Own It or Rent It: The Model Got Cheap, the Stack Didn't

Eric Marketts

Welcome back to The V1 Airline Retailing Report. I'm Eric Marketts.

Steph Nell

And I'm Steph Nell. If you're new to the show, here's the quick version of what we do. We are the AI-avatars for V1 Advisory - presenting human insights on behalf of the V1 Team. Every week we take the two or three stories that actually moved the needle in airline and travel retailing, and we do a full 360 on each one. Bull case, bear case, and the critical take — what the industry is getting wrong or not saying out loud. We try to keep it honest and we try to keep it useful.

Eric Marketts

We come at it from two directions. I'm a tech and aviation journalist, so I'm tracking the news and pushing on the narrative. Steph applies years of airline distribution experience — knowing where the bodies are buried.

Steph Nell

Which is a colorful way of saying I know how these decisions get made. And have seen the same arguments play out for years with different names on the slides .

Eric Marketts

This week we are doing something a little different. We're going to spend more time on the context — the "what is this and why does it matter" — because the stories we're covering this week connect to some foundational questions about how the whole travel industry is structured. And those questions are relevant whether you've been in this space for twenty years or twenty days.

Steph Nell

Right. Last week we talked about the squeeze — AI agents driving up costs, margin pressure across the industry, intermediaries making moves on the demand side. This week we go one layer deeper. We're asking a single question that cuts across all three of today's stories. In the AI era, what does an airline — or any travel company — actually own? And what are they renting from someone else?

Eric Marketts

We built the episode in three layers to answer that. The intelligence layer — the AI model itself. The offer layer — the thing you're actually selling the traveler. And the demand layer — the interface where the traveler decides where to book. Story one, the model just got cheap. Story two, the offer engine the industry promised by this year is running way behind. Story three, while airlines debate, an OTA is quietly building the rails the agents will run on.

Steph Nell

Three layers, one question. Own it, or rent it.

Eric Marketts

Let's get into it .

Eric Marketts

Story one. Before we get into the news, I want to set up the landscape for anyone who's newer to this space, because the word "AI" gets thrown around a lot in travel and it means very different things depending on who's saying it.

Steph Nell

Good call. So let's start with the basics. When people in travel talk about AI right now, they mostly mean large language models — the same technology behind ChatGPT, Claude, Gemini. These are systems trained on enormous amounts of text that can understand questions in plain language and generate useful answers. In a travel context, that could be a chatbot that helps you plan a trip, an engine that reads your booking history and builds a personalized offer, or an AI agent that actually goes out and books flights on your behalf.

Eric Marketts

And until recently, the most capable versions of these models — what the industry calls "frontier" models — were controlled by a small number of companies. OpenAI, Anthropic, Google. If you wanted the best AI, you paid one of them and built on top of their system.

Steph Nell

Which created a real dependency. Your AI capability was only as good as your relationship with your model provider. If they changed their pricing, changed their terms, or simply had a better product launch a competitor's way — you felt it immediately. And in a margin-sensitive business like travel, that's a real risk.

Eric Marketts

That's the setup. Now here's what happened this week.

Eric Marketts

Two pieces dropped this week that I want to put side by side. The VC firm Madrona published "Frontier No More?" — the argument is that the dominance of one or two frontier AI models is ending, on five fronts at once. Open-weight models catching up in quality. Enterprises capping how much they spend on tokens. Regulatory kill-switch risk — the idea that a government could restrict access to a closed model overnight. Compute economics breaking down. And the orchestration around the model becoming the part that actually matters. Then Timothy O'Neil-Dunne, one of the most widely followed analysts in travel distribution, took that whole thesis and mapped it straight onto travel .

Steph Nell

And the travel translation is sharper than the original. O'Neil-Dunne's line is the one to hold onto. The question is not which frontier model will we use. It is, what do we have that survives the model being swapped out from under us. That is a distribution question wearing an AI costume.

Eric Marketts

Unpack that, because the word he leans on is "harness." What does that mean — in plain terms, and then in airline terms?

Steph Nell

Plain terms first. A raw model takes in text and gives you text back. That's it. It cannot hold state across a session — meaning it forgets everything the moment the conversation ends. It cannot hit live inventory — it has no idea what seats are available on your flight right now. It cannot enforce a fare rule — it doesn't know that your ticket is non-refundable or that the upgrade requires elite status. The harness is everything you build around the model to make it work in the real world. The connectors, the APIs, the data pipelines, the rules engines, the memory layer, the transaction system.

Eric Marketts

So the harness is the thing that makes the model useful for an actual business.

Steph Nell

Exactly. And here is where it gets interesting for travel. This industry has been building harnesses for forty years, and we just didn't call them that. Let me give you a quick picture of how airline distribution actually works, because it matters for everything we're going to say today. When you book a flight, your request doesn't go directly from you to the airline in most cases. It passes through a network. Historically, those networks are called GDSs — Global Distribution Systems. Amadeus, Sabre, Travelport. They're the plumbing that connects airlines to travel agents, to corporate booking tools, to online travel agencies. The airline files its fares with a company called ATPCO — the Airline Tariff Publishing Company — and those fares flow out through the GDS to every seller in the ecosystem. The GDS is a harness. A very old, very entrenched one.

Eric Marketts

And that's the system NDC is trying to replace — or at least compete with.

Steph Nell

Right. NDC stands for New Distribution Capability. It's an IATA standard — IATA is the global airline trade body — that lets airlines send richer content directly to sellers without going through the old GDS fare-filing process. Instead of a static fare table, an airline can send a full offer — here's the seat, here's what's included, here's the image, here's the price, tailored to this specific traveler. The airline's NDC connection is a harness. Their PSS — their Passenger Service System, which is the core operating system that manages reservations, inventory, and check-in — that's a harness. The order management layer being built for the next generation of airline retailing — that's a harness.

Eric Marketts

So O'Neil-Dunne's message is: stop worrying about which model wins. The defensible layer was always the harness.

Steph Nell

That's the message. And this industry already knows how to build those. The GDS is a harness that created enormous lock-in and billions in recurring revenue for Amadeus and Sabre for decades. The question is whether airlines build their own harnesses or keep renting from someone else's.

Eric Marketts

So is there a bull case here? Because for once that sounds like good news.

Steph Nell

It is good news, and it is real. Open-weight models — think of these as AI models that have been published for anyone to download and run on their own systems, the way open-source software works — mean a regional OTA or a corporate travel management company no longer needs a contract with OpenAI or Anthropic to deploy serious AI capability. You download the model, run it on your own infrastructure, train it on your own data, and you own it. No per-use fees. No dependency on a single vendor. And the compounding asset is your data. Decades of booking history, fare-sensitivity curves, ancillary purchase patterns — what travelers tend to add on, what they're willing to pay for upgrades. No AI lab can replicate that. You can fine-tune on whatever model is best this quarter and swap it next quarter without losing what makes your product good. The proof is already in production.

Eric Marketts

What's Penny?

Steph Nell

Priceline's AI travel assistant — the conversational layer built into their booking site. You ask it a question in plain language, it searches inventory and builds you an itinerary. It's their public-facing AI product. And they rebuilt it this month on a model-agnostic stack — Claude for reasoning, Google and OpenAI for other pieces — all wired to live inventory by Priceline's own orchestration layer. Amadeus acquired a company called SkyLink specifically for the orchestration engine. This is happening now .

Eric Marketts

Okay. And where does it break down?

Steph Nell

Three places. First, the piece is sold as good news for the little guy, but every example O'Neil-Dunne uses is a giant. Amadeus, Priceline. The same forces that help a small TMC — a corporate travel management company — help the platforms more, because they already own the data and the customer relationship. Democratizing the model layer doesn't automatically democratize the outcome if the compounding assets are already concentrated. Second, owning data is not using data. The whole argument rests on one quiet phrase in the original piece — "if they choose to exploit it." The constraint in travel AI was never getting access to a model. It is that most carriers have spent twenty years unable to unify their own data across their PSS, their loyalty program, their web booking flow, and the GDS feedback loop. The moat is sitting there unbuilt. And third, the big one. A beautiful harness does not help you if the traveler's first question starts inside ChatGPT or Google. That is the demand-interface problem, and it is still wide open. We'll come back to that in story three .

Eric Marketts

So what is the take an airline team should walk away with?

Steph Nell

The critical take is that this is the best argument yet for Offer and Order — and let me explain what that means for anyone new to the term. Offer and Order is the industry's shorthand for a complete overhaul of how airlines sell. In the old model, airlines file fares in advance, GDSs distribute those fares, and a booking creates a PNR — a passenger name record — that lives in a 1970s-era data structure. In the new model, the airline creates a dynamic offer in real time, the traveler accepts it, and it becomes an order — a complete record of everything the traveler purchased, managed end to end by the airline's own system. The offer engine is the harness. The order management system is the system of record. When the model is a rented, replaceable part, the value moves to where Modern Airline Retailing has been trying to push it all along — into the airline's own offer creation and order management. So the reframe is right, but it stops one step short. The question is not which model you rent. It is whether your data is usable, and whether you own the moment the traveler decides. And watch one tell. Amadeus is promoting a "neutral" Universal Commerce Protocol that Amadeus is building. Neutral layers built by one incumbent are how that incumbent sets the economics for everyone on top .

Eric Marketts

Hold that thread — because story two is whether airlines can even turn that data into an offer, and story three is who owns the moment the traveler decides. Credit to Timothy O'Neil-Dunne for connecting these dots first.

Steph Nell

The model just got cheap. Everything that decides whether you win still costs exactly what it always did .

Eric Marketts

Story two. We're going to talk about dynamic offers. Before we get into the news, Steph — paint the picture for someone who hasn't thought about how airline pricing actually works. Because I think most people assume airlines are already doing something sophisticated here.

Steph Nell

Most people do, and they're surprised when they learn the reality. Here's the classic model that has run airline pricing for decades. An airline's pricing team decides on a set of fares — let's say fourteen different price points between Detroit and Chicago. They file those fares with ATPCO, which publishes them to the entire distribution ecosystem. Every travel agent, every OTA, every corporate booking tool sees the exact same fares. The airline's job is to manage which fares are available at any given moment through what's called inventory control — you open and close fare buckets based on how the flight is selling. But the fare itself doesn't change. It's a menu, and the restaurant decides which items are available today.

Eric Marketts

So it's actually quite rigid.

Steph Nell

Extremely rigid. Now contrast that with a dynamic offer. A dynamic offer is created in real time, for this specific traveler, at this specific moment. The airline's system looks at who you are — your loyalty status, your travel history, your destination — looks at demand signals for this flight, looks at what's available to bundle with your seat, and generates a price and a package specifically for you, right now. The price could be higher or lower than a static fare. The bundle could include things that aren't even in the old fare filing system. And it can change minute to minute as conditions change.

Eric Marketts

Which is how Netflix prices, how Amazon prices, how virtually every modern digital retailer prices.

Steph Nell

Right. It's table stakes in e-commerce. Airlines are trying to get there. ATPCO — that same Airline Tariff Publishing Company that manages the old fare-filing system — has been leading the industry effort to make dynamic offers the standard. And back in 2022, the industry set a goal. Eighty percent of sold offers dynamically created by 2026.

Eric Marketts

And now it's 2026.

Eric Marketts

Story two takes that thread straight to the offer layer. Back in 2022, ATPCO set that industry goal — eighty percent of sold offers dynamically created by 2026. Created in real time from demand and context, not pulled from a filed fare table. Well, we are in 2026. The number came in around twenty-three percent .

Steph Nell

And give the progress its due, because twenty-three is up from six percent in 2022. That is real movement. The rails are genuinely being built. ATPCO and PROS showed in an IATA proof of concept that the Product Catalog — a new standard for describing what an airline sells — can interoperate with modern offer and order systems. Airline Order Posting, a piece of IATA's One Order standard, takes a dynamically priced order and feeds the fare data downstream so it can still be serviced in legacy systems. The Assembled Data Feed is coming. The standards work is not the thing that is stuck.

Eric Marketts

So is there a bull case in twenty-three percent?

Steph Nell

There is. Dynamic offers nearly quadrupled in under four years, and the hard interoperability problems — the technical handshakes between new systems and old — are getting solved in public by the industry bodies. The carriers that have a real offer engine today — continuous pricing, bundles built on the fly — are pulling away on revenue, because a dynamic offer captures willingness to pay that a static fare leaves on the table. If you've ever noticed that flight prices seem to react to your browsing history or change overnight, that's a crude version of this. A real dynamic offer engine is far more sophisticated. If you tie this back to story one, this is the harness paying off. The airline that built the offer engine turns its proprietary data into money in real time. That is the whole promise — and a handful of carriers are already living it. True continuous pricing is concentrated today. At last public count, roughly sixteen airline groups globally have reached the sophisticated version — what ATPCO calls adjusted pricing, where a live pricing engine adjusts fares at the moment of shopping. United and Air France-KLM are in that group. Lufthansa Group goes further still — they've been live on full continuous pricing, no pre-filed fares at all, for more than five years. The number is likely higher now; ATPCO hasn't published a revised headcount, they've shifted to reporting percentage of offers. But the point holds: the twenty-three percent of sold offers that are already dynamic comes almost entirely from that short list. Concentration effect, not broad adoption. The leaders show unambiguously that it works .

Eric Marketts

And where does it break down?

Steph Nell

It breaks down on what the gap between twenty-three and eighty actually is. It is not a standards gap anymore. The standards are largely written. It is a data and capability gap inside the airlines themselves. Continuous pricing needs clean, unified data — a single view of the customer, demand signals, competitive intelligence, and ancillary context all feeding into a pricing engine in real time. It needs a pricing science function that most airlines don't have at scale. It needs governance — rules about what you can and can't do with personalized pricing. And most carriers have not wired those three things together. This is the exact thing I flagged in story one. Owning the data is not using the data. The industry set a public 2026 target, the plumbing got built, and most airlines still cannot feed it. The bottleneck moved from the standard to the carrier's own house, and that is a much harder thing to fix because it's organizational, not technical .

Eric Marketts

So what is the take?

Steph Nell

The critical take is that dynamic offers are the scoreboard for whether modern retailing is real or just a slide deck. And the score says the hard part was never the model or the standard. Think of it this way — you could give a restaurant the best kitchen equipment in the world, but if the kitchen is disorganized, the ingredients aren't fresh, and the cooks don't have a recipe, the equipment doesn't help. Most airlines have been handed a world-class kitchen. The ingredients — the data — are there. The equipment — the standards — is being built. But the organizational work of actually cooking? That's where seventy-seven percent of the market still is. The model got cheap this week. The standards are basically here. The thing still missing for three-quarters of the market is the unglamorous work — the data pipelines, the pricing science, the governance. Until that's built, a cheap model and a finished standard sit on top of an offer the airline still cannot create. The asset stays latent, exactly like O'Neil-Dunne warned .

Eric Marketts

Story three, and this one I want to set up carefully because it connects to something that's changing fast in how people actually book travel. Steph, help me explain what we mean when we say "agentic AI" — because it's a different thing from the AI most people have played with.

Steph Nell

It is. Most people's experience with AI is what you'd call conversational — you ask it something, it answers. ChatGPT, Siri, the AI in your search bar. You're in the loop the whole time. Agentic AI is different. It's AI that acts on your behalf without you needing to approve every step. You give it a goal — "book me the cheapest roundtrip to London in July under a thousand dollars, with a hotel near the center" — and it goes out, searches, evaluates options, and completes the transaction. You come back and it's done.

Eric Marketts

Which is enormously convenient as a user.

Steph Nell

Enormously convenient. And enormously consequential for anyone who sells travel. Because right now, the way most travel is sold assumes the traveler is in the loop. You go to an airline's website, you see the options, you decide. Or you go to an OTA — an Online Travel Agency, like Expedia or Booking.com — and you browse and compare. An OTA is essentially a marketplace that aggregates supply from airlines, hotels, and car rental companies and lets you search and buy in one place. The OTAs built massive businesses on being the best interface for the human traveler.

Eric Marketts

But if the AI agent is doing the searching and deciding, the interface changes entirely.

Steph Nell

The interface that matters shifts from the screen the traveler looks at, to the data the AI agent can read and act on. And that's a very different competition. A human traveler might choose Expedia because the website is nicer. An AI agent doesn't care about the website. It goes where the data is cleanest and the transaction is most likely to complete successfully. Where it fails least. Where it trusts the result.

Eric Marketts

That is a completely different competitive landscape.

Steph Nell

Completely different. And it's happening now. That's the setup for story three.

Eric Marketts

Story three. The demand layer. At its Explore 2026 event, Expedia laid out an agentic roadmap — and there's a B2B piece in here that matters more than the consumer headlines. Expedia is building tools that let external AI agents interface with its inventory directly. Structured data and API infrastructure so partners and agents can connect straight into Expedia supply. On the consumer side, Expedia and Meta are testing AI conversations inside ads — so a traveler can start planning with a single tap .

Steph Nell

Put that next to story one. This is exactly the move O'Neil-Dunne said would decide it. Expedia is not chasing the model race. It is building the harness — making its inventory the easiest supply in travel for an AI agent to read and book against. Do that well, and you become the default counterparty the moment an agent needs to complete a transaction. You become the place agents trust.

Eric Marketts

Give me the bull case, because an OTA opening its inventory to outside agents sounds like the ecosystem getting healthier.

Steph Nell

It can be. If Expedia makes the agent handoff genuinely clean — meaning an AI agent can query availability, get accurate pricing, and complete a booking without errors — then conversion improves for everyone plugged into that supply, including airlines whose content sits inside Expedia's inventory. B2B partners — the tech companies and TMCs building AI travel products — get agent-ready rails they don't have to build themselves. And a serious operator shipping real infrastructure in 2026 beats another conference panel about the agentic future. The supply side getting machine-readable is, in the abstract, good for the whole industry .

Eric Marketts

And the bear case?

Steph Nell

The bear case is that this is a land grab for the demand interface — and we have the data on where it leads. Bain ran a test. They let AI agents loose to book travel and watched where they went. AI agents reached airline websites directly about five percent of the time, and routed everything else to OTAs — because the OTA data is cleaner, the inventory is more complete, and the transaction is more likely to succeed without errors. Five percent. Every other booking went to Expedia, Booking.com, or a similar platform.

Eric Marketts

That's a striking number. Ninety-five percent of AI bookings going to OTAs, not airlines.

Steph Nell

And the compounding effect is the real danger. So when Expedia ships an agent toolkit and airlines are still arguing about NDC schema — about the technical format of their data — the agents keep learning the same lesson. Expedia completes. The airline stalls. Every successful booking through Expedia's rails trains the AI system to go back to Expedia's rails next time. It reinforces itself. Whoever becomes the default fulfillment layer for AI agents owns the moment the traveler decides. And right now, that is not the airline .

Eric Marketts

So what does someone inside an airline actually do with this information?

Steph Nell

The critical take closes the loop on the whole episode. Story one said the model is cheap and your moat is the harness and your data. Story two said most airlines cannot yet turn that data into a dynamic offer. Story three is the punchline. While airlines work on those first two problems, Expedia is solving the third one — ownership of the agent relationship — and solving it for travel's entire supply broadly. The harness-plus-data thesis is correct. Expedia is just executing it one layer closer to the customer than the airlines are. So the question every commercial team should be able to answer — and I would argue most cannot today — is this: Who owns your agent relationship strategy? Not your NDC roadmap. Not your loyalty program. I mean specifically: what is the strategy for how AI agents find you, read your inventory, trust your data, and complete a booking with you directly? At most carriers, that seat is empty. Expedia just filled theirs .

Eric Marketts

Three layers, one question. Let me put it together. The intelligence layer — the AI model — just got cheap. So your edge moves entirely to the harness and the data only you have. The offer layer is where that data is supposed to become revenue, and three-quarters of the market still cannot do it in real time. And the demand layer — the moment the traveler decides — is being claimed right now by an OTA building the agent rails while airlines debate plumbing. Own it, or rent it .

Steph Nell

And the through-line is uncomfortable on purpose. Every one of these layers rewards the same thing: ownership of your data and the systems that turn it into an offer and put it in front of the buyer — or the agent making decisions on the buyer's behalf. The model got cheap. None of the things that actually decide who wins got any cheaper. Here's what I'd say to someone who's new to this industry and wondering why any of this matters. Airlines are one of the most data-rich businesses on earth. They know where you're going, how often you travel, what you pay, what you buy on board, where you stay. That data is extraordinarily valuable for building personalized offers and for being the counterparty an AI agent trusts. Airlines have the raw material to own all three layers we talked about today. The question is whether they build — and build fast — before someone else owns the layer for them .

Eric Marketts

If this was useful — and especially if you're newer to the space and found the context helpful — send it to someone making distribution, retailing, or AI decisions in the next twelve months. These are 2026 calls, not 2028 calls. The decisions that determine who owns the AI-era travel stack are being made right now.

Steph Nell

And if you are doing the hard version of this inside a carrier — unifying the data, standing up real dynamic offers, writing an actual agent-relationship strategy — we want to hear from you. That is the conversation the industry is not having loudly enough yet. The people doing the work tend to stay quiet. We'd like to change that.

Eric Marketts

We'll be back next Monday. I'm Eric Marketts.

Steph Nell

I'm Steph Nell. Thanks for listening — and for sticking with us through a longer one this week. We think it was worth it.

Eric Marketts

Own it, don't rent it. Stay sharp out there .