
46% of the Traffic, Less Than 10% of the Revenue: How the Token Scissors Gap Is Tearing Apart the Global AI Market
English–Chinese Bilingual Industry Analysis|English First, 中文完整版见后半部分
English Version
American companies are now sending an extraordinary share of their AI workloads to Chinese models.
At peak periods, Chinese models reportedly processed as much as 46% of the Tokens used by American enterprises.
Yet Chinese model providers captured less than 10% of the corresponding revenue.
The contrast is difficult to ignore:
46% of the traffic, but less than 10% of the money.
This is not simply a story about Chinese models being cheaper.
Nor is it evidence that they are commercially unsuccessful.
It reveals a structural split in the global AI market: usage is moving toward the lowest-cost supply, while revenue remains concentrated among providers that can charge a premium for capability, brand, enterprise trust, and distribution.
The result is a widening Token scissors gap.
One side controls volume.
The other controls value.
Why is traffic moving toward Chinese models?
The answer begins with price.
DeepSeek V4 Flash reportedly charges RMB 0.02 per million input Tokens and RMB 2 per million output Tokens during standard-rate periods.
During peak hours, the output price rises to RMB 4 per million Tokens.
V4 Pro is priced at RMB 6 per million output Tokens during standard periods and RMB 12 during peak hours.
By comparison, OpenAI GPT-5.5 was priced at approximately USD 5 per million input Tokens and USD 30 per million output Tokens under the figures used in the original analysis.
After currency conversion, the difference is no longer a normal discount.
It becomes an entirely different economic category.
Depending on the comparison method, input costs may differ by dozens of times, while output costs may differ by more than one hundred times.
This is not merely “cheaper.”
For many workloads, the cost begins to feel almost negligible.
Justin Summerville, a data analyst at AI routing platform OpenRouter, has described Chinese open-source models as being roughly 60% to 90% cheaper than leading products from OpenAI and Anthropic.
Once the price gap reaches that scale, enterprises do not need to believe that the cheaper model is superior in every dimension.
They only need to conclude that it is good enough for a large percentage of their workloads.
That is exactly what is happening.
Enterprises are separating “must be best” from “good enough”
For years, companies treated AI model selection as a prestige decision.
They often chose the most capable model available and then attempted to control usage afterward.
That logic is becoming financially unsustainable.
A customer-support request does not always need the strongest reasoning model.
Neither does:
Document classification;
Information extraction;
Basic code completion;
Content rewriting;
Knowledge-base search;
Product tagging;
Routine translation;
Meeting-note processing;
Low-risk agent actions.
When a lower-cost model can complete these tasks reliably, sending every request to a premium frontier model becomes difficult to justify.
A 25-person AI startup in San Francisco, Lindy, reportedly shifted its traffic from Anthropic to DeepSeek V4 and reduced inference costs by approximately 90%.
Its founder said the decision could save millions of dollars.
The significance is larger than one startup’s cloud bill.
It demonstrates a new procurement principle:
Use the premium model only where the premium actually changes the outcome.
Everything else should be routed toward the lowest-cost model capable of completing the task.
This principle naturally moves traffic toward Chinese models.
The enterprise Token bill is becoming impossible to ignore
Agent-based workflows are accelerating this shift.
Traditional human–AI interactions may involve one prompt and one response.
An AI agent can:
Read documents;
Search databases;
Call tools;
Retry failed steps;
Generate intermediate plans;
Summarize results;
Verify its own work;
Ask other agents for help.
One user request can trigger dozens or hundreds of internal model calls.
OpenRouter data cited in the original discussion suggests that agent workflows can consume around 15 times more Tokens per request than ordinary chat interactions.
This changes the economics completely.
A model that appears affordable during a pilot can become extremely expensive when deployed across thousands of employees.
Uber’s experience illustrates the pressure.
After agentic coding tools spread across thousands of engineers, individual monthly API spending reportedly rose into the hundreds or even thousands of dollars.
An annual AI budget could disappear within months.
The problem is no longer whether AI creates value.
The problem is whether the organization can continue paying for that value at scale.
Lower Token prices do not necessarily produce lower AI bills
Enterprise procurement data also reflects this contradiction.
The average price per million Tokens across major providers may fall from around USD 10 to USD 2.50.
Yet total AI spending can continue to rise.
Why?
Because usage grows faster than unit prices fall.
A company that cuts the unit price by 75% but increases Token consumption tenfold still ends up with a much larger bill.
This is the new concern hanging over every CIO:
Tokens are cheaper, but the total invoice is becoming more frightening.
Agent adoption, longer context windows, multimodal workloads, tool calls, and persistent memory all increase consumption.
Cost sensitivity is therefore moving from the finance team into model architecture itself.
Companies are beginning to design systems around:
Model routing;
Prompt caching;
Batch processing;
Context compression;
Off-peak scheduling;
Dynamic escalation;
Task-level cost limits.
The cheapest Token is no longer a minor optimization.
It can determine whether the entire product is economically viable.
Why does the money not follow the traffic?
If Chinese models process such a large share of Tokens, why do they capture such a small share of revenue?
Because traffic and revenue are governed by different logic.
American frontier-model providers generally follow a capability-premium model:
We provide the strongest, safest, and most enterprise-ready system, so we charge the highest price.
Chinese model providers have largely followed a cost-efficiency model:
We can deliver competitive capability at a much lower cost, so we use price to expand adoption.
When these two models collide in the same market, the outcome is predictable.
Chinese providers exchange price for volume.
American providers exchange scarcity and perceived quality for margin.
One side wins Token share.
The other wins revenue per Token.
The platform numbers reveal the scissors gap
According to the OpenRouter data cited in the source material, Chinese models increased their Token-usage share on the platform from around 1% in early 2024 to more than 60% by May 2026.
Yet their revenue share remained far lower.
The original analysis places Chinese-model revenue share at approximately 17%, while higher-priced overseas models retained around 83%.
Under broader global estimates used in the article, Chinese models accounted for roughly 73% of model calls but only 18% of revenue.
These figures may vary depending on platform coverage and measurement method, but the commercial pattern is clear:
Chinese models generate far less revenue per Token.
The average commercial value of one Token can be only a fraction of that generated by a premium American model.
It resembles a company operating the world’s largest convenience store:
Millions of customers pass through every day, but the profit per transaction is tiny.
Next door, a boutique serves far fewer people but earns hundreds or thousands of times more from each purchase.
The convenience store controls foot traffic.
The boutique controls margin.
Low pricing is both a weapon and a constraint
The extremely low price of Chinese models is their strongest competitive weapon.
It is also what limits their revenue.
If one model charges one-hundredth of a competitor’s price, matching the competitor’s revenue requires roughly one hundred times the billable usage.
That is difficult even when Token share rises quickly.
The provider must also absorb:
Compute costs;
Network costs;
Research spending;
Model maintenance;
Customer support;
Capacity expansion;
Safety and compliance costs.
This creates a strategic dilemma.
If the supplier raises prices, developers may immediately move elsewhere.
If it keeps prices near the floor, traffic grows but margins remain thin.
If it relies on subsidies, the business becomes a test of capital endurance.
The provider wins adoption but may struggle to convert that adoption into sustainable cash flow.
That is why 46% of traffic can still produce less than 10% of the money.
Geography and trust deepen the revenue gap
Price is not the only reason revenue remains concentrated among American providers.
High-value enterprise contracts also depend on:
Data-residency commitments;
Security reviews;
Legal agreements;
Procurement relationships;
Regulatory comfort;
Brand recognition;
Local support;
Service-level guarantees;
Political and supply-chain risk.
An American enterprise may be willing to use a low-cost Chinese model for non-sensitive batch tasks while continuing to pay a premium American provider for:
Legal workflows;
Financial analysis;
Customer data;
Internal code;
Regulated workloads;
Executive decision support.
This creates a layered market.
Chinese models may win large volumes of lower-risk, price-sensitive traffic.
American providers retain high-margin workloads where trust and compliance matter more than raw Token cost.
The traffic gap and revenue gap are therefore reinforced by geography, regulation, and procurement behavior.
Agent workflows are widening the scissors
The rise of AI agents makes the imbalance more pronounced.
Agent systems do not simply consume more Tokens.
They also create more opportunities for model routing.
A single workflow may use:
A low-cost model for classification;
A stronger model for planning;
A coding model for implementation;
A fast model for tool calls;
A premium model for final verification.
As enterprises become more sophisticated, they stop treating model selection as an all-or-nothing decision.
They divide each workflow into stages.
This benefits low-cost models because they can absorb the high-volume middle of the workflow.
Premium models remain at the most valuable checkpoints.
In other words:
Chinese models may process most of the Tokens;
American models may still control the highest-value decisions.
That is the scissors gap in its most mature form.
Traffic flows toward efficiency.
Revenue concentrates around judgment.
Can the gap continue indefinitely?
Probably not in its current form.
Low-cost providers cannot expand infrastructure forever without improving monetization.
High-priced providers cannot assume that enterprises will continue using premium models for every task.
Both sides are moving toward the middle.
Chinese providers are beginning to introduce:
Enterprise service tiers;
Priority capacity;
Peak and off-peak pricing;
Private deployment;
Customization;
Reserved throughput;
Higher-value agent products.
American providers are introducing:
Smaller, cheaper models;
Batch discounts;
Caching;
Flexible processing;
Lower-priority inference;
Usage commitments.
The market is converging on a more complex structure.
Instead of one price per model, there will be multiple prices based on:
Urgency;
Quality;
latency;
Data sensitivity;
Capacity;
Task complexity;
Service guarantees.
The Token scissors gap may narrow, but it will not disappear automatically.
The central conflict remains:
Capability premium versus cost efficiency.
Unexhub: turning traffic into revenue more efficiently
This is where a platform such as Unexhub becomes strategically relevant.
Unexhub positions itself as a global Token aggregation, distribution, and compute-trading center.
According to its current platform narrative, it aims to connect major model ecosystems—including Gemini, ChatGPT, Claude, DeepSeek, and GLM—while also providing distribution opportunities for individual developers and small and medium-sized AI suppliers.
Its proposed model focuses on three capabilities:
Aggregated access;
Faster settlement;
API bidding and dynamic routing.
If executed effectively, these mechanisms could improve the conversion of Token traffic into supplier revenue.
Aggregation reduces the cost of reaching customers
A small model provider may have strong technology but weak distribution.
To sell APIs independently, it must build:
Developer documentation;
Billing;
Authentication;
Customer support;
Monitoring;
Fraud prevention;
Sales channels;
Contracts;
Collection systems.
These costs can consume the margin created by API usage.
A unified platform can reduce part of this burden.
Suppliers connect once and gain access to a broader pool of developers and enterprise buyers.
Developers gain access to multiple models through one interface.
This can improve the conversion chain:
Model capacity → platform exposure → developer traffic → billable usage.
For smaller suppliers, distribution efficiency may matter as much as model quality.
Faster settlement shortens the distance between usage and cash
Traffic does not help a supplier if revenue arrives months later.
Compute bills must be paid continuously.
GPU providers, cloud companies, engineers, and bandwidth suppliers do not wait for enterprise customers to complete a 90-day payment cycle.
Unexhub’s proposed usage-based settlement model aims to shorten the interval between API consumption and supplier payment.
The platform describes a structure in which suppliers can connect through a unified channel and receive revenue according to actual Token usage.
The practical benefits may include:
Faster cash recovery;
Lower accounts-receivable pressure;
Better utilization of idle capacity;
Less need to build an independent billing system;
Lower financial risk during rapid growth.
It would be more accurate to say that faster settlement can reduce financial pressure rather than eliminate it entirely.
But for smaller suppliers, even a shorter cash-conversion cycle can determine whether they survive long enough to benefit from rising traffic.
API bidding can improve the price of under-monetized traffic
The Token scissors gap exists partly because low-cost suppliers accept one fixed price for every request.
A bidding mechanism introduces more flexibility.
When supply is abundant and demand is low, developers can access lower prices.
When demand rises, qualified suppliers with available capacity may charge more.
Different workloads can also be valued differently.
For example:
A low-priority batch task may accept the lowest available bid;
A real-time customer interaction may pay more for low latency;
A regulated enterprise task may pay more for an approved processing route;
A complex reasoning task may pay more for a high-capability model.
This allows suppliers to stop treating every Token as identical.
Instead of selling all capacity at the lowest headline price, they can price according to service conditions.
That may help narrow the gap between traffic share and revenue share.
Millisecond routing makes the market useful to developers
Unexhub also emphasizes millisecond-level routing.
The importance of routing is often underestimated.
A developer does not simply need access to many models.
The developer needs the correct model selected for each request.
A routing layer can consider:
Price;
Latency;
Model capability;
Availability;
Context length;
Region;
Data sensitivity;
Failure history.
Routine requests can be sent to low-cost models.
Complex cases can be escalated.
Batch work can be delayed or directed to discounted capacity.
Sensitive requests can be restricted to approved endpoints.
This makes lower prices more useful because the developer does not need to manually redesign the application every time a model changes.
What “as low as 30% of market price” means in practice
According to Unexhub’s current positioning, selected eligible routes may be available at prices as low as approximately 30% of prevailing market or official rates.
The real value is not simply obtaining one discounted call.
For high-volume users, lower effective Token costs can mean:
More requests within the same budget;
More room for long-context workflows;
Lower-cost model testing;
Greater concurrency during traffic peaks;
More affordable multi-agent systems;
Reduced financial risk during product validation.
An enterprise does not need to send every task through the cheapest model.
It needs a portfolio of routes that keeps the total workflow cost under control.
The discount becomes most valuable when combined with intelligent routing.
The platform should not promise that every supplier loses all compliance burden
A marketplace can simplify access, settlement, and shared infrastructure.
It cannot automatically remove every supplier’s legal or operational responsibilities.
Whether a supplier still requires independent licenses, filings, security assessments, or contractual obligations depends on:
Who provides the service;
Who controls the model;
Who processes the data;
Who bills the customer;
Which jurisdiction applies;
How the commercial relationship is structured.
The strongest platform value is therefore not “compliance disappears.”
It is:
Repeated compliance, distribution, and settlement work can be standardized and shared.
That reduces the marginal cost of converting traffic into revenue.
Unexhub’s opportunity lies between the two sides of the scissors
On one side are developers and enterprises seeking:
Lower Token costs;
More model choices;
Easier integration;
Faster routing;
Less dependence on one provider.
On the other side are suppliers seeking:
More traffic;
Faster settlement;
Better utilization;
Broader distribution;
Higher effective revenue per unit of capacity.
Unexhub’s potential role is to connect these two sides.
If it can create sufficient market liquidity, transparent billing, stable capacity, and trustworthy settlement, it may help Chinese and smaller model suppliers turn high usage into more sustainable revenue.
It will not eliminate the structural difference between premium pricing and cost-efficiency pricing.
But it can reduce the friction that causes traffic to arrive without money following it.
The scissors gap is both victory and cost
Return to the original numbers:
46% of the traffic, less than 10% of the revenue.
This is not simply a failure of Chinese model providers.
It is also evidence of their success.
Their low prices have achieved something many competitors could not:
Massive usage;
Rapid developer adoption;
Global visibility;
Strong open-source ecosystems;
Integration into Western enterprise workflows.
But success at the usage layer comes with a cost.
When the price is only a fraction of the competitor’s, revenue cannot increase at the same speed as Token volume.
The provider acquires market share but sacrifices monetization.
That is the scissors gap:
One blade is falling unit price;
The other is rising usage;
Supplier margins are trapped in between.
The decisive question is no longer who processes the most Tokens
In the next stage of the market, Token share alone will not determine the winner.
The more important measures will be:
Revenue per successful task;
Gross margin per workload;
Settlement speed;
Customer-retention quality;
Enterprise-contract value;
Compliance readiness;
Routing efficiency;
Compute utilization.
A provider can process enormous volume and still lose money.
Another can process a small amount of high-value traffic and remain highly profitable.
The strongest businesses will combine both:
Low-cost models for volume;
Premium services for margin;
Routing for efficiency;
Settlement infrastructure for cash flow;
Compliance for enterprise access.
This is also the opportunity for Token marketplaces.
They can transform an industry obsessed with raw Token quantity into one focused on sustainable transaction value.
Final Conclusion
The global AI market is being split by two opposing pricing philosophies.
Chinese models are expanding through cost efficiency.
American frontier providers are protecting margins through capability premiums, enterprise trust, and high-value workloads.
This is why Chinese models can approach half of American enterprise Token usage while capturing less than one-tenth of the corresponding revenue.
Agent workflows will make the gap more visible, because they dramatically increase Token consumption while allowing premium models to remain concentrated at the most valuable decision points.
Platforms such as Unexhub are attempting to make the conversion between traffic and revenue more efficient.
According to its current positioning, Unexhub aims to provide:
Unified access to Gemini, ChatGPT, Claude, DeepSeek, GLM, and other model resources;
Distribution channels for independent and smaller suppliers;
Usage-based developer billing;
Eligible routes priced as low as approximately 30% of market or official rates;
Millisecond-level model routing;
Faster usage-based supplier settlement.
These benefits will matter only if the platform can also provide stable capacity, transparent measurement, auditable settlement, and credible governance.
The old AI market rewarded whoever built the strongest model.
The current market rewards whoever provides the cheapest usable model.
The next market will reward whoever can turn every Token of traffic into sustainable value.
The scissors gap is both a victory and a price.
Unexhub’s opportunity is not to make the gap disappear overnight.
It is to ensure that smaller suppliers do not burn out before their traffic finally becomes revenue.
中文版
46% 的流量,不足 10% 的收入:Token 剪刀差如何撕裂全球 AI 市场?
美国企业正在把越来越多的 AI 工作负载交给中国模型。
在部分统计周期内,中国模型处理的 Token 一度接近美国企业 AI 总用量的 46%。
然而,中国模型供应商从这些调用中获得的收入,却不足对应市场的 10%。
一边是接近半数的流量,另一边是不到一成的钱。
46% 的流量,不足 10% 的收入。
这不是简单的“中国模型价格更低”。
也不能说明中国模型在全球市场上失败了。
它揭示的是全球 AI 市场正在出现的一场结构性分裂:
调用量不断流向成本更低的供给,收入却依然集中在拥有能力溢价、企业信任、品牌影响力和成熟分销渠道的供应商手中。
一边控制流量。
另一边控制价值。
这就是不断扩大的 Token 剪刀差。
为什么流量正在涌向中国模型?
答案首先藏在价格里。
按照原文采用的计价口径,DeepSeek V4 Flash 每百万输入 Token 价格约为 0.02 元,每百万输出 Token 在平峰时段为 2 元,高峰时段为 4 元。
V4 Pro 的平峰输出价格为每百万 Token 6 元,高峰时段为 12 元。
同期,OpenAI GPT-5.5 的价格约为每百万输入 Token 5 美元、每百万输出 Token 30 美元。
经过汇率换算,这已经不是普通意义上的优惠。
而是两个完全不同的价格体系。
根据具体比较方式,输入成本可能相差数十倍,输出成本甚至可能相差上百倍。
这不只是“更便宜”。
对部分业务而言,调用成本已经低到近乎可以忽略。
AI 路由平台 OpenRouter 的数据分析师 Justin Summerville 曾表示,中国开源模型相较 OpenAI 和 Anthropic 的顶级产品,价格通常低 60% 至 90%。
当价格差达到这一程度,企业并不需要认定便宜模型在所有维度都更强。
只需要确认:
它能否完成足够多的日常任务?
事实证明,很多时候答案是肯定的。
企业正在把“必须最好”和“够用就行”分开
过去几年,企业选择 AI 模型时,往往优先采购能力最强的产品。
先上最好的模型,再想办法控制用量。
随着调用规模扩大,这套逻辑正在失效。
一次普通的客服请求,并不一定需要最强的推理模型。
以下任务通常也不需要每次都调用昂贵的前沿模型:
文档分类;
信息抽取;
基础代码补全;
内容改写;
知识库检索;
商品标签生成;
日常翻译;
会议纪要处理;
低风险智能体操作。
如果低价模型能够稳定完成这些任务,企业就很难继续解释为什么每一条请求都要发送给高价模型。
旧金山一家拥有 25 名员工的 AI 初创公司 Lindy,据称将流量从 Anthropic 大规模切换到 DeepSeek V4 后,推理成本下降了约 90%。
创始人表示,这一调整有望节省数百万美元。
这件事的意义不只是某家创业公司少付了一笔云账单。
它代表一种新的采购原则正在形成:
只有真正需要顶级能力的任务,才使用顶级模型。
其余请求,则交给能够可靠完成任务的最低成本模型。
这个原则天然会把大量流量推向中国模型。
企业已经无法忽视 Token 总账单
智能体工作流正在进一步加速流量迁移。
传统的人机对话,通常只有一次输入和一次输出。
AI 智能体却可能连续完成:
阅读文件;
搜索数据库;
调用工具;
重试失败步骤;
生成中间计划;
汇总结果;
自我校验;
请求其他智能体协助。
用户只发出一次请求,后台可能产生数十次乃至数百次模型调用。
原文引用的 OpenRouter 数据显示,AI 智能体工作流每次请求消耗的 Token,可能达到普通聊天交互的约 15 倍。
这会彻底改变企业的成本结构。
一个在试点阶段显得并不昂贵的模型,一旦部署到数千名员工和大量业务流程中,费用会迅速失控。
Uber 的案例说明了这种压力。
当代理编码工具逐步覆盖数千名工程师后,部分员工的人均月度 API 支出据称上升至数百乃至数千美元。
原本按全年规划的 AI 预算,可能在几个月内被耗尽。
问题已经不再是 AI 能不能创造价值。
而是企业能不能长期承担这种价值的成本。
Token 单价下降,AI 总账单却可能继续上升
企业采购数据也显示出同样的矛盾。
主流供应商每百万 Token 的平均价格,可能从约 10 美元下降至 2.5 美元。
单价明显下降,但企业 AI 总支出仍在增长。
原因并不复杂:
用量增长速度,远远快于单价下降速度。
如果单价降低 75%,而 Token 消耗增长十倍,最终账单依然会大幅上升。
这成为悬在每位 CIO 头顶的问题:
Token 越来越便宜,为什么总账却越来越吓人?
智能体普及、长上下文、多模态任务、工具调用和长期记忆,都在推动 Token 消耗增长。
因此,成本管理已经不再只是财务团队的工作,而是模型架构的一部分。
企业开始围绕以下能力重新设计系统:
多模型路由;
提示词缓存;
批量处理;
上下文压缩;
峰谷调度;
动态升级模型;
单任务成本上限。
最低成本的 Token,不再只是一个小型优化项。
它可能直接决定整个产品的商业模式能否成立。
为什么钱没有跟着流量一起过来?
如果中国模型已经处理了如此大量的 Token,为什么收入占比依然很低?
因为流量与收入,遵循的是两套不同的逻辑。
美国前沿模型供应商普遍采用 能力溢价逻辑:
我的模型更强、更稳定、更适合企业,所以我要收取更高价格。
中国模型供应商更倾向于采用 成本效率逻辑:
我的成本更低,也能完成大量任务,所以我通过低价扩大使用规模。
当这两种模式在同一市场碰撞,结果十分清晰。
中国模型用价格换取调用量。
美国模型用稀缺能力和信任换取高利润。
一边赢得 Token 份额。
另一边赢得每个 Token 的商业价值。
平台数据揭开了剪刀差
按照原文引用的 OpenRouter 数据,中国模型在该平台上的 Token 使用占比,从 2024 年初的约 1%,增长到 2026 年 5 月的 60% 以上。
但其收入占比依然明显较低。
原文估算,中国模型在该平台上的收入份额约为 17%,高价海外模型则占据约 83%。
放到更广泛的统计口径中,中国模型可能占据全球约 73% 的调用量,却只获得约 18% 的收入。
不同平台的覆盖范围和统计方式可能存在差异,但商业趋势已经非常清楚:
中国模型每个 Token 产生的收入明显更低。
一个中国模型 Token 的平均商业价值,可能只有美国高价模型的一小部分。
这就像经营着全球最大的便利店。
每天接待数百万顾客,每笔订单却只赚一分钱。
隔壁的精品店一天只服务几十位客户,但每笔订单都能获得数百甚至数千元利润。
便利店控制客流。
精品店控制毛利。
低价既是武器,也是收入天花板
极低的调用价格,是中国模型最强的竞争武器。
但它同样限制了收入增长。
如果一家模型的价格只有竞争对手的百分之一,那么要在收入上追平对方,理论上需要获得接近一百倍的付费调用量。
即使 Token 份额快速上涨,这个目标也很难实现。
供应商还需要承担:
算力成本;
网络成本;
研发投入;
模型维护;
客户支持;
容量扩张;
安全和合规成本。
于是形成一个战略困境。
涨价,开发者可能立刻迁移。
不涨价,流量继续增长,利润却十分有限。
长期依靠补贴,则会变成资产负债表的耐力赛。
供应商赢得了使用量,却未必能把使用量转化为可持续现金流。
这就是为什么 46% 的流量,依然可能只带来不足 10% 的收入。
地缘、信任与合规进一步扩大收入差距
价格并不是收入继续集中在美国供应商手中的唯一原因。
高价值企业合同还依赖于:
数据驻留承诺;
安全审查;
法律协议;
采购关系;
监管认可;
品牌可信度;
本地服务;
SLA 保障;
地缘与供应链风险。
一家美国企业可能愿意使用低价中国模型处理非敏感批量任务,却继续向美国供应商支付高价,用于:
法律工作流;
金融分析;
客户敏感数据;
内部代码;
受监管业务;
管理层决策支持。
这会形成一个分层市场。
中国模型获得大量低风险、价格敏感型流量。
美国供应商则保留隐私、信任和合规要求更高的高利润业务。
因此,流量与收入的错配还会受到地域、监管和企业采购习惯的进一步强化。
Agent 工作流正在继续拉大剪刀差
AI 智能体的普及会让这种结构更加明显。
智能体不仅消耗更多 Token,也为多模型路由创造了更多机会。
同一条业务流程可能同时使用:
低价模型做分类;
强模型做规划;
编程模型做实现;
快速模型做工具调用;
高价模型做最终校验。
当企业模型架构越来越成熟,就不会再把模型选择看成非此即彼。
它们会把工作流拆分成多个阶段。
这对低价模型非常有利,因为它们可以承接流程中数量最大的中间任务。
高价模型则继续守住最关键、最有价值的判断节点。
换句话说:
中国模型可能处理大部分 Token;
美国模型依然控制最昂贵的决策。
这正是 Token 剪刀差最成熟的形态。
流量流向效率。
收入集中在判断。
这道剪刀差能一直存在吗?
它很难永远维持现在的形态。
低价供应商无法在不提高变现效率的情况下无限扩张算力。
高价供应商也无法继续要求企业为每一个简单任务支付前沿模型费用。
双方都在向中间靠拢。
中国模型供应商开始推出:
企业服务档位;
优先容量;
峰谷计费;
私有部署;
定制服务;
预留吞吐;
更高价值的智能体产品。
美国供应商则不断增加:
更便宜的小模型;
批处理折扣;
缓存优惠;
弹性处理;
低优先级推理;
长期调用承诺优惠。
市场正在走向更加复杂的分层定价。
未来不再是一个模型只有一个价格,而是根据以下因素形成多个价格:
紧急程度;
能力要求;
响应时延;
数据敏感度;
可用容量;
任务复杂度;
服务保障。
Token 剪刀差可能缩小,却不会自动消失。
核心矛盾依然存在:
能力溢价,还是成本效率?
Unexhub:让流量更高效地变成收入
正是在这一背景下,Unexhub 的平台定位显得更加重要。
Unexhub 将自身定位为全球 Token 聚合、分发与算力交易中心。
按照平台当前的发展叙事,其目标是连接 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型生态,同时为个人开发者和中小型 AI 供应商提供分发渠道。
平台规划的核心能力包括:
聚合接入;
更快结算;
API 竞价与动态路由。
如果这些机制能够有效运行,就有机会提高 Token 流量转化为供应商收入的效率。
聚合,降低供应商获得客户的成本
一家中小模型供应商可能拥有不错的技术,却没有足够强的分销能力。
如果独立销售 API,就需要自行搭建:
开发者文档;
计费系统;
身份认证;
客户支持;
服务监控;
风险控制;
销售渠道;
合同体系;
收款与对账。
这些成本会不断吞噬 API 调用带来的微薄利润。
聚合平台可以减少其中一部分重复投入。
供应商完成一次接入,就有机会触达更广泛的开发者和企业用户。
开发者则通过一个接口获得多个模型。
这会缩短整个商业转化链路:
模型算力 → 平台曝光 → 开发者流量 → 可结算调用。
对于中小供应商而言,分销效率有时与模型能力同样重要。
更快结算,缩短用量与现金之间的距离
如果收入几个月后才能到账,再多流量也未必能够帮助供应商生存。
算力成本却是实时发生的。
GPU、云资源、工程人员和带宽供应商,不会等企业客户完成 90 天账期后再收钱。
Unexhub 规划中的按使用量结算机制,希望缩短 API 被调用与供应商获得收益之间的时间。
按照平台定位,供应商可以通过统一渠道接入,并根据实际 Token 使用量获得收益分配。
其实际价值可能包括:
更快回收现金;
降低应收账款压力;
提升闲置算力利用率;
减少自建计费系统的投入;
降低高速增长阶段的资金风险。
更准确地说,快速结算可以明显缓解财务压力,而不是绝对把资金压力降到零。
但对中小供应商而言,即使只是缩短现金回收周期,也可能决定它们能否撑到流量真正产生价值的那一天。
API 竞价,提高低价流量的变现效率
Token 剪刀差形成的一部分原因,是低价供应商用同一个固定价格出售所有请求。
竞价机制可以让价格变得更加灵活。
当市场供给充足、需求较低时,开发者可以获得更低价格。
当需求增加时,仍有可用容量的优质供应商可以获得更高收益。
不同任务也可以对应不同价格。
例如:
低优先级批处理任务,接受市场最低报价;
实时客服请求,为低时延支付更高价格;
受监管企业任务,为合规处理路线支付溢价;
复杂推理任务,为更强模型支付更高费用。
这样,供应商就不必继续把所有 Token 当作完全相同的商品出售。
它们可以根据服务条件形成差异化价格。
这有机会缩小流量份额与收入份额之间的差距。
毫秒级路由,让低价资源真正可用
Unexhub 还强调毫秒级路由能力。
路由的价值经常被低估。
开发者真正需要的,并不只是平台里有很多模型。
而是每一条请求都能被分配给合适的模型。
路由系统可以综合考虑:
价格;
时延;
模型能力;
可用性;
上下文长度;
服务地区;
数据敏感度;
历史失败率。
常规任务可以交给低成本模型。
复杂请求可以自动升级。
批量任务可以延迟到折扣资源。
敏感请求可以限定在经过批准的节点。
这样,开发者不必在每次模型价格或能力变化时重新开发整个应用。
低价才真正具备可使用价值。
“最低三折”对开发者真正意味着什么?
按照 Unexhub 当前平台定位,在符合具体模型、路由和服务条件的情况下,部分资源价格最低可接近市场或官方价格的三折。
它的实际价值,不只是某一次调用少花一些钱。
对于高频用户而言,更低的有效 Token 成本意味着:
同样预算支持更多请求;
长上下文业务拥有更大空间;
多模型测试成本降低;
流量高峰期可以保留更多并发;
多智能体系统更容易长期运行;
产品验证阶段的财务风险更小。
企业不需要把所有任务都发送给绝对最便宜的模型。
它需要的是一组能够控制整体工作流成本的路由组合。
当价格权益与智能路由结合时,折扣的价值才会被真正放大。
平台不能承诺供应商从此没有合规责任
聚合平台能够简化接入、结算和共享基础设施。
但它不能自动消除所有供应商的法律与运营责任。
供应商是否仍需独立取得许可、完成备案、接受安全测评或承担合同责任,取决于:
谁向最终用户提供服务;
谁控制模型;
谁处理数据;
谁向客户收费;
适用哪个司法管辖区;
双方商业关系如何设计。
因此,平台最合理的价值不是“加入以后合规义务全部消失”。
而是:
将大量重复的合规、分销与结算工作标准化,并由多方共享。
这可以降低流量转化为收入的边际成本。
Unexhub 的机会,正位于剪刀差的两端之间
剪刀的一端,是开发者和企业用户。
他们需要:
更低的 Token 成本;
更多模型选择;
更简单的接口;
更快的路由;
更低的单一供应商依赖。
剪刀的另一端,是模型供应商。
他们需要:
更多流量;
更快回款;
更高算力利用率;
更广泛的分销;
每单位容量获得更多有效收入。
Unexhub 的潜在角色,就是连接这两端。
如果平台能够建立足够的市场流动性、透明计费、稳定容量与可信结算,它可能帮助中国模型和中小供应商,把高使用量转化为更加可持续的收入。
它无法彻底消除能力溢价与成本效率之间的结构性差异。
但它能够降低“流量来了,钱却没跟上”的转化摩擦。
剪刀差既是胜利,也是代价
回到最初的数字:
46% 的流量,不足 10% 的收入。
这并不只是中国模型供应商的失败。
相反,它也证明了中国模型的成功。
极低的价格帮助它们获得了竞争对手很难复制的东西:
巨大调用量;
快速开发者采用;
全球市场关注;
强大的开源生态;
进入欧美企业工作流的机会。
但流量层面的成功,也有代价。
当价格只有竞争对手的一小部分时,收入不可能和 Token 数量同步增长。
供应商获得了市场份额,却牺牲了变现效率。
这就是剪刀差:
一边是不断下降的单价;
一边是不断增长的用量;
供应商的利润空间被夹在中间。
决定胜负的,不再是谁处理了最多 Token
进入下一个阶段后,Token 份额本身不会决定最终赢家。
市场会更加重视:
每个成功任务的收入;
单个工作流的毛利;
资金结算速度;
客户留存质量;
企业合同价值;
合规能力;
路由效率;
算力利用率。
一家供应商可以处理巨量请求,却持续亏损。
另一家供应商可能只处理少量高价值任务,却拥有稳定利润。
真正强大的商业模式,需要把两者结合:
用低价模型承接流量;
用高级服务创造利润;
用智能路由提高效率;
用结算基础设施保障现金流;
用合规能力进入企业市场。
这同样是 Token 交易平台的机会。
它们可以把一个只关注 Token 数量的行业,逐渐转变为关注可持续交易价值的市场。
最后的结论
全球 AI 市场正在被两种相反的定价哲学撕开。
中国模型通过成本效率扩大市场。
美国前沿模型则依靠能力溢价、企业信任和高价值业务保护利润。
这就是为什么中国模型能够接近美国企业 Token 用量的一半,却只能获得不足一成的收入。
智能体工作流会让这种差异更加明显。
因为它们会大幅增加 Token 消耗,同时让高价模型继续占据最关键、最昂贵的决策节点。
Unexhub 这样的聚合平台,正在尝试提高流量与收入之间的转化效率。
按照平台当前定位,Unexhub 希望提供:
对 Gemini、ChatGPT、Claude、DeepSeek、GLM 等模型资源的统一接入;
面向个人及中小供应商的分发渠道;
按实际 Token 使用量付费;
符合条件的路线最低接近市场或官方价格三折;
毫秒级模型路由;
更快的按量收益结算。
这些权益最终能否产生长期价值,仍然取决于平台能否提供稳定容量、透明计量、可审计结算和可信治理。
过去的 AI 市场,奖励的是谁能造出最强模型。
现在的市场,奖励的是谁能提供最便宜的可用模型。
未来的市场,将奖励谁能把每一份 Token 流量转化成可持续价值。
剪刀差既是胜利,也是代价。
Unexhub 的机会,不是让这道剪刀差一夜消失。
而是让中小供应商在流量真正变成收入之前,不至于先把自己烧死。