The cheapest thing in digital marketing right now is a million tokens. The most expensive thing is being believed.

That inversion explains nearly everything confusing about the current moment. The cost of querying a model at GPT-3.5-level performance collapsed from $20 per million tokens in November 2022 to $0.07 by October 2024 — a decline of more than 280-fold in under two years, according to Stanford’s AI Index. Andreessen Horowitz calls it “LLMflation”: roughly a tenfold price decline every year for a model of fixed capability. Intelligence, as a raw input, is approaching the price of electricity.

And yet the companies building this cheap intelligence have never spent more. Frontier labs are burning billions on compute. Chinese model companies are reporting fast-growing revenues alongside even faster-growing cloud bills. Something doesn’t add up — unless you understand where value is actually migrating.

I recently spent an evening in Wuzhen, at a closed-door session during the World Internet Conference, with people who build large models, deploy them industrially, and fund them. One question held the table for three hours: as models get stronger and tokens get cheaper, what does an AI company actually sell? The answer that emerged has less to do with model architecture than with three quieter assets — verified real-world data, sustainable compute economics, and measurable outcomes. If you run a marketing organization, each of the three is about to land on your desk.

The scarcity nobody budgeted for

Start with a sentence from a friend who has spent years inside a major model company: the more AI-generated content floods the internet, the more valuable data from the real world becomes.

This is no longer a hunch. In July 2024, researchers published a landmark paper in Nature showing that models trained recursively on model-generated content suffer “model collapse” — the rare, unusual, tail-end details of the original data disappear first, then the distribution narrows toward bland averages. Think of photocopying a photocopy. The first copy looks fine. By the tenth, the fine print is gone. Subsequent work offers a partial antidote — mixing real and synthetic data, anchored in accumulating real data, can avoid the collapse — but that caveat makes the underlying point sharper, not softer: real-world data is the anchor everything else depends on.

Now translate this into marketing terms. The industry’s model builders are quietly acquiring companies that sit on old, authentic, human-generated data — early SaaS firms with a decade of customer interactions, industrial companies with years of equipment logs. They understand something most marketing organizations haven’t priced in: the corpus your brand has published, and the behavioral data your customers generate, is becoming a contested input to the systems that will mediate your next customer relationship.

Here’s the part that should reframe your content strategy. When a procurement manager in Munich asks an AI assistant to compare industrial automation vendors, the answer is not assembled from your ad spend. It is assembled from what the model has learned about you — documentation, reviews, case studies, forum discussions, the texture of how real customers describe real problems. Synthetic noise degrades models; verified, consistent, human-authored signal about your brand trains them. Your content operation is no longer just publishing. It is feeding — or starving — the recommendation layer between you and your buyers.

The compute paradox hiding in your AI budget

The second conversation in Wuzhen was about a contradiction. From the user’s seat, AI keeps getting cheaper. From the builder’s seat, it has never been more expensive. Both are true, and the reconciliation matters for anyone signing AI-related budgets.

Unit prices are down 280-fold; total compute consumption is exploding. The reason is behavioral. Users no longer ask a model an occasional question. They run agents that plan, call tools, retry, and loop — a single delegated task can consume hundreds of model calls. Meanwhile the model companies themselves keep funding ever-larger training runs. One participant described the emerging industry structure bluntly: model companies supply technology, cloud platforms supply compute and distribution, and every dollar of AI service revenue gets split somewhere along that chain. MiniMax’s disclosed figures — roughly $117 million in first-half revenue, growing fast, against rising training and cloud costs — show what the pressure looks like from the inside.

The marketing translation: your AI vendor’s unit economics are now your business risk. The generous free tiers, the subsidized agent subscriptions, the platforms practically giving away tokens to acquire users — these are land grabs, not price levels. When the subsidy cycle ends, the repricing will hit marketing stacks first, because marketing is where token consumption is growing fastest: AI-generated campaigns, synthetic focus groups, agent-driven customer service, personalized content at scale. A marketing operation built on the assumption that intelligence stays at $0.07 per million tokens is making the same mistake as a logistics company that built its network assuming 2020 shipping rates.

The counter-move is not to avoid AI spend. It is to demand outcome pricing. Which brings us to the third conversation.

Tokens measure usage. They don’t measure value.

The last hour in Wuzhen came back to the oldest question in business: why does the customer pay?

Tokens — the basic units models process — have done something genuinely new. For the first time, intelligence consumption is meterable, recordable, billable. That matters. But metering is not value, and confusing the two is expensive. Picture two companies deploying AI against customer complaints. One burns through tokens generating an exhaustive thirty-page analysis. The other makes a fraction of the calls, isolates the root cause, and resolves the issue. The first consumed more intelligence. The second created more value. A customer has never once been satisfied by how many tokens were spent on them; they care whether the problem got solved, what it cost, and whether they can trust the result.

Apply that lens to your own funnel. Marketing has spent two decades optimizing proxies — impressions, clicks, MQLs — that drifted away from outcomes. The token economy threatens to repeat the mistake with a new unit. An AI content engine that produces ten thousand articles is a token-usage story. An AI system that shortens your sales cycle by twelve days is a value story. Only one of them survives a CFO review in a down quarter.

This is also where the China signal becomes a strategic preview rather than a local curiosity. China’s AI market — hyper-competitive, price-war-driven, with dozens of credible models — is simply further along this curve. What happens there first: free-token user acquisition wars, model companies folding into cloud ecosystems, and a rapid buyer shift from “how capable is your model” to “what did it actually do for me.” That repricing of the AI value chain will reach Western marketing stacks within eighteen to thirty-six months. The executives who treat China as a lens rather than a market will see it coming.

What a CEO does with this

Five moves, in rough order of urgency.

Audit your brand’s training footprint. Pull your public content — documentation, case studies, community answers, earned media — and ask a harder question than “is it good?”: is it verifiable, specific, and consistent enough to survive as training signal? Vague brand language teaches models nothing. Concrete, checkable claims — what broke, how it was fixed, what the numbers were — teach them to trust you.

Treat first-party customer data as a strategic asset, not an IT resource. If model companies are acquiring firms for their real-world data, the data your customers generate — service interactions, usage patterns, problem reports — has balance-sheet implications. Govern it accordingly, and before you feed it to any vendor’s model, know what you’re trading.

Reprice your AI contracts around outcomes. When a vendor pitches tokens, seats, or usage tiers, redirect the negotiation: cost per resolved ticket, per qualified opportunity, per retained customer. The technology now permits outcome metering. Vendors who resist it are telling you something.

Stress-test your marketing stack for the end of subsidies. Model your AI-dependent workflows at 5x and 20x current token prices. Anything that only works while intelligence is artificially free is a liability wearing a discount.

Build the provenance habit now. Data quality is not a label applied in advance — as my friend in Wuzhen put it, you only learn whether a dataset is high-quality after it has been validated inside a real task. The same is true of AI-assisted marketing decisions. Instrument the loop: what did the model recommend, what did we do, what happened. The companies that own their feedback loops will compound; the ones renting them will not.

Where this goes

Short term — the next twelve to twenty-four months — expect a data acquisition wave that looks irrational until you see it through the model-collapse lens, the first serious outcome-based AI pricing in marketing tech, and a quiet repricing as free-tier land grabs end. Watch Chinese model companies’ financials: the ratio of revenue growth to compute spend will tell you when the subsidy era peaks.

Longer term, three to five years out, the structure hardens. Provenance-verified real-world data becomes a priced asset class. Token metering matures into a settlement layer between enterprises — intelligence bought, sold, and audited like any other metered input. And the brands that dominate AI-mediated discovery will not be the ones that spent the most on content, but the ones whose real-world signal was consistent enough to train the machines that buyers trust.

A sentence from that evening in Wuzhen has stayed with me. Usage is not the destination. Tokens make intelligence measurable — but how many real problems you solve is what decides what intelligence is worth. Marketing spent twenty years learning to buy attention. The next twenty will be about earning a place in the data that teaches machines what to recommend. That budget line doesn’t exist yet. Create it before your competitor does.