
Kimi K3 Enters the Flagship Price Tier as Model Providers Rush Toward AI Compute Marketplaces for Survival
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 before the opening of the World Artificial Intelligence Conference, Moonshot AI officially unveiled Kimi K3.
The specifications were designed to command 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 score of 1,679 on Frontend Code Arena, placing it ahead of Claude Fable 5;
Third place on GDPval-AA v2, behind only Fable 5 Max and GPT-5.6 Sol Max;
A top score of 91.2 on BrowseComp.
The most dramatic demonstration was not a benchmark.
During a 48-hour autonomous run, Kimi K3 reportedly completed the full design process for a four-square-millimeter chip without continuous human intervention.
A powerful model, strong benchmark performance, and open-source availability—this should have been a straightforward victory story for China’s AI industry.
Then users opened the pricing page.
The narrative changed immediately.
Kimi K3 Has Entered the Flagship Price Tier
Kimi K3’s international API pricing is:
| Token type | Price per 1M tokens |
|---|---|
| Input | $3 |
| Output | $15 |
| Cached input | $0.30 |
Artificial Analysis described the model as relatively expensive, noting that its input pricing was approximately $1.75 above the market average and its output pricing about $8.40 higher.
Its estimated cost per task is around $0.94—close to GPT-5.6 Sol’s approximately $1.04 and roughly half the cost of Opus 4.8.
Domestic pricing is not especially low either:
Cached input: RMB 2 per million tokens;
Uncached input: RMB 20 per million tokens;
Output: RMB 100 per million tokens.
An open-source model has entered the pricing territory traditionally occupied by closed flagship systems.
That is one of the most important signals in the 2026 model market.
Open-source weights do not mean free inference.
A model may be freely downloadable while still requiring enormous infrastructure expenditure to train, deploy, optimize, and serve at scale.
Kimi K3’s pricing is therefore not simply a reflection of commercial ambition.
It is also a window into the real cost of operating frontier-scale models.
The Compute Trap Facing Model Providers
Kimi K3 is not an isolated case.
Throughout 2026, major Chinese AI companies—including Alibaba Cloud, Tencent Cloud, Baidu AI Cloud, Zhipu AI, and DeepSeek—have repeatedly adjusted model API and computing prices.
The reason is straightforward: demand is exploding.
Daily AI-token usage in China reportedly grew from roughly 100 billion tokens in early 2024 to more than 140 trillion by March 2026—an increase of over one thousand times in approximately two years.
During one week in early July 2026, Chinese large-model usage reportedly reached 23.45 trillion tokens, maintaining the country’s position as one of the world’s largest AI-consumption markets.
Demand appears limitless.
Model suppliers, however, are not necessarily living comfortably.
The First Barrier: Compute Costs
Training a 2.8-trillion-parameter model requires an extraordinary amount of computing power.
The cost does not end when training finishes.
Inference requires continuous spending on:
GPU clusters;
Networking;
Electricity;
Cooling;
Storage;
Operations;
Model optimization;
Redundant capacity;
Engineering teams.
A large computing center may theoretically generate hundreds of billions of tokens per hour.
Yet many domestic intelligent-computing centers reportedly operate at average cabinet-utilization rates of only 20% to 30%. Some enterprise facilities may fall below 10%.
This creates a painful contradiction.
On one side, shortages of advanced chips keep rental prices high.
On the other, expensive GPUs remain idle in underutilized data centers.
The industry is simultaneously experiencing compute scarcity and compute waste.
The real problem is not only the amount of hardware available.
It is whether supply can be scheduled efficiently and converted into standardized, sellable AI output.
The Second Barrier: Slow Payment Cycles
Cash flow is another survival issue.
Model and computing suppliers provide capacity immediately, but payment may arrive weeks or months later.
During that period, they must continue paying for:
Cloud infrastructure;
Electricity;
Bandwidth;
Hardware financing;
Maintenance;
Staff.
For large technology companies, long settlement cycles are inconvenient.
For smaller model providers and independent developers, they can become fatal.
The question is no longer:
How much profit can the supplier make?
It is:
Can the supplier survive until the next funding round or payment cycle?
A provider may have strong models and growing usage but still fail because cash cannot return quickly enough to support operations.
The Third Barrier: Structural Mismatch Between Supply and Demand
AI demand is highly uneven.
During peak periods, providers may struggle to supply enough inference capacity.
During off-peak hours, expensive infrastructure sits idle.
Traditional compute-leasing models often rely on:
Monthly reservations;
Fixed GPU contracts;
Long billing cycles;
Static pricing;
Capacity commitments.
These structures were designed for conventional cloud workloads.
They are poorly suited to modern model inference, where demand may change by the minute and the true product is no longer merely a GPU hour.
It is a completed unit of intelligent work—measured increasingly in tokens, latency, throughput, and service quality.
This is why AI computing markets are shifting from selling hardware capacity to selling standardized Token output.
Compute Marketplaces: From Selling GPUs to Selling Tokens
The pressure facing model suppliers explains why more companies are turning toward AI compute marketplaces.
In May 2026, Inner Mongolia launched the Ciyun platform, described as China’s first green-computing, full-stack AI trading platform.
It combines:
Compute scheduling and trading;
Model access;
Transaction settlement;
Domestic-chip compatibility;
Support for mainstream computing architectures.
Its initial ecosystem reportedly included major telecom operators, leading technology companies, and more than ten mainstream models.
China’s national computing platform has also expanded quickly.
It reportedly connects 15 provincial and municipal sub-platforms, with:
More than 130 computing-service providers;
Over 170 listed computing products;
More than 200 foundation and vertical models;
Participation from Huawei, Alibaba Cloud, Baidu, Tencent Cloud, and major telecom operators;
More than 10,000 registered enterprise users.
Other initiatives are moving toward Token-based standardization.
The Nanjing “Token Store” uses tokens as a unified unit for measuring computing output, connecting:
Compute supply;
Model development;
Industrial applications;
Financial services.
SiliconFlow positions itself as an independent token-supply platform, converting heterogeneous hardware into standardized model output through its inference engine and resource-orchestration system.
The underlying shift is clear:
Compute is changing from a fixed asset into a liquid, tradable service.
Model providers no longer need to own and maintain every data center themselves.
They can connect to a marketplace, obtain resources on demand, and settle according to actual Token usage.
Compute begins to resemble electricity or network bandwidth: consumed when needed, measured consistently, and distributed through a shared infrastructure layer.
Unexhub Is Taking the Model Further
If national and regional computing platforms are building the physical and institutional foundation, Unexhub is attempting to reconstruct the global distribution logic for Tokens.
Unexhub is building a computing-power trading center focused on Token aggregation and distribution.
Its purpose is to match fragmented supply from model and compute providers with demand from:
Enterprises;
Developers;
Agent platforms;
AI applications;
High-volume inference services.
The platform’s role is not merely to list models.
It aims to connect supply, routing, pricing, distribution, and financial settlement into one system.
Faster Settlement Gives Smaller Suppliers a Survival Mechanism
For smaller suppliers, delayed payment is often more dangerous than low demand.
Capacity may be consumed today, while revenue remains trapped in a long settlement process.
Unexhub’s financial clearing and settlement system is designed to return usage revenue more quickly:
When capacity is supplied and consumed, the corresponding payment can be distributed with a shorter delay.
For model and computing suppliers, faster settlement may provide several concrete benefits:
Reduced working-capital pressure;
More predictable cash flow;
Faster recovery of GPU and infrastructure investment;
Greater ability to maintain stable capacity;
Lower financial risk during expansion.
Unexhub is also preparing an early-supplier incentive program.
Eligible early participants may receive promotional commission rates as low as 0%, allowing them to retain more of their initial revenue while building demand and testing the platform.
The real value of this benefit is not simply “free onboarding.”
It can help early suppliers:
Lower channel-acquisition costs;
Reach enterprise and developer demand more quickly;
Shorten the time required to recover infrastructure investment;
Test different pricing strategies;
Preserve more capital for model improvement and capacity expansion.
For a small provider, these advantages may determine whether it can survive long enough to scale.
API Bidding Returns Token Pricing to the Market
One of Unexhub’s more ambitious plans is an API-bidding mechanism.
Traditional AI APIs use fixed list prices.
The vendor determines the rate, while developers decide whether to accept it.
This model does not fully reflect changing supply conditions.
During off-peak periods, providers may have large amounts of idle capacity.
During peak periods, the same resource becomes scarce.
Unexhub’s planned market mechanism would allow API pricing to change according to real supply and demand.
For developers and enterprises, this could mean:
Lower prices during off-peak periods;
More flexible scheduling of non-urgent workloads;
Access to multiple suppliers;
Reduced dependence on one model vendor;
Better control of high-volume Token budgets.
For suppliers, it could provide:
Higher utilization of idle capacity;
Access to broader global demand;
Better returns during peak periods;
Faster settlement;
Lower distribution costs.
This is more than a traditional discount campaign.
It attempts to shift part of Token pricing from fixed platform decisions toward market-based allocation.
From Peak-and-Off-Peak Pricing to Real-Time Competition
DeepSeek V4 reportedly introduced peak-and-off-peak API pricing in mid-July, with prices doubling during high-demand periods.
This reflects a broader industry trend.
As inference demand becomes more volatile, fixed pricing becomes less efficient.
Peak-and-off-peak pricing is the first step.
A bidding market goes further.
Instead of one platform defining two or three static price periods, multiple suppliers can compete to serve real demand based on:
Current capacity;
Model quality;
Latency;
Region;
Reliability;
Task type;
Price.
The pricing authority shifts from:
The platform sets the price.
to:
The market discovers the price.
That could transform Tokens from a simple billing unit into a genuine tradable commodity.
Why Kimi K3’s Price Is a Mirror for the Whole Industry
Kimi K3’s entry into the flagship price tier reflects more than Moonshot AI’s ambitions.
A 2.8-trillion-parameter model carries unavoidable costs.
Even after compression, optimization, and efficiency improvements, the provider must recover spending on:
Training;
Inference;
Hardware;
Energy;
Engineering;
Distribution;
Reliability;
Continued model development.
Its $3 input and $15 output pricing may already reflect significant internal optimization.
The important conclusion is not that Kimi K3 is overpriced.
It is that frontier-model economics are becoming increasingly difficult to hide.
Open-source availability does not eliminate operational costs.
High benchmark scores do not automatically produce a sustainable business.
A model can win leaderboards and still struggle to build profitable distribution.
The Survival Logic Behind Compute Marketplaces
Model providers are clustering around compute marketplaces for a practical reason:
They need liquidity.
Liquidity has several meanings in this market:
Idle GPU capacity must find buyers;
Developers must be able to find affordable models;
Suppliers must receive payment quickly;
Requests must be routed to available resources;
Prices must respond to demand;
Capacity must be transferable across regions and providers.
Without this liquidity, the market remains fragmented.
One provider has unused GPUs.
Another has customer demand but insufficient capacity.
A developer needs a cheaper model.
A small supplier needs immediate cash flow.
A trading center creates value by connecting these isolated positions.
This is not an ideological movement.
It is an industry survival mechanism.
PoloAPI Provides a Practical Multi-Model Access Layer
Unexhub focuses on the broader market infrastructure of Token aggregation, distribution, bidding, routing, and settlement.
At the application-access layer, platforms such as PoloAPI can help developers and enterprises use multiple mainstream models through one integration.
PoloAPI supports access to models including:
Gemini;
Claude;
ChatGPT;
DeepSeek;
Other domestic and international models.
This can reduce the need to maintain separate:
API integrations;
Authentication methods;
Billing accounts;
Balance systems;
Rate-limit logic;
Failover routes.
For developers, the immediate advantages may include:
One interface for several models;
Faster model switching;
Centralized Token and billing visibility;
Reduced integration and maintenance costs;
Backup routes when a primary provider is congested;
Support for high-concurrency workloads.
For enterprises, PoloAPI also provides customization services covering:
Model selection;
API integration;
Permission configuration;
Concurrency planning;
Cost management;
Business deployment.
Its service experience includes domestic and international companies, publicly listed enterprises, state-owned organizations, and other large clients.
Under selected resource plans, actual model-access prices may be as low as approximately 30% of official rates.
For high-volume users, this benefit is not merely a lower price for one request.
The same budget may support:
More user traffic;
More agent workflows;
Higher concurrency;
More experiments;
Additional verification;
A longer operating runway.
Naturally, price should not be the only consideration.
Enterprises should also evaluate:
Service stability;
Data security;
Billing transparency;
Model availability;
Failover capability;
Service-level commitments;
Technical-support responsiveness.
A low rate becomes valuable only when it is supported by reliable and controllable infrastructure.
Two Layers of the New AI Economy
The emerging AI market can increasingly be understood as two connected layers.
The application and model layer
This is where companies such as Moonshot AI, DeepSeek, Alibaba, Tencent, and other model developers compete through:
Benchmark performance;
Context length;
Multimodal capabilities;
Agent features;
Pricing;
Open-source strategies.
The trading and infrastructure layer
This is where compute and Token platforms compete through:
Resource aggregation;
Dynamic routing;
Price discovery;
Financial settlement;
Supplier incentives;
Standardized access;
Capacity liquidity.
The first layer answers:
What can the model do?
The second answers:
How can the market afford to use it?
Neither layer can succeed alone.
A powerful model without efficient distribution may remain commercially limited.
A sophisticated marketplace without useful models has nothing valuable to trade.
The industrial opportunity lies at the intersection.
The Challenges Have Not Disappeared
Compute marketplaces do not automatically solve every problem.
The industry must still address:
Model-quality consistency;
Data-security requirements;
Supplier reliability;
Cross-border compliance;
Price transparency;
Service-level guarantees;
Capacity verification;
Fraud prevention;
Dispute resolution.
Dynamic pricing can reduce costs, but it can also create volatility.
More suppliers can improve availability, but inconsistent quality can increase operational risk.
Fast settlement supports smaller providers, but it also requires robust clearing, auditing, and risk-control systems.
A mature Token marketplace must therefore provide more than a list of cheap APIs.
It needs:
Standardized quality metrics;
Transparent pricing;
Reliable routing;
Auditable usage records;
Supplier verification;
Stable settlement;
Strong security controls.
Only then can Tokens become a genuinely liquid production resource.
The Bottom Line
Kimi K3’s flagship pricing is not simply a sign of Moonshot AI’s ambition.
It is a mirror reflecting the entire AI industry’s struggle between computing costs and commercial returns.
Demand is exploding.
Model capability is improving.
Open-source systems are reaching frontier scale.
But suppliers still face:
Expensive compute;
Low hardware utilization;
Long payment cycles;
Volatile demand;
Difficult unit economics.
AI compute marketplaces and Token-trading platforms offer a possible survival path.
They can help capacity move more freely, make pricing more responsive, shorten supplier settlement cycles, and give smaller participants access to demand that was previously difficult to reach.
Unexhub is attempting to rebuild this Token-distribution logic at the market-infrastructure level.
PoloAPI offers a more practical access layer for developers and enterprises that need to use multiple models with lower integration and operating costs.
Model providers are not joining compute marketplaces out of sentiment.
They are joining because standing alone may no longer be economically sustainable.
Kimi K3’s price page tells the story clearly:
Benchmark scores are shown to the industry.
Infrastructure bills are paid by the provider.
The companies that build a complete loop between computing supply, Token distribution, model demand, pricing, and settlement will be the ones most likely to survive the next stage of competition.
中文版
Kimi K3 定价站进旗舰位,模型供应商扎堆 AI 算力交易中心求生
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。
从参数规模到评测表现,这款模型几乎每一项数据都足够引人注目:
2.8 万亿参数;
全球首个接近 3 万亿参数规模的开源模型;
100 万 Token 上下文窗口;
原生视觉理解能力;
Frontend Code Arena 得分 1,679,超过 Claude Fable 5 登顶;
GDPval-AA v2 排名第三,仅次于 Fable 5 Max 与 GPT-5.6 Sol Max;
BrowseComp 得分 91.2,取得最高成绩。
比跑分更具冲击力的,是一项自主执行实验。
Kimi K3 据称在连续运行 48 小时的情况下,独立完成了一枚 4 平方毫米芯片的完整设计流程。
能力强、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,当用户点开 API 定价页面后,故事的画风迅速发生变化。
Kimi K3 已经站进旗舰价位带
Kimi K3 的国际 API 定价为:
| Token 类型 | 每百万 Token 价格 |
|---|---|
| 输入 Token | 3 美元 |
| 输出 Token | 15 美元 |
| 缓存输入 | 0.30 美元 |
Artificial Analysis 对这一价格的评价相当直接:相对偏贵。
其输入价格较市场均值高出约 1.75 美元,输出价格高出约 8.40 美元。
单任务成本约为 0.94 美元,与 GPT-5.6 Sol 的约 1.04 美元已经非常接近,同时约为 Opus 4.8 的一半。
国内价格同样不算低:
缓存命中输入:2 元/百万 Token;
缓存未命中输入:20 元/百万 Token;
输出:100 元/百万 Token。
一款开源模型,正式进入了过去主要由闭源旗舰模型占据的价格区间。
这是 2026 年大模型市场中一个值得认真观察的信号。
开源权重,并不意味着免费推理。
模型可以免费下载,但训练、部署、优化以及大规模服务所需的基础设施投入,依然极其昂贵。
因此,Kimi K3 的定价不仅反映了商业雄心。
它也让前沿模型的真实运营成本浮出水面。
模型供应商面临的算力困局
Kimi K3 并不是孤例。
2026 年以来,阿里云、腾讯云、百度智能云、智谱 AI、DeepSeek 等厂商,均对大模型 API 与算力价格进行了不同程度的调整。
背后的原因非常直接:需求正在爆发。
中国 AI Token 日均使用量,据称已从 2024 年初的约千亿规模,增长至 2026 年 3 月的超过 140 万亿,两年内增长超过千倍。
仅 2026 年 7 月上旬一周,中国大模型用量据称便达到 23.45 万亿 Token,继续保持全球领先水平。
需求看起来近乎无限。
但模型供应商的日子并没有因此变得轻松。
第一关:算力成本
训练一个 2.8 万亿参数模型,需要规模惊人的计算资源。
而模型完成训练后,成本并不会停止。
持续推理仍然需要支付:
GPU 集群采购或租赁;
高速网络;
电力;
散热;
存储;
运维;
推理优化;
冗余算力;
工程团队。
大型智算中心理论上每小时可以生成数千亿 Token。
然而,国内许多智算中心的机柜平均利用率,仅在 20% 至 30% 左右徘徊,部分企业级中心甚至可能低于 10%。
这形成了一种非常矛盾的局面。
一方面,高端芯片短缺让算力租金居高不下。
另一方面,大量昂贵 GPU 却因为调度和需求匹配不足,在机房中闲置。
行业同时存在算力短缺与算力浪费。
问题不只是有没有硬件。
而是能否把分散资源有效调度,并转化成标准化、可以持续销售的 AI 产出。
第二关:回款周期
现金流,是另一道生死关。
模型与算力供应商立即提供资源,但回款可能要等待数周甚至数月。
在这段时间里,它们仍然需要持续承担:
云端资源费用;
电力与带宽;
硬件融资;
机房运维;
工程团队成本。
对大型科技公司而言,长账期是一种经营压力。
对中小模型厂商与独立开发者而言,却可能直接决定生死。
问题已经不是:
这项业务能赚多少钱?
而是:
供应商能否活到下一轮融资,或者下一次结算到账?
即使模型能力很强、调用量持续增长,只要资金回流速度过慢,供应商仍可能被现金流拖垮。
第三关:供需错配
大模型调用需求具有明显波动。
高峰期,模型供应商可能没有足够算力处理请求。
低谷期,大量昂贵资源又处于闲置状态。
传统算力租赁通常采用:
月度预留;
固定 GPU 合同;
长周期计费;
固定价格;
长期容量承诺。
这套模式最初是为传统云计算业务设计的。
但大模型推理需求可能每分钟都在变化,真正出售的也不再只是“一个 GPU 小时”。
而是一份完成的智能产出,需要同时衡量:
Token;
延迟;
吞吐量;
可用性;
服务质量。
因此,算力市场正在从“卖硬件”走向“卖 Token”。
算力交易中心:从卖 GPU 到卖 Token
模型供应商面临的压力,解释了为什么越来越多企业开始涌向 AI 算力交易中心。
2026 年 5 月,内蒙古茨云交易平台上线运营,被定位为国内首个绿色算力全栈 AI 交易平台。
平台集成:
算力调度与交易;
模型便捷调用;
交易结算;
国产芯片兼容;
主流算力架构支持。
首批生态据称已包括三大通信运营商、头部科技企业以及十余款主流模型。
中国算力平台也在快速扩张。
其已连接 15 个省市级子平台,并汇聚:
超过 130 家算力服务商;
170 余款算力产品;
200 余款基础与垂直模型;
华为、阿里云、百度、腾讯云及三大电信运营商;
超过一万家注册企业用户。
其他平台则开始推动 Token 标准化。
南京“Token 集合店”以 Token 作为统一的算力计量单位,连接:
算力供给;
模型开发;
产业应用;
金融服务。
硅基流动则定位为开放、独立的词元供给平台,通过自研推理引擎与资源编排系统,将异构算力转化为标准化 Token 供给。
这些平台背后的本质变化是:
算力正在从固定资产,转变为可以流动、交易和分配的服务。
模型供应商不必再独立建设并维护所有机房和集群。
接入平台后,可以根据实际 Token 使用获得资源并进行结算。
算力开始像水、电和网络带宽一样,按需使用、统一计量,并通过共享基础设施进行分发。
Unexhub 比传统平台再往前走了一步
如果说内蒙古茨云平台与中国算力平台正在铺设算力基础设施,那么 Unexhub 所做的,是进一步重构全球 Token 的分发逻辑。
Unexhub 正在建设面向 Token 聚合与分发的算力交易中心。
平台希望将分散的模型和算力供给,与以下需求连接起来:
企业;
开发者;
Agent 平台;
AI 应用;
高调用量推理业务。
它的角色不只是把不同模型列在一个页面中。
而是试图将算力供给、请求路由、价格发现、渠道分发与金融结算连接成一套完整系统。
快速结算,为中小供应商提供生存机制
对中小供应商而言,回款过慢往往比需求不足更加危险。
今天已经提供并被使用的算力,收入却可能长期停留在结算流程中。
Unexhub 的金融结算清算系统,目标是让调用收益更快返回供应端:
算力被提供并产生调用后,相应款项可以用更短周期完成分配。
对模型和算力供应商而言,快速结算可能带来:
缓解周转资金压力;
提高现金流可预测性;
更快回收 GPU 和基础设施投入;
提升持续供给能力;
降低扩张过程中的财务风险。
Unexhub 还计划推出早期供应商激励体系。
符合条件的早期参与者,可在推广阶段享受最低 0 佣金的平台费率,在积累需求和测试市场时保留更多初期收入。
这项权益的价值不只是“免费接入”。
它可以帮助早期供应商:
降低渠道获客成本;
更快连接企业和开发者需求;
缩短基础设施投资回收周期;
测试不同定价策略;
保留更多资金投入模型优化与算力扩容。
对中小供应商而言,这些条件可能直接决定其能否活过早期阶段。
API 竞价,让 Token 定价回归市场
Unexhub 更具想象力的一项机制,是 API 竞价模式。
传统 AI API 通常使用固定价目表。
供应商决定价格,开发者只能选择是否接受。
这套方式难以真实反映算力供给的实时变化。
在业务低谷期,供应商可能拥有大量闲置资源。
到了高峰期,同样的算力又会变得稀缺。
Unexhub 计划中的竞价机制,可以让 API 价格根据真实供需动态变化。
对开发者和企业而言,这可能意味着:
在闲时获得更低调用价格;
更灵活地安排非紧急任务;
接入更多模型与算力供应商;
降低对单一厂商的依赖;
更有效地控制高 Token 工作负载预算。
对供应商而言,则可能获得:
更高的闲置资源利用率;
触达更广泛的全球需求;
在高峰期获得更合理的收益;
更快完成收入结算;
降低渠道分发成本。
这并不只是一次普通促销。
它试图将部分 Token 定价过程,从平台固定报价转向市场化资源配置。
从峰谷价格走向实时竞价
DeepSeek V4 据称在 7 月中旬首次引入峰谷 API 定价,高峰时段价格上涨至低谷期的两倍。
这反映出一个越来越明显的行业趋势。
随着推理需求波动加剧,固定价格体系正在失去效率。
峰谷定价是第一步。
竞价市场则更进一步。
它不再只由某个平台规定两个或三个固定时段,而是让多家供应商根据以下因素竞争:
当前可用算力;
模型质量;
延迟;
地区;
稳定性;
任务类型;
价格。
定价权将由:
平台规定价格。
逐渐转向:
市场发现价格。
Token 也因此可能从单纯的计费单位,进一步成为可交易的算力商品。
Kimi K3 的价格,是整个行业的一面镜子
Kimi K3 进入旗舰价位带,反映的不只是月之暗面的野心。
2.8 万亿参数模型的成本客观存在。
即使经过量化、压缩、推理优化和效率提升,供应商仍然要回收:
训练费用;
推理成本;
硬件投入;
电力支出;
工程成本;
渠道分发;
稳定性保障;
后续模型研发。
每百万 Token 输入 3 美元、输出 15 美元,或许已经是经过大量内部优化后的结果。
真正值得得出的结论,并不是 Kimi K3 是否“定价过高”。
而是前沿模型的真实经济账,已经越来越难以被掩盖。
模型开源,不代表运营免费。
跑分领先,也不会自动带来可持续商业模式。
一款模型可以赢下排行榜,却依然难以建立能够盈利的分发体系。
模型供应商扎堆算力交易中心,是一种生存逻辑
模型供应商聚集到算力交易中心,并不是因为情怀。
而是因为它们需要流动性。
在这个市场中,流动性包括多重含义:
闲置 GPU 能够找到买家;
开发者可以找到价格合适的模型;
供应商可以更快获得回款;
请求可以被路由到可用资源;
价格能够反映真实需求;
算力可以在地区与供应商之间流动。
没有流动性,市场就会继续割裂。
一方拥有闲置 GPU。
另一方拥有需求,却没有足够算力。
开发者需要更便宜的模型。
中小供应商急需现金流。
交易中心的价值,就在于将这些原本割裂的位置连接起来。
这并不是理想主义运动。
而是一套产业求生机制。
PoloAPI 提供更加直接的多模型接入层
Unexhub 关注的是 Token 聚合、分发、竞价、路由与结算等市场基础设施。
在实际应用接入层,PoloAPI 等平台则可以帮助开发者和企业,通过一套接口使用多款主流模型。
PoloAPI 可接入:
Gemini;
Claude;
ChatGPT;
DeepSeek;
其他国内外模型。
这能够减少企业分别维护以下系统的负担:
API 接口;
鉴权方式;
计费账户;
余额体系;
限流逻辑;
备用通道。
对开发者而言,实际权益可能包括:
一套接口调用多款模型;
更快速地切换模型;
集中查看 Token 与费用;
降低接入和维护成本;
主通道拥堵时使用备用资源;
支持高并发调用场景。
PoloAPI 还提供企业级定制服务,覆盖:
模型选型;
API 集成;
权限配置;
并发规划;
成本管理;
业务部署。
其相关服务经验覆盖海内外公司、上市企业、国有机构及其他大型客户。
在部分资源方案中,实际模型调用价格最低可达到官方价格的约 30%。
对高调用量用户而言,这并不只是降低一条请求的费用。
同样预算可能支持:
更多用户请求;
更多智能体工作流;
更高业务并发;
更多产品实验;
更多结果复核;
更长的业务运营周期。
当然,价格不能成为唯一判断标准。
企业还应综合考察:
服务稳定性;
数据安全;
账单透明度;
模型可用性;
故障切换能力;
服务等级协议;
技术支持响应效率。
只有低价建立在可靠、稳定和可控的基础之上,才真正具有价值。
AI 新经济的两层结构
正在形成的 AI 市场,可以被理解为两个彼此连接的层级。
模型与应用层
月之暗面、DeepSeek、阿里、腾讯及其他模型厂商,围绕以下能力竞争:
跑分;
上下文长度;
多模态能力;
智能体功能;
定价;
开源策略。
交易与基础设施层
算力和 Token 平台则围绕以下能力竞争:
资源聚合;
动态路由;
价格发现;
金融结算;
供应商激励;
标准化接入;
算力流动性。
第一层回答:
模型能够做什么?
第二层回答:
市场如何负担并持续使用这些模型?
任何一层都无法单独成功。
强大的模型如果缺乏高效分发,商业规模会受到限制。
再先进的交易平台,如果没有真正有价值的模型,也无物可交易。
真正的产业机会,存在于二者交汇处。
挑战依然存在
算力交易中心并不能自动解决所有问题。
行业仍然需要处理:
模型质量一致性;
数据安全要求;
供应商可靠性;
跨境合规;
价格透明度;
服务等级保障;
算力真实性验证;
欺诈预防;
交易争议处理。
动态定价可以降低成本,却也可能带来价格波动。
供应商数量增加可以提升可用性,但质量不一致也可能增加业务风险。
快速结算能够支持中小供应商,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 列表。
它还需要:
统一质量指标;
透明定价;
可靠路由;
可审计的调用记录;
供应商验证;
稳定结算;
严格安全控制。
只有这样,Token 才可能成为真正具有流动性的生产资源。
写在最后
Kimi K3 的旗舰定价,并不只是月之暗面雄心的体现。
它更像一面镜子,照出了整个 AI 行业正在经历的算力成本与商业回报拉锯战。
需求在爆发。
模型能力不断提高。
开源模型正进入前沿规模。
但供应商依然面临:
昂贵算力;
硬件利用率不足;
回款周期漫长;
需求剧烈波动;
单位经济模型难以平衡。
AI 算力交易中心与 Token 市场,可能为供应商提供一条求生之路。
它们可以帮助算力更顺畅地流动,让价格更加贴近真实需求,缩短供应商结算周期,并让中小参与者接触到过去难以获得的市场需求。
Unexhub 正尝试从市场基础设施层重写 Token 的分发逻辑。
PoloAPI 则为需要使用多款模型的企业和开发者,提供更加直接的接入、路由与成本优化能力。
模型供应商扎堆算力交易中心,无关情怀。
而是因为在新的经济结构下,独自作战可能已经难以持续。
Kimi K3 的定价页,已经把问题写得很清楚:
跑分是给行业看的。
基础设施账单,却要由供应商自己支付。
谁能够率先打通算力供给、Token 分发、模型需求、动态定价与金融结算的完整闭环,谁才更有可能活到下一轮竞争。