Databricks Co-Founder Ion Stoica: Capital Is AI Labs' Only Moat

Summary

• Ion Stoica argues that capital is the leading moat at top AI labs, the one that has been demonstrated so far; capital and revenue together make a strong moat, while no technical moat is yet visible. Rankings shift about every 2 months: Gemini was widely seen as the strongest model at the end of last year, followed by Claude in the months after; he also says Opus 5.5 is very good at coding. Capital can attract talent, but he thinks it explains competition among labs better than talent does.
• Open-source models can win market share without catching up to the frontier because “every task has a threshold.” Stoica compares it to driving: if the task is getting from Santa Clara to San Francisco, a car that does 100 miles per hour is enough; a car that can go 300 miles is beside the point. He cites Vercel routing data showing that a few months ago about 70% of traffic went to proprietary models, while now about 70% goes to open-source models; he says he does not remember the exact figures.
• The opportunities for AI infra startups lie in emerging workloads: post-training, multimodal data, robotics stacks, and moving data across clusters. Post-training can substantially improve an existing model’s performance on specific tasks; strong open-weight models, many of them from China, provide a base for that work and could, once improved, outperform the best proprietary models on some tasks. When GPUs are scarce, large companies may source capacity from multiple cloud providers or build their own data centers; training across data centers on tens of trillions of tokens or more makes data movement and portability a challenge, which SkyPilot is working to address.
• Stoica agrees that taste matters, but does not think it has become a moat for AI labs. He compares it to explaining the same concept: some people can make it clear in 10–15 minutes, while others can talk for 3 hours and still leave you confused. If talent’s taste were a moat, model rankings would not shift about every 2 months; capital attracts talent, and talent moves between labs.
• He argues that AI security should improve by learning from attacks and steering systems toward desired outcomes. The internet and large distributed systems have also behaved in unexpected ways; but AI is ultimately a tool for raising productivity and improving lives.
• Arena’s vision is to have real people evaluate AI on real tasks. Its goal is to help people build AI that is more accurate, reliable, predictable, and cost-effective. Stoica does not dismiss the importance of human or static benchmarks, but argues that if systems are built for people, the ultimate evaluation should reflect people’s real preferences about model or agent performance.

Deep dive

1. AI infra is becoming a tightly coupled system

• Stoica’s opening argument: hardware and applications are getting more complex, and innovation is accelerating on both ends. With AI workloads extremely expensive and everyone pushing for maximum efficiency, “everything is tightly coupled, everything is integrated”—the whole stack has to be developed together as one large system.
• On AI attacks and cybersecurity, he argues for stepping back: “AI is ultimately a system, a tool.” Complex systems behaving in unexpected ways is nothing new; the internet and large distributed systems do it too. The goal should be to understand these attacks as much as possible, use them to improve security, and steer systems toward the outcomes we want—without forgetting that these tools are meant to raise productivity and improve lives.

2. Beyond vertically integrated giants, startup opportunities lie in emerging workloads

• He points to 4 areas: post-training, multimodal data processing, robotics stacks, and moving data across clusters. Post-training is growing because it can improve existing models and make them perform better on specific tasks. Strong open-weight models, many of them from China, provide a base for this work and, after improvement, can outperform the best proprietary models on some tasks.
• Robotics is still early, and many companies are still building models. An open-source stack suited to robotics applications could matter. Robots do not need trillion-parameter models, and the problem is “easier to tackle,” leaving room for startups.
• When GPUs and AI chips are scarce, even large companies may source capacity from multiple cloud providers or build their own data centers. The next challenge is pulling these resources together so training can span different data centers and clusters, while improving portability. SkyPilot is working on these problems, including moving data efficiently between sites; training runs may involve tens of trillions of tokens or more.

3. Capital is the labs’ main moat—and shifting rankings are his evidence

• Stoica first notes that he is a committed open-source advocate, and that this shapes his view. His assessment is that capital is the biggest moat for the largest labs. Competitors that can match their funding—or have less but still enough—can build competitive models. No lab has an absolute lead: one may be ahead by a few months, then the gap closes.
• He says Gemini was arguably the best model at the end of last year, with Claude leading in the months after. He also mentions names rendered in the transcript as Fabul [?] and Astra [?], and says Opus 5.5 is also very good at coding. Among Chinese models, DeepSeek surprised everyone in February last year; now GLM 5.3 and Qwen 3 are arguably better on at least some tasks.
• Stoica sees no visible technical moat at the labs yet, which is why open source still has an opportunity. He wants open source to become more widespread and narrow the gap, but says it does not have to close completely: every task has a threshold for “good enough,” and beyond that, being 10 times more capable may not matter. His car analogy: if all you need to do is drive from Santa Clara to San Francisco, 100 miles per hour is enough; 300 miles per hour does not matter unless you are racing.
• He cites Vercel’s report that its AI router has seen a shift in traffic over the past few months: a few months ago, about 70% went to proprietary models; now, about 70% goes to open-source models. He says he does not remember the exact figures.

4. Taste matters, but is not a visible moat for labs; open source spreads innovation

• 曹卿云 invokes Steve Jobs to ask whether taste is a moat. Stoica agrees that it matters: faced with the same difficult concept, one person can make it clear in 10–15 minutes, while another can talk for 3 hours and still leave you confused. How a model delivers an answer, responds to a task, and asks for clarification also matters. Interpersonal problems often come from poor communication; systems should account for people’s needs and “meet people where they are.”
• But when asked whether labs build a moat with more tasteful talent, he says that if they did, model rankings would not change about every 2 months. Google once had the edge, and Microsoft also began to gain one; talent across several labs is broadly comparable. He sees capital as more important because it attracts talent, and talent moves between labs.
• He argues that open source spreads innovation by letting people around the world participate; contributors at each lab are currently mostly confined to the lab itself. Citing Linux, he hopes AI becomes increasingly open—not just open weights, but also data, training software, and evaluations wherever possible—so more people can contribute and everyone can benefit.

5. Arena’s bet: humans are the ultimate benchmark

• Arena’s ultimate goal is to help people build better AI: more accurate, reliable, predictable, and cost-effective. That requires evaluation; otherwise, it is difficult to know which model is right for a given task.
• Arena believes good evaluation means having real people use real tasks to solve real problems, then express implicit or explicit preferences about how a model or agent performs. Stoica does not dismiss the importance of manual or static benchmarks. His central point is that if systems are built for people, the ultimate benchmark is people.