Kimi K3 2.8T 开源登顶,模型供应商把旗舰搬进 AI 算力交易中心摊成本

Kimi K3 reached 1,679 points on Frontend Code Arena, surpassing Claude Fable 5 and taking the top position. Across SWE-bench Verified, LiveCodeBench, Tau2, and AIME, it delivered leading open-source performance in code generation, agent execution, and mathematical reasoning.

Kimi K3’s 2.8T Open-Source Leap Tops the Charts—Now Model Providers Are Moving Flagships Into AI Compute Marketplaces to Share the Bill

Bilingual Article|English + Chinese
The complete English version appears first.
中文完整版位于文章后半部分,请继续向下阅读。

On July 16, 2026, just ahead of the World Artificial Intelligence Conference, Moonshot AI officially unveiled Kimi K3.

The specifications were built to attract attention:

  • 2.8 trillion parameters;

  • The world’s first open-source model approaching the three-trillion-parameter scale;

  • A one-million-token context window;

  • Native visual understanding;

  • A KDA hybrid linear-attention mechanism;

  • Attention-residual technology designed to improve scaling efficiency.

Its benchmark performance was equally striking.

Kimi K3 reached 1,679 points on Frontend Code Arena, surpassing Claude Fable 5 and taking the top position. Across SWE-bench Verified, LiveCodeBench, Tau2, and AIME, it delivered leading open-source performance in code generation, agent execution, and mathematical reasoning.

The most dramatic demonstration, however, did not come from a leaderboard.

During a 48-hour autonomous run, Kimi K3 reportedly completed an end-to-end chip-design workflow using open-source EDA tools and the Nangate 45nm process library. It did not merely produce a small piece of code. It designed, optimized, and validated a chip through a complete engineering process.

Powerful benchmarks, autonomous engineering, and open-source weights—this should have been a straightforward victory narrative for a Chinese foundation model.

Then users opened the pricing page.

The tone changed immediately.


Kimi K3 Has Entered the Flagship Price Tier

Kimi K3’s international API pricing is:

Token typePrice per 1M tokens
Input$3
Output$15

Its domestic pricing is similarly premium:

Token typePrice per 1M tokens
Cached inputRMB 2
Uncached inputRMB 20
OutputRMB 100

Artificial Analysis described the model as relatively expensive, estimating that its input price was approximately $1.75 above the market average and its output price around $8.40 higher.

Its estimated single-task cost is approximately $0.94—almost identical to GPT-5.6 Sol’s estimated $1.04.

An open-source model has entered the same price band as closed-source flagships.

That is one of the most revealing signals in the 2026 model market.

Open-source weights may reduce licensing restrictions, encourage ecosystem development, and make independent deployment possible.

They do not make training, inference, networking, electricity, optimization, or service delivery free.

Kimi K3 may be open-source.

The infrastructure behind it is not.


The Cost of 2.8 Trillion Parameters Cannot Be Carried by One Company Forever

Moonshot AI is not a public utility.

Training and serving a 2.8-trillion-parameter model require enormous capital.

Kimi K3 uses a mixture-of-experts architecture, activating 16 out of 896 experts during inference. Compared with K2, its overall scaling efficiency has reportedly improved by approximately 2.5 times.

The KDA hybrid linear-attention mechanism also helps reduce the compute burden of long-context workloads.

These are significant engineering improvements.

But efficiency gains do not eliminate the underlying cost.

The provider must still pay for:

  • Large-scale training clusters;

  • High-performance networking;

  • GPU procurement and leasing;

  • Electricity and cooling;

  • Storage and bandwidth;

  • Inference optimization;

  • Reliability and redundancy;

  • Engineering and operations;

  • Continuous model development.

Kimi K3’s $15-per-million-token output price reflects that reality.

The model may be more efficient than its predecessor, but operating a frontier-scale system remains expensive.

Moonshot AI has reportedly completed several funding rounds in 2026, with its valuation exceeding $31.5 billion.

Capital can fund the next training run and expand infrastructure.

It cannot permanently replace a sustainable business model.

Open-sourcing a flagship can accelerate adoption, build developer loyalty, and enlarge the ecosystem.

But open-source cannot mean serving every request at a loss.

This is not only Moonshot AI’s challenge.

It is the central economic problem facing almost every frontier-model supplier.


AI Compute Marketplaces: From Selling Models to Sharing Costs

AI compute trading centers are beginning to reshape the solution.

In May 2026, the Inner Mongolia Ciyun Trading Platform launched as a green-computing, full-stack AI marketplace.

In June, Token products from China’s three major telecom operators entered the China Compute Platform.

On July 15, the Southeast Token Operations Center officially began operations in Wenzhou.

These platforms are expanding beyond conventional GPU rental.

They increasingly provide:

  • Standardized Token services;

  • Access to open- and closed-source models;

  • Compute scheduling;

  • Model distribution;

  • Usage metering;

  • Unified billing;

  • Transaction settlement;

  • Multi-provider routing.

Many already support more than 100 mainstream models, including DeepSeek, Qwen, GLM, and a growing range of vertical systems.

The China Compute Platform has reportedly connected 15 provincial and municipal sub-platforms, onboarded more than 130 compute-service providers, listed over 170 computing products, and integrated more than 200 foundation and industry-specific models.

Suning’s Zhixin Token has gathered models such as:

  • Qwen;

  • DeepSeek;

  • Kimi;

  • GLM;

  • Doubao;

  • Seedance.

Flagship models are moving into shared marketplaces.

The reason is not simply wider exposure.

It is cost allocation.


Why Model Providers Need Shared Distribution

A model company traditionally has to build far more than the model itself.

It also needs:

  • A sales organization;

  • Developer relations;

  • API gateways;

  • Billing systems;

  • Usage metering;

  • Customer support;

  • Enterprise contracting;

  • Payment collection;

  • Capacity planning;

  • Failover infrastructure.

For a major cloud company, these systems are expensive but manageable.

For a startup or specialized model supplier, rebuilding the entire commercial stack can become overwhelming.

A compute marketplace offers an alternative.

The model provider connects once and gains access to:

  • An existing developer base;

  • Enterprise customers;

  • Standardized billing;

  • Shared routing infrastructure;

  • Usage-based settlement;

  • Additional compute suppliers;

  • Backup capacity.

The marketplace becomes the digital storefront.

The model becomes the product.

Tokens become the measurement and settlement unit.

This is similar to an e-commerce platform.

A brand does not need to build its own shopping mall before reaching customers. It can enter an established marketplace and share the platform’s traffic, payment, and distribution infrastructure.

For AI suppliers, this lowers the cost of going to market.


Better Utilization Is the First Form of Cost Sharing

The economics of frontier AI depend heavily on utilization.

Many intelligent-computing centers reportedly operate at only 20% to 30% average capacity. Some enterprise facilities may fall below 10%.

That creates a severe contradiction:

  • Advanced computing resources remain expensive;

  • High-demand periods still produce shortages;

  • Yet costly GPUs sit idle during quieter periods.

A model supplier operating its own isolated cluster may face both overcapacity and shortages at different times.

Token aggregation and intelligent routing can improve this balance.

When demand from multiple users, industries, and regions is combined, the platform can assign idle capacity to available workloads.

Higher utilization spreads fixed infrastructure costs across more completed tasks.

The logic is simple:

The more consistently the hardware runs, the lower the effective cost per Token can become.

This is the first meaning of “sharing costs.”

The provider is no longer relying only on its own customer traffic to keep the cluster busy.

The marketplace pools demand across the ecosystem.


Unexhub Pushes Token Cost Sharing Further

If regional compute platforms and the China Compute Platform are building foundational marketplaces, Unexhub is attempting to extend the logic into a global Token aggregation and distribution system.

Unexhub is building a computing-power trading center designed to match fragmented model and compute supply with demand from:

  • Enterprises;

  • Developers;

  • AI applications;

  • Agent platforms;

  • High-volume inference businesses.

Its ecosystem can include suppliers at different scales, from major model providers to startups, small companies, and independent developers.

Available channels may cover models such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Other domestic, international, and specialized AI models.

The objective is not merely to place several APIs behind one page.

It is to connect:

  • Supply;

  • Distribution;

  • Routing;

  • Pricing;

  • Usage metering;

  • Financial settlement.

That is where Token cost sharing becomes a market mechanism rather than a simple infrastructure arrangement.


Faster Settlement Protects Smaller Suppliers’ Cash Flow

High computing costs are not the only problem facing smaller suppliers.

The payment cycle can be equally dangerous.

A supplier may deliver inference capacity immediately, while waiting weeks or months to receive payment.

During that time, it must continue paying for:

  • GPUs;

  • Cloud resources;

  • Electricity;

  • Bandwidth;

  • Operations;

  • Engineering teams.

A business can show revenue growth and still fail because cash returns too slowly.

Unexhub’s settlement and clearing system is designed to shorten the delay between resource consumption and supplier payment:

When computing capacity is consumed, the corresponding revenue can be settled more quickly.

For suppliers, this can translate into:

  • Lower working-capital pressure;

  • More predictable cash flow;

  • Faster recovery of infrastructure investment;

  • More stable service capacity;

  • Greater ability to reinvest in models and hardware.

Eligible early suppliers may also receive promotional commission rates as low as 0%.

The practical value is broader than “free onboarding.”

Early suppliers may gain:

  • Lower distribution costs;

  • Faster access to enterprise demand;

  • More retained revenue during the growth stage;

  • Greater flexibility to test pricing;

  • A shorter infrastructure-payback period.

For a smaller supplier, these benefits may determine whether it can survive long enough to scale.


API Bidding Turns Token Prices Into a Market Signal

Unexhub’s planned API-bidding model takes the cost-sharing logic further.

Traditional model APIs rely on fixed list prices.

The supplier sets the rate, while users decide whether to buy.

But AI demand varies sharply.

During off-peak periods, significant capacity may remain unused.

During demand peaks, the same capacity becomes scarce.

A bidding system allows API prices to move with actual supply and demand.

Developers and enterprises may benefit from:

  • Lower prices during off-peak periods;

  • Flexible scheduling of non-urgent workloads;

  • Access to multiple suppliers;

  • Less dependence on one provider;

  • More control over large Token budgets.

Suppliers may benefit from:

  • Higher utilization of idle capacity;

  • Access to broader demand;

  • Better returns during peak periods;

  • Faster usage settlement;

  • Lower distribution costs.

This is not simply another discount program.

It allows the price of Token capacity to reflect current market conditions.

The Token becomes more than a billing unit.

It becomes a price signal for allocating AI resources.


Millisecond Routing Makes Dynamic Pricing Usable

Dynamic prices are valuable only when requests can be routed quickly.

Unexhub emphasizes millisecond-level routing according to factors such as:

  • Model capability;

  • Token price;

  • Response latency;

  • Regional availability;

  • Current capacity;

  • Task type;

  • Service quality.

A simple classification request should not automatically consume the same premium resources as a difficult coding or reasoning task.

A non-urgent batch job can be moved to lower-cost off-peak capacity.

A latency-sensitive agent can use a faster premium route.

A high-risk answer can be sent to a second model for independent review.

A practical routing framework might look like this:

WorkloadSuitable resource
Classification and extractionLightweight, low-cost model
Batch generationDiscounted off-peak capacity
General enterprise tasksBalanced model
Coding and tool orchestrationCoding-focused model
Complex reasoningPremium frontier model
High-risk outputMulti-model verification

This allows teams to reserve flagship resources for tasks that genuinely need them.

It also helps spread the cost of frontier models across a broader mix of workloads.


Pay Only for the Tokens Actually Consumed

Unexhub supports usage-based payment according to actual Token consumption.

Developers and smaller enterprises do not need to purchase large prepaid packages from every provider before testing a model.

They can:

  • Begin with a limited budget;

  • Compare several models;

  • Scale according to actual demand;

  • Avoid unused prepaid capacity;

  • Switch routes more easily;

  • Match spending to real business volume.

Under selected supply conditions and promotional plans, effective prices may be as low as approximately 30% of prevailing market rates.

For individual developers, this lowers the cost of experimentation.

For high-volume businesses, the benefit becomes much larger.

The same budget may support:

  • More user requests;

  • More agent workflows;

  • Higher concurrency;

  • Additional model experiments;

  • More retries;

  • Independent verification;

  • A longer operating runway.

The discount is valuable not merely because one call becomes cheaper.

It increases how much useful work can be completed within the same budget.


Kimi K3 and Compute Marketplaces Are Two Sides of the Same Shift

Kimi K3’s 2.8-trillion-parameter open-source release and the expansion of compute trading centers may appear to be separate stories.

They are part of the same economic transformation.

On one side:

  • Model capabilities are improving rapidly;

  • Open-source systems are reaching frontier scale;

  • Autonomous engineering tasks are becoming possible;

  • Context windows and multimodal abilities are expanding.

On the other:

  • Training costs are rising;

  • Inference remains expensive;

  • GPU utilization is uneven;

  • Distribution is fragmented;

  • Payment cycles create pressure;

  • Suppliers need sustainable unit economics.

Compute marketplaces are being built to amortize, distribute, and market-price those costs.

Model suppliers are moving flagship systems into shared marketplaces not because the marketplaces are fashionable.

They are doing so because carrying the full cost alone is increasingly difficult.

If costs cannot be distributed, open-source service cannot remain sustainable.

If infrastructure cannot maintain utilization, Token prices cannot fall.

If suppliers cannot receive payment quickly, smaller participants cannot survive.


Two Layers of the Emerging AI Economy

The new AI economy increasingly contains two connected layers.

The model layer

Companies such as Moonshot AI compete through:

  • Parameter scale;

  • Benchmark performance;

  • Multimodal capabilities;

  • Agent functions;

  • Open-source strategy;

  • Context length.

This layer answers:

What can the model do?

The market-infrastructure layer

Compute and Token platforms compete through:

  • Resource aggregation;

  • Intelligent routing;

  • Price discovery;

  • Usage settlement;

  • Supplier incentives;

  • Shared distribution;

  • Capacity utilization.

This layer answers:

How can the market afford to use the model?

Neither layer can succeed alone.

A frontier model without efficient distribution may struggle commercially.

A compute marketplace without strong models has little of value to distribute.

The industrial opportunity lies where model capability and market liquidity meet.


Challenges Still Remain

Compute trading centers cannot eliminate every problem.

The industry must still address:

  • Supplier verification;

  • Model-quality consistency;

  • Capacity authenticity;

  • Data security;

  • Cross-border compliance;

  • Pricing transparency;

  • Service-level guarantees;

  • Usage auditing;

  • Settlement risk;

  • Dispute resolution.

More suppliers can improve availability, but may also introduce inconsistent quality.

Dynamic pricing can lower costs, but may create volatility.

Faster settlement supports smaller suppliers, but requires strong clearing, auditing, and risk-control systems.

A mature Token marketplace cannot be only a list of discounted APIs.

It must also provide:

  • Transparent quality indicators;

  • Reliable routing;

  • Auditable usage records;

  • Stable financial settlement;

  • Security controls;

  • Supplier governance;

  • Clear service standards.

Only then can Tokens become a liquid and trusted production resource.


The Bottom Line

Kimi K3’s 2.8-trillion-parameter open-source release is a bold technical milestone.

Its next test, however, is commercial rather than academic.

The central question is not whether the model can top another benchmark.

It is whether the cost of training, inference, distribution, and service delivery can be shared across a sustainable ecosystem.

AI compute trading centers offer one possible answer.

They can aggregate demand, improve utilization, reduce duplicated commercial infrastructure, and spread fixed costs across more users.

Unexhub is attempting to push that model further through:

  • Global Token aggregation;

  • Faster supplier settlement;

  • API bidding;

  • Millisecond-level routing;

  • Usage-based payment;

  • Early-supplier incentives.

The logic is straightforward:

Model capability is becoming more concentrated and expensive.

Distribution must become more open and efficient.

Model suppliers are moving flagship systems into compute marketplaces not out of excitement, but out of necessity.

Without cost allocation, frontier open-source models cannot be served sustainably.

Without sustainable service, open-source leadership remains only a benchmark story.

Benchmarks are shown to the industry.

The infrastructure bill is paid by the provider.

The supplier that first completes a viable loop across computing supply, Token distribution, pricing, routing, and settlement will have the strongest chance of surviving the next round.


中文版

Kimi K3 以 2.8 万亿参数开源登顶,模型供应商正把旗舰模型搬进 AI 算力交易中心分摊成本

2026 年 7 月 16 日,世界人工智能大会开幕前夕,月之暗面正式发布 Kimi K3。

无论参数规模还是评测表现,这款模型都足够吸引行业目光:

  • 2.8 万亿参数;

  • 全球首个接近 3 万亿参数规模的开源模型;

  • 采用 KDA 混合线性注意力机制;

  • 引入注意力残差技术;

  • 原生支持视觉理解;

  • 拥有 100 万 Token 上下文窗口。

在 Frontend Code Arena 中,Kimi K3 以 1,679 分超过 Claude Fable 5,登上榜首。

在 SWE-bench Verified、LiveCodeBench、Tau2 与 AIME 等评测中,它也在代码生成、智能体任务执行和数学推理等方向取得了领先的开源模型表现。

比跑分更具冲击力的,是一次长达 48 小时的自主芯片设计实验。

Kimi K3 据称利用开源 EDA 工具与 Nangate 45nm 工艺库,独立完成了芯片的设计、优化与验证。

这并不是让模型“写一段代码跑一跑”。

而是把一项完整的端到端工程项目交给了模型。

性能强、跑分高,而且开源。

这个故事原本应该是一场标准的国产大模型胜利叙事。

然而,当用户打开 API 定价页面时,画风迅速发生变化。


Kimi K3 已经进入旗舰价格区间

Kimi K3 的国际 API 定价为:

Token 类型每百万 Token 价格
输入 Token3 美元
输出 Token15 美元

国内价格同样进入高端区间:

Token 类型每百万 Token 价格
缓存命中输入2 元
缓存未命中输入20 元
输出 Token100 元

Artificial Analysis 将其评价为相对昂贵,估算 Kimi K3 的输入价格较市场均值高出约 1.75 美元,输出价格则高出约 8.40 美元。

其单任务成本约为 0.94 美元,与 GPT-5.6 Sol 约 1.04 美元的水平已经非常接近。

一款开源模型,正式站进了闭源旗舰模型所在的价格区间。

这是 2026 年大模型市场中一个非常值得关注的信号。

开源权重可以降低授权限制,推动生态建设,也允许企业进行独立部署。

但开源并不意味着训练、推理、网络、电力、优化和服务交付全部免费。

Kimi K3 是开源的。

支撑它运行的基础设施不是。


2.8 万亿参数的成本,不可能永远由一家企业独自承担

月之暗面不是公益机构。

训练并持续服务一个 2.8 万亿参数模型,需要规模巨大的资本投入。

Kimi K3 采用 MoE 混合专家架构,在推理过程中从 896 个专家中激活 16 个。

相比 K2,其整体规模扩展效率据称提高了约 2.5 倍。

KDA 混合线性注意力机制,也有助于降低长上下文场景中的计算压力。

这些都是非常重要的工程进步。

但效率提高,并不会让底层成本彻底消失。

模型供应商仍然需要承担:

  • 大规模训练集群;

  • 高性能网络;

  • GPU 采购和租赁;

  • 电力与散热;

  • 存储与带宽;

  • 推理优化;

  • 冗余与稳定性保障;

  • 工程与运维团队;

  • 后续模型研发。

每百万 Token 输出 15 美元,正是这种现实压力的体现。

Kimi K3 或许比上一代模型更加高效,但前沿大模型的运营成本仍然十分昂贵。

月之暗面在 2026 年据称已完成多轮融资,估值超过 315 亿美元。

资本可以支持下一轮训练,也可以扩建基础设施。

但资本无法永久替代可持续的商业模式。

开放旗舰模型,可以加速开发者采用、扩大生态并提升品牌影响力。

但开源不能等于长期亏损服务。

这不只是月之暗面的问题。

也是几乎所有前沿模型供应商面对的核心经济难题。


AI 算力交易中心:从“卖模型”转向“分摊成本”

AI 算力交易中心,正在重新塑造这道难题的解法。

2026 年 5 月,内蒙古茨云交易平台正式上线,定位为绿色算力全栈 AI 交易平台。

6 月,三大通信运营商的 Token 产品接入中国算力平台。

7 月 15 日,东南 Token 运营中心在温州正式启动。

这些平台已经不再只是传统的 GPU 租赁市场。

它们逐渐提供:

  • 标准化 Token 服务;

  • 开源与闭源模型接入;

  • 算力调度;

  • 模型分发;

  • 调用计量;

  • 统一计费;

  • 交易结算;

  • 多供应商路由。

部分平台已支持超过 100 款主流模型,包括 DeepSeek、通义千问、GLM 以及越来越多的垂直模型。

中国算力平台据称已联通 15 个省市级子平台,入驻超过 130 家算力服务商,上架 170 余款算力产品,并接入 200 余款基础和行业模型。

苏宁“智新 Token”也已汇聚:

  • 通义千问;

  • DeepSeek;

  • Kimi;

  • GLM;

  • 豆包;

  • Seedance。

旗舰模型正在集体进入共享卖场。

背后的原因,不只是获得更多曝光。

更重要的是分摊成本。


为什么模型供应商需要共享分发体系?

传统模式下,一家模型公司需要建设的,远远不只是模型。

它还要建设:

  • 销售团队;

  • 开发者生态;

  • API 网关;

  • 计费系统;

  • 用量统计;

  • 客户支持;

  • 企业签约体系;

  • 回款与结算机制;

  • 容量规划;

  • 故障切换基础设施。

对大型云厂商而言,这些系统虽然昂贵,但尚可承受。

对初创企业和专业模型供应商而言,重新建设完整商业体系可能成为沉重负担。

算力交易中心提供了另一条路径。

模型供应商完成一次接入,便可能获得:

  • 现成的开发者流量;

  • 企业客户资源;

  • 标准化计费;

  • 共享路由体系;

  • 按量结算;

  • 额外算力供应;

  • 备用资源通道。

平台成为数字卖场。

模型成为商品。

Token 成为计量与结算单位。

这与电商平台的逻辑非常相似。

品牌方不需要先建设自己的商场,才能触达消费者。

它可以进入成熟平台,共享流量、支付和分发基础设施。

对 AI 模型供应商来说,这显著降低了商业化成本。


提高利用率,是分摊成本的第一步

前沿 AI 的经济性,很大程度上取决于算力利用率。

国内大量智算中心的平均利用率,据称仅在 20% 至 30% 左右,部分企业级设施甚至可能低于 10%。

这形成了一组严重矛盾:

  • 高端算力资源依然昂贵;

  • 高峰期仍然可能算力不足;

  • 低谷期大量 GPU 却处于闲置状态。

模型供应商如果只依赖自己的客户流量维持独立集群,就可能同时面临低谷过剩与高峰短缺。

Token 聚合与智能路由,可以改善这种失衡。

当来自不同行业、地区和用户的需求被汇聚在一起时,平台可以将闲置资源分配给正在等待的工作负载。

利用率越高,固定基础设施成本便能被更多任务共同分摊。

逻辑非常简单:

硬件持续运转的时间越长,单个 Token 的实际成本才越有下降空间。

这就是“成本分摊”的第一层含义。

模型供应商不再只依赖自己的客户,把集群长期跑满。

平台可以从整个生态中聚合需求。


Unexhub:将 Token 成本分摊逻辑再向前推进

如果区域算力平台与中国算力平台正在建设基础卖场,Unexhub 则试图将这套逻辑扩展为全球 Token 聚合与分发系统。

Unexhub 正在建设面向 Token 聚合和分发的算力交易中心,将分散的模型和算力供给,与以下需求进行匹配:

  • 企业;

  • 开发者;

  • AI 应用;

  • 智能体平台;

  • 高调用量推理业务。

其生态可以接入不同规模的供应商,从大型模型服务商,到初创公司、中小企业和独立开发者。

可用模型渠道包括:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 其他国内外通用及行业模型。

平台的目标并不是简单把多个 API 放在同一个页面。

而是连接:

  • 算力供给;

  • 模型分发;

  • 请求路由;

  • 动态定价;

  • 用量计量;

  • 金融结算。

这时,Token 成本分摊就不再只是基础设施层面的合作。

而开始成为市场机制。


快速结算,保护中小供应商现金流

高算力成本并不是中小供应商面对的唯一问题。

回款周期同样危险。

供应商可能立即提供推理算力,却要等待数周甚至数月才能收到款项。

与此同时,它们仍然需要持续支付:

  • GPU 成本;

  • 云端资源;

  • 电费;

  • 带宽;

  • 运维;

  • 工程团队开支。

一家企业即使账面上有收入,也可能因为现金回流太慢而无法继续经营。

Unexhub 的结算清算系统,试图缩短算力使用与供应商回款之间的周期:

算力被调用后,对应收入可以更快完成结算。

对供应商来说,这可能带来:

  • 缓解周转资金压力;

  • 提高现金流可预测性;

  • 更快回收基础设施投入;

  • 保持稳定服务能力;

  • 将更多收益继续投入模型和硬件。

符合条件的早期供应商,还可以享受最低 0 佣金的平台费率。

这项权益的价值并不只是“免费入驻”。

早期参与者还可能获得:

  • 更低的分发成本;

  • 更快触达企业需求;

  • 在增长阶段保留更多收入;

  • 更灵活地测试市场价格;

  • 缩短基础设施投资回报周期。

对中小模型供应商而言,这些条件可能直接决定其能否活到规模化阶段。


API 竞价,让 Token 价格成为市场信号

Unexhub 计划中的 API 竞价模式,将成本分摊逻辑进一步市场化。

传统模型 API 主要采用固定价目表。

供应商制定价格,用户决定是否接受。

但 AI 需求一直在波动。

低谷时,可能有大量算力未被使用。

高峰时,同样的资源又会变得稀缺。

竞价机制允许 API 价格根据真实供需变化。

对开发者和企业而言

可能获得:

  • 更低的闲时价格;

  • 更灵活安排非紧急任务;

  • 接入更多供应商;

  • 降低对单一厂商的依赖;

  • 更好控制大规模 Token 预算。

对供应商而言

可能获得:

  • 提升闲置算力利用率;

  • 触达更加广泛的需求;

  • 在高峰期获得更合理回报;

  • 更快完成调用收益结算;

  • 降低渠道分发成本。

这并不只是一个新的优惠方案。

而是让 Token 算力价格开始反映实时市场状况。

Token 不再只是计费单位。

它也开始成为分配 AI 资源的价格信号。


毫秒级路由,让动态价格真正可用

动态价格必须配合快速路由,才能产生实际价值。

Unexhub 强调根据以下因素进行毫秒级请求路由:

  • 模型能力;

  • Token 价格;

  • 响应延迟;

  • 地区可用性;

  • 当前算力容量;

  • 任务类型;

  • 服务质量。

简单分类任务,不应该默认消耗与复杂代码或推理任务相同的高端资源。

非紧急批量任务,可以被分配到更便宜的闲时通道。

对延迟敏感的智能体,可以使用更快的优先资源。

高风险结果,则可以交给第二款模型独立验证。

一个实用的路由结构可以是:

工作负载更适合的资源
分类与提取轻量低成本模型
批量生成闲时折扣算力
通用企业任务均衡型模型
编程与工具编排代码型模型
复杂推理高端前沿模型
高风险输出多模型复核

这样可以把旗舰模型留给真正需要高端能力的任务。

同时也能把前沿模型的成本,分摊到更加合理的工作负载结构中。


按实际 Token 消耗付费

Unexhub 支持按照真实 Token 使用量结算。

开发者与中小企业在测试模型前,不需要分别向每一家供应商购买大额预付套餐。

它们可以:

  • 从较小预算开始;

  • 同时对比多款模型;

  • 根据真实需求扩容;

  • 避免预付资源闲置;

  • 更方便地切换通道;

  • 让支出与业务量匹配。

在部分供应条件与推广方案中,实际价格最低可达到市场价格的约 3 折。

对个人开发者而言,这可以降低试错成本。

对高调用量业务而言,价值会进一步放大。

同一预算可能支撑:

  • 更多用户请求;

  • 更多智能体工作流;

  • 更高并发;

  • 更多模型实验;

  • 更多失败重试;

  • 更多独立验证;

  • 更长的业务运营周期。

这项优惠的价值,不只在于单次 API 调用更便宜。

而在于同样预算可以完成更多真正有价值的工作。


Kimi K3 与算力交易中心,是同一场变化的两面

Kimi K3 的 2.8 万亿参数开源发布,与算力交易中心的扩张,看上去是两条不同新闻。

实际上,它们属于同一场经济结构变化。

一边是:

  • 模型能力快速提升;

  • 开源模型进入前沿规模;

  • 自主工程任务逐渐成为现实;

  • 上下文与多模态能力不断扩张。

另一边是:

  • 训练成本持续上升;

  • 推理依然昂贵;

  • GPU 利用率不均衡;

  • 分发体系割裂;

  • 回款周期带来压力;

  • 供应商需要可持续的单位经济模型。

算力交易中心正在尝试分摊、分发,并通过市场化方式重新定价这些成本。

模型供应商把旗舰模型放进共享卖场,并不是因为算力交易中心足够热闹。

而是因为独自承担全部成本,已经变得越来越困难。

成本无法分摊,开源服务就难以持续。

算力无法保持利用率,Token 价格就缺少下降空间。

供应商无法快速回款,中小参与者便难以生存。


AI 新经济的两层结构

新的 AI 经济,正逐渐形成两个相互连接的层级。

模型层

月之暗面等模型公司围绕以下方向竞争:

  • 参数规模;

  • 跑分;

  • 多模态能力;

  • 智能体功能;

  • 开源策略;

  • 上下文长度。

这一层回答的是:

模型能够做什么?

市场基础设施层

算力与 Token 平台则围绕以下能力竞争:

  • 资源聚合;

  • 智能路由;

  • 价格发现;

  • 调用结算;

  • 供应商激励;

  • 共享分发;

  • 算力利用率。

这一层回答的是:

市场如何负担并持续使用模型?

任何一层都无法独立成功。

前沿模型如果缺乏高效分发,商业化会受到限制。

算力交易平台如果没有足够优秀的模型,也无物可分发。

真正的产业机会,出现在模型能力与市场流动性结合的地方。


挑战依然存在

算力交易中心无法自动解决所有问题。

行业仍然需要处理:

  • 供应商认证;

  • 模型质量一致性;

  • 算力真实性;

  • 数据安全;

  • 跨境合规;

  • 价格透明度;

  • 服务等级保障;

  • 调用审计;

  • 结算风险;

  • 争议处理。

更多供应商可以提高资源可用性,也可能带来质量差异。

动态价格可以降低成本,也可能导致波动。

快速结算能够帮助中小供应商,却需要更成熟的清算、审计和风控体系。

真正成熟的 Token 市场,不能只是一张低价 API 清单。

它还必须提供:

  • 透明的质量指标;

  • 可靠的请求路由;

  • 可审计的调用记录;

  • 稳定的金融结算;

  • 安全控制;

  • 供应商治理;

  • 清晰的服务标准。

只有做到这些,Token 才可能成为真正具有流动性和可信度的生产资源。


写在最后

Kimi K3 以 2.8 万亿参数开源,是一项大胆的技术里程碑。

但它接下来真正要面对的考验,并不是学术问题,而是商业问题。

核心问题已经不再只是模型能否继续刷新跑分。

而是训练、推理、分发与服务成本,能否通过一套可持续生态进行分摊。

AI 算力交易中心提供了一种可能的解法。

它们可以聚合需求、提高利用率、减少重复商业基础设施,并让固定成本由更多用户共同承担。

Unexhub 则试图在此基础上继续加入:

  • 全球 Token 聚合;

  • 更快的供应商结算;

  • API 竞价;

  • 毫秒级路由;

  • 按量付费;

  • 早期供应商激励。

逻辑十分清晰:

模型能力越来越集中,也越来越昂贵。

分发体系必须变得更加开放与高效。

模型供应商把旗舰模型搬进算力交易中心,不是因为追逐风口。

而是因为现实需要。

没有成本分摊,前沿开源模型就难以持续提供服务。

没有可持续服务,开源领先也只能停留在跑分故事中。

跑分是给行业看的。

基础设施账单,却要由供应商自己支付。

谁能率先跑通算力供给、Token 分发、动态定价、请求路由与金融结算的成本分摊闭环,谁才更有机会活到下一轮竞争。