Thread by 空谷 Arvin Xu (@arvin17x)

[1/9] As of yesterday, a monlobechat.coma launch, Lobe Chat Cloud (https://t.co/z4k5TITVKc) has reached its first milestone of $1,000+ MRR, bringing its revenue to 30,000 RMB ($4,000+). Let's share some new learnings and insights from this month's practice:

[2/9] First, let's talk about the revenue data in detail. After just over 30 days, our MRR officially broke through the $1,000 mark, with total revenue of approximately $4,000+, or just over 30,000 RMB. I feel like we're off to a decent start. What was quite surprising was that we only needed 58 subscribers to reach $1,000 MRR. Looking back at this data, I suddenly felt the veil of grand narratives being pierced. In the traditional internet landscape, it seems only products that reach tens or even hundreds of millions of users are considered successful, while those reaching millions or even hundreds of thousands of users are barely even considered successful. But for us, we don't really need to worry about achieving millions or tens of millions of monthly active users. Simply having a thousand to two thousand users willing to pay for our services consistently is enough to keep our team running smoothly over the long term. Open sourcing LobeChat is no longer just a hobby fueled by passion, but can become a truly committed and sustained endeavor. Although we have heard many stories about open source, this feeling is even stronger when we work in this direction and start to get positive feedback.

[3/9] After discussing the optimistic future, let's talk about the challenges we face now: Our current ROI is not good. In the past month, the number of registered users for Cloud has reached over 7,000, but there are only over 60 paying subscribers. This may be because there are too many competitors in the Chat market, and everyone has a lot of options to choose from, and our Cloud is not very competitive at the moment. It may also be that most of the users in the first month came from the open source community, and everyone signed up for Cloud just to show their support 🤣 In short, the current paid conversion rate is less than 100%, which is the biggest problem we are facing. Although LobeChat's overall product experience ranks first in the open source version, our Cloud is still a younger brother compared to mature commercial products such as ChatGPT/Claude/Poe, and there is much to learn.

[4/9] Of course, since we've open-sourced all of LobeChat's code, a low paid conversion rate is probably to be expected. Actually, when we decided to open source, we assumed that people looking for open-source solutions on GitHub were likely freeloaders (including ourselves 🤡), and that it would be difficult to monetize them. 🤑 Furthermore, our recent observations show that paying Cloud users are largely not from the open-source community, which aligns with our initial assumptions. So, from a functional perspective, open-sourcing core features likely won't significantly impact the Cloud version's revenue, as the user base is completely different. Therefore, adhering to this principle, the soon-to-be-released "File Upload/Knowledge Base Conversation" feature (spy photos here~) will also be fully open-sourced. The search plugin currently exclusive to the Cloud version will also be open-sourced to the community version at an appropriate time. We will also continue to differentiate Cloud, focusing on providing value-added services and operating in the assistant market, thereby forming a progressively integrated model with the community open-source version.

[5/9] If we were to point out the pitfalls we encountered after launching the product, perhaps the biggest one was the adoption of a fixed-usage subscription model. Before launching, we thought that since fixed-amount subscriptions are the mainstream approach for AI chat products, we could easily copy them. However, the reality is that fixed-usage subscriptions don't fit in with our current token-based pricing model. After launch, we saw that some heavy users who purchased the Basic plan consumed all their tokens within a few days, becoming unable to continue using the product and forcing them to upgrade. Once their usage reached the Professional plan, they could no longer use the product for the entire month. So, while our product capabilities already support granular token-based billing, a fixed-usage subscription model doesn't fully leverage this advantage. Therefore, our next step is to optimize the subscription model to a relatively low basic service subscription fee combined with on-demand token purchases, reducing the pressure of subscription and usage.

[June 9] It was after a complete review of the front-end, back-end, and payment chain that we discovered numerous flaws in LobeChat's product design. For example, with a token-based payment model, token usage accumulates over multiple conversations, a fact often unaware to new users. While each user's question might not consume many tokens, the accumulated context could accumulate to a significant amount, ultimately leading to over 100,000 tokens consumed in a single conversation request. The result was that even though we'd marked the advanced model as valid for 3,000 conversations, users might experience a limit limit after just a few dozen conversations, leading to a poor experience and a feeling of being cheated. These issues didn't exist in the open source version, as the product service wasn't closed-loop. Users simply entered their own API key and paid for their usage and consumption, with no feedback from us. However, this became a significant issue in our Cloud version, exacerbated by the fixed-consumption subscription model. Therefore, when users inquired about these issues stemming from our product design flaws, we immediately reset their limits. If you have subscribed to Cloud and encounter this problem in your daily use, please feel free to contact me to help you reset your quota. These additional usage losses should be borne by us.

[7/9] Many domestic online retailer platforms now use OneAPI or NewAPI for API key management, and then create a variety of web UIs for users to choose from. However, we wanted users to have a consistent product experience, with a complete, closed-loop user experience that would appear more legitimate. Therefore, during our Cloud implementation, we invested considerable effort in implementing Stripe integration ourselves, and our payment strategy proved highly effective (https://t.co/ooSkW6rNqE). After a round of research, we chose LiteLLM for its better fit and more powerful features for AI gateway and statistics. However, we discovered that it wasn't compatible with our backend management interface. Therefore, we began building Cloud Admin to meet our daily Cloud management needs. This helped us successfully prevent fraudulent practices early on. We will continue to enhance Cloud Admin's capabilities as our management needs evolve. Beyond basic user management, we'll also integrate subscription management, token usage statistics and analysis, GUI-based AI interfaces, load balancing, and other configuration capabilities. Once we've perfected this solution on the Cloud, it has the potential to become a one-stop solution for conversational AI applications. Is anyone interested? These are also interesting areas I wouldn't have touched upon when I was just working on a pure front-end Chat WebUI, but I've discovered interesting areas while working on the Cloud.

[8/9] In the final section, let's talk about costs. While $1,000 in MRR might seem like a lot, so far we've barely broken even, with a slight profit margin. The AI ​​API costs alone account for over half of our total, making us feel like we're working for a large model vendor. 🤣 The operating costs of the various infrastructure needed to maintain our business run into several hundred dollars per month, so overall, we're not making much profit. We focus on MRR because the AI ​​API costs are non-negligible. Monthly token usage is substantial, and only MRR can objectively reflect our profitability. (Even though our current monthly gross revenue is over $4,000, that's spread over the monthly API costs.) We have to mention our partner, AiHubMix @akakenle. We use their services directly on Cloud, and they're 100% official, so the quality and stability of their interfaces are guaranteed. Their API is probably one of the few on the market that directly supports Sonnet 3.5 Tools Calling, and the overall experience is perfect when paired with LobeChat. And even though it was an official transfer, they still gave us a discount, albeit a small amount, which allowed us to make a small profit.

[9/9] Due to space limitations, I'm not covering technical topics this time. If you're interested, we'll open a dedicated thread to discuss this topic after we release the knowledge base. For example, I saw @idoubicc's exorbitant Vercel bill a few days ago, while our other preview site only costs around $30 with the same amount of traffic. I feel like these optimization experiences are worth discussing in detail. So, that's what I've learned so far this month—it's been a ton of experience. We also welcome everyone to try LobeChat, whether it's the open source version or the Cloud version. We welcome suggestions and criticism and strive to make it even better. This is also our product journey from scratch, and I hope you'll witness it together.