Going to #HITEC? Ask about Tokenmaxxing!

The technology industry spent two quarters paying staff to burn AI tokens — then spent the next two nursing the hangover. Now token prices are collapsing, and that should worry you, not reassure you: cheap tokens are how the bill blows up, not how it shrinks. Hospitality is arriving at the party just as tech files out. Here is the one word that protects you on the HITEC floor, and the single number that turns it into a discipline.

A Word the Technology Industry Already Regrets

Walk the HITEC floor this year, and nearly every booth will say the same thing. AI. On the banners, in the demos, stitched into the lanyard talk. The word has stopped discriminating — everyone has it — so you need a sharper instrument. One question does the work. Ask them about tokenmaxxing.

Tokenmaxxing means using or consuming as many AI tokens as possible, often to signal heavy AI adoption or productivity. Still, it can be misleading because high token use does not automatically mean useful output, better decisions, or business value. In hospitality, the smarter approach is to measure outcomes — faster service, fewer errors, better guest responses, improved revenue, or lower workload — rather than simply counting how many tokens were burned.

Tokenmaxxing is a real, well-documented phenomenon in the first half of 2026 — and the firms at its center are already retreating from it. The recipe was simple: push staff to use AI as hard as possible. Weekly targets. Internal leaderboards. Bonuses tied to consumption. Amazon, Meta, and OpenAI all stood up rankings to see which engineers burned the most tokens. Meta went furthest, making AI use a core expectation in performance reviews and running a leaderboard it nicknamed "Claudeonomics" that piled up 73.7 trillion tokens in roughly a month.

Then the invoices arrived. Meta now warns of billions in internal AI costs and has pivoted from tokenmaxxing to token management — budgets, allocations, and a central dashboard to track spend. Amazon scrapped its leaderboards with a one-line correction: don't use AI for its own sake. Disney told engineers to raise their velocity, but warned them off tokenmaxxing by name. Microsoft's Satya Nadella called the behavior "addictive." Paramount imposed monthly per-user token caps. Uber torched its entire annual token budget in four months and conceded it could not draw a clean line between the spend and the benefit. Hype, bonanza, hangover — the full arc, compressed into two quarters.

Matan Grinberg, who builds AI tooling at Factory, sorts that arc into three phases worth carrying onto the show floor. Phase one: the board corners the chief executive and demands an AI strategy. Phase two: AI at all costs — usage written into performance reviews, adoption chased for its own sake. That is tokenmaxxing, and it arrived faster than anyone planned. Phase three is the hangover, when somebody finally opens the invoice. One CIO he spoke to found the firm spending hundreds of thousands a month routing trivia — the day's weather, a calorie count — through a frontier model. You do not, as Grinberg puts it, need "the frontier of human intelligence" to tell you it's raining.

Why It Was Always Going to Fail

None of this would surprise anyone who has opened a management textbook. In 1975, the economist Charles Goodhart noted that a statistical regularity buckles the instant you make it a control target. The popular restatement, credited to the anthropologist Marilyn Strathern, lands the same blow: the moment a measure becomes the thing you are judged on, it stops measuring anything worth knowing. Reward the proxy, and people optimize the proxy — not the goal it stood in for.

Token consumption is a textbook-perfect bad target. It takes an abstract ambition — "become an AI-led business" — and collapses it into a concrete, countable number, which is precisely the condition researchers call surrogation: the metric quietly becomes the mission. Staff duly obliged — writing deliberately baroque prompts, leaving agents grinding overnight, spinning up purposeless side-projects whose only job was to consume. Palantir's Shyam Sankar reduced it to five words: "More tokens means more slop."

We have watched this film before. Software once measured productivity in lines of code, until it was obvious that more lines mostly meant more bloat — Bill Gates likened it to judging an aircraft's progress by its weight. The AI-era remake is already reflected in the numbers: one large study of GitHub developers found that coding agents produced a 741% increase in lines of code, but only a 20% increase in actual software releases. Activity soared; output barely twitched. The macro scoreboard is no kinder. MIT's NANDA initiative found that 95% of generative AI pilots produced no measurable business impact, and McKinsey logged 88% adoption, with just 39% reporting any earnings effect. Adoption tells one story. Outcomes tell another entirely.

Why Hospitality Should Lean In, Not Exhale

It is tempting to file all this under someone else's problem: big tech, big budgets, big egos. Resist the temptation.

Look at what our own trade press is toasting right now: AI everywhere — booking engines, payments, personalization, "intelligent" everything. All real, all promising, and all framed in exactly the adoption-first language the technology sector has just spent two quarters regretting. Hospitality runs a cycle behind. We are about to stroll, beaming, into a room the technology industry is already filing out of.

And our exposure is shaped differently — arguably worse. Hotels don't build AI. We buy it. That shifts the misaligned incentive from inside the building to across the negotiating table. In the enterprise version, the company wanted output while the employee gamed consumption. In the hospitality version, your vendor's revenue often is your consumption. The market is converging on usage-based pricing as the default — "usage is probably going to be the predominant monetization strategy in the agentic world," as one marketplace chief put it — while building token meters to bill for it.

And the meter is getting cheaper — which is the trap, not the reprieve. The price war has already arrived: Anthropic has cut its flagship Opus rate by roughly two-thirds, OpenAI is undercutting it, and both are racing to offer discounts ahead of public listings. Lower unit prices sound like relief. They are the accelerant. This is Jevons' paradox in a hotel uniform — make a resource cheaper, and you don't spend less, you consume so much more that total spend climbs anyway. Cheap tokens are exactly what tempt a property to push every interaction through a reasoning model, to stack multi-agent workflows.

Because it can now afford to stop asking whether the work was needed at all, the analysts tracking the same price war in healthcare have already caught the sleight of hand: one new model generation quietly raised token counts by up to a third for identical text. This hidden volume premium swallowed much of the headline cut. A cheaper token still bills by the token. When the meter is the business model, "use more AI" stops being neutral advice. It is the pitch.

That is the trap, re-tailored for hospitality. Not your staff chasing a leaderboard, but a procurement decision that quietly rewards the supplier every time the system does more work than the guest actually needed.

The Instrument: Measure the Guest, Not the Tokens

The answer is not to fear AI or buy less of it. The firms that recovered didn't renounce AI; they moved from tokenmaxxing to what is now being called efficiency-maxxing — the most value wrung from each token, rather than the most tokens burned. They stopped counting activity and started counting outcomes.

And it is not abstract. The levers are known: route routine work to the cheapest model that clears the bar, rather than paying frontier rates to ask the weather; cache what repeats; reach for an open-weight model — or no model at all — when the frontier is overkill. Grinberg frames the real C-suite task of the next two years as resource allocation across three currencies at once — tokens, dollars, and people — aimed at the few outcomes that move the business, not the intermediate metrics that flatter it. Shipping four features instead of three was never the point. Serving the guest was.

Hospitality already knows this move cold. We don't judge the chiller plant by kilowatt hours consumed; we measure energy cost per occupied room, and USALI gives it a line. We don't praise a kitchen for buying the most food; we watch the cost per cover. The discipline is native to us. It has never been pointed at intelligence.

That is the entire purpose of TCPG — Token Cost Per Guest. Total AI token cost, divided by guests served. One number that converts an abstract, vendor-flattering input into a concrete, operator-owned outcome. It is the anti-tokenmaxxing metric by construction: the only way to improve it is to serve the guest better for less, which is the whole reason you bought the technology. Give it a home in the accounts, lobby it into USALI, and "how much AI are we using?" finally becomes the right question — answered in guests rather than tokens.

Other sectors are already there under their own names — healthcare analysts now model cost per patient encounter down to a fraction of a cent: same instinct, different guest.

And here is the part no meter can see: some of your best work is not token denominated at all. The welcome that lands. The problem quietly fixed before checkout. The concierge who remembers. No reasoning model improves those, and pointing tokens at them only adds cost. The discipline is not merely spending fewer tokens — it is knowing which moments deserve none.

The industry has named the disease for us. It is called tokenmaxxing. The cure still needs selling.

Six Questions for the HITEC Floor

So when a vendor leans across the booth and tells you the platform is AI-powered, agentic, frontier-grade — smile, and ask the questions that separate value from vanity.

1. Does your pricing bill me based on tokens consumed or on outcomes delivered? Whose incentive does the meter serve — yours, or mine?

2. Show me the cost per guest, not usage. If your dashboard can't express it, why not?

3. Do you route the routine work to the cheapest model that clears the bar — or am I paying frontier rates to ask the weather?

4. When your model or tokenizer changes, does my token volume jump for the same task — and who absorbs that cost, you or me?

5. Name a customer whose per-guest spend fell while service held steady or improved. That is efficiency-maxxing — prove you can do it.

6. Where does this sit in my USALI accounts — and can you render it as Token Cost Per Guest?

A vendor who can answer those is selling you a tool — a vendor who can't is selling you a meter.

The One Word

The booths will all say AI. The good ones will tell you what it costs per guest — and prove the number is falling even as they sell you more capability. That is the whole difference between a partner and a passenger on your P&L.

Take one word onto the floor this year. Ask about tokenmaxxing — and watch who flinches.

The intelligence may be artificial. But the experience is human.

Sources & Attribution

RTÉ News — "Token-maxxing: How tech firms' AI staff push backfired," 13 June 2026.

The Decoder — "Meta shifts from 'tokenmaxxing' to token managing as internal AI costs reportedly hit billions," June 2026.

Business Insider — "Disney Wants Tech Staffers to Move Fast With AI Without 'Tokenmaxxing'," June 2026.

Quartz — "A 50-year-old economics law explains why AI token maxing was always going to fail," June 2026.

Matan Grinberg (CEO, Factory), in conversation with Harry Stebbings — interview, June 2026; three-phase framing and the frontier-on-trivia account drawn from this source.

Nelson Advisors/healthcare.Digital — "What would an Anthropic v OpenAI Token Price War mean for HealthTech?", June 2026; source for price-war figures and the tokenizer-density premium.

Tech Brew (Morning Brew) — "How low can tokens go?" 12 June 2026.

Channel Dive — "Pax8 builds token tracking to help MSPs bill for AI," 13 June 2026.

Tech Bullion — "How hotel commerce technology is transforming revenue generation in 2026," 12 June 2026.

Frameworks referenced: Goodhart's Law (Charles Goodhart, 1975) and its Strathern restatement (1997); Campbell's Law (Donald T. Campbell); surrogation (Choi, Hecht & Tayler); vanity vs. actionable metrics (Eric Ries); the multitasking model (Holmström & Milgrom, 1991); Jevons' paradox (William Stanley Jevons, 1865). Primary reporting is cited above; the synthesis and the TCPG framing are the author's.

Made with the help of various AI tools, but with a HITL.