# The Rise of Inference Capital Markets
**作者**: 0xJeff
**日期**: 2026-07-21T12:32:11.000Z
**来源**: [https://x.com/0xJeff/status/2079929079684587603](https://x.com/0xJeff/status/2079929079684587603)
---
The demand for intelligence is swallowing the industry whole. AI providers don't have enough compute to service the demand, leaving the gap for Inference providers to step in and service those demand.
> **0xJeff@0xJeff**: [原文链接](https://x.com/0xJeff/status/2079544969065545892)
>
> Man... AI is cheap but it's so expensive
>
> Cost to get started (inference per 1M token) went from $18.40 to $6.07 (3x cheaper in a year)
>
> But the cost to continuously run it (agentic workflows) is 23.4x more expensive
>
> On top of this, more users are adopting AI, and each user
>
> 
This insatiable need for intelligence gave rise to a vibrant onchain inference capital markets. A place where demand can be serviced and settled onchain. A place where idle resources are utilized, offering cheaper inference alternatives to what one could find on the market.
> **0xJeff@0xJeff**: [原文链接](https://x.com/0xJeff/status/2077684198811553892)
>
> One of the Undertapped Markets = Inference Capital Markets
> > Primary market well covered on compute, inference as emerging market
> > GPU financing as rapidly growing market
> > Standardization & settlement remains the biggest gap
> > Reference pricing emerges as #1 value accrual
>
> 
> 
Inference Capital Markets consist of 5 different layers. Layer 1 / Primary Market is the most mature so far, and the most vibrant segment is Inference segment where more than 10+ new onchain AI players are actively competing for market share.
This particular segment is what we'll be focusing on in this article.
## The Origin of Onchain Inference
The narrative itself started early this year when Venice AI (a privacy-focused & consumer-focused inference platform) found PMF. Users doubled within a few months vs. half a year to a year. Revenue followed suits which trickled down to the value of buyback & burn narrative of the VVV token.
The clever design of VVV <> DIEM tokenomics where VVV can be staked and used to mint DIEM, a tokenized platform credit that expires & replenishes itself daily — further fueled reputation of the project as well as the attractiveness of its token.
(up until the point they announced a large equity fundraising round, but we won't talk about that this time).
Venice, along with all other Web2 inference players (Baseten, Fireworks, Together AI) who also raised hundreds of millions were able to rapidly grow, thanks to the insatiable demand of token-hungry agents.
General harnesses (OpenClaw, Hermes) and coding harnesses (Codex, Claude Code) consume 10-100x more tokens compared to chat-based/conversational AI and enterprises are pouring millions into it.
Both to improve the internal productivity of engineers (ship better, ship faster) + to offer autonomous agents that can solve more problems & automate more workflows for users. Both of which consume a lot of inference.
As a result, Inference costs have become the #1 highest variable cost that scale dynamically regardless of whether revenue is generated.
Before onchain inference narrative, we only had decentralized compute as the segment that generates significant demand. But since demand shifted from training to inference, value is rapidly flowing towards inference providers.
## The Landscape
The landscape is largely split into 3 categories, each with differing business model and value proposition
- 01. Sport Markets & Marketplaces — Enables NEW Markets
- 02. Gateways & Routers — Caters to most IMMEDIATE demand
- 03. Inference Networks — HIGHEST Value Proposition
The Demand Side showcases top sources that potentially consume inference from the first 3 categories. In reality, the top usage likely goes to builders & power users of agents.
Many of the builders today tend to cover more than 1 category here as part of their GTM strategy — own the supply chain (make money from directly selling inference) + own the demand amplifier (increase the monetization surface area with harnesses, frameworks, toolings, and agent products).
## 01 Spot Markets & Marketplaces
Because inference has become a tradable asset with standard market price (in/out per 1M tokens), secondary markets are starting to emerge, catering to 2 type of users
- Sellers: sell unused inference credits or spare GPU capacity
- Buyers: buy inference tokens at 20-80% discount from market price
And there are 3 types of sellers
- (i) Subscribers of AI models with unused credits
- (ii) Retail or datacenters with idle GPU capacity
- (iii) Unused tokenized inference "DIEM" sellers
Most AI power users subscribe to Anthropic, OpenAI, Zai, Moonshot, and other subscription packages. Many ended up not using all of their quotas.
So where do the leftover quotas go?
Nowhere.
If these users want to maximize their profit, they can sell leftover quotas for cheap on secondary marketplaces — to the people who're willing to pay for them.
However, this breaks the ToS of virtually all AI providers (because AI providers save a lot money from unused subscription/unused subsidy). Marketplaces that facilitate this type of transaction risk getting sued by AI providers (if they get big enough).
The second type of sellers are pretty common. Datacenters and inference providers tend to have spare capacity because they're intentionally over-provisioned to handle demand spike during peak hours. During idle time, spare capacity goes into secondary markets (so that at least they're earning something during the period).
The third type recently emerged, building on top of Venice's DIEM tokenized inference credits. 1 DIEM gives $1 of platform credit per day. Many bought DIEM only to speculative on its future value instead of utilizing the $1/day credit.
Secondary markets enable users to stake DIEM or sell the unused credit at a discount. Thus, earning yields 10-20% APR on DIEM. Morpho has this wstDIEM/DIEM where wstDIEM (LSD version of DIEM) can be used as a collateral to borrow DIEM which essentially allow users to leverage farm DIEM (up to ~56% APR according to Liquid protocol, the team behind wstDIEM). Bear in mind the liq is very thin on wstDIEM/DIEM market so this is not doable at scale at the moment.
> **liquid @_proxystudio**: [原文链接](https://x.com/_proxystudio/status/2075743588424593627)
>
> wstETH vs wstDIEM leveraged yield farming #'s
> wrapping & staking ETH on lido (for example) and depositing it to morpho to lev yield is one of the most dependable and liquid lev yield farming strategies in defi. people frequently loop wstETH 10x to achieve a 5-8% APR
> 10x lev
>
> 
## Top 3 Marketplaces Ranking
[If you like this piece, feel free to continue reading on our Substack HERE]
It's where we'll cover
- The State of 02 Gateways & Routers and 03 Inference Networks segments and where value is accruing
- Top 3 players in each category (01 Marketplaces, 02 Gateways & Routers, 03 Inference Networks) ranked by product attractiveness (and alignment)
- Why we're about to witness the biggest wave of Onchain AI
## 相关链接
- [0xJeff](https://x.com/0xJeff)
- [@0xJeff](https://x.com/0xJeff)
- [14K](https://x.com/0xJeff/status/2079929079684587603/analytics)
- [Jul 21](https://x.com/0xJeff/status/2079544969065545892)
- [25K](https://x.com/0xJeff/status/2079544969065545892/analytics)
- [Jul 16](https://x.com/0xJeff/status/2077684198811553892)
- [80K](https://x.com/0xJeff/status/2077684198811553892/analytics)
- [Jul 11](https://x.com/_proxystudio/status/2075743588424593627)
- [11K](https://x.com/_proxystudio/status/2075743588424593627/analytics)
- [HERE](https://defi0xjeff.substack.com/p/the-rise-of-inference-capital-markets)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:58 PM · Jul 22, 2026](https://x.com/0xJeff/status/2079929079684587603)
- [14.2K Views](https://x.com/0xJeff/status/2079929079684587603/analytics)
- [View quotes](https://x.com/0xJeff/status/2079929079684587603/quotes)
---
*导出时间: 2026/7/23 13:38:25*
---
## 中文翻译
# 推理资本市场的崛起
**作者**: 0xJeff
**日期**: 2026-07-21T12:32:11.000Z
**来源**: [https://x.com/0xJeff/status/2079929079684587603](https://x.com/0xJeff/status/2079929079684587603)
---
对智能的需求正在吞噬整个行业。AI 提供商没有足够的算力来满足需求，这为推理提供商介入并服务这些需求留出了空间。
> **0xJeff@0xJeff**: [原文链接](https://x.com/0xJeff/status/2079544969065545892)
>
> 天啊……AI 很便宜，但也太贵了
>
> 启动成本（每 100 万 token 的推理费用）从 18.40 美元降至 6.07 美元（一年内便宜了 3 倍）
>
> 但持续运行它的成本（代理工作流）要贵 23.4 倍
>
> 除此之外，越来越多的用户正在采用 AI，而且每个用户
>
> 
这种对智能的渴求催生了一个充满活力的链上推理资本市场。这是一个需求可以被链上服务和结算的地方。这也是一个利用闲置资源的地方，提供比市场上现有选择更便宜的推理方案。
> **0xJeff@0xJeff**: [原文链接](https://x.com/0xJeff/status/2077684198811553892)
>
> 未被充分发掘的市场之一 = 推理资本市场
> > 算力一级市场已被充分覆盖，推理作为新兴市场
> > GPU 融资作为快速增长的市场
> > 标准化与结算仍是最大的缺口
> > 参考定价成为首要的价值积累点
>
> 
> 
推理资本市场由 5 个不同的层级组成。第 1 层 / 一级市场是目前最成熟的，其中最活跃的细分市场是推理市场，有 10 多个新的链上 AI 参与者正在积极争夺市场份额。
我们将重点关注本文的这一特定细分市场。
## 链上推理的起源
这一叙事始于今年早些时候，当时 Venice AI（一个专注于隐私和消费者的推理平台）找到了产品市场契合点（PMF）。用户数在几个月内就翻了一番，而以前通常需要半年到一年。收入也随之增长，进而传导至 VVV 代币的回购与销毁叙事的价值。
VVV 与 DIEM 代币经济学的巧妙设计，即 VVV 可以被质押并用于铸造 DIEM——一种每天过期并自动补充的代币化平台积分——进一步推动了项目的声誉及其代币的吸引力。
（直到他们宣布了一大笔股权融资轮，但我们这次不谈这个）。
Venice 以及所有其他筹集了数亿美元资金的 Web2 推理参与者（Baseten, Fireworks, Together AI）之所以能够快速增长，都要归功于那些对 token 饥渴难耐的代理的无尽需求。
通用模型和编码模型相比基于聊天的对话式 AI 要消耗 10-100 倍的 token，企业正在为此投入数百万美元。
既是为了提高工程师的内部生产力（构建更好的产品，更快地交付），也是为了提供能够为用户解决更多问题并自动化更多工作流的自主代理。这两者都消耗大量的推理资源。
因此，推理成本已成为第一大可变成本，无论是否产生收入，它都会动态扩展。
在链上推理叙事出现之前，我们只有去中心化算力这一产生显著需求的细分市场。但由于需求已从训练转向推理，价值正迅速流向推理提供商。
## 行业格局
该格局主要分为 3 个类别，每个类别都有不同的商业模式和价值主张
- 01. 现货市场与交易平台 — 开启新市场
- 02. 网关与路由器 — 满足最直接的需求
- 03. 推理网络 — 最高的价值主张
需求端展示了可能从前 3 个类别消耗推理的主要来源。实际上，最主要的使用者可能是代理的构建者和高级用户。
如今许多构建者倾向于作为其 GTM（市场进入）策略的一部分覆盖不止一个类别——既拥有供应链（通过直接销售推理赚钱）又拥有需求放大器（通过模型、框架、工具和代理产品增加变现面）。
## 01 现货市场与交易平台
由于推理已成为一种具有标准市场价格（每 100 万 token 的输入/输出）的可交易资产，二级市场开始涌现， catering to 2 type of users
- 卖家：出售未使用的推理额度或闲置 GPU 算力
- 买家：以低于市场价 20-80% 的折扣购买推理 token
主要有 3 种类型的卖家
- (i) 拥有未使用额度的 AI 模型订阅者
- (ii) 拥有闲置 GPU 算力的个人或数据中心
- (iii) 未使用的代币化推理“DIEM”卖家
大多数 AI 高级用户都订阅了 Anthropic、OpenAI、Zai、Moonshot 和其他订阅套餐。许多人最终并没有用完所有的配额。
那么，剩余的配额去哪儿了？
无处可去。
如果这些用户想最大化他们的利润，他们可以在二级市场上廉价出售剩余配额——卖给愿意为此付费的人。
然而，这几乎违反了所有 AI 提供商的服务条款（因为 AI 提供商从未使用的订阅/未使用的补贴中节省了大量资金）。促进此类交易的市场平台面临被 AI 提供商起诉的风险（如果他们做得足够大的话）。
第二类卖家相当常见。数据中心和推理提供商往往会有闲置算力，因为它们故意超额配置以在高峰时段处理需求激增。在空闲时间，闲置算力会进入二级市场（这样至少在此期间他们还能赚点钱）。
第三类卖家最近才出现，是在 Venice 的 DIEM 代币化推理额度的基础上构建的。1 个 DIEM 每天提供 1 美元的平台额度。许多人购买 DIEM 只是为了投机其未来价值，而不是利用每天 1 美元的额度。
二级市场允许用户质押 DIEM 或折价出售未使用的额度。从而，在 DIEM 上获得 10-20% 的 APR 收益。Morpho 上有 wstDIEM/DIEM 市场，其中 wstDIEM（DIEM 的 LSD 版本）可以用作抵押品来借入 DIEM，这本质上允许用户杠杆耕作 DIEM（根据 wstDIEM 背后的团队 Liquid 协议，APR 高达约 56%）。请注意，wstDIEM/DIEM 市场的流动性非常稀薄，因此目前无法大规模操作。
> **liquid @_proxystudio**: [原文链接](https://x.com/_proxystudio/status/2075743588424593627)
>
> wstETH vs wstDIEM 杠杆收益耕作数据
> 在 Lido 上封装并质押 ETH（例如）并将其存入 Morpho 进行杠杆收益，是 DeFi 中最可靠和流动性最强的杠杆收益耕作策略之一。人们经常循环 wstETH 10 倍以实现 5-8% 的 APR
> 10 倍杠杆
>
> 
## 前 3 大交易平台排名
[如果你喜欢这篇文章，欢迎在我们的 Substack 上继续阅读 HERE]
在那里我们将涵盖
- 02 网关与路由器和 03 推理网络细分市场的现状以及价值积累之处
- 每个类别（01 交易平台，02 网关与路由器，03 推理网络）排名前 3 的参与者，按产品吸引力（以及一致性）排名
- 为什么我们即将见证链上 AI 的最大浪潮
## 相关链接
- [0xJeff](https://x.com/0xJeff)
- [@0xJeff](https://x.com/0xJeff)
- [14K](https://x.com/0xJeff/status/2079929079684587603/analytics)
- [Jul 21](https://x.com/0xJeff/status/2079544969065545892)
- [25K](https://x.com/0xJeff/status/2079544969065545892/analytics)
- [Jul 16](https://x.com/0xJeff/status/2077684198811553892)
- [80K](https://x.com/0xJeff/status/2077684198811553892/analytics)
- [Jul 11](https://x.com/_proxystudio/status/2075743588424593627)
- [11K](https://x.com/_proxystudio/status/2075743588424593627/analytics)
- [HERE](https://defi0xjeff.substack.com/p/the-rise-of-inference-capital-markets)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:58 PM · Jul 22, 2026](https://x.com/0xJeff/status/2079929079684587603)
- [14.2K Views](https://x.com/0xJeff/status/2079929079684587603/analytics)
- [View quotes](https://x.com/0xJeff/status/2079929079684587603/quotes)
---
*导出时间: 2026/7/23 13:38:25*