What AI really costs in travel — and the architecture that survives it

Earlier this year Jensen Huang said he’d be “deeply alarmed” if a top engineer on $500K a year wasn’t burning at least $250K a year in AI tokens. The market, meanwhile, has been trained to think AI costs $20 to $100 a month. Those aren’t two estimates of the same thing — they’re two different conversations, and in travel the gap between them is where budgets quietly go to die.

The bill stops being predictable

When I started building Bitravel I made the same assumption everyone makes: monthly subscription, manageable overhead, a line you could forecast. Then I ran real travel data through real models — flight inventory, hotel rates, fare rules, supplier exceptions, raw GDS responses, straight into the prompt. The bill was not predictable anymore.

Travel is unusual here. One search can throw off more data than many industries touch in a whole workflow. Send all of it to the most expensive model every time and you’re not managing cost — you’re watching it compound. Uber’s CTO said publicly they’d burned through their entire annual token budget. If it caught them out, it will catch out a travel team.

The dimension nobody’s pricing yet

When you replace a person — a salary, a fixed annual line — with an AI agent on usage-based pricing, finance loses solid ground. A salary is forecastable. AI cost scales with activity, and in travel activity is seasonal and spiky: a fare war, a schedule change, a region that suddenly books out. The P&L has no model for an expense that behaves like that yet. It’s a finance problem dressed up as an engineering one.

The question that actually fixes it

For a while I was asking “how do we spend less.” Wrong question — it just degrades the answers travelers get. The right question is narrower: what does the model actually need to see?

• Strip the data before it reaches the model — send only what this step needs.
• Fragment the workflow — a booking is a chain of small reasoning tasks, not one giant one.
• Route by complexity — cheap models for mechanical steps, expensive reasoning only where it’s genuinely needed.

Not glamorous. It’s data discipline. But it’s the difference between an AI travel agent you can run at scale and a demo that stops being viable the moment real volume hits.

Precision beats spend

Jensen is right that serious AI costs serious money. But the edge won’t go to whoever spends the most — it’ll go to whoever understands their own data well enough to spend with precision, and to whoever can hand finance a number that doesn’t lurch every peak season.

This is the kind of thing we work through building Bitravel — an AI travel agent built to handle real travel data at real volume, accountably. Book a 30-minute call for a live look, no deck.