
From Nvidia’s Sell-Off to A-Share Optical Modules: How DojoAgents Turns Financial Research into an Automated Agent Loop
English Version First|中文完整版见后
Connecting “Nvidia’s overnight decline” with “A-share optical-module stocks” is a classic financial research problem.
The underlying logic is not especially difficult. Nvidia is widely viewed as a bellwether for AI computing demand, while optical modules are critical components of the high-speed networks connecting AI servers and data centers. Changes in GPU demand, cloud capital expenditure, or AI infrastructure expectations can therefore transmit across the supply chain.
The real challenge lies elsewhere.
When an investor notices Nvidia falling sharply in the middle of the night, the market does not wait patiently for a complete research report. The investor may need to move repeatedly between different terminals and data sources:
Check overnight prices and trading volume
Read breaking news and company disclosures
Search analyst reports
Reconstruct the AI infrastructure supply chain
Identify related A-share and Hong Kong-listed companies
Calculate the exposure of the current portfolio
By the time these fragments are assembled into a coherent view, the market may already have completed its first round of repricing.
That is the problem DojoAgents is trying to solve.
Just over a week after its open-source release, DojoAgents surpassed 1,000 GitHub stars. Its appeal does not come from being another chatbot that can explain why a stock moved.
Its deeper value is that it turns financial research from manual information assembly into an automated, tool-driven workflow.
Chatbot vs. Financial Agent: The Missing Main Loop
Over the past two years, large language models have become increasingly capable of producing structured company analyses, summarizing earnings reports, and explaining how a news event might affect the market.
But in real financial workflows, producing an answer is only the final step.
The fundamental difference is this:
A chatbot focuses on:
How should I answer this question?
A financial agent must also determine:
What should I investigate next, which tool should I call, and how should I verify the result?
Without a loop, an LLM call is essentially an advanced calculator: it receives an input, generates an output, and stops.
The core of DojoAgents is its Agent Loop, an execution engine that decomposes a financial question into a repeated sequence of research actions:
Think → Act → Observe → Think Again
After receiving a question, the agent first evaluates what information is missing. It then calls the appropriate data or analytical tool, observes the returned results, updates its understanding, and decides what to investigate next.
The loop continues until the available evidence is sufficient to support a structured conclusion—or until the system identifies uncertainties that still require human review.
This is what separates a one-time response from an actual research process.
How One Market Question Is Decomposed
Consider the question:
Nvidia fell overnight. Will A-share optical-module, PCB, and semiconductor-equipment stocks be affected today?
A conventional chatbot may immediately generate a plausible explanation based on general knowledge.
DojoAgents approaches the problem as a research chain.
Step 1: Collect Market Data and News
The agent first retrieves Nvidia’s overnight price performance, trading volume, related news, company announcements, and broader technology-sector movements.
It then attempts to determine the nature of the decline:
Was it driven by a change in AI demand?
Did earnings or guidance disappoint?
Was it a valuation adjustment?
Did supply-chain expectations change?
Was the movement caused by policy concerns?
Or was it largely a market-sentiment event?
This distinction is essential.
A decline caused by company-specific legal risk may have a very different impact on Chinese AI infrastructure suppliers than a decline caused by weaker data-center capital expenditure.
Step 2: Map the Industry Chain
The agent then identifies Nvidia’s position within the AI computing ecosystem and traces its upstream and downstream relationships.
A simplified transmission path may look like this:
AI accelerators → servers → data centers → high-speed interconnection → optical modules → PCB and semiconductor equipment
Optical modules are essential to high-bandwidth communication inside AI clusters. Their demand may be influenced by GPU shipments, data-center expansion, network architecture upgrades, and cloud capital expenditure.
The agent must therefore determine whether Nvidia’s decline changes the underlying demand outlook or merely reflects short-term market positioning.
Step 3: Map the Impact Across Markets
The next step is to translate a U.S. market event into possible effects on A-shares and Hong Kong stocks.
The agent may search for:
Companies with direct supply-chain exposure
Businesses whose valuations are closely linked to AI infrastructure sentiment
Suppliers affected by cloud capital-expenditure expectations
Stocks that have historically shown a high correlation with Nvidia or the broader semiconductor sector
This is not a simple one-to-one mapping.
A company can have real operating exposure, thematic exposure, or merely short-term sentiment exposure. These relationships need to be separated.
Step 4: Diagnose Portfolio Exposure
Finally, the agent can compare the identified transmission chain with the user’s current holdings.
It may examine:
Direct exposure to optical modules or AI servers
Indirect exposure through PCB, cooling, power, or semiconductor equipment
Concentration in one AI infrastructure theme
Cross-market correlation
Potential downside sensitivity
Whether the portfolio already reflects the relevant risk
The user therefore sees more than a conclusion.
They can also see:
Which data was retrieved
Which tools were called
How the industrial chain was extended
Which assumptions were verified
Which uncertainties remain unresolved
How the final judgment was assembled
For financial research, this evidence trail is often more valuable than a polished paragraph alone.
The Tool Layer: The Hands and Feet of the Agent Loop
An Agent Loop cannot operate through language reasoning alone.
It needs tools.
DojoAgents converts many of the tasks traditionally performed manually by researchers into callable capabilities, including:
Market scanning
News retrieval
Company analysis
Financial-data processing
Industry-chain mapping
Portfolio construction
Exposure diagnosis
Risk analysis
Performance attribution
These tools allow the agent to move beyond “what it remembers” and interact with updated data, internal systems, and analytical environments.
The framework can be understood as a layered architecture composed of:
Infrastructure layer — runtime, deployment, storage, and model access
Data layer — market prices, financial statements, news, portfolios, and private datasets
Tool layer — callable research and analytical capabilities
Agent Loop layer — planning, execution, observation, and repeated reasoning
Interaction layer — chat gateway, web dashboard, and command-line interface
This structure separates data access, tool execution, agent reasoning, and user interaction.
As a result, teams can replace or expand individual components without rebuilding the entire system.
MCP: A Standard Interface for Financial Data and Private Tools
The DojoAgents tool layer can be extended through MCP, or Model Context Protocol.
Introduced by Anthropic in 2024, MCP provides a standardized way for AI models and agent systems to connect with external data sources and tools.
It is often described as a “USB-C interface for AI” because it attempts to reduce the fragmentation created by different databases, plugins, applications, and internal systems.
For a financial institution, this matters because relevant information may be distributed across:
Public market-data services
Internal research databases
Portfolio-management systems
Risk engines
Proprietary factor libraries
Compliance systems
Local files
Cloud services
Once these resources are exposed through an MCP server, an agent-side MCP client can interact with them through a standardized communication layer.
This makes it easier to add proprietary tools without rewriting the core agent framework.
For example, a brokerage or asset manager could connect:
An internal earnings-estimate database
A private analyst-report archive
A portfolio risk engine
A compliance-approved news source
A proprietary industry-classification system
The result is not merely a general financial chatbot with more information.
It is an agent connected to the institution’s actual research environment.
Flexible Model Access: Cloud APIs and Local Deployment
At the model layer, DojoAgents can connect to several model ecosystems through compatible APIs and adapters, including:
OpenAI
Anthropic
Gemini
DeepSeek
GLM
It also supports local inference endpoints through frameworks such as:
Ollama
vLLM
This flexibility is particularly important in financial environments, where different tasks may require different deployment strategies.
A team may use:
A frontier cloud model for complex reasoning
A lower-cost model for repetitive classification
A local model for sensitive portfolio data
A fast model for market scanning
A higher-accuracy model for final risk interpretation
The framework does not force every task through the same model.
It allows the model layer to become another part of the workflow that can be routed and optimized.
From One Prompt to a Continuous Agent Loop
The traditional approach to a question such as “How will Nvidia’s decline affect A-share optical modules?” is to write an increasingly long prompt.
The prompt may include:
Industrial-chain relationships
Required data sources
Research methodology
Output structure
Risk considerations
Cross-market logic
The hope is that the model will process everything correctly in one attempt.
But financial research is not a one-time event.
Markets change. Data updates. New disclosures appear. Original assumptions become invalid. Correlations strengthen or break. The causal relationship itself may need to be reconstructed.
DojoAgents takes a different approach:
Do not provide the answer. Provide the process.
Instead of attempting to place every piece of knowledge inside one prompt, the framework allows the agent to determine:
What information is required
Where that information should be retrieved
Which tool should process it
How the result should affect the next step
Whether the evidence is sufficient
What still needs to be verified
The real purpose of an Agent Loop is therefore not to automate writing.
It is to automate the research process itself.
Compute Cost: The Final Barrier to Scaling Agent Loops
The Agent Loop turns manual research assembly into an automated workflow.
But each iteration consumes model capacity.
A single financial question may trigger repeated operations:
Retrieve market data
Read news
Analyze company fundamentals
Map the supply chain
Compare related securities
Diagnose portfolio exposure
Review conflicting evidence
Generate a structured conclusion
Each step may involve one or more model calls.
When the system is used for occasional demonstrations, the cost may appear manageable. When financial institutions begin deploying agents continuously across markets, sectors, portfolios, and client accounts, Token consumption can expand rapidly.
The technical question is no longer only:
Can the agent complete the task?
It becomes:
Can the institution afford to keep the agent running continuously?
That is where Unexhub enters the infrastructure layer.
Unexhub: Making Token Capacity More Market-Oriented
Unexhub aims to transform Token capacity from an isolated resource controlled by individual model providers into infrastructure that can be aggregated, distributed, routed, priced, and settled through a compute trading center.
Its proposed model contains three core mechanisms.
1. Token Aggregation and Distribution
Model capacity is distributed across large AI companies, regional computing centers, specialist model providers, smaller enterprises, and independent developers.
Many smaller suppliers possess high-quality models or unused inference capacity but lack:
Stable customer traffic
Global distribution channels
Intelligent routing
Unified billing
Market visibility
Pricing power
Unexhub aggregates these resources into a unified supply pool and distributes them to enterprises and developers through standardized API access.
For suppliers, this creates an additional route to market.
For users, it creates more choices among models, endpoints, service levels, and prices.
2. Faster Settlement
Traditional enterprise payment cycles can place considerable pressure on smaller compute suppliers.
Unexhub proposes settlement based more closely on actual API usage, enabling suppliers to view revenue and receive their share on a shorter cycle.
Rather than waiting through lengthy reconciliation periods, suppliers can gain clearer visibility into:
Token consumption
Revenue by model
Route performance
Customer demand
Settlement progress
Qualified early suppliers may also receive promotional commission rates as low as 0%.
The value is not limited to saving platform fees.
It creates a lower-cost period in which suppliers can test demand, build initial usage, validate their service quality, and establish a distribution channel before entering a standard commission structure.
3. API Bidding and Dynamic Pricing
Unexhub also proposes an API bidding mechanism.
Instead of relying entirely on a fixed official price, effective model costs can respond to:
Available capacity
Current demand
Model performance
Response latency
Service reliability
Peak and off-peak periods
Competition among suppliers
Developers may gain access to lower-cost routes when capacity is available, while suppliers may earn stronger returns when demand rises.
The Token price therefore becomes a market-matching result rather than only a static number published by one provider.
When the Agent Loop Meets a Compute Trading Center
DojoAgents and Unexhub operate at different layers of the same system.
DojoAgents reconstructs the financial research workflow at the application layer.
It converts market scanning, news analysis, industry-chain mapping, portfolio diagnosis, and risk evaluation into a repeatable agent process.
Unexhub attempts to reconstruct the underlying cost structure.
It aggregates model providers, routes requests between available endpoints, introduces supply-side competition, and accelerates settlement.
One answers:
How can financial research be automated?
The other answers:
How can continuous automated research remain affordable?
When a DojoAgents loop continues running in the background—collecting data, interpreting news, validating hypotheses, and updating portfolio analysis—Unexhub can route those model requests toward more cost-efficient capacity.
Developers can pay according to actual Token usage rather than purchasing an oversized package in advance.
Depending on model availability, route conditions, supplier competition, and promotional terms, selected services may be available at prices as low as approximately 30% of comparable market reference rates.
For a single research query, the difference may seem limited.
For an institution running continuous market monitoring, cross-market analysis, and portfolio-risk agents, lower unit costs can accumulate into meaningful monthly and annual savings.
Unexhub’s Value for Developers and Model Suppliers
Unexhub is building a Token aggregation and distribution network designed to connect enterprise users, developers, model providers, and compute suppliers.
The platform is intended to aggregate access to mainstream domestic and international AI ecosystems, including services associated with:
Gemini
ChatGPT
Claude
DeepSeek
GLM
Other specialist and independent models
Through a unified interface, developers can reduce the burden of maintaining separate:
API accounts
Authentication systems
Request formats
Billing dashboards
Supplier contracts
Routing rules
Its proposed capabilities include:
Multi-model API aggregation
Pay-as-you-use Token billing
Millisecond-level intelligent routing
Dynamic supplier and endpoint selection
API bidding
Centralized usage management
Faster supplier revenue settlement
Enterprise integration and customization
For developers, the benefit is not simply access to a lower headline price.
They can select models according to task difficulty, latency requirements, privacy needs, and available budget—while paying only for actual consumption.
For suppliers, one-click or standardized access to a distribution channel can reduce the cost of finding customers, building billing infrastructure, and collecting revenue.
The platform’s early-stage zero-commission policy can also give qualified suppliers additional room to validate demand before taking on normal channel costs.
The Real Upgrade: Automating Research, Not Just Answers
The link between Nvidia and A-share optical modules is only one example.
The more important shift is in how the question is processed.
In the old workflow, the researcher manually moves between data terminals, news platforms, analyst reports, industrial-chain databases, and portfolio systems.
In the new workflow, an agent decides:
What to retrieve
Which tool to use
How to interpret the result
What to investigate next
When the evidence is sufficient
Which risks remain unverified
DojoAgents turns that process into an Agent Loop.
Unexhub attempts to make the Token consumption behind the loop more economically sustainable.
One transforms manual research into an automated pipeline.
The other helps prevent that pipeline from becoming too expensive to operate at scale.
The next generation of financial AI will not be judged only by the quality of a single answer.
It will be judged by whether it can continuously research, verify, update, explain, and operate within a sustainable cost structure.
That is the real transition:
From Prompt to Agent Loop—and from an agent that can run to an agent that institutions can afford to keep running.
从“英伟达下跌”到“A 股光模块”:DojoAgents 如何把金融研究变成自动运行的 Agent Loop
中文完整版
把“英伟达隔夜下跌”和“A 股光模块板块”联系起来,是金融行业中非常典型的一类研究问题。
这条逻辑本身并不难理解。
英伟达被广泛视为 AI 算力需求的风向标,光模块则是连接 AI 服务器与数据中心的关键网络组件。GPU 出货、云厂商资本开支和 AI 基础设施建设预期发生变化,都可能沿产业链传导至光模块、PCB、服务器和半导体设备等环节。
真正困难的,并不是知道二者“可能有关”。
而是如何在市场完成第一轮定价之前,把所有证据快速串联起来。
一个半夜看到英伟达大跌的投资者,往往需要在多个系统之间反复切换:
查隔夜行情和成交变化
阅读突发新闻与公司公告
翻阅机构研报
梳理 AI 算力产业链
寻找 A 股与港股相关标的
计算当前持仓的风险敞口
等到这些零散信息被拼成一张完整的图,市场可能早已完成第一轮反应。
这正是 DojoAgents 想要解决的问题。
DojoAgents 在 GitHub 开源仅一周多,便突破了 1000 Star。它受到关注,不是因为市场上又多了一个能够解释股票涨跌的聊天机器人。
它真正的价值,是把金融研究从手动拼图,转化为一条由数据和工具驱动的自动化流水线。
聊天机器人与金融智能体,差的是一个“主循环”
过去两年,大模型已经可以生成结构完整的公司分析,也能解释一条新闻可能对市场产生怎样的影响。
但进入真实金融场景后,仅仅能够“给出答案”远远不够。
聊天机器人关心的是:
这个问题应该怎么回答?
金融智能体还必须知道:
下一步应该查什么、调用什么工具,以及怎样验证结果?
没有循环机制的 LLM 调用,本质上仍然像一个高级计算器:接收一次输入,生成一次输出,然后结束。
DojoAgents 真正的核心,是 **Agent Loop(智能体循环)**引擎。
它会把一个金融问题拆解为不断重复的研究过程:
思考 → 行动 → 观察 → 继续思考
智能体拿到问题后,首先判断当前缺少哪些信息;随后调用相应的数据或分析工具,观察返回结果,更新对问题的理解,再决定下一步继续查什么。
这个循环会持续运行,直到现有证据足以支持一份结构化结论,或者系统明确识别出仍然需要人工判断的不确定项。
这正是一次性问答与真正研究流程之间的区别。
一个问题如何被完整拆解
以这个问题为例:
英伟达隔夜下跌,今天会不会影响 A 股的光模块、PCB 和半导体设备板块?
传统聊天机器人可能会直接基于已有知识,生成一段听起来合理的解释。
DojoAgents 则会把它拆解成一条完整的研究链。
第一步:采集行情与新闻
智能体首先获取英伟达隔夜的股价走势、成交量、相关新闻、公司公告,以及整个科技板块的变化。
随后判断这次下跌究竟属于哪一种性质:
AI 需求预期是否发生变化
财报或业绩指引是否低于预期
是否只是估值回调
供应链预期是否改变
是否出现新的政策风险
还是市场情绪主导的短期波动
这一步非常关键。
如果下跌来自公司自身的诉讼风险,对中国 AI 基础设施供应链的影响,可能远小于云厂商资本开支下降带来的冲击。
第二步:映射产业链
智能体接下来需要定位英伟达在 AI 算力生态中的位置,并沿着上下游继续延伸。
一条简化后的传导路径可能是:
AI 加速芯片 → 服务器 → 数据中心 → 高速互联 → 光模块 → PCB 与半导体设备
光模块是 AI 集群内部实现高速数据通信的重要组件,其需求可能受到 GPU 出货量、数据中心扩张、网络架构升级与云厂商资本开支的共同影响。
因此,智能体不能只看到“英伟达跌了”,还要判断这次下跌是否改变了产业需求预期,还是只反映短期资金情绪。
第三步:完成跨市场传导
下一步,是把美股事件映射到 A 股与港股市场。
智能体可能进一步寻找:
与英伟达或 AI 基础设施存在直接供应关系的企业
估值高度依赖 AI 算力主题的公司
受云计算资本开支影响较大的供应商
历史上与英伟达或半导体指数联动明显的标的
这不是简单的一一对应。
一家企业可能拥有真实的业务敞口,也可能只有主题敞口,甚至只是短期情绪相关。智能体需要把这些关系分开。
第四步:诊断持仓风险
最后,智能体可以把产业链传导关系与用户当前持仓进行对照。
它可能分析:
是否直接持有光模块或 AI 服务器公司
是否通过 PCB、散热、电源或半导体设备间接暴露
是否过度集中于同一 AI 基础设施主题
不同市场持仓之间是否高度相关
组合对相关事件的下行敏感度
当前股价是否已经提前反映风险
用户看到的不再只有最终结论,还能够进一步了解:
智能体调用了哪些数据
使用了哪些工具
产业链是怎样逐步延伸的
哪些假设已经得到验证
哪些信息仍然存在不确定性
结论是如何一步步形成的
对于金融研究而言,这条证据链通常比一段语言流畅的答案更有价值。
工具层:Agent Loop 的“手”和“脚”
Agent Loop 不能只依靠语言推理运行。
它还必须拥有工具。
DojoAgents 把许多原本需要研究员手动完成的任务,拆解成智能体可以自主调用的能力组件,包括:
市场扫描
新闻检索
公司分析
财务数据处理
产业链映射
组合构建
持仓敞口诊断
风险分析
业绩归因
这些工具让智能体不再只依赖模型“记得什么”,而是可以与最新数据、企业内部系统和专业分析环境进行交互。
整个框架可以理解为一套分层架构:
Infrastructure 基础设施层:负责运行环境、部署、存储与模型接入
Data 数据层:负责行情、财报、新闻、持仓和私有数据
Tools 工具层:提供可调用的金融研究与分析能力
Agent Loop 智能体循环层:负责规划、执行、观察和持续推理
Interaction 交互层:包括聊天网关、Web 仪表盘和命令行入口
这种结构将数据接入、工具执行、智能体推理和用户交互分离开来。
团队可以单独替换或扩展其中某个组件,而不需要重建整套系统。
MCP:连接金融数据与私有工具的标准接口
DojoAgents 的工具层可以通过 **MCP(Model Context Protocol,模型上下文协议)**进行扩展。
MCP 由 Anthropic 于 2024 年推出,目的是为模型和智能体系统连接外部数据与工具提供更加标准化的方式。
它经常被形容为“AI 世界的 USB-C 接口”,因为它试图减少数据库、插件、应用和企业系统之间的接入碎片化。
对于金融机构来说,这一点尤为重要。
真实研究所需的信息可能分散在:
公共行情服务
内部研究数据库
组合管理系统
风险引擎
私有因子库
合规系统
本地文件
云端服务
当这些资源通过 MCP Server 暴露之后,智能体侧的 MCP Client 就可以使用相对统一的通信方式与其连接。
这让机构更容易加入自己的私有工具,而不必修改整个智能体框架。
例如,一家券商或资管机构可以接入:
内部盈利预测数据库
私有研报资料库
组合风险分析引擎
经过合规审核的新闻源
自有行业分类体系
这样得到的就不再只是一个“知道更多内容”的通用金融机器人。
而是一个真正连接到机构研究环境中的智能体。
灵活的模型接入:云端 API 与本地部署并存
在模型层,DojoAgents 可以通过兼容接口与适配器接入多个模型生态,包括:
OpenAI
Anthropic
Gemini
DeepSeek
GLM
同时,它也支持通过以下框架连接本地模型:
Ollama
vLLM
这种灵活性对金融场景非常重要。
不同任务可能需要不同的模型配置:
复杂推理使用高能力云端模型
重复分类使用低成本模型
敏感持仓数据交给本地模型
市场扫描使用高速模型
最终风险解释使用高准确率模型
DojoAgents 不需要强迫所有任务都经过同一款模型。
模型本身也可以成为工作流中被路由和优化的一部分。
从一个 Prompt,走向持续运行的 Agent Loop
过去处理“英伟达下跌如何影响 A 股光模块”这类问题,常见做法是不断堆长 Prompt。
Prompt 里可能包含:
产业链关系
数据源要求
分析框架
输出结构
风险提示
跨市场传导逻辑
然后期待模型一次性给出正确答案。
但金融研究从来不是一次性事件。
市场在变化,数据在更新,新公告不断出现,原有假设可能失效,相关性也可能增强或断裂,甚至因果关系本身都需要重新构建。
DojoAgents 换了一种方式:
不给固定答案,而是提供一套流程。
它不再试图把所有知识都塞进一个 Prompt,而是让智能体自己判断:
当前需要什么信息
应该去哪里获取
使用什么工具处理
返回结果如何影响下一步
现有证据是否已经充分
还有哪些内容需要继续验证
Agent Loop 真正自动化的,并不是“写答案”。
而是“做研究”这个过程本身。
算力成本:Agent Loop 规模化的最后一道门槛
Agent Loop 把手动拼图变成了自动化流水线。
但每一次循环都需要消耗模型算力。
一个金融问题可能触发多轮操作:
获取市场数据
阅读相关新闻
分析公司基本面
映射产业链
比较相关标的
诊断持仓敞口
复核冲突证据
生成结构化结论
每一个步骤都可能包含一次甚至多次模型调用。
当系统只用于偶尔演示时,这部分成本可能并不明显;但当金融机构准备让智能体跨市场、跨行业、跨组合持续运行时,Token 消耗会快速扩大。
问题不再只是:
智能体能不能完成任务?
而会变成:
企业能不能长期承担智能体持续运行的成本?
这正是 Unexhub 从基础设施层切入的位置。
Unexhub:让 Token 成为可市场化配置的资源
Unexhub 希望把原本由不同模型厂商分别控制的 Token 能力,转化为可以通过算力交易中心进行聚合、分发、路由、定价和结算的基础设施资源。
其模式可以拆解为三套核心机制。
第一层:Token 聚合与分发
模型算力分散在大型 AI 企业、区域算力中心、垂直模型供应商、中小企业和个人开发者手中。
许多中小型供应方拥有优质模型或闲置推理容量,却缺少:
稳定用户流量
全球分发渠道
智能路由能力
统一计费系统
市场曝光
定价与议价能力
Unexhub 将这些资源聚合进统一供给池,再通过标准化 API 提供给企业用户和开发者。
对供应方而言,这增加了一条商业化渠道。
对使用者而言,则意味着拥有更多模型、节点、服务等级和价格选择。
第二层:更快结算
传统企业账期可能给中小型算力供应方带来较大的资金压力。
Unexhub 希望按照真实 API 调用量,更及时地统计收益和完成调用分成,使供应方能够更快看到:
Token 使用量
不同模型的收入表现
路由运行情况
客户需求变化
当前结算进度
符合条件的早期供应方,还可享受最低 0% 佣金的阶段性权益。
这项权益不只是节省渠道费用。
它还为供应方提供了一段成本更低的市场验证期,可以先测试需求、积累调用量、检验服务质量,再进入常规佣金模式。
第三层:API 竞价与动态定价
Unexhub 还提出 API 竞价机制。
模型的实际调用价格不再只能依赖固定的官方报价,而可以根据以下因素动态变化:
当前可用容量
实时调用需求
模型性能
响应延迟
服务稳定性
高峰与低谷时段
不同供应方之间的竞争
开发者可以在容量充足时获得更具成本优势的线路,供应方也可以在需求高峰期获得更合理的收益。
Token 价格开始从单一厂商公布的静态数字,转变为市场供需匹配后的结果。
当 Agent Loop 遇上算力交易中心
DojoAgents 与 Unexhub 处在同一套系统的不同层级。
DojoAgents 重构的是金融研究的应用层。
它把市场扫描、新闻分析、产业链映射、持仓诊断和风险判断,转化成一条可以反复执行的智能体工作流。
Unexhub 试图重构的是底层成本结构。
它聚合模型供应方,在不同节点之间路由请求,引入供给侧竞争,并缩短收益结算周期。
一个回答:
金融研究怎样实现自动化?
另一个回答:
持续运行自动化研究,怎样才能在成本上可承受?
当 DojoAgents 的 Agent Loop 在后台持续采集数据、解读新闻、验证假设和更新组合分析时,Unexhub 可以尝试将这些模型请求路由至更具成本竞争力的算力线路。
开发者可以按照实际 Token 使用量付费,无需预先购买远超需求的固定套餐。
根据模型供应、线路条件、节点竞争和阶段性权益的不同,部分服务最低可享受约为市场参考价格 3 折的成本方案。
对于一次研究查询而言,这种价格差异可能并不明显。
但对于持续运行市场监控、跨市场分析与组合风险智能体的机构,单位成本下降会逐步累积为更加可观的月度和年度预算节省。
Unexhub 为开发者与模型供应方提供什么
Unexhub 正在建设一个连接企业用户、开发者、模型供应方与算力供应方的 Token 聚合和分发网络。
平台致力于聚合海内外主流 AI 模型生态,包括:
Gemini
ChatGPT
Claude
DeepSeek
GLM
其他专业模型与独立模型
通过统一接口,开发者可以减少对以下系统的重复维护:
多套 API 账户
不同鉴权方式
分散请求格式
独立计费后台
多家供应商合同
人工路由规则
其核心能力包括:
多模型 API 聚合
按 Token 实际用量付费
毫秒级智能路由
动态供应方与节点选择
API 竞价
集中化用量管理
供应方收益快速结算
企业级接入与定制服务
对开发者而言,价值不只是看到一个更低的宣传价格。
他们可以根据任务难度、延迟要求、隐私需求和预算,灵活选择不同模型,并只为实际消耗付费。
对供应方而言,通过标准化方式接入分发渠道,可以降低获客、计费和回款成本。
平台面向符合条件的早期供应方提供的零佣金权益,也能够帮助他们在承担常规渠道成本之前,先验证真实市场需求。
真正的升级:自动化的不是答案,而是研究
“英伟达与 A 股光模块之间有什么关系”,只是一个例子。
真正重要的变化,是这个问题被处理的方式正在改变。
过去,研究员需要手动穿梭于行情终端、新闻平台、研报数据库、产业链工具与持仓系统之间。
现在,智能体开始自主判断:
需要获取什么
应该调用哪个工具
返回结果如何解释
下一步继续调查什么
什么时候证据已经充分
哪些风险仍然没有得到确认
DojoAgents 将这个过程变成 Agent Loop。
Unexhub 则试图让支撑这一循环的 Token 成本更加可持续。
一个把手动研究转化为自动化流水线。
另一个避免这条流水线在规模化运行后变得过于昂贵。
下一代金融 AI 的竞争标准,不会只看它能否生成一份听起来专业的答案。
还要看它是否能够持续研究、反复验证、动态更新、解释过程,并在可持续的成本结构中长期运行。
这才是真正的变化:
从 Prompt 走向 Agent Loop,也从“智能体能够运行”,走向“机构可以长期负担智能体运行”。