Summary

Zilliz’s growth shows that vector databases are emerging from the “desert,” but this was not an overnight boom. The company has been operating for 8 years, with around 130 employees globally and 3.3x revenue growth over the past 12 months; specific revenue and profit figures were not disclosed. 星爵 sees Nvidia’s mention of Zilliz at GTC 2023 as a high point for the category, not a turning point in the company’s fate, because infra is built through “compounding every day, every year.”
The real demand inflection is the expansion of unstructured data from tens of millions or hundreds of millions of records to tens or even hundreds of billions, across model inference, training-data curation, and industry applications. Vector databases now support more than RAG: they are used for multimodal autonomous-driving data, e-commerce recommendations and fraud detection, AI drug discovery, and genomic screening. As scale rises, cutting vector-database costs has become urgent.
Zilliz’s next phase is to expand from a standalone online vector database into a data platform combining a database with Vector Lake. The online system will continue to handle high-accuracy, low-latency queries, while the offline system scans tens to hundreds of billions of records on a daily, weekly, or monthly basis. A customer among the world’s largest IT companies wants to convert every webpage on the internet into vectors, giving large models “internet-scale semantic search.”
“We had no competitors” refers to 2018, when nobody else was in the field—not to an absence of competition today. Pinecone is valued at $750M and Zilliz at $600M. 星爵 says Zilliz currently leads on performance by 3–5x, but stresses that “technology will eventually level the playing field.” The durable advantages are developer feedback from open source, roadmap clarity, and iteration speed.
Open source is not a low-cost shortcut; it is a second mountain that must be crossed alongside commercialization. Since 2018–2019, Zilliz has pursued a dual-core model: the open-source and closed-source engines remain compatible in experience, interfaces, and product form, while the commercial engine is rewritten and kept dynamically 12–18 months ahead. Seamless migration and better performance answer the question of “why pay,” at the cost of two teams, two products, and a lead that is always moving.
One immediate operating pressure is whether Zilliz’s AI customers themselves will survive to the next funding round. In October–November 2024, many of the first wave of AI startups collapsed in a concentrated wave after failing to find PMF, offering undifferentiated products, and being unable to raise again. Zilliz also lost customers among its top 10 over the past 18 months, including a leading US AI company. It has had to replace the gap left by vanished customers while winning new ones to sustain growth—“changing the engine while flying the plane.”
星爵’s worldview has shifted from pure technical idealism toward “be a little ahead, and good enough,” but he still sees idealism as the source of product advantage. His AI investment focus is on public clouds, leading foundation-model companies, and strong AI applications with access to energy and infrastructure. The AI tools he uses most include ChatGPT, DeepSeek, and Cursor; he has even replaced Word with Cursor for writing and knowledge-base management.

Deep dive

1. The Vector-Database High Point Wasn’t the Turning Point—Eight Years of Compounding Were

星爵’s operating snapshot: Zilliz has been operating for 8 years, with around 130 employees globally and 3.3x revenue growth over the past 12 months. The company chose not to disclose specific revenue or profit figures.
Asked about Nvidia singling out Zilliz at GTC 2023, his first thought was not that the company’s fate had changed, but that it was “more a high point for the vector-database category.” When Zilliz entered the market in 2018, the market, users, and use cases were all uncertain; it was not even clear whether vector databases could become a large market.
There is no shortcut in infra that lets one flash of algorithmic insight leapfrog competitors. 星爵 describes patiently refining the product and letting “every day, every year compound.” Gen AI innovation continues to accelerate, but in October–November 2024, many first-wave startups collapsed in a concentrated wave after failing to find real PMF, offering undifferentiated products, and being unable to raise more money.

2. Vectors Are the Data Unit of Neural Networks, Not a Gen AI Add-On

星爵 explains databases through the metaphor of a library: in the age of text, libraries organized information; in the IT era, relational databases stored the numbers and characters used in finance, ERP, and e-commerce. In the AI era, deep learning converts language, images, and video into embeddings—that is, feature vectors.
A vector database stores those vectors at scale and lets developers retrieve unstructured data through natural-language and semantic queries. It predates generative AI because embeddings are the basic format used inside neural networks and in their exchanges with external data; they “aren’t unique to large language models.”
Zilliz was already serving the previous generation of AI companies in 2018, when convolutional neural networks and RNNs were among the dominant approaches. Gen AI pushed vector databases into the spotlight, but it did not create the underlying demand; it expanded the number of vectors and the range of applications.

3. Tens of Billions to Hundreds of Billions of Records Are Turning Databases into Vector Lake

Five or 6 years ago, tens of millions or even hundreds of millions of vectors counted as large scale. Customers now talk about tens or even hundreds of billions. Applications have expanded from knowledge bases and RAG to training-data curation, multimodal autonomous-driving data, e-commerce recommendations and fraud detection, and the screening of 3D protein structures and genetic sequences.
Zilliz’s next step is not simply to scale up the online database, but to combine the vector database with Vector Lake, a data lake for unstructured data. The former handles low-latency, high-accuracy queries; the latter performs offline analysis of tens or hundreds of billions of records on a daily, weekly, or monthly basis, addressing the challenge of querying every record in real time.
星爵 cites a use case at one of the world’s largest IT companies: converting every webpage on the internet into a vector to provide AI search with “internet-scale semantic search.” For large-model queries to be as accurate as possible, the ideal state would be real-time retrieval across all information on the internet.
Enterprise knowledge bases present a different scaling problem: each customer may have only 10,000 records, but 10,000 customers create 100M records and 100,000 customers create 1B. At that point, the challenge is not just capacity, but tenant isolation, data independence, and security.

4. “No Competitors” Belonged to the Desert Phase; Today’s Winner Is the Faster Iterator

星爵 added a qualifier to the provocative headline: “We had no competitors.” In 2018, the concept of a vector database did not yet exist globally. But if nobody followed you for long, “the odds are you took the wrong path,” so the arrival of competitors ultimately made him happy.
Competition with Pinecone is “very tight”: Pinecone is valued at $750M and Zilliz at $600M. The core strategic difference is closed source versus open source. 星爵 says Zilliz currently leads on performance by 3–5x, but does not treat that as a permanent moat because “technology will eventually level the playing field.” The lead still comes from the feedback loop created by the open community.
For open-source projects such as Qdrant and Weaviate, 星爵 summarizes Milvus’s advantages as performance, scalability, and TCO. FAISS is not a comparable database, but a collection of vector-search algorithms. Zilliz describes itself as the largest participant and contributor outside Facebook, and Milvus also uses FAISS extensively as its algorithmic foundation.
Traditional databases adding vector plugins look to 星爵 like “range-extender cars”—a gasoline car with a battery. They can work when data volumes are small and use cases narrow, but as critical workloads scale, customers will still migrate to native systems. LangChain and LlamaIndex are application-layer frameworks, operating at a different layer from databases; the two sides are strategic partners in practice.

5. Milvus’s Product Advantage Ultimately Comes Down to Full-Lifecycle TCO

Hardware costs start with performance: the fewer machines required for the same task, the lower the upfront investment. Over the past 7 or 8 years, Milvus has also continuously adapted to AI frameworks and large language models, reducing development costs through ready-made integrations.
The product boundary now extends beyond vector nearest-neighbor search. Milvus supports scalar filtering, hybrid scalar-and-vector queries, clustering, classification, and re-ranking, as well as more data types and query methods, reducing the amount of component assembly developers must do themselves.
Post-launch costs also matter in the competition. Zilliz provides visualization and operations toolchains and integrates with data collection, permission management, and enterprise access-control systems. That is why 星爵 does not focus on any particular rival; what he really worries about is whether the company can “innovate at a faster pace.”

6. Open Source Is No Shortcut; Dual Core Means Climbing Two Mountains

Even starting over, 星爵 would still choose open source. Responding to Reynold Xin’s view that the company should have started closed source, he said that without open source there would have been no Databricks. The early open-source community helped Databricks raise money and win its first users, while the developer ecosystem could ultimately give it more room to grow than Snowflake.
But “this definitely isn’t a shortcut.” Open source lets a product enter developers’ stacks at low cost and allows users to download and inspect the implementation, but finding commercial PMF can feel like starting a company for the second time. 星爵 argues that the harder the two mountains are to climb, the harder the path is for competitors to copy.
Traditional open core adds enterprise services around an open-source core. It delivers high R&D efficiency, but cannot escape the customer’s “soul-searching question”: what exactly is the difference between the paid and free versions? Databricks’s dual core maintains two engines compatible in interfaces, experience, and product form, while its commercial execution engine is rewritten in C++ and its open-source execution engine uses Java.
The commercial core must stay ahead in design, functionality, and performance—and the lead must be “dynamic.” As the open-source version continues moving forward, the closed-source version must preserve a rolling 12–18-month lead. Zilliz made the decision to follow this path in 2018–2019. The cost is two teams, two products, and much higher demands on engineering, product, and organizational execution.

7. Open Source Competes for Transparency, Exit Rights, and Developer Time

星爵 believes the core value of open source today is not necessarily attracting large numbers of outsiders to write core code, but “laying the technology open” and avoiding a black box. Engineers can study the architecture and implementation details, and if they stop working with the company, they can turn to the community or build a team to maintain and upgrade the system themselves.
Overseas enterprises often adopt open-source projects not to avoid paying, but to prevent lock-in. Even if they stop working with the commercial company in the future, they can return to the community or build an internal team to maintain and upgrade the system. That ability to exit is itself a procurement benefit.
The Milvus community has more than 300 developers, around 20% of them at Zilliz, yet those Zilliz developers contribute 80%–90% of the code. External developers tend to fix bugs, build adjacent tools, add functionality, and handle integrations. 星爵 accepts the reality that a commercial company leads the project, because database core developers require a long time to mature.
DeepSeek’s open-source strategy serves a different purpose: helping a late entrant quickly capture user mindshare and “the most valuable thing—developer time.” 星爵’s view is that once a developer or company installs DeepSeek, they are unlikely to install another provider’s model. Open source therefore becomes a competitive strategy for claiming a position.

8. Pioneers Need 1,000 Attempts; Innovation “Cannot Be Managed”

星爵 says success can only be defined in hindsight, when he retires. Zilliz must become not only a pioneer in unstructured data and vector databases, but also the “integrator” and ultimate winner. He acknowledges that his deepest fear is becoming the pioneer who never captures the final prize.
An innovator may try 1,000 approaches and eliminate 999, while a follower only needs to reproduce the successful answer once. In an AI environment of accelerating iteration, sustained leadership depends on continuous innovation and rapid product iteration, not on having entered the market first.
星爵’s cultural judgment is blunt: “Innovation cannot be managed; the moment you manage it, it dies.” The company’s answer is not to design processes that manufacture innovation, but to hire people who are intrinsically innovative and willing to iterate quickly.

9. Commercialization Hurts More Than Technology; Lost Customers Can Punch Straight Through Growth

If he could speak to himself 8 years ago, 星爵 says he “might have advised himself not to start a company.” Problems never run out. If you choose this path, you have to treat it as a lifestyle and something you want to do for the rest of your life, or “you may break down.”
The hardest period has been roughly the past 2 years. The team had previously focused on product, open source, and technology—all within engineers’ comfort zone. Commercialization was a first: the company lacked experience, set itself an aggressive growth target, and had to build its organization, architecture, and processes as it went.
During the market reset in 2024, one of the leading AI customers in the US quickly ran into trouble; the loss had nothing to do with the quality of the database product. After the correction of the past 18 months, Zilliz lost customers among its top 10. The team had to fill the hole left by disappearing customers while winning more customers to sustain growth—“changing the engine while flying the plane, and assembling it at the same time.”
星爵’s operating playbook is not complicated: always spend more time hiring the right people, recover quickly from mistakes, rebuild morale by “winning battles,” and avoid repeating errors—especially errors of principle. He ultimately reduces competition to a test of endurance: “Your competitor may have made 2 more mistakes than you,” provided you first make peace with your own imperfections.

10. Idealism Fades to Gray; AI Investment Goes to Clouds, Models, and Applications

Eight years of entrepreneurship took 星爵 from a “colorful outer layer” to “gray underwear.” He once believed teams should be completely transparent, that the best companies did not need management, and that management was synonymous with bureaucracy. As the company grew, he acknowledged that transparency depends on the audience, context, and stage, and that “management really is a science.”
His technical aesthetic has also shifted from endlessly pursuing something better to the business-world standard of “good enough is fine”: be somewhat better than the competitor, but not necessarily much better. Citing Intel and Nvidia’s incremental “toothpaste-squeezing” approach, he argues that excellent companies understand how to coordinate the pace of innovation with the pace of commercialization.
This realism does not negate the starting point. The pure pursuit of technical excellence is the root of Zilliz’s current product and technology advantages. Even as the company becomes more business-savvy, 星爵 still believes it matters which side he started from.
星爵 singled out Amazon’s cloud platform as certain to see better growth and argued that AI has reached a point where competition is increasingly about energy and infrastructure. Large data centers will become cloud infrastructure, potentially benefiting resource-rich companies and public clouds. Next come leading foundation-model companies and stronger AI application companies. The AI tools he uses most include ChatGPT, DeepSeek, and Cursor; he even uses Cursor to manage his knowledge base, write articles, and replace Word.

20VC: Inside Carnegie Mellon's $4BN Endowment with Miles Dieffenbach The AI Revolution in Education with Shawn Jansepar, Director of Engineering at Khan Academy