Much has already been written about AI transforming hotel operations. That’s great and amazing, but much less discussed is the economic reality underpinning these new tools.
Artificial intelligence isn’t free, and unlike traditional software licensing, its primary operating cost can be variable. Every prompt, workflow, agent, translation, report and recommendation quietly consumes ‘compute’ which is mostly energy but also factors in other resource costs like debt service on building all those shiny new datacenters. Our concern is that hotels aren’t properly stress-testing this rapidly growing budgetary item.
Let’s first take a big step back. Information technology is first and foremost about value creation, not just budget.
Value Creation Thinking
For instance, suppose I said I can source a bushel of 100 red apples for $50 instead of your normal price of $100, which you then sell retail at $2 per apple. That’s a direct savings or added net value of $50, with each commoditized unit interchangeable for no price delta accounted here (and not factoring in wastage or spoilage for this example).
Now suppose I told you about this hot new type of golden apple that’s selling retail for $5 per unit (2.5x that of red apples) but cost $200 for a bushel of 100 ($2 per apple). If you switch, you’re doubling your total cost per bushel, but there’s the potential for $200 in net new value creation (calculated as profit from golden apples minus profit from red apples).
Sometimes you need to spend more to make more. Yes, it’s still ‘potential’ rather than offering more certainty like the straight cost savings, but even when tested it may be a worthwhile venture to invest in some golden apples.
Enter Token Creep
Like this simple example, AI has the potential to 10x your value creation in many areas over the next decade. That all said, IT budgets are very real and we have to work towards them. This is where the concept of ‘token creep’ becomes critical.
We’ve previously argued that hotel technology should increasingly become a zero-sum exercise, where every new platform ought to replace an existing system rather than simply becoming another dashboard, subscription and implementation project. AI introduces another layer to this discussion because many of its costs don’t arrive as fixed SaaS licenses. They accumulate invisibly in the background through token consumption.
The challenge is that token creep may not manifest itself in just one way. Instead, there are two very different mechanisms that hotel owners and IT leaders should begin monitoring over the next several years. All of these harks back to the core macroeconomics of compute or tokens equating to energy, and how those resources are going to increase in cost over the next few years.
Public Inflation: This is the straightforward scenario whereby AI providers openly increase token pricing, API/MCP rates or subscription tiers as operating costs rise. Enterprise customers receive updated pricing schedules, budgets adjust accordingly and procurement teams negotiate new contracts. While nobody enjoys paying more, at least the economics are transparent.
Secret Shrinkflation: Rather than charging more, providers simply deliver less compute for the same monthly subscription or enterprise allocation. Token quotas quietly become tighter. Your monthly allowance runs out on the 15th each month instead of the 20th. Higher-performing reasoning models become premium upgrades. What previously ran comfortably within an enterprise subscription now requires another add-on module or a move to the next pricing tier. Or worst of all, lower-tier subscriptions become deprioritized, and you get the spinning wheel of death during high-traffic periods.
Consumers have seen this playbook before. Chocolate bars became significantly smaller during the pandemic. Potato chip bags now contain more air. Laundry detergent bottles quietly lose a few milliliters while maintaining the same shelf price. AI may increasingly adopt the same strategy because reducing allocations often creates less customer backlash than explicitly raising prices.
The Macroeconomic AI Arms Race
Training and serving frontier AI models requires extraordinary computational infrastructure. Every inference generated by a LLM consumes electricity across thousands of GPUs housed within increasingly expensive hyperscale datacenters.
Those facilities themselves require billions of dollars of capital investment, sophisticated cooling systems, networking infrastructure and access to reliable power grids that, in many regions, are already struggling to keep pace with demand.
At the same time, the frontier model developers remain locked in an arms race. Whether it’s GPT-5 and beyond, Claude’s Fable, Gemini, Llama or the next generation of open-weight reasoning models, every company is competing to build more capable systems with larger context windows, stronger reasoning and increasingly autonomous agentic capabilities. It’s more, more, more!
Despite any compute-saving techniques like ‘mixture of experts’ (MoE), better models generally require more computation during both training and inference. More computation means greater energy consumption. Greater energy consumption ultimately means higher operating costs that must eventually be recovered somewhere within the value chain.
Also add to this that the financing of datacenters and GPUs is very opaque right now. If there are IPOs then those balance sheets become public, forcing the leading technology companies in this field into liquidity events that may mandate higher pricing to service debt and generate free cash flow for shareholders.
Second also, think a bit more about the total, global energy consumption as it relates to the time it takes to roll out more energy infrastructure. If the war in the Middle East has taught us anything, it’s that the whole world is intricately connected, and supply crises in one region can ricochet outward to many other markets. So, be it solar (which is variable power), more efficient or transitionary carbon fuels, nuclear, small modular reactors, thorium power, or building datacenters in space, all these massive infrastructure projects take years to approve and complete. How’s your energy bill going to increase in the meantime? And how will costs for datacenter compute be paid forward to customers?
Third also, consider the impact on politics and AI guardrails. The one thing about computers is that they don’t vote, and a lack of jobs or upward career mobility can tend to make people angry, thereby incentivizing politicians to act on policies that curtail AI companies in various ways. One such method would be taxation, and you can get your bottom dollar that the bucks these AI companies would have to pay would be passed on to businesses using their tokens.
While these three also matters are speculation, they nevertheless show that AI usage is a pervasive socioeconomic challenge, with many signals pointing towards increasing costs to the end users.
The Hotel Budget Impact
Hotels may not notice these pressures immediately because many AI capabilities are currently embedded inside existing enterprise software. PMS vendors, CRMs, RMS platforms, booking engines and customer service applications increasingly advertise AI as part of broader subscriptions. Today, many of these features feel complimentary or modestly priced.
That may not last. As enterprise vendors begin absorbing their own AI infrastructure costs, many may inevitably return to customers during renewal negotiations with revised commercial models.
Premium AI modules may become separate line items. Existing functionality could be usage-capped. Three-year agreements may include aggressively escalating AI service fees or token consumption thresholds that didn’t exist during the initial contract.
The important point is that hotels shouldn’t think of tokens as merely an IT expense. They are becoming an operating input, much like cloud infrastructure did over the past decade. Except this time, consumption scales with every workflow employees automate, every AI assistant they consult and every intelligent process embedded in the organization.
Our immediate recommendation isn’t to slow AI adoption. Quite the opposite! It’s value creation, after all, and not just reining in budget.
AI will almost certainly become foundational across hotel operations, commercial strategy and guest engagement. But adoption should now be accompanied by ‘compute governance’. Just as finance departments stress-test labor costs, utilities and insurance premiums against multiple future scenarios, hotel technology teams should begin modeling token sensitivity over the next three to five years.
What happens if enterprise AI costs increase 20% annually? What if embedded AI modules become separately licensed? What if token allocations remain unchanged while employee usage doubles without a resultant bump in value creation?
These questions may seem premature today, but good strategists must think ahead. The hospitality industry has spent decades learning how to budget for software licenses, cloud infrastructure and cybersecurity. AI introduces a new variable where consumption, economics and competitive pressures all influence future operating costs simultaneously.
Token creep is unlikely to become tomorrow’s headline. It will more likely emerge gradually with one renewal, one usage cap and one premium feature at a time. That’s how shrinkflation works. The hotel companies that recognize this macroeconomic shift early will be far better positioned to embrace AI confidently rather than discovering, too late, that their technology budgets have quietly become hostage to compute.
