DeepSeek峰谷定价引爆争议:token交易市场的模型供应商如何用"峰谷+批量"把企业AI账单打下来

AI 算力正在从固定价格的商品,变成具有时间价值的动态资产。


The Butcher Puts Down the Knife; the Calculator Takes Over

At midnight on August 17, the new API pricing for the DeepSeek V4 series officially took effect. The company once nicknamed the AI industry's “price butcher” is beginning to put down its knife.

This is not simply a price increase. For V4 Pro, the peak-period price for cache-hit input jumped from RMB 0.025 to RMB 0.30 per million tokens—an 1,100% increase. Output pricing rose from RMB 6 to RMB 27 per million tokens—an increase of 350%. The same API can now cost twice as much depending on when it is called.

The reaction was immediate. One community comment was repeatedly reposted:

“DeepSeek doesn't have to be the strongest. But it can't be expensive. Once it's expensive, that's the original sin.”

But the story is more complicated than that.

Peak-and-off-peak pricing is not simply about charging more. It signals something much bigger: AI compute is moving from a fixed-price commodity toward a time-sensitive asset.


Peak vs. Off-Peak: Time-of-Use Pricing for AI Compute

DeepSeek defines peak hours as 9:00–12:00 and 14:00–18:00 Beijing time, precisely when enterprise and developer traffic is typically concentrated.

The remaining 17 hours are off-peak, with prices cut by roughly half.

For V4 Pro, for example:

PricingPeakOff-Peak
Cache-miss inputRMB 9/M tokensRMB 4.5/M tokens
OutputRMB 27/M tokensRMB 13.5/M tokens
Cache-hit inputRMB 0.30/M tokensRMB 0.15/M tokens

DeepSeek's stated logic is straightforward: use market-based pricing to encourage enterprises and developers to shift workloads away from peak hours and stabilize compute demand.

The mechanism is almost identical to time-of-use electricity pricing.

Electricity costs more during periods of heavy demand and less overnight—not simply because power companies want to charge more, but because price signals encourage consumers to move flexible demand into periods when infrastructure is underutilized.

The same principle is now being applied to AI inference.

For enterprises, the implication is simple:

The same workload can cost dramatically less if you choose the right time to run it.

That changes the way AI infrastructure should be budgeted. Instead of asking only “Which model is cheapest?”, companies increasingly need to ask:

“When is the cheapest time to use this model?”


Who Gets Hurt—and Who Benefits?

The impact of the pricing change is far from uniform.

1. Continuous Agent Workloads Feel the Most Pressure

Individual developers and teams running AI agents continuously are among the most exposed.

Their workloads are persistent and interactive, making it difficult to move everything into off-peak windows.

OpenCode previously shared an example of a typical V4 Flash request containing approximately 410 new input tokens, 71,300 cached tokens, and 310 output tokens.

Under the previous pricing, each request cost roughly $0.000344. Under the new pricing, the estimated cost rises to approximately $0.000794 during off-peak hours and $0.001588 during peak hours—around 2.31× and 4.62× the previous level.

For high-frequency agent workloads, that difference can accumulate quickly.


2. Batch Inference Gets a New Opportunity

Enterprise batch workloads are in a completely different position.

Suppose an enterprise consumes 100 million output tokens during peak hours every day. Under the old pricing, the daily cost would be around RMB 600. Under the new peak pricing, that could rise to RMB 2,700.

But if the same workload can be moved into off-peak hours, the cost can be cut substantially.

Tasks such as:

  • Large-scale data labeling

  • Document processing

  • Code refactoring

  • Model evaluation

  • Offline benchmarking

  • Dataset generation

do not necessarily require an immediate response.

For these workloads, a scheduler can simply adopt an “off-peak first” strategy.

The model does not change.
The output quality does not change.
The workload does not change.

Only the timing changes—and the bill can change with it.


3. Aggregation Platforms Add Another Layer of Price Competition

Large enterprises also have another option: use model aggregation and routing platforms to access alternative inference supply.

Platforms such as OpenRouter connect the same model to multiple inference providers. As of August 18, DeepSeek V4 Pro was reportedly available through 11 providers, with many standard input and output prices clustered around $1.32 and $3.96 per million tokens.

The important point is not the exact number.

It is that an official price increase does not necessarily translate into an identical price increase across every distribution channel.

Aggregation platforms can use provider competition, routing strategies, capacity availability, and volume-based negotiation to absorb part of the price pressure.

This is where the AI market begins to move beyond a simple vendor price list.


The Token Trading Market: Turning Peak and Off-Peak Pricing into a Permanent Mechanism

DeepSeek's peak-and-off-peak pricing may only be the beginning.

A new intermediary layer is emerging around token distribution—connecting upstream model providers with downstream developers, enterprises, and individual users.

At its core, this is becoming a liquidity infrastructure for the global movement of AI tokens, connecting wholesale compute supply with retail-level demand.

The problem is straightforward:

AI compute has idle periods and congested periods, but traditional API pricing is usually fixed.

DeepSeek has begun breaking that model with time-based pricing.

A compute trading center aims to take the idea one step further and make dynamic pricing a permanent market mechanism.


Unexhub: From Price Lists to Market Pricing

This is where Unexhub positions itself differently from a conventional API aggregator.

Unexhub is designed as a token aggregation and distribution compute trading center, with three core mechanisms.

1. Real-Time Settlement

Traditional AI suppliers often face a significant cash-flow problem.

They deploy models, provide inference capacity, accumulate usage, and then wait for settlement.

Unexhub's model is designed around:

“Supply more, earn more—settle in real time.”

Instead of “work first, get paid later,” suppliers can move toward “deliver capacity, receive settlement.”

For smaller AI providers, this can be particularly meaningful because faster settlement reduces working-capital pressure and makes it easier to monetize idle inference capacity.


2. Early-Stage Zero Commission

Unexhub also uses an early-stage 0% commission strategy to attract suppliers.

Compared with the common 5%–5.5% platform-fee range associated with some aggregation platforms, zero commission means participating suppliers can retain 100% of their inference revenue during the qualifying period.

That benefit is not merely attractive to suppliers.

If distribution costs fall, suppliers gain more room to compete on API prices, which can ultimately create better pricing opportunities for developers and enterprise users.

In other words:

Lower platform friction → stronger supplier participation → more competition → greater room for lower user prices.


3. API Competitive Bidding

This is the most important part of the model.

Traditional API pricing works like a catalog:

The provider sets the price.
The customer pays the price.

A market-oriented model works differently.

When supply exceeds demand, providers can compete for workloads by offering more attractive prices.

When demand surges and capacity becomes scarce, prices can rise accordingly.

For developers, this creates the possibility of accessing lower-cost inference during periods of excess capacity.

For suppliers, it creates an opportunity to monetize idle capacity instead of leaving GPUs unused.

The underlying idea is simple:

Turn AI inference from a fixed price list into a real-time supply-and-demand market.

DeepSeek's peak/off-peak system is still primarily a price schedule defined by the model provider.

A compute trading center goes one step further:

Let market competition help determine the price.


Compute Is Becoming a Tradable Asset

On August 11, 2026, CME announced plans to launch the first global compute futures contracts on October 5, linked to GPU rental price indexes for H100 and B200 systems.

Whether viewed from the perspective of infrastructure, cloud computing, or financial markets, the direction is becoming increasingly clear:

Compute is moving toward standardization, price discovery, and financialization.

This is the transformation that many commodities have gone through.

First comes customized bilateral trading.

Then comes standardized pricing.

Then marketplaces emerge.

Then derivatives and financial instruments appear.

AI compute is beginning to follow the same path.

DeepSeek's peak/off-peak pricing represents the micro-level version of this transformation at the API layer.

It turns “when you use compute” into a price signal.

Token trading markets can potentially amplify that signal into a broader market mechanism:

aggregate demand → match supply → discover prices → settle transactions.

For enterprises, AI infrastructure procurement is therefore moving from:

“Look at the price list.”

to:

“Look at the market.”


The Real Shift: From the Price Tag to the Calculator

The controversy surrounding DeepSeek's new pricing is not ultimately about whether the company has become “expensive.”

It is about the emergence of a new way to calculate AI infrastructure costs.

Yesterday, enterprises asked:

“How much does one million tokens cost?”

Tomorrow, the more important questions may be:

  • Is the input cached?

  • Is the workload urgent?

  • Is it running during peak or off-peak hours?

  • Which provider has available capacity?

  • Which routing path is cheapest?

  • Can suppliers compete for the workload?

  • Can settlement happen immediately?

The token price itself becomes only one variable in a much larger equation.

And this is precisely where the concept of a compute trading center becomes important.

The objective is not simply to advertise a lower price.

It is to make price discovery, routing, settlement, and supply competition part of the infrastructure itself.


Conclusion: The Butcher Has Put Down the Knife. The Calculator Has Just Started Running.

DeepSeek's peak-and-off-peak pricing marks a turning point.

The era of the AI “price butcher”—where the simplest way to win users was to keep cutting prices—is gradually giving way to a more sophisticated era of capacity management and market-based pricing.

The future competition may no longer be:

Who can make AI cheaper?

It may become:

Who can allocate AI compute more efficiently?

For enterprises, the opportunity is equally clear.

Flexible workloads can be scheduled during lower-cost periods. Batch workloads can be routed toward cheaper capacity. Multiple providers can compete for the same demand. And users can increasingly pay according to actual token consumption rather than committing to large fixed infrastructure investments.

That is where Unexhub's value proposition becomes more concrete.

Unexhub is building a token aggregation and distribution compute trading center, bringing together AI suppliers of different sizes—from major model providers such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, to high-quality models supplied by individuals and smaller AI companies.

Instead of asking developers to manually compare dozens of providers, Unexhub aims to aggregate supply, enable competitive bidding, and use millisecond-level routing to connect demand with available inference capacity.

For developers, the model is straightforward: pay by actual token usage, with eligible routes potentially reaching as low as 30% of prevailing market prices, depending on real-time supply, demand, and provider offers. For suppliers, the platform provides a one-click distribution channel and real-time settlement of usage revenue.

The real value is therefore not simply another discount.

It is the possibility of turning discounts into a market mechanism.

DeepSeek has shown that time can change the price of compute.

Aggregation has shown that channels can change the price.

Competitive bidding may ultimately show that the market itself can change the price.

The butcher has put down the knife.

Now, the calculator takes over.


屠夫收刀,算盘登场

8 月 17 日零点,DeepSeek V4 系列 API 新价格正式生效。

这家曾被业内戏称为大模型“价格屠夫”的公司,正在逐渐收起它的屠刀。

但这一次,并不能简单理解为“涨价”。

V4 Pro 高峰时段缓存命中输入价格从 0.025 元上涨至 0.30 元/百万 Token,涨幅达到 1100%;输出价格则从 6 元上涨至 27 元/百万 Token,涨幅达到 350%。

同一个 API,仅仅因为调用时间不同,就可能产生巨大的价格差异。

社区因此迅速出现争议:

“DeepSeek 可以不强,但不能贵。贵了就是原罪。”

但如果把视角拉远,会发现这件事真正值得关注的地方,并不是 DeepSeek 到底“贵不贵”。

而是:

AI 算力正在从固定价格的商品,变成具有时间价值的动态资产。


峰谷定价:AI 算力版的“夜间电价”

DeepSeek 将北京时间每天 9:00–12:00、14:00–18:00 划定为高峰时段。

这恰好覆盖了企业和开发者最集中的工作时间。

其余 17 个小时,则属于闲时,部分价格直接下降约一半。

以 V4 Pro 为例:

计费项目高峰时段闲时时段
缓存未命中输入9 元/百万 Token4.5 元/百万 Token
输出27 元/百万 Token13.5 元/百万 Token
缓存命中输入0.30 元/百万 Token0.15 元/百万 Token

DeepSeek 的逻辑其实非常清楚:

通过市场化价格,引导企业和开发者主动错峰调度任务,从而缓解高峰期算力压力。

这和电力行业的峰谷电价几乎是同一个逻辑。

白天用电高峰价格更高,晚上用电低谷价格更低,并不是单纯为了“多收钱”,而是通过价格信号,把一部分可以延后的需求从拥堵时段转移到基础设施空闲时段。

AI 算力正在出现同样的机制。

对于企业而言,这意味着:

同样的任务,选对时间调用,账单就可能明显下降。

于是,企业采购 AI 算力时,未来需要问的问题不再只是:

“哪个模型最便宜?”

而是:

“什么时候调用这个模型最便宜?”


谁受伤,谁受益?

峰谷定价对不同用户群体的影响完全不同。

1. 连续运行的智能体工作负载压力最大

最容易受到冲击的,是个人开发者以及持续运行 AI Agent 的团队。

这类任务具有明显的连续性和交互性,很难把所有请求统一挪到闲时。

OpenCode 曾披露过一个典型 V4 Flash 请求:

  • 新输入 Token:约 410

  • 缓存 Token:约 71,300

  • 输出 Token:约 310

按照此前价格,每次请求成本约为 0.000344 美元。

新价格下,闲时约为 0.000794 美元,高峰则约为 0.001588 美元。

也就是说,成本大约变成原来的 2.31 倍和 4.62 倍。

对于高频 Agent 工作负载而言,这种差异经过大量调用之后,会迅速累积成一笔可观的账单。


2. 批量推理反而迎来机会

企业批量推理的情况则完全不同。

假设一家企业每天在高峰时段消耗 1 亿输出 Token。

按照此前价格,每天成本约为 600 元;按照新的高峰价格,则可能上升到 2700 元。

但如果把同样的任务转移到闲时执行,成本就能够显著下降。

例如:

  • 大规模数据标注

  • 文档批量处理

  • 代码重构

  • 模型评估

  • 离线 Benchmark

  • 数据集生成

这些任务并不一定要求秒级返回。

因此,企业完全可以通过调度器建立:

“闲时优先”的 AI 算力策略。

模型没有换。

输出质量没有降低。

任务本身也没有改变。

改变的只是调用时间,而账单也随之改变。


3. 聚合平台带来新的价格竞争空间

大型企业还有另一个选择:

通过模型聚合与路由平台寻找不同的推理供应。

OpenRouter 等平台将同一个模型接入多个推理服务商。

截至 8 月 18 日,DeepSeek V4 Pro 已接入多家供应商,部分普通输入和输出价格明显低于官方渠道。

真正值得关注的并不是某一个具体数字。

而是:

官方价格上涨,并不意味着所有渠道都会按照同样幅度上涨。

聚合平台可以利用供应商之间的竞争、路由策略、实时算力余量以及批量采购能力,吸收一部分价格压力。

AI 推理市场因此开始从简单的“厂商价目表”,走向更加复杂的市场化定价。


Token 交易市场:把“峰谷价格”变成常态

DeepSeek 的峰谷定价可能只是一个开始。

围绕 Token 分发,一个新的中间市场正在形成。

它连接上游模型供应商与下游开发者、企业以及个人用户。

从本质上看,这是一个连接全球 AI 算力供给与需求的 Token 流动性基础设施。

它解决的问题非常直接:

算力存在高峰和闲时,但传统 API 往往只有一个固定价格。

DeepSeek 开始用峰谷定价打破这种固定模式。

而算力交易中心则试图进一步把这种动态机制变成市场的常态。


Unexhub:从固定价目表走向市场定价

Unexhub 正在构建的是一个 Token 聚合与分发的算力交易中心。

它并不是简单地把几十个模型的价格放在一起供用户选择,而是围绕三个核心机制构建交易基础设施。

第一,实时结算清算

传统模式下,中小 AI 供应商需要部署模型、提供推理服务,然后等待结算。

这意味着供应商需要承担一定的资金周转压力。

Unexhub 的实时结算机制强调:

供多少,收多少。

从“先干活、后收款”,逐步转向:

干多少,结多少。

对于中小供应商而言,更快的资金回收能够降低运营压力,也能让闲置推理能力更快变现。


第二,早期零佣金

Unexhub 通过早期 0 佣金策略吸引供应商加入。

相比部分聚合平台常见的 5%–5.5% 平台服务费,符合条件的早期供应商能够保留 100% 的推理收入。

这个权益的价值并不只体现在供应商端。

平台分发成本降低之后,供应商拥有更大的价格竞争空间,最终也可能将更多成本优势传导给开发者和企业用户。

简单来说:

平台摩擦更低 → 供应商参与更多 → 市场竞争增强 → 用户价格拥有更大下降空间。


第三,API 竞价模式

这是整个机制中最核心的一环。

传统 API 定价像一本固定价目表:

厂商定价,用户付款。

而竞价模式则完全不同。

当市场供给大于需求时,供应商可以通过更具竞争力的报价争取订单。

当需求突然增加、算力供给紧张时,价格则可能上涨。

对于开发者而言,这意味着在算力供给充足的时段,有机会获得更具吸引力的 Token 价格。

对于供应商而言,则意味着可以把原本闲置的算力转化成收入。

核心逻辑就是:

把 AI 推理从固定价格目录,变成实时供需市场。

DeepSeek 的峰谷定价,是模型厂商根据时间主动制定价格。

算力交易中心进一步做的是:

让市场竞争参与决定价格。


算力正在成为可交易资产

2026 年 8 月 11 日,芝加哥商业交易所宣布计划于 10 月 5 日推出首批与 H100、B200 GPU 租赁价格指数相关的算力期货合约。

无论从基础设施、云计算还是金融市场角度来看,一个趋势都越来越清晰:

算力正在走向标准化、价格发现和金融化。

很多大宗商品都经历过类似过程:

从定制化一对一交易开始;

然后形成标准化报价;

随后出现交易市场;

最终形成金融衍生品。

AI 算力正在沿着类似路径发展。

DeepSeek 的峰谷定价,就是这一趋势在 API 层面的微观体现。

它把:

“什么时候使用算力”

变成了一个价格信号。

而 Token 交易市场,则有可能进一步把这个价格信号放大为完整的市场机制:

聚合需求 → 撮合供给 → 价格发现 → 实时结算。

企业采购 AI 算力,也正在从:

“看价目表。”

逐渐转向:

“看市场。”


真正的变化:不是价格表变了,而是算盘变了

DeepSeek 这一次引发争议的真正价值,并不在于它是不是已经“贵了”。

真正的变化,是 AI 算力的成本计算方式正在发生变化。

过去企业计算 API 成本,只需要问:

“每百万 Token 多少钱?”

未来可能需要同时考虑:

  • 是否命中缓存?

  • 是否需要实时响应?

  • 当前是否处于高峰?

  • 哪家供应商有闲置算力?

  • 哪条路由价格最低?

  • 是否可以让供应商竞价?

  • 是否支持实时结算?

Token 单价,只是成本公式中的一个变量。

真正决定最终账单的,是时间、缓存、路由、供需和交易机制共同构成的成本结构。

这正是算力交易中心存在的意义。

它并不是单纯宣传一个“更低价格”。

而是试图把:

价格发现、动态路由、供应商竞争和实时结算

全部变成基础设施的一部分。


结语:屠夫收刀,算盘登场

DeepSeek 的峰谷定价,正在成为 AI 算力市场的一道分水岭。

过去的“价格屠夫”时代,竞争方式很简单:

谁能把价格打下来,谁就能吸引用户。

但未来的竞争可能变成:

谁能让算力被更高效地使用?

对于企业而言,这意味着新的成本优化空间正在出现。

可以把灵活任务安排到低价时段;

可以把批量任务交给低成本算力;

可以让多个供应商竞争同一批需求;

也可以根据真实 Token 用量付费,而不必为了峰值需求提前购买大量固定资产。

这正是 Unexhub 想解决的问题。

Unexhub 正在构建 Token 聚合与分发的算力交易中心,汇聚不同规模的 AI 供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型生态,同时也吸引个人开发者、中小企业提供的优质 AI 模型和推理能力。

平台通过聚合供应、API 竞价以及毫秒级路由,让企业和开发者能够更高效地匹配不同供应商的实时算力。

对于开发者而言,核心权益非常直接:

按实际 Token 使用量付费,而不是为闲置算力买单。

在符合具体供应与竞价条件的情况下,开发者最低可享受市场价格约 3 折的调用机会;对于供应商,则可以通过一键接入渠道获得新的需求来源,并按照实际调用情况进行实时结算分成。

因此,Unexhub 的价值并不只是“再便宜一点”。

更重要的是,它试图把:

“优惠价格”变成“市场机制”。

DeepSeek 告诉市场:

时间可以改变算力价格。

聚合平台告诉市场:

渠道可以改变算力价格。

而 API 竞价机制最终可能告诉市场:

市场本身,就可以决定算力价格。

屠夫已经收刀。

真正的算盘,才刚刚拨响。