The Token Economy is Maturing

Image Source | Internet (Please contact us for removal if infringement occurs) Partially AI-generated

At this time last year, I was joking with a friend working on AI infrastructure: You're just water sellers in the AI era—every token rings, and gold flows. He just smiled and didn't respond.

When we chatted again this year, he showed me some numbers: The company's revenue had increased more than sixfold, but the gross margin had plummeted to negative 24%. The more they sold, the heavier their losses became.

This is probably the most absurd yet real aspect of the token business. In 2024, the average daily token consumption in China was 100 billion; by March of this year, it had surpassed 140 trillion, a thousandfold increase in two years.

Logically, with such an explosion in token usage, companies selling tokens should have ridden the wave and made a fortune.

However, companies that truly treat tokens as an independent commodity to sell are generally not as prosperous as imagined.

So, where did the problem lie? Based on current market dynamics, it's not a lack of demand—the issue lies with the very concept of 'selling tokens.'

Tokens Have Become a 'Sell More, Lose More' Business

Silicon Flow, one of the closest independent players to a 'pure token factory' narrative in China, saw its revenue surge from over 7 million to over 55 million in 2025, a 653% year-over-year increase.

However, its gross margin plummeted from positive 39.4% to negative 24%, with a full-year net loss of 345 million and an adjusted loss of nearly 200 million.

Ironically, its public cloud pay-as-you-go business, which most resembles a 'token factory,' had a gross margin of negative 119%; in contrast, its less glamorous privatized deployment projects achieved a gross margin of 82.5%.

DataCanvas generated 1.098 billion in revenue in 2025—not a small figure—but over 90% came from traditional intelligent computing infrastructure solutions, with true intelligent computing cloud services accounting for only 7.1%.

Runjian Co., Ltd.'s token factory secured a $200 million cross-border annual contract, a highly symbolic deal, but the company stated bluntly in its interim report: The token factory's gross margin is low, and depreciation during the fixed asset production period is excessively high.

When these fragments are pieced together, a counterintuitive picture emerges: The exponential growth in token usage is real, but 'charging per token' as an independent business model does not generate healthy profits for most players.

China International Capital Corporation (CICC) estimates that the theoretical gross margin for token factories could reach 60%, assuming self-built computing power and sales at hypothetical prices.

In reality, the gross margin for token factories procured collectively by domestic service model manufacturers is only 5% to 6%, while third-party service providers generally fall below 20%—a price war quickly erodes profitability.

Tokens are not without value, but standardized tokens are inherently difficult to sell as a 'commodity.'

They are measurable, replaceable, and have numerous suppliers, meaning pricing power has never been in the hands of sellers.

When upstream model manufacturers lower prices and downstream clients compare prices based on volume, heavy-asset players in the middle are squeezed between depreciation and pricing, becoming increasingly passive as they expand production.

Three Tiers of Players, Three Fates

Companies still 'selling tokens' today operate roughly on three different levels.

Profitability varies at each level, but the ceiling height and the location of load-bearing walls are determined by the same factor: The closer you are to 'tokens' as a unit of measurement, the farther you are from 'what clients truly want.'

The bottom tier is the gray arbitrage layer, essentially API resellers. The extreme profitability of this business once existed—for example, investors have seen reselling projects with monthly turnovers of 5 million and gross margins near 50%.

Profits come not from technological innovation but from information asymmetry and regulatory gaps: bulk purchase discounts, overseas account pools, splitting corporate free quotas, and even model substitutions that don't match advertised specifications.

A $200 monthly account could be resold for $2,000 to $3,000—sounding like easy money.

However, this window is closing at a visible pace. Claude is tightening third-party harness policies, while domestic models consistently rank high on overseas leaderboards via OpenRouter, gradually squeezing arbitrage opportunities at the supply end.

More critically, resellers' 'clients' are inherently unstable: C-end users follow trends, drifting away once hype fades—after OpenClaw's popularity peaked, consumption dropped back to previous levels.

Compliance costs further erode profits: Non-compliant resellers don't last long, while compliant ones see margins fall below 15%.

This tier won't disappear entirely, but it will never grow into a 'company'—just a series of transient deals.

The middle tier consists of heavy-asset token factories. Listed companies like SoftStone, Changshan Beiming, Runjian Co., Ltd., and DataCanvas all follow the same logic: Servers have been purchased, depreciation is ongoing, and rather than letting them sit idle by the card-hour, they repackage computing power into inference services sold by the token, at least boosting utilization.

This essentially uses 'selling tokens' as a new outlet for computing inventory, rather than tokens themselves creating a new profit pool.

Runjian's $200 million cross-border contract matters not because of high unit prices but because it locks in future consumption demand with minimal procurement commitments—in heavy-asset businesses, locking in demand is more critical than gross margin figures.

The top tier comprises companies that don't rely on token price differentials but on something beyond tokens.

OpenRouter is an extreme example: It doesn't build models, buy GPUs, or operate data centers, yet it achieves $140 million in annualized revenue with a 71% gross margin.

Because it doesn't sell tokens—it sells 'choice' and 'routing efficiency.' Fireworks maintains a 50%-level gross margin not through tokens themselves but by compressing latency to 1/3 to 1/12 of conventional solutions via its self-developed inference engine, generating more valid tokens per card.

Linghong Intelligence pivoted from data annotation to medical corpora and synthetic data, targeting the highest-compliance, least-standardized vertical token datasets, with cumulative orders exceeding 80 million—relying on hospital authorizations and compliance capabilities, not API pricing.

When these three tiers are viewed together, the underlying logic becomes clear: Tokens are a unit of measurement, not an anchor of value.

The closer a company is to the measurement unit and relies solely on it for fees, the more easily it becomes commoditized; the closer it is to clients' specific problems and embeds tokens within efficiency or compliance, the more pricing power it holds.

The Moat Lies Beyond 'Tokens'

If 'selling standard tokens' is destined to be a path toward zero gross margins, companies that want to stay in this game must build moats beyond the word 'token' itself.

Four clear directions emerge from players that are thriving—or at least have a rationale for surviving:

First, turn inference efficiency into a structural cost advantage rather than a one-time optimization. Everyone does inference optimization, and open-source frameworks keep improving, so single-point optimizations are quickly matched.

The real barrier lies in binding 'client-specific models + deep inference acceleration': The model belongs to the client, and the inference engine is deeply adapted to it. Switching suppliers means redoing end-to-end tuning, with migration costs themselves forming a moat.

Fireworks and Together AI's paths prove that independent MaaS providers don't profit from token markups but from 'producing more valid tokens with the same GPU.'

Second, drill down to the chip layer or heterogeneous scheduling layer to alter the cost structure itself. Cerebras' MaaS service achieved a 53% gross margin in Q1 2026 not through operational efficiency but by rewriting the financial model of inference with its self-developed chips.

Wuwen Xinqiong positions itself as a computing power intermediary between chips and models, doing heterogeneous scheduling—essentially occupying space in the 'underlying pipeline of token production.'

This path requires heavy investment and long cycles but shifts competition from 'who sells cheaper' to 'whose cost curve grows differently.'

Third, dive into vertical scenarios' compliance and data barriers. Medical tokens, legal tokens, high-quality datasets for embodied AI—these share a common trait: Real data is constrained by privacy and regulation, while synthetic data is still in validation. Whoever secures compliance authorizations and establishes closed loops for cleaning and stratified governance controls supply-side scarcity.

Linghong's early advantage in medical corpora and OpenRouter's network effects at the aggregation layer essentially mean 'pricing based on things beyond tokens.'

Fourth, use cross-border or overseas expansion to sell tokens into markets with higher pricing power. Domestic token services have gross margins below 20%, while overseas MaaS hovers around 50%—this price gap itself presents a structural opportunity.

Runjian's $200 million contract keeps computing power domestic while selling APIs overseas, following this logic.

However, going overseas isn't just about slapping an English interface on an API. Cross-border compliance, data residency, model admission, settlement channels—each layer is a barrier, and these barriers protect early entrants.

The Window for 'Pure Token Sellers' is Closing

From current market developments, the token economy is not a bubble. After all, daily invocation volumes surged from 100 billion to 140 trillion, with intelligent agents genuinely consuming tokens to execute tasks. As the unit of measurement and settlement for the AI industry, tokens will persist long-term.

China Telecom, China Unicom, and China Mobile all emphasize 'token operations,' with major players treating token services as a new gateway to cloud businesses—confirming tokens' Stable (wěngù, stable) position as infrastructure-level billing units.

But 'exclusively selling tokens' and 'tokens being widely used' are two different matters.

Major players can squeeze token profits to near-zero because they want clients to stay on their clouds, earning money from storage, databases, computing power, and other services.

Independent players that merely 'transport and markup standard tokens' face an almost endless downward price curve.

The gray resellers' extreme profitability came from regulatory gaps, which are narrowing; heavy-asset token factories that only sell computing inventory by the token will see depreciation eat away all profits.

Future survivors in this track (sài dào, arena) will likely stop calling themselves 'token sellers.'

They'll either be efficiency companies, driving unit token costs to unbeatable levels with inference engines and chips;

Or scenario companies, embedding tokens within medical, legal, or embodied AI deliverables where clients pay for results, not measurement units;

Or routing companies, like OpenRouter, earning money through aggregation and choice rather than price differentials;

Or cross-border companies, selling tokens produced from domestic computing and power advantages into higher-margin markets.

Ultimately, tokens are valuable, but you can't survive by selling 'value' itself—you survive by solving problems with your tokens.

To date, some dedicated token sellers are still enduring losses, while others are quietly removing 'tokens' from their business plans, replacing them with 'inference efficiency,' 'vertical data,' and 'cross-border services.'

After all, the era of profiting solely from tokens themselves is fading into the past.