From 40% to 65%: How Domestic AI Clusters Are Squeezing More Potential from Every Chip

The industry is facing both insufficient computing capacity and significant structural idleness.


From 40% to 65%: How Domestic AI Clusters Are Squeezing More Potential from Every Chip

According to data from the China Academy of Information and Communications Technology (CAICT), the average GPU utilization rate across China’s intelligent computing centers is only around 30%, with some small and mid-sized clusters falling below 20%.

China’s intelligent computing capacity has already reached 1,882 EFLOPS, yet actual utilization at many intelligent computing centers remains between 20% and 40%. In some domestic-chip clusters, idle capacity can reach 50%–80%.

That leads to a paradox:

The industry is facing both insufficient computing capacity and significant structural idleness.

The problem is not simply that there is not enough computing power.

It is that too much of the available power is being used inefficiently.


Computing Optimization: From “Adding Hardware” to “Squeezing Out Efficiency”

The AI industry is gradually shifting its focus from raw performance per chip to cost and efficiency per unit of computation.

The question is no longer simply how many GPUs a company owns. It is becoming:

How many more tokens can we generate from every kilowatt-hour, every second of GPU time, and every dollar of infrastructure spending?

That question has become a critical competitive advantage for AI infrastructure providers.

Tencent Cloud: Making Existing Hardware Work Harder

Tencent Cloud TencentOS provides a useful example.

Through fine-grained qGPU partitioning, enterprises can reportedly reduce GPU hardware procurement costs by more than 60%. With offline workload co-location, server CPU utilization can rise from 15% to 45%, reducing overall usage costs by roughly 50%.

Combined with the tidal scheduling mechanism of the TI platform, overall cluster utilization can increase from approximately 30% to as high as 90%—effectively tripling utilization.

Beijing Data Group: Turning Computing into a “Token Factory”

The “Jingsuan Token Factory” launched by Beijing Data Group follows a similar philosophy.

Through heterogeneous computing resource pooling and scheduling, utilization can rise from the industry’s typical 30%–40% range to more than 80%.

With intelligent PD-separated scheduling, computing throughput can increase by up to 90%, while inference latency in large-model scenarios can improve by more than 70%.

For long-context workloads, the combination of full-stack inference acceleration and PD separation can reportedly reduce the cost of generated tokens by more than 50%.

The broader trend is clear:

Hardware optimization can push cluster utilization from roughly 30%–50% toward 60%–70%, while software optimization can increase token-generation efficiency by more than 30% without necessarily sacrificing model quality.

And some cases are even more aggressive.

Wuwen Xinqiong has demonstrated hybrid training across six different domestic and international chip types, achieving more than 97% computing utilization.

In another optimization effort involving Zhonghao Xinying and ZTE, TPU cluster efficiency reportedly improved by more than 10×.

From 40% to 65%.

From 30% to 90%.

From 50% to 97%.

Behind these numbers lies a simple conclusion:

The chips have already been purchased. The real competitive advantage is learning how to use them efficiently.


The Token Trading Market: Turning “Idle Capacity” into “Liquid Capacity”

Low utilization is ultimately a scheduling problem.

Computing resources are often locked into specific projects, specific time windows, or specific clusters. When demand changes, those resources cannot easily move to where they are needed most.

An integrated computing network can change that dynamic.

Cross-center optimization, peak-load shifting, workload queuing, and unified scheduling can push the utilization of computing resources significantly higher.

More importantly, tokens can serve as a standardized unit for measuring and settling computing resources across different GPU configurations.

This changes the role of computing power.

It is no longer merely a collection of physical hardware assets.

It becomes a programmable and tradable digital resource.

In that model, users can consume computing resources much like electricity:

use what you need, when you need it, and pay according to actual consumption.

The deeper value of the token economy is standardization.

Complex infrastructure and software frameworks remain behind the scenes, while standardized APIs expose a common interface to the market.

Upstream data-service providers can settle revenue according to effective token usage. Computing and model providers can price services according to token consumption. Downstream applications can calculate their AI costs and ROI using the same basic unit.

A standardized measurement framework can reduce negotiation complexity, reconciliation costs, and custom integration work across the ecosystem.

In other words, tokens can become not only a measurement unit, but also a common language connecting computing supply with AI demand.


Unexhub’s Approach: Turning “Computing Utilization” into “Computing Liquidity”

This is where the combination of computing optimization + token trading becomes particularly interesting.

Unexhub is building a token aggregation and distribution-oriented computing trading center, with the goal of turning computing resources from fixed assets into tradable market resources.

Its model focuses on several layers.

1. Real-Time Settlement

For small and medium-sized computing suppliers, capital turnover can be a major constraint.

Unexhub’s real-time settlement mechanism is designed to synchronize token supply with revenue settlement, helping suppliers recover funds faster and reducing the pressure of long payment cycles.

2. Early-Stage Zero Commission

The early supplier zero-commission policy is more than a simple discount.

It is designed to encourage more computing and model providers to enter the network and increase market liquidity.

For suppliers, it lowers the cost of testing a new distribution channel.

For buyers, more suppliers mean more available capacity and greater pricing competition.

3. Dynamic API Bidding

The most important layer is the API bidding mechanism.

Traditional API pricing is generally fixed. Peak congestion and off-peak idle capacity can end up being priced almost identically.

Unexhub’s model allows API prices to respond dynamically to supply and demand.

When demand rises, suppliers can capture higher prices.

When capacity is underutilized, prices can fall, giving developers and enterprises access to more economical computing resources.

This creates a market feedback loop:

high demand → higher prices → more supply → improved capacity allocation

and

low demand → lower prices → increased utilization → less idle capacity.

That is the difference between simply improving computing efficiency and creating a market that can continuously redistribute computing resources.


From “Computing Utilization” to “Computing Trading Efficiency”

Computing optimization answers one question:

How can we make each chip do more work?

A token trading market addresses another:

How can we make idle computing capacity available to someone who actually needs it?

The first is technical efficiency.

The second is market efficiency.

Together, they form a more complete cost-reduction mechanism.

When domestic computing clusters push utilization from 40% toward 65%, and token markets allow idle capacity to move between suppliers and users, the cost structure of AI inference begins to change fundamentally.

The decline in cost per token does not necessarily have to come from a temporary price promotion by one model provider.

It can come from a broader transformation across the infrastructure chain:

from adding more hardware → to maximizing utilization → to making computing resources tradable.


65% Is Not the Finish Line

Neither 65% nor 97% should be considered the final target.

The real goal is much simpler:

Every chip should be working where it creates the most value.

Every kilowatt-hour should produce as much useful intelligence as possible.

And every token should represent measurable, transparent computing value.

That is the direction in which the AI infrastructure market is evolving.

Unexhub positions itself around this transition by building a global token aggregation and distribution computing marketplace. The platform brings together AI providers across different scales, including major model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as high-quality models offered by individual developers and smaller businesses.

For enterprises and developers, the value is not simply “more models.”

It is more choice, competitive pricing, and faster routing across available AI capacity.

Unexhub states that developers can pay according to actual token usage, with eligible access potentially reaching as low as 30% of prevailing market prices, while suppliers can connect their services through a unified channel and receive real-time settlement for usage-based revenue sharing.

The underlying idea is straightforward:

Don’t just build more computing power. Make the computing power you already have work harder—and make unused capacity easier to trade.


中文版

从 40% 到 65%:国产集群如何榨干每一颗芯片的潜力?

根据中国信通院数据,全国智能计算中心的 GPU 平均利用率目前只有 30% 左右,部分中小型集群甚至低于 20%。

与此同时,中国智能算力规模已经达到 1882 EFLOPS,但不少智算中心的实际利用率仍然只有 20%–40%。更值得注意的是,部分国产芯片集群的闲置率甚至达到 50%–80%。

这就形成了一个看似矛盾、实际上非常典型的局面:

一边喊算力不够,一边又有大量算力闲置。

所以问题并不只是“算力不足”。

更重要的问题是:

已有的算力,没有被高效利用。


算力精调:从“堆硬件”到“榨效率”

AI 行业正在从过去单纯关注“单卡性能”,逐渐转向更加现实的指标:

单位算力到底能创造多少价值?

企业真正关心的,已经不只是采购了多少 GPU,而是:

每一度电、每一秒 GPU 时间、每一美元基础设施投入,究竟能够生成多少有效 Token?

这正在成为 AI 基础设施竞争的核心。

腾讯云:让已经买回来的 GPU 干更多活

腾讯云 TencentOS 提供了一个典型案例。

通过精细化 qGPU 切分,企业能够节省 60% 以上的 GPU 硬件采购成本。

通过离线混部技术,服务器 CPU 利用率可以从 15% 提升到 45%,整体使用成本降低约 50%。

再配合 TI 平台的潮汐调度机制,集群整体算力利用率可以从约 30% 提升至 90%,相当于提升约三倍。

这意味着:

同样一批芯片,不一定需要继续购买更多硬件,也可以通过调度优化释放出新的产能。

北京数据集团:把算力变成“Token 工厂”

北京数据集团推出的**“京算 Token 工厂”**也体现了类似思路。

通过异构算力池化调度,资源利用率可以从行业普遍的 30%–40% 提升至 80% 以上。

采用 PD 分离智能调度架构后,算力吞吐量最高提升 90%,大模型场景下推理时延速度提升 70% 以上。

在长文本场景中,全栈推理加速与 PD 分离技术协同,还可以让生成 Token 的成本降低 50% 以上。

整个行业正在形成一个越来越清晰的趋势:

硬件侧优化,可以把集群利用率从原来的 30%–50% 推向 60%–70%;软件侧优化,则可以在保证模型精度和质量的前提下,让 Token 生成效率提升 30% 以上。

还有更加激进的案例。

无问芯穹已经实现 6 款国内外不同芯片的混合训练,算力利用率超过 97%。

在中昊芯英与中兴通讯的联合优化中,TPU 集群效率更实现了超过 10 倍的提升。

从 40% 到 65%。

从 30% 到 90%。

从 50% 到 97%。

这些数字背后其实说明了同一件事:

芯片买回来只是第一步,真正的竞争壁垒,是如何把每一颗芯片用到极致。


Token 交易市场:把“闲置算力”变成“流动算力”

算力利用率低,本质上是一个调度问题。

大量算力被固定在特定项目、特定时间和特定集群里。

当需求发生变化时,这些资源却很难迅速流向真正需要它们的地方。

一体化算力网络正在改变这一点。

通过跨中心调度、削峰填谷、任务排队以及统一资源管理,可以进一步提高整个算力网络的利用效率。

而 Token 的出现,又提供了一个新的可能:

用统一的词元计量方式,对不同规格 GPU 所提供的计算能力进行标准化描述和结算。

这意味着算力正在发生角色变化。

过去,它主要是一堆固定的硬件资产。

未来,它可以成为一种可调度、可交易的数字化资源。

用户可以像使用电力一样使用算力:

需要多少就买多少,需要什么时候用就什么时候用,并根据实际使用量付费。

Token 经济真正重要的地方,也正在于此。

复杂的软件框架、算力调度和模型服务隐藏在基础设施内部,而对外则通过标准化 API 提供统一的服务接口。

上游数据服务商可以按照有效 Token 用量获得收益;

中游算力和模型供应商可以根据 Token 消耗进行计价;

下游 AI 应用则可以依据 Token 使用量核算投入产出。

统一的计量口径,可以进一步减少商务谈判、账单核对以及定制化适配所产生的成本。

Token 不只是一个计量单位,它正在尝试成为连接算力供给与 AI 需求的一种共同语言。


Unexhub 的算盘:把“算力利用率”变成“算力交易率”

当“算力精调”和“Token 交易”开始形成双轮驱动,Unexhub 的定位也变得更加清晰。

Unexhub 正在搭建一个 Token 聚合与分发的算力交易中心,核心思路是:

让算力从固定资产变成可以进入市场流通的资源。

第一层:实时结算

对于中小型算力供应商来说,资金周转往往是一个现实问题。

Unexhub 的实时结算机制,旨在让供应商提供的 Token 数量与收益结算更加同步,从而缩短资金回笼周期,降低长期垫资压力。

第二层:早期入驻零佣金

早期供应商零佣金并不仅仅是一项优惠。

它更像是一种市场流动性激励机制。

对于供应商来说,可以降低尝试新分销渠道的成本;

对于开发者和企业用户来说,更多供应商进入市场,则意味着更多算力选择、更充分的价格竞争。

最终形成的是一个正向循环:

供应商更多 → 供给增加 → 价格竞争增强 → 用户选择更多。

第三层:API 动态竞价

真正值得关注的是 API 竞价机制。

传统 API 往往采用固定价格。

高峰期拥堵时是一个价格,凌晨算力闲置时还是一个价格。

Unexhub 的思路则是让 API 价格根据市场供需动态变化。

需求旺盛时,价格上升,供应商获得更高收益;

资源闲置时,价格下降,开发者和企业则可以用更低成本获得算力。

这实际上是在算力市场中增加一套“价格信号灯”。

需求高 → 价格上涨 → 吸引更多供给。

需求低 → 价格下降 → 刺激更多使用。

这与单纯提升 GPU 利用率存在本质区别。

前者解决的是技术效率。

后者解决的是市场效率。


从“算力利用率”到“算力交易效率”

算力精调解决的是一个问题:

如何让每一颗芯片多干一些活?

Token 交易市场解决的则是另一个问题:

如何让闲置算力找到真正需要它的人?

一个是技术调度。

一个是市场调度。

两者结合,才构成更加完整的 AI 基础设施降本逻辑。

当国产集群把利用率从 40% 推向 65%,当 Token 交易市场让闲置算力能够在不同供应商和用户之间流动,AI 推理的成本结构也会随之发生变化。

单位 Token 成本的下降,不一定只来自某一家模型厂商的降价促销。

它还可能来自整个产业链的一次系统性转型:

从“堆更多硬件” → 到“提高硬件利用率” → 再到“让算力进入市场流通”。


65% 不是终点,97% 也不是

无论是 65%,还是 97%,都不应该成为最终目标。

真正的目标其实很简单:

让每一颗芯片,都用在最有价值的地方。

让每一度电,都尽可能产生更多有效智能。

让每一个 Token,都对应更加透明、可衡量的计算价值。

这也是 AI 基础设施正在发生的真正变化。

Unexhub 正围绕这一趋势搭建全球 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,同时也接入个人开发者和中小企业提供的优质 AI 模型。

对于企业和开发者而言,它提供的价值并不只是“模型更多”。

更重要的是:

选择更多、价格更具竞争力,以及更快速的算力路由。

根据 Unexhub 的平台定位,开发者可以按照实际 Token 使用量付费,符合条件的服务最低可享受市场价格约 3 折的成本优势;供应商则可以通过统一渠道快速接入,并按照实际调用获得实时结算与分成。

换句话说:

真正的降本,不只是买更便宜的芯片,也不是简单等待模型厂商降价,而是让已经存在的算力被更充分地利用,让闲置资源更容易进入市场。

从 40% 到 65%,再到 97%。

这场关于 AI 算力效率的竞争,真正比拼的已经不是“谁拥有更多芯片”,而是:

谁能让每一颗芯片产生更多价值。