ChatGPT Agent 多步烧 Token 倒逼「Token交易市场」成型,OpenAI 拉谷歌云入供应商名单的真实算盘

As agents move from demonstrations into production, token consumption is emerging as one of the most difficult constraints on large-scale adoption.

When AI Agents Start Burning Tokens, the Rules of the Compute Market Change

In agentic coding benchmarks, GPT-5.6 Sol reportedly produces 54% fewer output tokens than comparable competing models while completing tasks in less than half the time.

It is easy to understand why OpenAI would put numbers like these at the center of its messaging. In the age of AI agents, intelligence alone is no longer enough. The amount of compute, time, and tokens required to complete a task is becoming just as important as the quality of the final answer.

As agents move from demonstrations into production, token consumption is emerging as one of the most difficult constraints on large-scale adoption.

The Token Black Hole Behind Multi-Step Agents

A conventional chatbot may answer one question with one model call. An AI agent behaves very differently.

It reads files, searches databases, invokes tools, checks results, corrects mistakes, requests additional context, and repeats the process until the task is complete. Every additional step consumes more input and output tokens.

In high-demand scenarios, the annual token cost of a single software-engineering agent can reach $3,833. A legal contract agent may cost approximately $2,409 per year, while a customer-service agent may consume around $2,190. Even a relatively lightweight HR onboarding agent can generate annual token expenses of roughly $1,095.

The absolute numbers are already significant. The growth rate is even more striking.

China’s average daily token consumption reportedly increased from approximately 100 billion tokens in early 2024 to more than 140 trillion by March 2026. In just over two years, usage expanded by more than 1,000 times.

The rapid rise of AI agents is one of the major forces behind this nonlinear growth.

Model prices continue to fall, but lower unit prices do not automatically produce lower total costs. When companies deploy agents across coding, customer support, research, sales, legal review, and internal operations, the increase in token consumption can easily outpace price reductions.

This is why some SaaS providers have begun to face negative gross margins on services designed for their heaviest AI users.

The more customers use the service, the more money the provider can lose.

GPT-5.6 Ultra mode illustrates the trade-off clearly. It can coordinate four agents in parallel by default, exchanging greater token consumption for stronger task quality and higher completion rates. GPT-5.6 Terra may cost $2.50 per million input tokens and $15 per million output tokens, but once multiple agents begin planning, executing, reviewing, and retrying together, token usage can grow surprisingly quickly.

Token is therefore no longer just a technical unit on an API invoice.

It is becoming a new accounting layer through which enterprises allocate software budgets, labor costs, automation capacity, and workflow access.

OpenAI Turns to Google Cloud: Not Choosing Sides, but Securing Supply

On April 27, 2026, OpenAI and Microsoft announced revisions to their partnership, shifting Microsoft’s previously exclusive rights toward a non-exclusive arrangement. GPT models could subsequently be distributed through additional cloud platforms, including AWS and Google Cloud.

This was more than a routine commercial adjustment.

Google Cloud also appeared on OpenAI’s list of subprocessors, signaling that the relationship extended into the infrastructure layer.

The deeper reason lies in compute supply.

As early as June 2025, reports suggested that OpenAI was considering renting chips through Google Cloud to support part of the inference workload behind products such as ChatGPT. One attraction was the opportunity to use Google’s TPU infrastructure rather than relying entirely on high-cost GPU capacity such as NVIDIA H100 and B200 clusters.

The objective was not simply to replace one chip provider with another. It was to reduce dependency on a single hardware and cloud supply chain.

OpenAI’s demand for compute has grown to a scale that no single provider can easily satisfy. In February 2026, Amazon reportedly agreed to invest up to $50 billion in OpenAI, with AWS becoming the exclusive third-party cloud distribution provider for OpenAI’s Frontier enterprise platform. At the same time, OpenAI continued expanding the availability of its models through Google Cloud.

This is not a matter of choosing between Microsoft, Amazon, and Google.

It is a matter of securing capacity wherever sufficient capacity exists.

The appetite of large-scale AI inference is pushing model companies away from single-source procurement and toward distributed, multi-cloud compute supply.

That shift is also creating the foundation for a much larger market: the trading and distribution of tokenized compute.

The Token Market Could Become the Crude-Oil Market of AI

Tokens now span text, images, audio, video, and multimodal reasoning. They are increasingly becoming the basic unit used to measure the consumption and output of AI services.

Oil is priced by the barrel. Electricity is settled by the kilowatt-hour. AI inference is gradually being measured by the token.

At GTC 2026, Jensen Huang introduced the idea of the “token factory.” In this vision, the data center of the future is no longer merely a warehouse that stores files. It is a factory that continuously converts electricity, chips, data, and models into intelligent output.

The competitive advantage of such a factory is determined by three things:

How many useful tokens it can produce, how quickly it can produce them, and how much each token costs.

But production is only the first step.

Once tokens are produced, how should they be distributed? How should their value be calculated? How should buyers and sellers discover the right price?

Today’s API market still relies primarily on static price lists. OpenAI publishes one price, Anthropic publishes another, and Google publishes its own. Each ecosystem operates independently, with limited liquidity between providers.

When demand rises, the published price often remains unchanged. When servers are underutilized, idle capacity cannot easily be offered to buyers elsewhere. A token generated during a quiet period is usually priced in the same way as a token generated during peak demand.

That is not yet a mature market. It is closer to a collection of disconnected spreadsheets.

What the industry ultimately needs is a token-matching marketplace: a system in which compute can be priced competitively like equities, dispatched dynamically like electricity, and exchanged across providers with the flexibility of foreign currencies.

Turning Compute From a Fixed Asset Into a Liquid Asset

OpenAI’s use of Google Cloud is only an early signal.

AWS, Microsoft Azure, and Google Cloud are all accumulating enormous amounts of compute capacity. Yet much of that capacity remains enclosed within separate platforms, pricing systems, interfaces, and model ecosystems.

A company purchasing AWS capacity generally consumes it within the AWS environment. Google TPU resources are primarily delivered through Google’s own infrastructure and inference services. Developers must maintain separate accounts, integrations, billing systems, and operational processes.

Such fragmentation creates waste.

One provider may have idle capacity while developers elsewhere face queues, rate limits, or peak-hour prices. Smaller model suppliers may possess valuable inference capacity but lack enough distribution channels to keep it fully utilized.

A functioning token marketplace could connect these two sides directly.

When compute is underutilized, suppliers could release additional token capacity into the market. When demand is low, developers could purchase that capacity at a lower price. During peak periods, prices could rise, encouraging more suppliers to bring capacity online.

This would transform compute from a fixed asset that sits inside one data center into a liquid asset that can be priced, routed, settled, and distributed according to real-time demand.

Dynamic API bidding would create an even more efficient mechanism.

Prices could move with supply, demand, latency requirements, model quality, regional availability, and service-level guarantees. Suppliers would gain higher returns during periods of strong demand, while developers could schedule flexible workloads during off-peak hours to reduce costs.

Instead of one company deciding the price for the entire ecosystem, the market itself would participate in resource allocation.

Tokens Are Becoming an Enterprise Profit Unit

Jensen Huang has described multiple commercial layers for AI services, ranging from free access to premium products priced as high as $150 per million tokens.

The important point is not the exact number. It is the changing role of the token.

Tokens are becoming measurable units of production, consumption, cost, and revenue. Enterprises may soon allocate token budgets to AI agents in much the same way that they allocate salaries to employees or cloud budgets to engineering teams.

A coding agent may receive a monthly token allowance. A customer-service agent may be evaluated according to the token cost per resolved ticket. A research agent may be measured by the cost of producing each verified report.

Once these measurements become standardized, businesses can compare models, suppliers, agents, and workflows on a common economic basis.

That is why a standardized token market could become as important to the AI economy as standardized electricity markets were to industrial development.

Without a reliable electricity market, factories could not purchase power efficiently at scale. Without a liquid and standardized token market, AI agents risk remaining expensive experiments that become harder to sustain as usage grows.

The next stage of the compute economy begins when tokens can be routed freely, prices can respond to market conditions, and suppliers can transact directly with developers and enterprises.

Unexhub: Building an Aggregation, Distribution, and Settlement Layer for Tokens

The industry does not necessarily lack compute. What it lacks is a market mechanism capable of making that compute flow efficiently.

Unexhub is being developed around this missing layer: a compute marketplace for aggregating and distributing global token supply.

The platform brings together AI resources across different scales, including supply connected to mainstream model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as models and inference capacity provided by individual developers, specialized teams, and small and medium-sized AI companies.

Instead of requiring developers to integrate and manage every provider separately, Unexhub aggregates available model APIs behind a unified access and distribution layer.

Its intelligent routing system can evaluate price, model availability, latency, and task requirements in real time, then direct each request toward an appropriate supplier. Millisecond-level routing helps reduce unnecessary waiting and allows applications to switch between available resources more efficiently.

For developers and enterprise users, the benefits are practical:

They can pay according to actual token consumption rather than committing to large blocks of unused capacity. Depending on the model, workload, and available supply, selected routes may cost as little as 30% of the prevailing market price. Flexible workloads can also take advantage of lower-cost capacity during off-peak periods.

For suppliers, Unexhub lowers the barrier to distribution.

A supplier can connect its model or inference capacity through a unified channel, gain access to additional developer and enterprise demand, and receive usage-based revenue settlement. Real-time settlement of API revenue can shorten payment cycles and reduce the working-capital pressure commonly faced by smaller providers.

The result is not simply another API directory.

Unexhub is working toward an integrated marketplace that combines token aggregation, intelligent routing, competitive pricing, distribution, metering, and financial settlement.

Developers gain access to more models and more transparent prices. Suppliers gain additional traffic, higher utilization, and faster revenue recovery. Enterprises gain a unified way to control token spending across multiple models and workloads.

As agentic applications continue to consume tokens at an accelerating rate, these capabilities will become increasingly valuable.

OpenAI’s expansion across Google Cloud, AWS, and other infrastructure providers shows that even the world’s largest AI companies need diversified compute supply. The next step is to make that supply liquid, measurable, and tradable.

Compute itself is not necessarily scarce.

What remains scarce is an efficient market capable of moving it to the right model, the right application, and the right user at the right moment.

That market is now beginning to take shape.


当智能体开始疯狂消耗 Token,整个算力市场的交易逻辑都将被改写

在智能体编程测试中,GPT-5.6 Sol 相较同类竞品,据称能够减少 54% 的输出 Token,同时将任务耗时缩短一半以上。

OpenAI 将这样的数据放在宣传核心位置,并不难理解。在 AI 智能体时代,模型是否足够聪明已经不再是唯一标准。完成同一个任务需要消耗多少 Token、占用多少算力、花费多长时间,正在变得与最终结果同样重要。

当智能体从演示项目进入真实业务,Token 消耗正成为规模化落地过程中最难绕开的瓶颈之一。

多步骤智能体背后的 Token 黑洞

传统聊天机器人可能只需要调用一次模型,就能完成一次问答。

智能体却完全不同。

它需要读取文件、检索数据库、调用工具、检查执行结果、修正错误、补充上下文,并不断重复这一过程,直到任务真正完成。每增加一个步骤,都会带来新的输入和输出 Token 消耗。

在高需求场景下,一个软件工程智能体的年 Token 成本可达到 3,833 美元;法律合同智能体约为 2,409 美元,客服智能体约为 2,190 美元;即使是相对轻量的 HR 入职智能体,年 Token 成本也可能达到 1,095 美元。

这些绝对数字已经不低,更值得警惕的是增长速度。

国内 Token 日均调用量据称已经从 2024 年初的约千亿,增长至 2026 年 3 月的 140 万亿以上。短短两年多,使用量增长超过 1,000 倍。

AI 智能体应用的快速普及,正是推动这种非线性增长的重要力量。

模型单价确实在不断下降,但单价下降并不等于企业 AI 总成本必然下降。当企业同时在编程、客服、研究、销售、法务和内部运营中部署大量智能体时,Token 用量的增长完全可能超过模型降价的幅度。

这也是为什么部分 SaaS 厂商在服务重度 AI 用户时,已经面临毛利率下降甚至转负的问题。

用户用得越多,平台反而可能亏得越多。

GPT-5.6 Ultra 模式很好地体现了这种取舍。它默认可以并行调度 4 个智能体,通过更多的 Token 消耗换取更高的任务质量和完成率。GPT-5.6 Terra 的输入价格为每百万 Token 2.5 美元,输出价格为每百万 Token 15 美元,但一旦多个智能体开始共同规划、执行、复核和重试,Token 消耗依然会迅速增加。

因此,Token 早已不只是 API 账单上的一个技术计量单位。

它正在成为企业重新分配软件预算、人力成本、自动化能力和工作流权限的一本全新成本账。

OpenAI 引入谷歌云:不是选边站,而是寻找更多“粮仓”

2026 年 4 月 27 日,OpenAI 与微软宣布调整合作协议,将微软此前的部分独家授权转向非独家模式。此后,GPT 系列模型可以通过 AWS、谷歌云等更多云平台进行分发。

这并不是一次普通的商业合作调整。

谷歌云同时出现在 OpenAI 的次级处理器供应商名单中,说明双方的合作已经延伸至底层基础设施。

更深层的原因,隐藏在算力供给之中。

早在 2025 年 6 月,就有报道称 OpenAI 正计划通过谷歌云租用芯片,为 ChatGPT 等产品承担部分推理任务。谷歌云的吸引力之一,在于其 TPU 基础设施能够帮助 OpenAI 降低对英伟达 H100、B200 等高成本 GPU 集群的单一依赖。

这并不是简单地用一种芯片替换另一种芯片,而是降低对单一硬件和单一云供应链的依赖。

OpenAI 对算力的需求已经增长到单一供应商难以完全承接的规模。2026 年 2 月,亚马逊据称同意向 OpenAI 注资最多 500 亿美元,并由 AWS 担任 OpenAI 企业平台 Frontier 的独家第三方云分发供应商。与此同时,OpenAI 也继续推动模型在谷歌云等平台上的部署和分发。

这不是在微软、亚马逊和谷歌之间选边站。

而是哪里有足够的算力,就去哪里寻找供应。

大模型推理对算力的旺盛需求,正在推动模型公司从单一采购转向多云、多来源、分布式的算力供应体系。

而这场变化,也正在为一个更大的市场奠定基础:Token 化算力的聚合、交易与分发。

Token 交易市场,可能成为 AI 时代的“原油市场”

Token 已经横跨文本、图像、语音、视频和多模态推理,逐渐成为衡量 AI 服务消耗与产出的基础单位。

原油按桶计价,电力按千瓦时结算,AI 推理则越来越多地按 Token 计量。

黄仁勋在 2026 年 GTC 大会上提出了“Token 工厂”的概念。在这一构想中,未来的数据中心不再只是存储文件的仓库,而是持续将电力、芯片、数据和模型转化为智能输出的工厂。

一家 Token 工厂是否具备竞争力,主要取决于三个问题:

能够生产多少有价值的 Token、生产速度有多快,以及每个 Token 的成本有多低。

但生产出来之后呢?

Token 应该如何流通?如何定价?买卖双方又该如何发现合理价格?

目前的 API 市场仍然以静态价目表为主。OpenAI 制定自己的价格,Anthropic 制定自己的价格,谷歌也有独立的计费体系。不同平台之间缺少真正的流动性。

需求上涨时,公开价格未必会及时变化;服务器闲置时,多余算力也无法轻易出售给其他平台的开发者。闲置时段生产的 Token,往往与高峰时段生产的 Token 使用同一套定价。

这还算不上一个成熟市场,更像是一组彼此隔离的 Excel 价目表。

行业真正需要的,是一个 Token 撮合交易市场:让算力可以像股票一样竞价、像电力一样动态调度,并像外汇一样在不同供应商之间灵活流通。

将算力从“固定资产”变成“流动资产”

OpenAI 引入谷歌云,只是一个开始。

AWS、微软 Azure 和谷歌云都在持续积累庞大的算力资源,但这些资源仍然被封闭在不同的平台、接口、价格体系和模型生态中。

企业购买 AWS 的算力,通常需要在 AWS 生态内使用;谷歌 TPU 主要通过谷歌自己的基础设施和推理服务提供。开发者则需要维护不同的账号、接口、账单和运维流程。

这种碎片化会带来大量资源浪费。

某个供应商可能存在闲置算力,而另一端的开发者却正在面对排队、限流和高峰价格。许多中小模型供应商拥有质量不错的模型与推理资源,却因为缺少分发渠道,难以获得持续稳定的调用量。

真正的 Token 交易市场,可以直接连接这两端。

算力闲置时,供应商可以释放额外的 Token 产能;需求处于低谷时,开发者可以用更低的价格购买这些资源;需求高峰到来时,价格适度上涨,吸引更多供应商释放算力。

算力由此不再只是锁定在某个数据中心里的固定资产,而是可以根据实时需求完成定价、路由、结算和分发的流动资产。

进一步引入 API 竞价机制后,资源配置效率还可以继续提升。

API 价格可以根据市场供需、响应延迟、模型质量、区域资源和服务等级动态变化。繁忙时段,供应商能够获得更高收益;闲置时段,开发者可以用更低成本运行弹性任务。

价格不再完全由某一家平台单方面决定,而是由真实供需共同形成。

Token 正在成为企业的利润核算单位

黄仁勋曾将 AI 服务划分为多个商业层级,从免费服务到每百万 Token 150 美元的高价值产品。

真正重要的并不是某一个具体价格,而是 Token 的角色正在发生变化。

Token 正在同时成为生产单位、消耗单位、成本单位和收入单位。未来,企业可能会像给员工发放工资、给工程团队分配云预算一样,为不同智能体配置 Token 预算。

编程智能体可以拥有每月 Token 配额;客服智能体可以按照“每解决一个工单消耗多少 Token”进行考核;研究智能体则可以按照“生成一份经过验证的报告需要多少成本”进行评估。

一旦计量标准逐渐统一,企业便可以在同一套经济体系下比较不同模型、供应商、智能体和工作流。

这正是标准化 Token 市场可能产生的价值。

电力市场为工业化提供了稳定、可计量、可交易的能源供应;Token 市场则可能为智能体经济提供稳定、透明、可流通的智能算力供应。

如果缺少统一而高效的 Token 市场,AI 智能体就很容易停留在“展示效果很好、规模化成本失控”的阶段。

只有当 Token 可以灵活路由、价格能够随市场变化、供应商可以与开发者和企业直接撮合时,算力经济的下一阶段才会真正开启。

Unexhub:构建 Token 聚合、分发与结算基础设施

AI 行业未必真的缺少算力。

更稀缺的,是一个能够让算力高效流动起来的市场机制。

Unexhub 正围绕这一缺失环节,打造面向全球市场的 Token 聚合与分发型算力交易中心。

平台汇聚不同规模的 AI 供应资源,包括覆盖 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型生态的服务能力,也包括由个人开发者、专业团队以及中小型 AI 企业提供的优质模型和推理算力。

开发者不必再分别对接和维护每一家模型供应商。Unexhub 将分散的模型 API 聚合至统一的访问与分发体系中,让用户能够通过一个入口调用多种模型资源。

平台的智能路由系统可以根据实时价格、模型状态、响应速度和任务需求,为每一次请求匹配更合适的供应渠道。毫秒级路由能够减少不必要的等待,并在部分资源出现拥堵或波动时,更灵活地切换可用算力。

对于开发者和企业用户而言,最直接的价值是更加灵活、透明的成本控制。

用户可以按照实际消耗的 Token 付费,无需提前购买大量可能无法充分使用的固定算力。在不同模型、任务类型和实时供应条件下,部分调用线路的价格最低可至市场常规价格的 3 折。对于不要求立即完成的弹性任务,用户还可以优先匹配闲时资源,以更低成本完成批量推理。

对于模型和算力供应商而言,Unexhub 则降低了进入市场和获取流量的门槛。

供应商只需通过统一接口接入模型或推理资源,即可触达更多开发者和企业用户,并根据实际调用量获得收益。API 调用分成支持实时结算,有助于缩短传统业务中的回款周期,缓解中小供应商的现金流和运营压力。

因此,Unexhub 并不只是一个模型 API 列表。

它所构建的是一个结合 Token 聚合、智能路由、竞价分发、统一计量与金融结算的完整交易体系。

开发者能够获得更多模型选择和更透明的价格;供应商能够获得额外流量、更高的算力利用率和更快的资金回收;企业则能够通过统一入口管理不同模型和业务场景下的 Token 支出。

随着智能体应用继续加速消耗 Token,这些能力将变得越来越重要。

OpenAI 同时连接谷歌云、AWS 等多个基础设施供应商,已经说明即使是全球最大的 AI 公司,也需要更加多元的算力来源。

接下来的关键,是让这些算力变得可计量、可流通、可交易。

算力本身未必稀缺。

真正稀缺的,是一个能够在正确的时间,将正确的算力分配给正确模型、正确应用和正确用户的高效市场。

而这个市场,正在逐渐成型。