Booking Holdings CFO Says Hyperscalers Spending Hundreds of Billions on AI Don't Know Their Return

Ewout Steenbergen runs the finance side of Booking Holdings, the company behind Booking.com, Priceline, Agoda and OpenTable. Speaking Thursday at Fortune's AIQ Summit at the New York Stock Exchange, he said the hyperscalers plowing hundreds of billions into large language models don't actually know what they're going to get back.

"I even think that the hyperscalers that spend hundreds of billions on the development of their large language models, they don't really know what is going to be the ROI," Steenbergen told Fortune senior writer Sheryl Estrada. "They have maybe some assumptions and some hypothesis, but the whole point is: you don't want to fall behind."

Heard secondhand from a CFO whose company doesn't build the models, this explanation also clarifies why Booking isn't trying to.

The Math Behind the Spending Spree

JPMorgan estimates AI capital expenditure will reach $800 billion in 2026 and climb to $1.1 trillion in 2027, according to figures cited by Crypto Briefing. That spending is projected to consume roughly 93% of hyperscalers' operating cash flow in 2026, up from just 33% in 2023.

Morningstar has floated a payback period of three to five years for AI infrastructure, meaning the investments made today would need to start generating returns before 2030 to pencil out, Crypto Briefing reported. These are forecasts from two research firms, not confirmed outcomes, and nobody has published hard numbers showing the hyperscalers have hit that bar.

The defenders of this spending have a real argument. If you believe large language models are the next computing platform, sitting out while a handful of competitors grab the infrastructure and the customer lock-in is the bigger risk. Steenbergen said as much himself: the point of the spending is not falling behind, even without a clean ROI model. Demand for chips, data centers and compute is real revenue flowing to Nvidia, the cloud providers and the broader supply chain right now, not a hypothetical.

But that argument cuts both ways. Spending at 93% of cash flow leaves almost nothing for shareholder returns and almost no cushion if demand doesn't show up on schedule. This is private capital, not taxpayer money, so if the bet goes bad it's shareholders eating it, not the public. That's how it should work.

What Booking Is Actually Doing Instead

Booking isn't building foundation models. It's renting them, and Steenbergen laid out a far more measurable picture of what that buys.

On customer service, bookings are up at a high-single-digit rate while service costs are slightly down, Steenbergen said, which means cost per booking has fallen "by a lot." Customer satisfaction rose at the same time, he said.

On engineering, Booking's roughly 9,000 engineers are getting about 30% more code into production, counting only merge requests that pass testing and quality control, according to Steenbergen.

To keep the AI bill from spiraling, Booking uses what Steenbergen called "effective model cost routing": cheap, open-source models handle simple tasks, pricier models get reserved for complex ones. Engineering teams are judged on total IT cost per merge request, a blended figure covering both human labor and AI token spend. Token costs can rise, he said, as long as the cost of getting each piece of code into production keeps falling.

On the customer-facing side, the numbers are still tiny. Steenbergen said on Booking's Q2 earnings call in August that referrals from large language models, paid and unpaid combined, were under 1% of total room nights booked. The company's near-term AI story is mostly an internal efficiency story, not yet a customer acquisition one.

Steenbergen was clear that he doesn't think AI changes the fundamentals of travel itself. "The hotel is still the hotel and the airline is still the airline and the rental car is still the rental car," he said. What changes, in his view, is the planning and booking experience. Travelers currently check about five different platforms on average before booking, he said, and Booking's bet is that AI collapses that research process and helps manage disruptions, like a delayed flight that wrecks a restaurant reservation, under what the company calls its "connected trip."

Booking still spends $8 billion to $9 billion a year on paid customer acquisition, through search, social media and metasearch sites, with roughly a third of its customers arriving through those paid channels. Whether AI-driven discovery tools cut into that bill, or just add another channel on top of it, is a question Booking's own numbers don't yet answer. The next earnings call will be the place to look for whether the sub-1% LLM referral share has moved.