How Much Gold Is There?

A few weeks ago, the AI scientist and critic Gary Marcus described a small, telling moment: Eric Schmidt, the former chief executive of Google, was booed while delivering a commencement address. Booed — for being too bullish on artificial intelligence. Three years ago, that would have been unthinkable. The chatbot was a miracle, and we were all believers.

Marcus has a phrase for where the money has gone. The chipmakers, he says, are selling shovels in a gold rush. And then he asks the question that should keep every operator awake at night: how much gold is there?

The numbers behind the question are staggering. The largest technology companies are pouring something approaching a trillion dollars into AI infrastructure this year, and Nvidia’s own chief executive [Jensen Huang] has suggested the figure is heading toward three or four trillion. Against that, a widely cited MIT study of more than 300 enterprise AI deployments found that 95% produced no measurable return. Enormous spending. Almost no proven gold.

It is tempting to file this under “Silicon Valley’s problem.” It is not. Because how much gold is there is not only a question for hyperscalers and venture capitalists. It is a question every hotel will answer — quietly, expensively, and one guest at a time.

From a metric to a discipline

Two weeks ago, I introduced a measure I called TCPG — Token Cost Per Guest — and argued that it belongs in our accounts alongside the costs we already track, and in time within USALI [Uniform System Accounting Lodging Industry] itself. That paper answered a simple question: what is it? Total AI token cost, divided by guests served. The new variable cost of intelligence, sitting beside the cost of distribution, we learned to respect only after OTA commissions had quietly crept up on us.

This piece answers a harder question — the one Marcus has forced into the open. Not what TCPG is, but what it's for when the money gets nervous?

We are keeping score wrong

Every AI conversation in our industry is a numerator. Look what it can do. Look how many hoteliers are piloting it. Capability, adoption, demonstration — all numerators, all the time. Almost no one states the denominator.

That is strange, because we are among the most disciplined per-unit thinkers in business. We measure RevPAR, ADR, and cost per occupied room. We measure the kilowatt-hour, the labor minute, and, in some hotels, the bar of soap. We have metered everything in the building — except the newest and fastest-growing variable cost.

The token. It hides inside your PMS invoice and your GRMS subscription, marked up by a vendor who would rather you never broke it out. 

TCPG’s first job is the humblest and the most important: to make that invisible cost visible. And visible properly — fully loaded. Not the sticker price on a vendor’s slide, but the real figure once you add orchestration, retries, storage, and the human who checks the output. That number is routinely several times the one you were quoted.

The operating system arrives — and the meter goes missing

Watch the language change this year. Software vendors have stopped selling applications. At a recent major industry showcase in Amsterdam, a leading property-management company took the stage and called itself the operating system for hospitality — everything under one contract, one data layer, one throat to choke. For some properties, the consolidation is real and worth having. But look underneath the branding, because that is where TCPG earns its keep.

The quieter shift is in how these platforms now make their money. A growing share of platform revenue in our sector comes not from software subscriptions but from the payment rail itself — interchange earned on every transaction that flows through the system. It is the same move that turned a leading restaurant-technology company from a software business into a payments business with software attached. The category changed under the noses of people still calling it the PMS market.

Here is why that should trouble any operator trying to understand the cost of their intelligence. When a platform earns its margin on payment volume, it has no reason to meter — or even to know — what its AI costs to run per guest. The token cost no longer merely hides in the invoice. It disappears into a business model pointed somewhere else entirely. The meter you were asking for was never installed because the vendor’s own accounts do not require it.

The capital flows say the same thing. The largest funding rounds in hospitality technology last year went to the platform layer, not to the AI features sitting atop it. The money is going into the plumbing; the trade press keeps photographing the taps. Marcus’s trillion-dollar question lands, unchanged, on your property — when the intelligence is bundled inside a payments platform, are you buying value, or quietly subsidizing someone else’s bet with your guests’ transactions? You cannot answer that without a meter of your own.

The metric that wins either way

Here is why TCPG matters more in a nervous market, not less.

Marcus believes the bubble deflates. He may be right; he may be wrong. To the operator who is metering, it scarcely matters.

If the boom runs on, TCPG gives you a hurdle rate. You spend against value rather than hype, and you are not quietly folded into someone else’s billion-dollar bet.

If it bursts, token prices collapse — and the operator who knows their TCPG harvests that collapse straight to margin. The operator who bought the bundle captures nothing because they never knew the number in the first place.

A metric that pays out whether the music keeps playing or stops is a rare thing to own. I cannot name another line on the P&L that does it.

Ninety-five and five

Return to that MIT figure. Ninety-five percent of deployments showed no measurable return. The lesson is not that AI does not work. The lesson is that almost no one measured. Five percent are extracting real value; ninety-five percent are unexamined write-offs. 

The single difference between the two groups is the meter. The five percent know their cost-to-value. TCPG is how a hotel crosses from one camp into the other.

One caution is that it is the easiest way to misuse the idea. A cost with nothing standing beside it is merely cost-cutting — and in our industry, cutting toward the cheapest possible model is a slow way to destroy the product, because the experience is the product. So TCPG is never a single number. It is a ratio: cost on one side, the guest’s experience on the other. Read alone, it misleads. Read as a ratio, it disciplines.

Look down

Marcus reaches for an old cartoon to describe the industry — the coyote who runs off the cliff and stays airborne, right up until he looks down. For everyone betting the trillion dollars, that is a warning: don’t look down.

For the hotelier with a meter, it is the opposite. Looking down is not falling. It is landing on your feet. The property that knows its Token Cost Per Guest does not fear the deflation Marcus predicts; it is the only one positioned to profit from it.

When the people pouring the concrete are still arguing about a trillion-dollar revenue hole, your edge was never going to be picking the cleverest model. Your edge is being the one who measured.

The option you are quietly buying

There is one more reason to keep that meter close, and it is arriving faster than the keynotes admit. The newest pitch is agentic AI — software that acts on its own. It prices the room. It answers the guest at midnight. It settles the folio with no human in the loop. Some of this is genuinely live; a guest can already complete a hotel booking within a conversational AI assistant today, so this is no longer speculation for 2027. But a great deal of what is being sold is still roadmap, presented as though it were already delivered.

That gap is the most expensive thing in the room, and almost no one is pricing it. The narrow agent that already runs — the one that nudges a rate overnight — is cheap; it spends a handful of tokens to reach a decision. The conversational concierge in the demonstration is a token-hungry thing by comparison, and it runs on every guest interaction, not once a night. These are not the same cost per guest. The brochure rarely separates them. TCPG forces the separation.

So let me forecast, because that is what a meter is for. Over the next twelve to eighteen months, I expect agentic AI in our sector to split cleanly in two. The cheap, narrow kind — pricing, reconciliation, routing — will embed quietly and pay for itself, until we stop noticing it the way we stopped noticing the spell-checker. The expensive, conversational kind will be rationed: moved behind premium tiers, capped, throttled, or left politely coming soon — not because it failed, but because its cost per guest never cleared the value it returned. The market will discover its TCPG the hard way, one surprised finance director at a time.

And watch the cross-subsidy, because it is the part nobody says out loud. Today, the real token bill is partly hidden by healthy payment margins. Those margins are themselves now under pressure, as new payment rails begin to erode the interchange that funds the bundle. When that subsidy thins — and it will — the true cost of the intelligence surfaces on the hotel’s books for the first time, in a year when margins can least absorb a surprise.

That’s the moment your TCPG metric stops being a nice idea and becomes the number that exposes the whole thing.

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

TCPG is simply how we keep the artificial part cheap and accountable — so the human part stays funded.

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

Footnotes

1. Abode Worldwide, Hospitality Tech Investment Index 2026. Across the 40 companies that raised more than US$1 billion between April 2025 and March 2026, property-management systems were the single largest category, with seven companies raising a combined US$408.1 million — more than any other segment. See abodeworldwide.com/hospitality-tech-investment-index and Hospitality Net, 13 April 2026. 

2. Card networks are already moving to defend interchange against lower-cost rails. A February 2026 Citrini Research note warned that cost-minimising AI agents could route transactions around the 2–3% card interchange toward stablecoins, where settlement costs are fractions of a cent; Mastercard’s acquisition of stablecoin-infrastructure firm BVNK and Stripe’s Machine Payments Protocol are among the early responses. See Fortune, 11 May 2026, and Blockonomi, 20 March 2026.