How AI Observability Platforms Make Money -- Langfuse and Dify

The Layer Above Frameworks

In the previous post, we analyzed how LangChain, LlamaIndex, and CrewAI give away frameworks for free and charge on the operations layer.

This time, we look at two companies that made that “operations layer” itself their core product. Langfuse (LLM observability platform) and Dify (LLM app development platform). Both are open source, both allow self-hosting, and both generate revenue despite that.

graph TB
    subgraph LAYER1["Layer 1: Frameworks"]
        A["LangChain / LlamaIndex / CrewAI"]
    end

    subgraph LAYER2["Layer 2: Observation & Operations Platforms"]
        B["Langfuse — Observability"]
        C["Dify — App Builder"]
    end

    subgraph LAYER3["Layer 3: Infrastructure"]
        D["HuggingFace / Qdrant / Weaviate"]
    end

    A -->|"Send tracing data"| B
    A -->|"Build apps no-code"| C
    B --> D
    C --> D

    style LAYER2 fill:#fff3e0,stroke:#ff9800

---

Langfuse: MIT + Enterprise Edition Model

Positioning

Langfuse is a dedicated observability platform for LLM applications. It competes directly with LangSmith, but with one critical difference: it’s not tied to any specific framework.

LangSmith is optimized for the LangChain ecosystem. Langfuse traces LLM calls from LangChain, LlamaIndex, the OpenAI SDK, the Anthropic SDK — anything.

Comparison	LangSmith	Langfuse
Framework dependency	Optimized for LangChain	Framework-agnostic
Self-hosting	No	Yes (Docker Compose)
Open-source core	No (closed source)	Yes (MIT)
Tracing	Yes	Yes
Evaluations (Evals)	Yes	Yes
Prompt management	Yes	Yes
Cost tracking	Basic	Yes (auto-calculated per model)

License Structure: MIT + EE

Langfuse’s core strategy is a MIT + Enterprise Edition (EE) dual license.

graph LR
    subgraph MIT["MIT License (Free)"]
        A1["Tracing"]
        A2["Evaluations"]
        A3["Prompt Management"]
        A4["Cost Tracking"]
        A5["Self-Hosting"]
    end

    subgraph EE["Enterprise Edition (Paid)"]
        B1["SSO / SAML"]
        B2["Fine-grained RBAC"]
        B3["Audit Logs"]
        B4["SLA Guarantees"]
        B5["Dedicated Infrastructure"]
    end

    MIT -->|"Enterprise scale-up"| EE

    style MIT fill:#e8f5e9,stroke:#4caf50
    style EE fill:#fff3e0,stroke:#ff9800

The MIT core is broad. Tracing, evaluations, prompt management, cost tracking — every core feature is MIT. Self-hosting is fully supported. This isn’t “limited features free + full features paid.” It’s all features free + enterprise management features paid.

Pricing Structure

Plan	Monthly Cost	Included Observations	Additional
Hobby	Free	50K/month	—
Pro	$59/month	100K/month	$10/additional 100K
Team	$499/month	1M/month	Volume discounts
Enterprise	Custom	Unlimited	SSO, SLA, dedicated infra
Self-hosted	$0	Unlimited	You operate the infra

Why Self-Hosting Doesn’t Cannibalize Revenue

Allowing self-hosting naturally raises the question: “Won’t everyone just use it for free?” In practice, they don’t:

User Type	Choice	Reason
Individual / side project	Self-host	Cost savings, learning purposes
Startup (5-20 people)	Cloud Pro/Team	No infra ops staff, fast start
Mid-size company (50-200)	Enterprise Cloud	Need SSO, audit logs, SLA
Large enterprise (200+)	Enterprise self-host	Data sovereignty, internal security policies

Self-hosting users are future paying customers. When an individual learns Langfuse through self-hosting and recommends it at their company, the company converts to Cloud or Enterprise. Self-hosting is a zero-cost customer education channel.

Langfuse Growth Metrics

Metric	Figure (2026 Q1 estimate)
GitHub Stars	8K+
Monthly Cloud Observations	Billions (estimated)
Self-hosted instances	Thousands (estimated from Docker pulls)
Core team size	~15 (Berlin)
Funding	Series A (amount undisclosed)

---

Dify: Conditionally Open Source

Positioning

Dify is an LLM app builder platform. If Langfuse is about “observing apps you’ve already built,” Dify focuses on “building LLM apps themselves using no-code/low-code.”

Feature	Description
Visual Workflow	Drag-and-drop LLM pipeline construction
RAG Engine	Upload documents, auto-index, connect to search
Agent Builder	Define tool-using agents through UI
Prompt IDE	Write, test, and version-control prompts
Auto API Generation	Instantly deploy built apps as REST APIs
Observability / Logs	Built-in tracing, cost tracking

License: Open Source with Conditions

Dify’s license is more restrictive than Langfuse’s (MIT).

Aspect	Langfuse (MIT)	Dify
Code viewing	Yes	Yes
Internal self-hosting	Yes	Yes
Modification and deployment	Yes, unlimited	Yes, with conditions
Reselling as multi-tenant SaaS	Yes	No, requires separate license
Logo/branding removal	Yes	Requires paid license
Commercial use	Yes, unlimited	Yes, internal use is free

Dify’s license says: “Source code is open, but don’t take it and build a competing SaaS.” More restrictive than MIT, but a practical compromise:

• Internal self-hosting: Sufficient for most users
• Resale restriction: Prevents AWS/Azure from cloning Dify into a Managed Dify service
• Most usage is effectively free: Using Dify for your own company’s LLM apps is completely free

Pricing Structure

Plan	Monthly Cost	Message Credits	Team Members	App Count
Sandbox	Free	200/session	1	10
Professional	$59/month	5,000/month	3	50
Team	$159/month	10,000/month	Unlimited	Unlimited
Enterprise	Custom	Unlimited	Unlimited	Unlimited

Pricing axis: Message credits + team members

Dify uses a hybrid pricing model similar to LangChain, but the difference is that its positioning as an app builder makes the pricing feel natural. “My app was called 100 times, so I used credits” — directly tied to business value.

---

Langfuse vs Dify: Same Layer, Different Strategies

graph TB
    subgraph LANGFUSE["Langfuse Strategy"]
        LF1["Fully MIT"]
        LF2["Unlimited self-hosting"]
        LF3["Enterprise features for revenue"]
        LF4["Framework-independent"]
    end

    subgraph DIFY["Dify Strategy"]
        DF1["Conditionally open source"]
        DF2["Self-hosting possible (limited)"]
        DF3["Message credit pricing"]
        DF4["All-in-one platform"]
    end

    LF1 -.->|"Wider community"| LF4
    DF1 -.->|"Resale defense"| DF4

Comparison	Langfuse	Dify
Core value	”Observe apps you’ve built"	"Platform that builds apps for you”
Target users	Developers (code writers)	Developers + non-developers
Competitors	LangSmith, Arize	Flowise, n8n, Zapier AI
License	MIT (unrestricted)	Conditional (resale restricted)
Self-hosting strategy	Future customer education channel	Internal adoption channel
Paid conversion trigger	Team scale-up	Message volume growth
GitHub Stars	8K+	55K+

Positioning Determines Pricing

• Langfuse: “Observability is needed by every LLM app” — maximize universality — MIT for minimum entry barrier — charge on Enterprise features
• Dify: “Anyone can build LLM apps” — maximize usability — charge on message credits for usage-based revenue — restrict license against resale

---

Why the Observability Layer Has Better Revenue Structure Than Frameworks

Comparing the monetization structures of frameworks and observation/operations platforms, the structural advantages of the observability layer become clear:

Factor	Frameworks	Observability/Ops Platforms
Switching costs	Low (can swap by modifying code)	High (data and workflow migration is difficult)
Competitive alternatives	Many (LangChain vs LlamaIndex vs …)	Few (dedicated observability tools are rare)
Data accumulation effect	None	Present (history has value)
Usage frequency	Only during development	Continuous throughout operations
Pricing justification	Low (free alternatives exist)	High (time saved = cost reduced)

Frameworks are “tools for building.” Observability platforms are “tools for operating.” Building is a one-time activity; operations continue indefinitely. Continuous usage = continuous pricing = recurring revenue.

Implications for Solo Builders

From MMU’s positioning perspective:

• MMU CLI is a building tool — used intensively before launch, then stops
• Playbook Pack is an operational guide — content referenced even after launch
• AI Coach is an observability layer — continuously diagnosing checklist status

The lesson from Langfuse/Dify: Tools used once are harder to monetize than tools used continuously. MMU’s long-term strategy is shifting the revenue axis from CLI (one-time) to AI Coach (recurring).

---

Summary

Aspect	Langfuse	Dify
License	MIT (fully open)	Conditional (resale restricted)
Pricing axis	Observations + EE	Message credits + team members
Self-hosting	Unlimited (future customer channel)	Internal use free
Key lesson	Observability is universal — open it as wide as possible	App builders need competitive defense

The next post analyzes the infrastructure layer — Hugging Face, Qdrant, Weaviate — and their monetization strategies. We’ll examine how companies that build “self-hostable by anyone” infrastructure still generate revenue.