GPU 排队一年、芯片占推理成本 6-7 成!Token 交易市场如何破解"算力主锚"困局?

让 GPU 背后的算力更容易流动,让 Token 更容易交易,让价格更加灵活,也让每一份已经存在的算力得到更充分的利用。


One H100 GPU, Eighteen Months Later

The 2026 AI computing market is beginning to look almost absurd.

NVIDIA's H100 inventory is effectively exhausted, with the reported shortage reaching roughly 430,000 units and delivery lead times stretching to as long as 18 months. Capacity for the next-generation Blackwell family has already been heavily committed by major customers such as Google and Meta.

The economics are moving just as quickly.

A one-year rental contract for an H100 SXM5 reportedly climbed from $1.70 per GPU-hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%. Even more striking, B200 rental prices reportedly rose from $2.63 to $5.10 when contracts were renewed—an increase of roughly 94%.

For customers ordering 1,000 GPUs, delivery schedules are increasingly being pushed into Q2 2027, creating 12-to-15-month waiting periods.

This isn't simply a temporary market fluctuation.

It reflects a deeper computing-capacity bottleneck that could reshape the economics of AI.

And there is an especially ironic part:

The GPU you're waiting 18 months to receive may also be the single largest component of your AI inference cost structure.


The Cost Anchor: Why GPUs Still Dominate the Inference Bill

Let's start with the economics.

A 1 GW data center can have a total four-year cost of ownership approaching $55 billion under some infrastructure assumptions.

GPU procurement can account for roughly $25 billion, or about 45% of the total. Power and cooling may represent another $11 billion, networking around $5 billion, and storage approximately $4 billion.

Electricity?

Around $2.75 billion—only about 5% of the total in this example.

In other words, cheap electricity isn't necessarily the factor that determines the economics.

The much bigger questions are:

Which GPUs are you using? How many do you need? And how efficiently can you connect and operate them?

At the same time, inference is becoming the dominant workload.

Yang Yuanqing, Lenovo's chairman and CEO, has described the structural shift this way: AI computing demand previously leaned heavily toward training, at roughly 80% training and 20% inference. Today, the ratio is closer to 50/50, while the longer-term structure could move toward 80% inference and 20% training.

That changes everything.

Training is a large upfront investment.

Inference is a continuous operating expense.

Every user request, agent action, API call, and generated token keeps consuming computing resources.

That makes the GPU infrastructure layer a critical cost anchor for the entire AI inference economy.

You can optimize algorithms, quantize models, compress weights, and improve distillation.

Those techniques can reduce marginal costs.

But GPU procurement, depreciation, and cluster construction are expenses that often appear on the balance sheet before the first production request is ever processed.

And that anchor is becoming increasingly expensive.


The Scarcity Multiplier: How GPU Queues Distort AI Economics

An 18-month GPU wait isn't merely inconvenient.

It can distort the economics of an AI project in at least three ways.

1. Scarcity Premiums Eat Into Margins

High-end GPUs such as H100 and H200 can trade at substantial premiums over official pricing.

Cloud providers with utilization rates approaching 85%-98% have limited spare capacity, while rental commitments may require customers to sign long-term contracts.

If you can't wait for capacity, you pay for access to someone else's capacity.

Speed becomes a premium product.

2. Vendor Concentration Reduces Negotiating Power

NVIDIA remains overwhelmingly dominant in the global AI accelerator market.

Training workloads prioritize compute density.

Inference workloads increasingly prioritize throughput, latency, and energy efficiency.

But regardless of the workload, the number of truly interchangeable alternatives remains limited.

When customers cannot easily switch suppliers, pricing power naturally concentrates on the supply side.

3. Idle Capacity Creates Invisible Costs

Buying GPUs doesn't mean every GPU-hour will be productive.

Underutilized clusters still generate depreciation, infrastructure, cooling, networking, and management costs.

This creates a particularly frustrating situation:

You can simultaneously have a shortage of AI computing capacity and a large amount of idle capacity.

The problem isn't always the total number of GPUs.

It is where those GPUs are, who owns them, and when they are available.


Token Markets: Loosening the Grip of the "Cost Anchor"

The deeper problem isn't simply that GPUs are expensive.

It's that pricing power is concentrated.

If you have limited suppliers, limited alternatives, and limited flexibility, you have limited negotiating power.

This is where token markets become interesting.

China's National Data Administration has formally referred to Token as "词元" (token) and positioned it as a settlement unit connecting technology supply with commercial demand.

Meanwhile, China's daily token consumption has reportedly grown from roughly 100 billion in early 2024 to 140 trillion by March 2026.

The scale of token consumption is transforming the economics of AI infrastructure.

Unexhub is pursuing a more market-oriented approach by building a token aggregation, distribution, and computing-power trading center.

The central idea is straightforward:

Instead of requiring every developer to own or reserve physical computing infrastructure, make computing capacity accessible as a tradable digital resource.

That changes the economic model.

Real-time settlement can help small and medium-sized AI suppliers recover revenue based on actual token usage rather than waiting through long payment cycles.

Early suppliers can receive 0% commission incentives, lowering the cost of entering the distribution ecosystem and encouraging additional computing supply to enter the marketplace.

And through API bidding, pricing can respond dynamically to market supply and demand.

When demand is high, suppliers can capture higher returns.

When capacity is abundant, developers can potentially access lower prices.

The objective isn't to eliminate the GPU as the physical foundation of computing.

It is to separate access to computing capacity from ownership of the hardware.


The Cost Anchor Is Starting to Loosen

If inference eventually accounts for more than 70% of AI computing demand, the economics of AI infrastructure will change dramatically.

Training is largely an upfront investment.

Inference is continuous.

Once an AI system operates around the clock, token consumption becomes an increasingly important part of the marginal cost structure.

That means the market doesn't necessarily need to eliminate the GPU cost anchor.

It needs to make that anchor less restrictive.

This is the deeper promise of a token-based computing marketplace:

Turn computing from an asset that must be reserved months or years in advance into a resource that can increasingly be purchased according to actual demand.

You don't necessarily need to wait 18 months for a GPU if the market can provide access to inference capacity today.

The underlying hardware scarcity still exists.

But its cost can potentially be distributed across a broader marketplace rather than imposed directly on every individual developer.


Can a Token Market Really Solve the GPU Bottleneck?

That's the real question.

The answer depends on one crucial variable:

Can the market move faster than the hardware supply chain?

If GPU manufacturing remains constrained, token trading cannot manufacture additional chips.

But it can potentially make better use of the computing capacity that already exists.

That means the next stage of AI infrastructure may not be about owning more GPUs.

It may be about allocating the GPUs that already exist more efficiently.

And that's where the combination of model routing, token pricing, and computing-power marketplaces becomes increasingly important.


Unexhub: Turning Computing Capacity Into a Tradable Resource

Unexhub is positioning itself as a global token aggregation and distribution platform and computing-power trading center, bringing together AI suppliers of different scales.

Its ecosystem is designed to include major models such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, while also connecting AI services supplied by individual developers and small and medium-sized companies.

The platform's value proposition is built around several practical benefits:

  • Multi-model aggregation — access multiple AI providers through a unified ecosystem

  • Millisecond-level routing — dynamically route requests to appropriate model resources

  • Competitive market pricing — allow supplier-side competition to influence prices

  • Pay-by-token usage — pay according to actual consumption

  • 0% commission incentives for eligible early suppliers — lower the barrier to joining the distribution network

  • Real-time settlement — help suppliers receive revenue from usage more quickly

For developers, the most tangible benefit is cost flexibility.

Under applicable market conditions, Unexhub says developers can access eligible services at as little as 30% of prevailing market prices.

The value isn't simply "cheap AI."

It's the ability to compare multiple models and suppliers, choose the right capability for a specific workload, and avoid committing large amounts of capital to infrastructure that may sit idle.

For suppliers, the equation is equally straightforward:

Integrate once, gain access to additional distribution channels, and receive usage-based revenue through real-time settlement.

The ultimate goal is not to replace GPUs.

It is to make the computing power behind those GPUs more liquid, more accessible, and more efficiently allocated.


中文版|中文版从这里开始

一块 H100,等一年半

2026 年的 AI 算力市场,正在上演一出越来越荒诞的现实剧。

英伟达 H100 的库存基本耗尽,市场报告显示,单款 H100 的缺口一度达到 43 万张,交付周期最长甚至拉到 18 个月。

下一代 Blackwell 系列的产能,则早已被 Google、Meta 等大型客户大量锁定。

价格同样在快速变化。

H100 SXM5 的一年租赁合同价格,据报道从 2025 年 10 月的每卡每小时 1.70 美元,上涨至 2026 年 3 月的 2.35 美元,涨幅接近 40%。

更夸张的是,B200 GPU 在合同续约时,租赁价格据报道从 2.63 美元上涨到 5.10 美元,涨幅约 94%。

如果一次采购 1000 张 GPU,交付周期甚至可能被推迟到 2027 年第二季度。

换句话说:

等一批 GPU,可能需要 12~15 个月。

这已经不只是普通的市场波动。

它正在演变成一场影响 AI 成本结构的算力供应危机。

而最讽刺的地方在于:

你等了一年半才拿到的那块 GPU,很可能恰恰就是整个 AI 推理成本结构里最重要的一块。


算力主锚:为什么 GPU 仍然是推理账单的"定海神针"?

先算一笔基础设施账。

按照部分基础设施测算,一座 1GW 数据中心四年的总拥有成本可能接近 550 亿美元。

其中 GPU 采购约 250 亿美元,占比约 45%;供电与散热约 110 亿美元;网络约 50 亿美元;存储约 40 亿美元。

电费是多少?

约 27.5 亿美元,仅占约 5%。

这意味着,在整个 AI 数据中心的经济模型里,真正决定成本结构的,并不只是电价。

更重要的是:

你用什么 GPU、需要多少 GPU,以及这些 GPU 能否被高效组织成集群。

与此同时,AI 推理正在成为越来越重要的算力需求来源。

联想集团董事长兼 CEO 杨元庆曾对 AI 算力结构变化做出判断:此前 AI 算力投入大约是 80% 用于训练、20% 用于推理;目前已经接近 50% 对 50%;未来合理的结构甚至可能变成:

80% 用于推理,20% 用于训练。

这意味着 AI 产业的成本重心正在发生变化。

训练是一笔巨大的前期投入。

推理则是一笔持续发生的运营成本。

用户每发起一次请求、Agent 每执行一步、API 每调用一次,都会继续消耗算力。

因此,GPU 基础设施正在成为 AI 推理经济中的一个关键成本主锚。

算法优化、模型量化、压缩和蒸馏可以降低边际成本。

但 GPU 采购、折旧以及集群建设费用,却往往在项目上线之前就已经写进预算。

而现在,这个"主锚"还在继续收紧。


稀缺放大器:GPU 排队如何扭曲 AI 成本结构?

GPU 排队一年半,并不只是"时间太长"。

它至少通过三种方式进一步放大 AI 成本压力。

第一,稀缺溢价吞噬利润

H100、H200 等高端 GPU 的市场价格可能显著高于官方价格。

部分云厂商的 GPU 利用率已经非常高,剩余容量有限,同时租赁往往还要求较长的合约周期。

等不起怎么办?

只能花更多钱购买别人已经拥有的算力。

时间,本身开始变成一种溢价商品。

第二,供应商集中降低议价空间

英伟达仍然在全球 AI 加速器市场占据绝对主导地位。

训练看重算力密度。

推理越来越看重吞吐、延迟与能效。

但无论是哪一种场景,真正能够大规模替代的方案仍然有限。

当用户无法轻易切换供应商时:

定价权自然会向供给侧集中。

第三,闲置算力制造隐性成本

买 GPU 并不意味着每一张 GPU 都会持续高效运行。

即使 GPU 没有满负荷工作,折旧、机房、散热、网络和管理成本依然存在。

于是出现了一个非常矛盾的现象:

市场一边喊"算力不够",另一边却存在大量没有被充分利用的算力。

真正的问题有时并不是全球 GPU 总量不足。

而是:

算力在哪里、属于谁、什么时候可以使用。


Token 交易市场:让"主锚"不再锁死所有人

所以问题的本质并不只是"GPU 太贵"。

真正的问题是:

算力定价权过于集中。

供应商有限、替代方案有限、调度空间有限,自然意味着用户的议价能力有限。

这也是 Token 交易市场值得关注的原因。

国家数据局已经正式将 Token 命名为"词元",并将其定位为连接技术供给与商业需求的重要结算单位。

与此同时,中国日均词元调用量已经从 2024 年初约 1000 亿,增长到 2026 年 3 月的 140 万亿。

Token 消耗规模正在快速扩大。

而 Unexhub 正在尝试进一步推进一个市场化方向:

建立 Token 聚合、分发与算力交易中心。

核心逻辑是:

不再要求每一个开发者都必须拥有或者提前锁定实体 GPU,而是让算力以可交易数字资源的形式进入市场。

这会改变算力的使用方式。

实时结算可以让中小 AI 供应商按照实际 Token 调用量更快获得收入,降低传统长账期带来的资金压力。

对于符合条件的早期供应商,0 佣金入驻权益能够降低进入分销生态的成本,同时帮助平台吸引更多算力供应。

而 API 竞价机制,则让价格能够根据市场供需动态变化。

需求旺盛时,供应商获得更高收益。

资源充足时,开发者有机会获得更低价格。

这并不是要消灭 GPU。

GPU 仍然是算力的物理基础。

真正要改变的是:

让"拥有 GPU"与"使用算力"不再完全绑定。


算力主锚的松动,才刚刚开始

如果未来 AI 推理占据超过 70% 的算力需求,那么整个 AI 基础设施的成本结构都将发生巨大变化。

训练是一笔前期投资。

推理则是持续发生的运营成本。

当一个 AI 系统 24 小时运行,每秒都在消耗 Token 时,Token 所代表的边际成本会越来越重要。

因此,Token 交易市场并不是要消灭 GPU 这个"主锚"。

而是试图给这个主锚增加一个松动装置:

把原本需要提前数月甚至一年半购买的固定算力,逐渐转化成可以按照实际需求进行交易的流动资源。

如果市场今天就能够提供推理算力,那么开发者就不一定必须等 18 个月才能获得 AI 能力。

GPU 的稀缺依然存在。

但稀缺产生的成本,可以通过更大的市场进行分摊,而不是让每一个开发者独自承担。


Token 交易市场真的能解决 GPU 瓶颈吗?

这才是最值得追问的问题。

答案取决于一个关键变量:

市场运行速度,能不能跑赢硬件供应链?

如果 GPU 生产依然受到限制,Token 交易市场当然不可能凭空制造更多芯片。

但它可以尝试让已经存在的算力被更高效地利用。

于是 AI 基础设施的下一阶段,可能不再只是:

"谁拥有更多 GPU?"

而是:

"谁能让已经存在的 GPU 被更高效地分配?"

这也是模型路由、Token 定价以及算力交易市场越来越重要的原因。


Unexhub:让 AI 算力变成可以交易的资源

Unexhub 正在构建面向全球的 Token 聚合与分发平台及算力交易中心,连接不同规模的 AI 供应商。

平台生态覆盖 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型,同时也吸纳个人开发者、中小企业提供的 AI 服务。

其核心价值可以概括为几个方面:

  • 多模型聚合:减少开发者分别对接多个 AI 服务商的成本

  • 毫秒级路由:根据模型与资源情况快速分配请求

  • 市场化价格竞争:让供应商之间形成更灵活的价格机制

  • 按 Token 实际付费:根据真实调用量控制成本

  • 早期供应商 0 佣金权益:降低接入分销生态的门槛

  • 实时结算:帮助供应商更快获得调用分成

对于开发者来说,最直接的价值是更大的成本弹性。

在具体市场条件和相应服务范围内,Unexhub 的部分服务价格最低可达到市场价格的 3 折。

真正值得关注的并不只是"便宜"两个字。

而是开发者可以在一个市场里比较不同模型、不同供应商以及不同价格,在满足任务效果的前提下,尽可能避免为闲置算力和不必要的模型能力买单。

对于供应商而言,则可以通过一次接入获得更多分发渠道,并根据实际调用量进行实时结算。

最终,Unexhub 并不是要取代 GPU。

它真正想做的是:

让 GPU 背后的算力更容易流动,让 Token 更容易交易,让价格更加灵活,也让每一份已经存在的算力得到更充分的利用。

一块 H100,可能还要等一年半。

但获取 AI 能力,未必还需要等那么久。