Lucía wants to take a Mediterranean cruise in June. Before booking, she compares cruise lines, dates and cabin types. She looks for flights to the departure port and a couple of nights in a hotel in Barcelona. She opens three websites, changes the dates and checks again the next day in case the price has dropped. In the end, she books just once.
For the travel industry, that single booking has triggered a cascade of queries, and only one of them ends in a sale. The industry has a name for this ratio: look-to-book, the number of searches needed to generate one booking. It is one of the metrics that most concerns airlines, travel agencies and technology providers today. So much so that IATA, the International Air Transport Association, devoted a specific report to it in October 2025, produced together with airlines, technology providers and sellers.
The figure has skyrocketed. In the early days of online sales, airlines handled between 100 and 200 searches for every ticket sold. Today the figure stands between 10,000 and 20,000, and the consultancy OAG estimates that, with AI-powered trip planning, it could reach 200,000.
Each of those searches costs money, energy and computing power. And here the travel industry faces a problem that other industries don’t. A flight search is not like loading a web page, which can be cached and served a thousand times without being recalculated. Prices and seat availability change constantly. That is why airfares cannot be cached like the rest of the content on the internet, and each query largely requires everything to be calculated again.
The scale is huge. Amadeus, the Madrid-based technology company that connects much of the industry, processes around 3 billion flight searches a day. For each origin and destination, it calculates thousands of routes, billions of fares and trillions of possible offers per passenger. And all of it must be resolved in seconds, while the user waits in front of the screen.
Behind those figures lies very concrete infrastructure. The Amadeus data center in Erding, near Munich, which the company itself calls “the brain of travel”, processes more than 1.6 billion transactions a day. It is one of the largest private data centers in Europe and is constantly being expanded and upgraded.
Packages and cruises add another layer of complexity. A booking like Lucía’s combines inventories that live in different systems: the cabin, the flight, the hotel, the transfers and, sometimes, the excursions. Each one has its own availability and its own rules, and all of them must be locked at the same time until payment is confirmed. If one fails, the whole package fails. And if two customers try to book the last sea-view cabin in the same second, the system has to decide, with no margin for error, who gets it.
Now come AI agents. A human traveler gets tired, makes a decision and books. Until now, every search required a person in front of a screen, with an intention and a click. Agents remove that limitation. An agent checks again every time a price changes, compares across services, negotiates on the traveler’s behalf and can launch hundreds of searches to plan a single trip. And if a user relies on several agents, or agents query other agents, the effect multiplies.
OAG’s CEO summed it up with a very vivid example: it makes no sense for answering a query to cost 200 dollars when the ticket costs 100. It is a deliberate exaggeration, but it points to a real risk: that the cost of searching grows faster than the business of selling.
For infrastructure, this means a fundamental shift. Behind every search there are processors at work, energy being consumed and heat that has to be removed. The industry’s data centers are sized for searches, not for sales. And that decouples the cost of the infrastructure from the revenue it generates: traffic can grow tenfold without a single extra ticket being sold.
The answer lies in efficiency, and here AI plays on both sides. Amadeus uses machine learning models to discard unhelpful fare combinations before calculating them, with around 2.5 billion inferences a day in 2025. Does it make sense to explore a stopover in Toronto and another in Rio de Janeiro to fly from New York to Nice? Probably not, and the model rules it out before spending resources on it. In four years, the company has tripled the efficiency of its applications: servers three times faster that need three times less computing for the same volume of queries.
The paradox is an interesting one. AI creates a large part of the problem and, at the same time, is part of the solution.
For those who operate the infrastructure, the nature of the challenge changes. Having enough capacity for peak-season spikes is no longer enough. You need to:
- Know how much each workload really consumes.
- Anticipate the growth of machine-generated traffic, which does not follow the patterns of human traffic
- Plan power and cooling with real data, not estimates.
- Measure efficiency per query, not just total capacity.
The next peak season won’t be set by travelers alone. It will also be set by the machines searching on their behalf.
And when Lucía finally boards, another story will begin: the story of the data center that travels with her.