI-related revenue right now, and its annualized revenue has been pegged at a mere $3.4 billion.
- Even his generous predictions of $5-$10 billion in annual revenue from major tech companies — from Google and Meta to Tencent and Tesla — still leaves a giant hole of $500 billion in revenue just to make up for 2024's infrastructure investment, according to his calculations.
Reality check: Barclays estimates that AI capex by 2026 will be sufficient to support 12,000 AI products of the scale of ChatGPT.
- "We do expect lots of new services that will bring some of this bull case to light, but probably not 12,000 of them," write Barclays analysts.
- Meanwhile, Goldman's Covello points out that even Salesforce, which has been aggressively spending on AI, showed little revenue boost in its Q2 financials.
Zoom out: Other unknowns include whether AI tech will become cheap enough to generate significant cost savings, whether it'll solve the kind of highly complex problems that would make it worth the price, and whether we'll be able to supply the energy needed to keep up with AI's growth.
The other side: Tech leaders don't see "overbuilding" as a dirty word.
- They remember how the dot-com bubble overbuilt telecom capacity before the bust of 2000-2002 wiped out legions of investors. But within a handful of years, all that capacity — and plenty more — was put to good use.
The bottom line: Generative AI's topline benefits are not arriving anytime soon, realists argue.
