Amadeus has signed a memorandum of understanding with NVIDIA after early tests of GPU acceleration for flight search, and one of its own executives has said search is moving toward real-time computing as AI demand grows.
On 21 October Amadeus and NVIDIA signed a memorandum of understanding to deepen work they had already started on flight search. The published detail is limited. Amadeus used NVIDIA’s CUDA platform to adapt one part of the flight-search process so that many calculations run at the same time on GPUs, and early tests on NVIDIA’s Hopper and Blackwell chips showed what the release calls promising potential for response times and scalability. The companies also say they are assessing travel-specific foundation models and cloud efficiencies. The release gives no benchmark figures, no timeline and no customers, and TDN is treating it as exploratory work rather than a product.
The scale explains the interest. Amadeus has said it processes about 3 billion flight search requests a day, and that it ran 2.5 billion machine-learning inferences a day in 2025 to optimise those searches. Both figures come from Amadeus and have not been independently audited. In a case study published by Google Cloud, Luc Viguié, Amadeus’s director of strategic partnerships, said personalization and more sophisticated airline retailing are reducing how far Amadeus can rely on caching, so more results have to be computed in real time, while search volumes keep growing. That is Amadeus’s explanation of its own workload, not a verdict on caching across the industry. The case study adds that AI assistants querying travel systems on behalf of users will probably accelerate the growth.
The same case study shows Amadeus exploring more than one hardware route. Its shopping engine is planned to run on Google’s Axion Arm processors. Early benchmarking indicated approximately 10 percent lower backend response times than previous Arm hardware, and the case study’s summary lists a throughput gain of up to about 20 percent as anticipated, within a phased rollout. The NVIDIA work is a parallel exploration, and nothing published says the two share a production workload or architecture. TDN reads both as hardware responses to a demand problem, though Amadeus has not framed it that way.
Amadeus is also working on the other side of the equation. Its Advanced Airline Profile uses machine learning to screen out irrelevant NDC shopping requests from travel sellers before they reach airline systems. Amadeus said early results showed that the tool could reduce unproductive traffic by more than 70 percent in some airline profiles, without compromising offer quality. Maxime Boussard, NDC program director at Air France-KLM, separately said it had blocked more than 70 percent of irrelevant traffic to the airline’s system. Amadeus said at the time that it expected deployment across its NDC airlines to be completed by the first quarter of 2026, and TDN has not confirmed completion. Both claims come from the company or its customer rather than independent confirmation, and the filter protects airline systems rather than reducing the compute behind Amadeus’s own search engine, so it sits at a different layer from the NVIDIA work.
Sabre is emphasising a different part of the problem. Its SabreMosaic Cache-powered Intelligent Shopping, announced in February 2026, uses predictive algorithms to choose between live airline polling and validated cached results, aiming to cut unnecessary requests without sacrificing the accuracy of bookable offers. Sabre says measured deployments have seen up to a 28 percent reduction in look-to-book ratios, up to 95 percent accuracy against live polling, and response times under 500 milliseconds for traditional flight content. Those are vendor-reported figures with “up to” qualifiers, and the release does not say how many deployments they cover. Wego, the pilot partner, said it began co-developing the product two years earlier because it expected agentic shopping to push look-to-book ratios up asymptotically. Sabre also pitches the product as a way for agencies to avoid airline penalties, which shows where look-to-book pressure lands.
Travelport’s public response has emphasised speed and agent productivity. In July 2025 it reported average air search results of 0.84 seconds on Travelport+, 23 percent faster than before, even though shopping volumes had risen 200 percent since 2023. Travelport based the figure on its own internal testing, and it predates the NVIDIA announcement, so it is not a current like-for-like benchmark. Its AI-powered Content Curation Layer addresses a related but separate problem, helping agents navigate fragmented content.
Read together, the three providers are working on different combinations of filtering, caching and compute acceleration rather than choosing one strategy each. None has published what a search costs under its approach. The word cost covers at least three things that are easy to confuse: the infrastructure needed to process a search, the distribution cost of querying airline systems including look-to-book penalties, and the total cost per completed booking. GPU acceleration, caching and filtering could each lower one of these while leaving the others unchanged, and the published material does not say which.
TDN has not independently tested any of these claims, and everything above comes from company releases, technical material and case studies whose testing methods differ too much for a direct comparison. Amadeus has not said what its early GPU results mean in numbers, which part of the search process benefits most, how the economics compare with its Axion-based CPU work, or who the “other actors” in the foundation-model effort are.
In TDN’s view, the MoU matters less for what it delivers today than for the problem it exposes. Amadeus describes itself in the release as the neutral execution layer between providers, sellers and digital assistants, and as search grows more complex, the question of who pays for that work becomes harder to avoid. The useful measure is not which system returns results fastest, but which approach delivers an accurate, bookable result at the lowest total cost, and whether the savings survive when AI assistants generate more queries per completed booking. Smaller agencies and carriers, including many in the markets TDN covers, may have the least room to absorb rising costs, though no published figure yet shows how exposed they are. Until a provider publishes comparable numbers on cost, response time, accuracy and conversion together, the industry can compare speed claims but cannot say which approach is cheapest.
