高盛:Token 价格战尾声,DeepSeek 峰谷定价让 Token交易市场回归真实成本

2026 年 6 月 29 日晚,DeepSeek 向 API 开发者发送通知:V4 正式版计划于 7 月中旬上线,同时引入国内大模型行业较为少见的峰谷分时计费机制。

价格战打不动了:DeepSeek 峰谷定价,Token 市场开始回归真实成本

2026 年 6 月 29 日晚,DeepSeek 向 API 开发者发送通知:V4 正式版计划于 7 月中旬上线,同时引入国内大模型行业较为少见的峰谷分时计费机制。

在工作日两个主要调用高峰——9:00 至 12:00、14:00 至 18:00——部分 API 调用价格将调整为闲时价格的两倍。这两个时间段几乎覆盖了国内开发者最集中的工作时段。

根据公布的价格:

模型闲时输出价格高峰输出价格
V4 Pro6 元/百万 Token12 元/百万 Token
V4 Flash2 元/百万 Token4 元/百万 Token

消息公布后,开发者社区迅速出现分歧。

有人认为,高峰时段覆盖了大部分正常工作时间,这种定价方式对国内开发者并不友好;也有人指出,即使高峰价格翻倍,DeepSeek 与部分同级模型相比,仍然具有一定的价格竞争力。

但真正值得关注的,并不只是“涨没涨价”。

更重要的问题是:为什么大模型厂商不再继续单纯降价,而是开始用峰谷价格调节算力需求?

从高盛相关研究观点来看,这可能意味着持续数月的大模型价格战正在触及边界。Token 的价格体系,正从“低价换规模”逐步回到算力、电力和基础设施所决定的真实成本。


高盛的判断:价格战接近拐点,Token 开始回归成本

据文中所引高盛 7 月 3 日研究观点,DeepSeek 推出峰谷定价,并不意味着市场需求下降。

恰恰相反,这种变化可能说明 AI 调用需求仍在增长,而高峰时段的 GPU、电力和数据中心资源正在趋于紧张。

高盛估算,调整后的 V4 Pro 平均价格约为 0.35 美元/百万 Token,V4 Flash 平均价格约为 0.12 美元/百万 Token。

相比具体价格,更值得注意的是其背后的行业判断:

自 2026 年 4 月底以来不断加剧的大模型价格战,可能正在接近阶段性终点。

过去一段时间,大模型厂商为了争夺开发者和 API 市场份额,持续压低 Token 价格。

DeepSeek 曾大幅下调部分模型价格,部分厂商的降幅甚至接近 99%。当价格被压到接近零毛利,甚至需要依靠亏损换取调用量时,这种竞争就不再只是正常促销,而是在透支未来的研发与基础设施投入。

大模型服务并不是没有成本。

每一次 API 调用背后,都对应着 GPU 折旧、电力消耗、带宽、存储、机房散热、工程运维和模型迭代。当 Token 售价长期低于真实推理成本,厂商可以短期获得流量,却很难建立可持续的商业模式。

因此,DeepSeek 的峰谷定价更像一次价格纠偏,而不只是简单涨价。

行业竞争的核心,也正在从“谁卖得更便宜”,逐步转向:

  • 谁的推理成本更低;

  • 谁的 GPU 利用率更高;

  • 谁的模型能够用更少 Token 完成任务;

  • 谁能在高峰期保持稳定服务;

  • 谁能建立长期可持续的商业闭环。

价格战并没有完全消失,但仅靠低价已经越来越难以建立真正的竞争壁垒。


为什么价格战打不下去了?

答案首先藏在能源成本里。

按照相关行业预测,2026 年全球数据中心用电量可能达到 565 TWh。其中,AI 优化服务器的耗电量预计将从 95 TWh 增长至 175 TWh,增幅约为 84%。

新一代 GPU 的功耗也在快速增加。

单块 Blackwell 架构 GPU 的功耗可达到约 1 kW,相比部分 Hopper 架构产品约 700 W 的功耗水平,增长接近 40%。

当数千甚至数万张 GPU 同时运行时,API 调用的高峰,本质上也是 GPU、电力、冷却系统和数据中心容量的高峰。

因此,算力的时段分布,最终会映射到电力和基础设施的时段分布上。

这也是 DeepSeek 峰谷定价与电网峰谷电价具有相似逻辑的原因:

  • 高峰时段提高价格,减少非必要调用;

  • 低谷时段降低价格,引导批量任务错峰运行;

  • 实时业务保留在白天;

  • 非实时任务转移至夜间;

  • 平台通过价格杠杆提高设备利用率。

从这个角度看,峰谷定价并不只是涨价,而是一种需求疏导机制。

算力没有突然变贵,真正变贵的是在最拥堵的时间使用算力。


从统一定价到分时收费:Token 第一次拥有“时间价格”

过去,大模型 API 通常采用统一定价。

无论用户在凌晨两点调用,还是在工作日下午三点调用,同一模型的 Token 价格通常没有区别。

这种定价方式简单,也方便企业做预算,但它无法反映高峰与低谷之间真实的资源差异。

DeepSeek 引入峰谷价格后,Token 第一次被明确加入了时间维度。

同一个模型、同一种任务、同样数量的 Token,仅仅因为调用时间不同,最终费用就可能相差一倍。

这意味着,Token 不再只有一张静态价目表,而开始拥有类似电力资源的“时间价格”。

DeepSeek 的策略看似前后矛盾,实际上有着清晰逻辑。

此前,V4 Pro 部分价格曾从 24 元/百万 Token 下调至 6 元,降幅达到 75%。降低闲时基础价格,是为了让更多开发者能够使用;提高峰值时段价格,则是为了引导用户合理安排调用时间。

两种动作共同服务于一个目标:

先降低使用门槛,再提高算力利用效率。


开发者真正需要升级的,是任务调度能力

当 Token 价格随时间变化,企业控制 API 成本的方式也会发生改变。

过去,开发团队主要通过以下方法节省费用:

  • 缩短 Prompt;

  • 减少重复上下文;

  • 开启 Prompt Caching;

  • 使用价格更低的模型;

  • 限制最大输出长度;

  • 减少不必要的重试。

未来,还要增加一个重要变量:任务何时执行。

企业可以根据业务时效性,将工作负载划分为不同等级。

必须实时执行的任务

例如:

  • 在线客服;

  • 实时翻译;

  • 风险识别;

  • 用户交互;

  • 实时代码补全;

  • 生产环境告警处理。

这些任务对延迟敏感,需要优先保证稳定性,即使处于高峰时段,也不能简单推迟。

可以延迟执行的任务

例如:

  • 批量内容生成;

  • 数据清洗;

  • 大规模文档摘要;

  • 离线模型评测;

  • 商品描述生成;

  • 日报、周报整理;

  • 非紧急 Agent 工作流。

这些任务可以进入队列,在夜间或价格更低的时段自动执行。

同样一批任务,仅仅通过改变执行时间,就可能明显降低成本。

因此,未来企业之间的成本差距,不只取决于谁拿到了更低的 API 报价,还取决于谁拥有更成熟的调度系统。

模型选择决定能力上限,调度策略决定成本下限。


Unexhub:从“何时使用”进一步走向“如何买卖”

DeepSeek 的峰谷定价,主要解决“什么时候调用更划算”的问题。

Unexhub 尝试解决的,则是更进一步的问题:

  • Token 应该从哪里采购?

  • 不同供应商如何参与供给?

  • API 价格如何形成?

  • 供应商如何更快回款?

  • 闲置算力如何进入流通市场?

Unexhub 希望通过构建 Token 聚合与分发的算力交易中心,将分散在不同模型厂商、GPU 服务商和中小供应商手中的算力资源统一接入,再面向企业与开发者进行分发。

其核心逻辑可以分为三个部分。


第一层:即时结算,释放中小供应商现金流

传统算力合作往往存在较长账期。

中小供应商已经完成模型推理或算力交付,却可能需要等待一个月甚至更长时间才能收到款项。

在此期间,供应商仍要持续承担:

  • GPU 和服务器成本;

  • 数据中心机柜费用;

  • 电力与散热费用;

  • 网络带宽费用;

  • 运维人员成本;

  • 设备扩容投入。

收入尚未到账,支出却每天发生,这会对中小供应商形成明显的资金压力。

Unexhub 推动金融级即时结算,希望让供应商在完成算力供给后,更快获得对应款项。

“供应多少、结算多少”的模式,能够缩短资金周转周期,让收入不再长期停留在账期中。

对于供应商而言,即时结算带来的价值不仅是“到账更快”,还包括:

  • 减少长期垫资压力;

  • 更快支付基础设施费用;

  • 提高资金周转效率;

  • 更灵活地扩充算力;

  • 降低因长账期产生的经营风险。

这并不是让所有成本真正归零,而是尽可能降低由于回款延迟造成的财务压力。

当资金流转速度提高,更多中小供应商才有动力持续释放算力资源。


第二层:早期零佣金,降低供给进入门槛

一个新的 Token 交易市场要真正运转,首先需要足够多的供应商参与。

供应商数量不足,开发者就没有足够的模型、价格和线路可以选择;开发者需求不足,供应商也没有持续提供资源的动力。

为了吸引早期供给,Unexhub 计划为符合条件的早期结算供应商提供最低可至 0 抽佣的激励政策。

零佣金的意义,不只是少收一笔平台服务费。

对早期供应商而言,它意味着可以在更低成本下:

  • 测试不同模型的真实需求;

  • 提高闲置 GPU 的利用率;

  • 获得新的开发者客户;

  • 减少渠道拓展成本;

  • 尽可能保留早期调用收入;

  • 验证长期商业模式。

对于平台而言,早期零佣金则有助于降低供给进入市场的摩擦。

供应商愿意提供资源,开发者才有更多选择;供给越充足,价格越透明,市场才越容易形成真正的流动性。


第三层:API 竞价,为 Token 建立价格发现机制

Unexhub 更具想象空间的一步,是筹备 API 竞价模式。

在传统 API 市场中,价格通常由模型厂商单方面决定。

开发者只能决定是否购买,却很难参与价格形成过程。即使某个时间段存在大量闲置算力,价格也未必会立即下降;即使高峰期资源紧张,供应商也很难通过市场报价获得额外收益。

API 竞价模式试图让价格根据实时供需变化:

  • 算力充足时,供应商可以通过更有竞争力的价格获得调用量;

  • 高峰需求上升时,能够稳定提供资源的供应商可以获得更高回报;

  • 开发者可以在价格合适时采购更多 Token;

  • 平台可以综合价格、延迟和稳定性进行动态路由。

这相当于为 Token 建立一套价格发现机制。

Token 不再只有固定标签,而是开始像电力现货、云计算竞价实例或其他标准化数字资源一样,根据市场供需形成价格。

所谓“像股票一样交易 Token”,更适合作为市场机制的比喻,而不是金融属性上的直接等同。

Token 并不天然等同于证券,也不意味着交易平台一定涉及金融投机。

真正值得关注的是,Token 开始具备:

公开报价、供需匹配、动态成交和快速结算。


时间维度与市场维度正在同时进入 Token 定价

DeepSeek 的峰谷定价与 Unexhub 的算力交易中心,看起来是两件不同的事情。

但将两者放在一起,可以看到 Token 定价逻辑正在发生的完整变化。

过去,Token 主要由厂商定价:

  • 模型厂商制定单价;

  • 开发者按固定价格购买;

  • 不同时间段价格相同;

  • 市场供需难以及时反映;

  • 用户缺乏更多采购渠道。

现在,DeepSeek 引入了时间维度:

同一种 Token,在上午 10 点和晚上 10 点,价格可能不同。

Unexhub 引入了市场维度:

同一种 Token,在供给充足和需求旺盛时,价格也可能不同。

时间维度加上市场维度,意味着 Token 正从单一的“厂商报价”,逐步走向更灵活的“市场定价”。

这并不意味着厂商将失去定价权,而是说,Token 的最终成交价格会越来越多地受到调用时间、资源供给、服务质量、路由稳定性和结算条件的共同影响。


B2B 清结算模式也在被重新定义

更深层的变化,发生在企业级算力供应与结算环节。

在传统模式下,算力供应商通常通过长期出租 GPU 获得收入。

这种商业模式往往具有以下特点:

  • 按月或按年签订合同;

  • 资源颗粒度较大;

  • 定价相对固定;

  • 账期较长;

  • 回款速度较慢;

  • 临时闲置资源难以变现。

Token 化交易则可以把大规模算力拆分成按实际调用量计费的标准化单位。

供应商不再只能出售整台服务器或整个月的 GPU 使用权,而是可以将闲置算力转化为更细颗粒度的 API 服务。

如果再叠加即时结算和动态竞价,供应商的商业模式就会从“出租固定资产”,转向“持续出售算力产出”。

由此带来的变化包括:

  • 闲置算力可以按调用量变现;

  • 低谷资源也能产生收入;

  • 高峰资源可以获得更高回报;

  • 资金回流速度得到提升;

  • 供应商可以动态调整供给;

  • GPU 资产周转效率提高。

算力不再只是被动出租的设备,而是逐渐成为一种能够被实时计量、分发、定价和结算的数字生产资料。


对开发者而言:可以用代码对冲价格波动

当 Token 价格开始随时间和市场变化,企业不一定只能被动接受账单上涨。

开发者可以通过系统设计主动控制成本。

一个更加成熟的 AI 应用,可以根据任务特征自动判断:

  • 任务是否必须立即执行;

  • 是否可以延迟至低谷时段;

  • 当前应该调用哪个模型;

  • 是否可以复用缓存结果;

  • 是否需要完整上下文;

  • 是否可以采用 Batch API;

  • 当前价格是否超过预算阈值;

  • 是否需要切换备用供应商。

例如,一家电商平台每天需要生成 10 万条商品描述。

如果所有任务都安排在工作日下午执行,就可能全部按高峰价格计费;如果将其中 80% 的非紧急任务转移到夜间,整体 Token 成本就可能明显下降。

实时客服保留在白天,批量内容生成则放在闲时。

企业不一定需要增加预算,仅通过任务分级和调度策略,就可能获得更大的 AI 使用规模。

这就是未来 AI 工程的重要能力:

用软件架构的灵活性,对冲 Token 价格的波动。


PoloAPI:在价格回归成本后,提供更灵活的采购路径

当大模型价格战逐渐回归理性,企业面对的问题就不再只是“谁的官方价格最低”。

更重要的是:

  • 如何统一接入多个模型;

  • 如何比较不同渠道的实际成本;

  • 如何减少重复开发;

  • 如何在高峰期切换备用线路;

  • 如何避免被单一供应商绑定;

  • 如何获得适合企业业务的定制支持。

PoloAPI 这样的多模型聚合平台,为企业和开发者提供了更灵活的采购方式。

通过统一接口,PoloAPI 可以接入 Gemini、Claude、ChatGPT、DeepSeek 等海内外主流模型。开发团队无需分别维护多套 SDK、鉴权方式、余额账户和账单系统,即可在一个 API 接入层中调用不同模型。

在符合条件的模型与服务方案下,部分调用价格最低可达到相应官方价格的约 30%。

对于批量内容生成、智能客服、代码辅助、数据处理和高并发 Agent 等长期消耗 Token 的场景,这种价格空间可以降低持续调用带来的预算压力。

真正有价值的,不只是单次请求更便宜,而是企业可以在相同预算下:

  • 承载更大的调用量;

  • 测试更多模型;

  • 扩展更多 AI 业务;

  • 为高峰期预留备用线路;

  • 将高性能模型与低成本模型组合使用;

  • 更精细地控制单任务成本。

PoloAPI 同时提供企业级定制服务,覆盖:

  • 多模型统一接入;

  • 原有 API 系统迁移;

  • 企业网关与权限适配;

  • 模型选型与动态路由;

  • 并发、限流及稳定性方案;

  • 调用账单与成本优化;

  • 专属部署和技术支持。

平台已积累国内外企业的大模型接入与定制服务经验,相关场景覆盖大型公司、上市公司及国有企业等不同类型客户。

在 Token 价格逐渐回归真实成本的阶段,聚合平台的价值并不是继续推动无底线价格战,而是通过统一采购、智能路由和规模化服务,帮助企业获得更可控的综合成本。


从购买单一模型,走向管理算力组合

过去,企业使用大模型时,通常先选择一家供应商,然后围绕单一模型搭建全部业务。

但当峰谷定价、API 竞价和多模型聚合逐渐成熟后,这种模式可能会被“算力组合管理”取代。

企业可以根据任务价值建立不同层级:

  • 复杂推理使用高性能模型;

  • 日常业务使用均衡型模型;

  • 批量任务使用低成本模型;

  • 离线任务安排在低谷时段;

  • 实时请求保留高峰资源;

  • 多条供应线路互为备用。

这种“组合”并不是金融投资,而是一种算力资源配置方式。

企业不再需要让一个昂贵模型处理所有工作,而是可以根据任务难度、时效、质量要求和当前价格,选择更合适的模型与调用渠道。

未来的 AI 成本管理,不再只是看一张 API 价目表。

它更像一套完整的算力供应链管理系统。


工程师视角:价格战结束之后,竞争才真正开始

高盛将 DeepSeek 峰谷定价视为 AI 云服务走向成熟的重要信号。

更准确地说,Token 正在经历电力市场曾经走过的过程:

  • 从统一价格走向峰谷价格;

  • 从固定报价走向动态报价;

  • 从单一供应走向多渠道采购;

  • 从长期租赁走向按需调用;

  • 从月度账期走向更快结算;

  • 从被动消费走向主动调度。

价格战之所以难以继续,并不是市场失去了竞争,而是长期亏本出售 Token 的模式不可持续。

峰谷定价让价格更接近算力的真实成本,交易平台让闲置资源进入流通,多模型聚合平台则帮助企业在不同模型和渠道之间进行选择与调度。

Token 正越来越像 AI 时代的“电费”。

它有高峰,也有低谷;有固定价格,也有动态价格;有直接采购,也会出现更多市场化交易方式。

对于每天查看 API 账单的工程师而言,未来需要回答的,将不再只是:

“哪个模型效果最好?”

还包括:

  • 什么时候调用更划算?

  • 哪些任务可以延迟?

  • 从哪个渠道采购更灵活?

  • 高峰期应该切换到哪个模型?

  • 如何使用缓存、队列和批处理降低成本?

  • 如何将 Token 价格变化写进系统架构?

模型能力依然决定产品体验。

但在下一阶段,企业管理、采购和调度 Token 的能力,也将成为竞争力的重要组成部分。

价格战正在退潮,而真正围绕成本效率、资源调度和商业化能力的竞争,才刚刚开始。


English Edition Below

Global Readers: Continue Scrolling for the Full English Version

The Price War Has Hit Its Limit: DeepSeek’s Peak Pricing Signals a Return to the Real Cost of Tokens

On the evening of June 29, 2026, DeepSeek notified its API developers that the official V4 model was scheduled to launch in mid-July. Alongside the release, the company introduced a peak-and-off-peak billing mechanism that remains unusual in China’s large-model API market.

During two major weekday demand windows—9:00 a.m. to 12:00 p.m. and 2:00 p.m. to 6:00 p.m.—some API prices would rise to twice their off-peak rates.

These periods cover most of the standard working hours used by domestic developers.

According to the announced pricing:

ModelOff-peak output pricePeak output price
V4 ProRMB 6 per million tokensRMB 12 per million tokens
V4 FlashRMB 2 per million tokensRMB 4 per million tokens

The announcement quickly divided the developer community.

Some argued that placing most of the working day inside the peak window was unfriendly to domestic users. Others noted that even after doubling, DeepSeek’s rates remained competitive compared with some models in the same category.

But the most important issue is not whether the price is simply higher.

The real question is: Why are large-model providers moving away from continuous price cuts and beginning to use peak pricing to manage computing demand?

According to the Goldman Sachs research view cited in the original material, this may indicate that the aggressive large-model price war is approaching its practical limit. Token pricing is beginning to move away from growth-at-any-cost discounts and back toward the real expenses created by computing infrastructure, electricity, and model serving.


Goldman Sachs’ View: The Price War Is Nearing an Inflection Point

The Goldman Sachs research view cited in the article suggests that DeepSeek’s peak pricing does not indicate weak demand.

Instead, it may reflect the opposite: AI usage is still rising, while GPU capacity, electricity, and data-center resources are becoming tighter during peak periods.

Goldman Sachs estimated that the adjusted average price for V4 Pro would be approximately $0.35 per million tokens, while the V4 Flash average would be around $0.12.

More important than the individual figures is the broader industry conclusion:

The aggressive large-model price war that intensified after late April 2026 may be approaching a turning point.

Over the previous months, AI companies continuously reduced token prices in an attempt to acquire developers and API market share.

DeepSeek sharply lowered the price of some services, while certain competing models experienced reductions approaching 99%. When prices are pushed toward zero margins—or even require providers to operate at a loss—the market is no longer experiencing ordinary competition. It is consuming the capital required for future model development and infrastructure investment.

Large-model services do not operate without costs.

Every API request consumes GPU capacity, electricity, network bandwidth, storage, cooling, engineering support, and ongoing model development. If tokens are sold below inference cost for an extended period, a provider may gain traffic in the short term but struggle to build a sustainable business.

From this perspective, DeepSeek’s peak pricing is better understood as a pricing correction than as a simple increase.

Competition is shifting from “Who can charge the lowest rate?” toward a broader contest involving:

  • Inference cost;

  • GPU utilization;

  • Token efficiency;

  • Peak-period service stability;

  • Sustainable commercialization;

  • Long-term investment capability.

The price war has not disappeared completely, but price alone is becoming less effective as a lasting competitive advantage.


Why Can the Price War No Longer Continue?

The answer begins with energy.

According to relevant industry forecasts, global data-center electricity consumption may reach 565 TWh in 2026. Electricity used by AI-optimized servers is projected to increase from 95 TWh to 175 TWh, an increase of approximately 84%.

The power requirements of newer GPUs are also increasing rapidly.

A Blackwell-architecture GPU may consume around 1 kW of power, compared with approximately 700 W for certain Hopper-generation products—an increase close to 40%.

When thousands or tens of thousands of GPUs run simultaneously, peak API usage becomes a peak not only in computing demand, but also in electricity, cooling, network, and data-center capacity.

The distribution of computing workloads over time therefore mirrors the distribution of electricity demand.

This is why DeepSeek’s peak pricing follows logic similar to peak-and-off-peak electricity tariffs:

  • Higher prices reduce unnecessary peak-period demand;

  • Lower prices encourage batch workloads to move off-peak;

  • Real-time applications remain active during the day;

  • Nonurgent workloads shift to quieter periods;

  • Providers use price signals to improve infrastructure utilization.

Viewed this way, peak pricing is not only a price increase. It is a demand-routing mechanism.

Computing power has not suddenly become expensive. What has become expensive is using it at the most congested time.


From Uniform Pricing to Time-Based Pricing

Historically, large-model APIs have generally used uniform pricing.

The same request submitted at 2:00 a.m. or 3:00 p.m. typically carried the same token price.

This approach was simple and predictable, but it did not reflect the true difference between peak and off-peak resource availability.

By introducing time-of-use pricing, DeepSeek has added an explicit time dimension to tokens.

The same model, the same workload, and the same token volume may now cost twice as much because of when the request is sent.

Tokens no longer have only a static list price. They now begin to carry a time-dependent value similar to electricity.

DeepSeek’s earlier and current pricing actions may appear contradictory, but the underlying logic is consistent.

Before introducing peak rates, the company had reduced some V4 Pro pricing from RMB 24 per million tokens to RMB 6, a decline of 75%. Lower off-peak prices make the model accessible to more developers, while higher peak prices encourage users to schedule their workloads more efficiently.

Both moves support the same objective:

Lower the entry barrier first, then improve the efficiency of computing-resource consumption.


The Real Upgrade Developers Need Is Workload Scheduling

As token prices begin changing by time, API cost management will also change.

Previously, development teams mainly reduced costs by:

  • Shortening prompts;

  • Removing repeated context;

  • Using prompt caching;

  • Selecting less expensive models;

  • Limiting output length;

  • Reducing unnecessary retries.

In the future, another variable will become equally important:

When the task is executed.

Enterprises can classify workloads by urgency.

Tasks that must run in real time

Examples include:

  • Online customer service;

  • Live translation;

  • Risk detection;

  • User-facing interaction;

  • Real-time code completion;

  • Production incident response.

These workloads are latency-sensitive and cannot simply be delayed, even during peak periods.

Tasks that can be delayed

Examples include:

  • Batch content generation;

  • Data cleaning;

  • Large-scale document summarization;

  • Offline model evaluation;

  • Product-description generation;

  • Daily and weekly report preparation;

  • Nonurgent agent workflows.

These jobs can be placed in queues and executed automatically at night or during lower-priced periods.

The same workload may cost significantly less simply by changing its execution time.

The cost difference between enterprises will therefore depend not only on who negotiates a lower API rate, but also on who builds the more capable scheduling system.

Model selection defines the ceiling of capability. Scheduling strategy defines the floor of cost.


Unexhub: Moving From “When to Use” to “How to Buy and Sell”

DeepSeek’s peak pricing mainly addresses the question of when API usage is most economical.

Unexhub is attempting to address a broader set of questions:

  • Where should tokens be purchased?

  • How can different suppliers participate?

  • How should API prices be formed?

  • How can suppliers receive payment faster?

  • How can idle computing capacity enter circulation?

Unexhub aims to build a token aggregation and distribution center that connects computing resources from model providers, GPU operators, and smaller suppliers, then distributes that capacity to enterprises and developers.

Its model can be divided into three parts.


Layer One: Faster Settlement and Supplier Cash Flow

Traditional computing agreements often involve long payment cycles.

Smaller suppliers may complete inference jobs or provide computing capacity, yet wait one or two months before receiving payment.

During that period, they must continue paying for:

  • GPUs and servers;

  • Data-center racks;

  • Electricity and cooling;

  • Network bandwidth;

  • Engineering and maintenance;

  • Infrastructure expansion.

Revenue remains delayed while expenses continue every day, creating significant cash-flow pressure.

Unexhub is developing a financial-grade real-time settlement model intended to help suppliers receive funds more quickly after delivering computing services.

A “supply and settle” structure can shorten capital turnover and reduce the amount of revenue trapped inside long payment cycles.

For suppliers, the value extends beyond faster payment. It may help them:

  • Reduce working-capital pressure;

  • Pay infrastructure expenses sooner;

  • Improve capital efficiency;

  • Expand capacity more flexibly;

  • Reduce the operational risk created by long settlement periods.

This does not mean that all business costs literally disappear. It means reducing the financial pressure caused by delayed payment as much as possible.

When capital moves faster, smaller suppliers have a stronger incentive to release more computing resources into the market.


Layer Two: Early Zero-Commission Incentives

A new token marketplace requires enough suppliers before it can offer meaningful choice.

Without sufficient supply, developers cannot compare enough models, prices, and service routes. Without demand, suppliers have little reason to continue listing their capacity.

To attract early providers, Unexhub plans to offer eligible early settlement suppliers commission rates as low as zero commission.

The value of zero commission is not simply the removal of a platform fee.

For early suppliers, it creates a lower-cost opportunity to:

  • Test real demand for different models;

  • Improve idle GPU utilization;

  • Reach new developer customers;

  • Reduce channel-acquisition expenses;

  • Retain a larger portion of early revenue;

  • Validate the long-term business model.

For the platform, zero commission reduces the friction of bringing new supply into the market.

When suppliers are willing to provide resources, developers gain more choices. As supply expands, pricing becomes more transparent and the market can begin to build genuine liquidity.


Layer Three: API Bidding and Token Price Discovery

The most ambitious part of Unexhub’s concept is an API bidding mechanism.

In the traditional API market, model providers set the price unilaterally.

Developers can decide whether to buy, but they have little influence over how the price is formed. Even when large amounts of capacity are idle, rates may not fall immediately. During peak demand, suppliers may also have limited ability to earn additional returns.

API bidding attempts to make prices respond to real-time supply and demand:

  • When capacity is abundant, suppliers may offer more competitive prices to win traffic;

  • During peak demand, stable providers may receive higher returns;

  • Developers may purchase more tokens when prices are favorable;

  • The platform may route traffic according to price, latency, and reliability.

This creates a potential price-discovery mechanism for tokens.

Instead of carrying a permanent fixed label, tokens begin to behave more like spot electricity, cloud-computing instances, or other standardized digital resources whose prices reflect market conditions.

The phrase “trading tokens like stocks” is best understood as a market-structure analogy rather than a direct financial classification.

Tokens are not automatically securities, and such a marketplace does not necessarily imply speculative finance.

The more important development is that tokens may gain:

Visible pricing, supply matching, dynamic transactions, and faster settlement.


Time and Market Variables Are Entering Token Pricing Together

DeepSeek’s peak pricing and Unexhub’s computing marketplace may appear unrelated.

Viewed together, however, they reveal a broader transformation in token pricing.

Previously, token prices were mainly provider-defined:

  • Model companies set the price;

  • Developers purchased at the listed rate;

  • Prices remained similar across time periods;

  • Supply and demand were not reflected quickly;

  • Users had limited procurement channels.

DeepSeek has introduced the time dimension:

The same token may cost differently at 10:00 a.m. and 10:00 p.m.

Unexhub introduces the market dimension:

The same token may also cost differently when supply is abundant or demand is high.

When time and market variables are combined, token pricing begins moving from a single provider quotation toward more responsive market pricing.

This does not mean model providers will lose control of pricing. It means that final transaction costs may increasingly reflect timing, resource supply, service quality, routing stability, and settlement conditions.


B2B Clearing and Settlement Are Being Redefined

A deeper change is taking place in enterprise computing supply.

Under the traditional model, suppliers often earn revenue by leasing GPUs under long-term agreements.

These arrangements commonly involve:

  • Monthly or annual contracts;

  • Large resource units;

  • Relatively fixed prices;

  • Long payment terms;

  • Slow cash recovery;

  • Limited monetization of temporary idle capacity.

Token-based transactions can divide large blocks of computing power into standardized units billed according to actual API usage.

Suppliers no longer need to sell only an entire server or a full month of GPU access. They can turn temporary spare capacity into more granular API services.

When real-time settlement and dynamic bidding are added, the supplier model shifts from “renting fixed assets” to “continuously selling computing output.”

This can produce several changes:

  • Idle compute can be monetized according to usage;

  • Off-peak resources can still generate revenue;

  • Peak capacity can earn higher returns;

  • Capital can return faster;

  • Suppliers can adjust capacity dynamically;

  • GPU asset turnover can improve.

Computing infrastructure stops being only passively rented equipment and begins to function as digital production capacity that can be measured, distributed, priced, and settled continuously.


For Developers: Using Software to Hedge Price Volatility

As token prices change with time and market conditions, enterprises do not need to accept rising bills passively.

Developers can use system design to control costs.

A mature AI application can automatically determine:

  • Whether a task must run immediately;

  • Whether it can be delayed until off-peak hours;

  • Which model should be used;

  • Whether cached results can be reused;

  • Whether the full context is necessary;

  • Whether a batch API is appropriate;

  • Whether the current price exceeds a budget threshold;

  • Whether traffic should switch to another supplier.

Consider an e-commerce platform that generates 100,000 product descriptions every day.

If all tasks run during weekday afternoons, they may all be billed at peak prices. If 80% of the nonurgent workload is shifted overnight, the total token cost may fall significantly.

Real-time customer service stays active during the day, while batch content production moves off-peak.

The enterprise does not necessarily need a larger budget. Better workload classification and scheduling may allow it to process more AI tasks with the same spending level.

This is becoming a core capability in AI engineering:

Using software flexibility to offset token-price volatility.


PoloAPI: A More Flexible Procurement Path as Pricing Returns to Cost

As the large-model price war becomes more rational, enterprises must consider more than which provider lists the lowest official rate.

They also need to determine:

  • How to connect multiple models through one interface;

  • How to compare the actual cost of different channels;

  • How to reduce repetitive integration work;

  • How to switch to backup capacity during peak demand;

  • How to avoid dependence on a single provider;

  • How to obtain enterprise-specific implementation support.

Multi-model aggregation platforms such as PoloAPI provide a more flexible procurement path.

Through a unified API, PoloAPI can connect major domestic and international models, including Gemini, Claude, ChatGPT, and DeepSeek. Development teams do not need to maintain separate SDKs, authentication systems, prepaid balances, and billing platforms for every provider.

For eligible models and service plans, some rates may be as low as approximately 30% of the corresponding official price.

For workloads that consume tokens continuously—such as batch content generation, intelligent customer service, coding assistance, data processing, and high-concurrency agent systems—this pricing flexibility can reduce long-term budget pressure.

The value is not limited to cheaper individual requests. Under the same budget, enterprises may be able to:

  • Support larger request volumes;

  • Test more models;

  • Expand more AI use cases;

  • Reserve backup routes for peak periods;

  • Combine high-performance and lower-cost models;

  • Control task-level spending more precisely.

PoloAPI also provides enterprise customization services covering:

  • Unified multi-model access;

  • Migration from existing API systems;

  • Enterprise gateways and permission controls;

  • Model selection and dynamic routing;

  • Concurrency, rate-limit, and reliability design;

  • Billing and cost optimization;

  • Dedicated deployment and technical support.

The platform has accumulated experience in large-model integration and customization for domestic and international organizations, including large companies, publicly listed enterprises, and state-owned businesses.

As token pricing returns toward real costs, the value of aggregation platforms is not to continue an unsustainable race to the bottom. It is to help enterprises achieve more controllable total costs through unified procurement, intelligent routing, and scaled service delivery.


From Buying One Model to Managing a Compute Portfolio

Historically, enterprises selected one model provider and built most of their applications around that service.

As peak pricing, API bidding, and multi-model aggregation mature, this approach may be replaced by compute-portfolio management.

Enterprises can build different layers for different tasks:

  • High-performance models for difficult reasoning;

  • Balanced models for routine business workloads;

  • Lower-cost models for batch processing;

  • Off-peak execution for offline tasks;

  • Peak capacity reserved for real-time demand;

  • Multiple providers used as backup channels.

The word “portfolio” does not imply financial investment. It describes a resource-allocation strategy.

Companies no longer need to force one expensive model to handle every task. They can select the right model and procurement channel according to task complexity, urgency, quality requirements, and current pricing.

Future AI cost management will involve more than reading an API price sheet.

It will resemble a complete computing-supply-chain management system.


An Engineer’s View: The Real Competition Begins After the Price War

Goldman Sachs reportedly views DeepSeek’s peak pricing as an important sign that AI cloud services are maturing.

A more precise interpretation is that tokens are beginning to follow the path once taken by electricity markets:

  • From uniform rates to peak-and-off-peak pricing;

  • From fixed quotations to dynamic pricing;

  • From single-source supply to multi-channel procurement;

  • From long-term leasing to on-demand usage;

  • From monthly settlement to faster clearing;

  • From passive consumption to active scheduling.

The token price war is becoming difficult to sustain not because competition has disappeared, but because selling inference below cost cannot continue indefinitely.

Peak pricing makes prices more reflective of real computing expenses. Trading platforms can bring idle resources into circulation. Multi-model aggregation platforms can help enterprises choose and route among different models and channels.

Tokens are increasingly becoming the electricity bill of the AI era.

They have peaks and valleys, fixed rates and dynamic prices, direct procurement and increasingly market-oriented transaction mechanisms.

For engineers who monitor API bills every day, the future question will no longer be only:

“Which model performs best?”

It will also include:

  • When is the most economical time to run the task?

  • Which workloads can be delayed?

  • Which procurement channel is more flexible?

  • Which model should be used during peak periods?

  • How can caching, queues, and batch processing reduce costs?

  • How should token-price changes be designed into the system architecture?

Model capability will continue to shape product quality.

But in the next stage of AI competition, the ability to procure, manage, and schedule tokens will become equally important.

The price war is fading. The real competition in cost efficiency, resource orchestration, and commercialization is only beginning.