Ion Stoica on Open vs. Closed Source and Infrastructure Startups

Summary

• AI infrastructure startups can win by identifying the core problems in emerging workloads and solving them exceptionally well. Ion Stoica sees hardware, applications, and efficiency optimization becoming increasingly intertwined, but startups can still enter through post-training, multimodal data, robotics, and cross-cloud training and data movement: “Solve these problems and excel at them.”
• Post-training could keep expanding because it can materially improve existing models on specific tasks. A large pool of competitive open models—many of them from China—provides ready-made foundations; after post-training, they could even outperform the best proprietary models on certain tasks.
• Compute scarcity is turning multicloud portability and data movement into infrastructure opportunities. Even large companies may need to stitch together GPUs from multiple cloud providers and their own data centers; when training data reaches “several trillion or more,” SkyPilot-style scheduling across locations and data transfer both matter.
• For closed-source labs, the clearest competitive factor today is capital, not any other technical factor that has been proven. No company has yet shown it can build this class of model through something other than capital; capital is not enough—revenue is also needed. Model rankings rotate every few months: at the end of last year it might have been Gemini, followed by Claude, leading Ion to conclude that “no lab has fully dominated.”
• Open models do not need to match Frontier Lock models across the board; they only need to clear the usability threshold for a given task. Ion uses a car analogy: if all you need is a car that can safely drive between Santa Clara and San Francisco at 100 miles per hour, its ability to go 300 miles does not matter unless the task itself is a race; more tasks will therefore be covered by open models that are good enough—indeed, more than good enough.
• A model’s “taste” does affect the user experience, but it looks more like UX and communication quality than a durable lab moat. For the same concept, one person may make it click in 10–15 minutes while another can spend 3 hours explaining it and still leave people confused; but if a lab’s edge really came down to a handful of people with exceptional taste, model rankings would not reshuffle roughly every two months.
• Arena’s vision is to use human preferences from real-world tasks to build better AI, rather than relying solely on static leaderboards. Ion wants AI to be more accurate, reliable, predictable, and cost-effective; static benchmarks still matter, but “if we are building systems for humans, the ultimate measure is humans themselves.”

Deep dive

1. AI Infrastructure Is Becoming a Whole-System Problem

• Ion Stoica’s first-order view is that innovation in both hardware and applications is accelerating, while the cost of AI workloads is forcing every layer to pursue maximum efficiency. Components are therefore becoming “tightly coupled and highly integrated,” increasingly requiring co-development as one large system.

• Even amid this complexity, the opportunity for startups remains in “identifying core problems, solving them, and excelling at them,” particularly by focusing on emerging problems and emerging workloads.

• As AI pushes into every kind of system, cybersecurity, and even human-risk scenarios, he argues that people should first “calm down”: AI is a complex tool that can exhibit unexpected behavior, much like the internet and large distributed systems. The right response is to learn from attacks, improve security, and guide these systems—without forgetting that these tools are ultimately meant to lift productivity and improve lives.

2. Post-Training, Robotics, and Data Movement Are Near-Term Startup Wedges

• One of the emerging workloads Ion is most bullish on is post-training. It can effectively improve an existing model’s performance on specific tasks, while a supply of competitive open models provides ready-made foundations for the work.

• Chinese open models have expanded this opportunity in particular: companies can post-train them and may surpass the best proprietary models on specific tasks.

• Multimodal data is a second bottleneck. As models process video, images, text, voice, and audio simultaneously, handling this data better will become a larger infrastructure challenge.

• Robotics links the three issues of models, data, and the open stack. The industry is still early, and many companies are still building models; Ion argues that an open stack suited to robotics applications, without requiring the construction of models with several trillion parameters, is “very important—and an opportunity.”

3. Open Source Does Not Need to Win the Overall Leaderboard—Only Clear the Task Threshold

• 曹卿云 pressed Ion on the long-term gap between Frontier Lock models and open models. Ion’s answer was clear in its stance but grounded in reality: “I firmly believe in the open-source spirit”（我坚信开源精神）; a competitor may still be able to build a competitive model with less capital.

• The rapid turnover in model rankings weakens the case for a permanent technical moat. Ion said Gemini may have been the best at the end of last year, only for the lead to shift clearly to Claude a few months later; some models may be slightly ahead on coding, while Oppus 5.5 is also very good. No lab has dominated for long.

• More important is not whether open source can match across every dimension, but whether the task has a capability threshold. Ion’s car analogy is the clearest: “If you need a car to drive safely at 100 miles per hour,” its ability to go 300 miles does not matter unless the task itself is a race.

• 曹卿云 relayed Versel’s observation, explicitly noting that he did not remember the exact number: a few months ago, proprietary models accounted for roughly 70%; more recently, open models also accounted for roughly 70%. The point was to show that the model-selection landscape is changing.

4. Taste Shapes the Experience, but Is Not Yet a Durable Moat

• Ion accepts that “taste” matters to the model user experience, but grounds it in communication and UX: how a model structures an answer and when it asks the user to clarify the task determine whether users can tap more of its capabilities. “If you’re building systems for humans, you have to account for human needs.”

• 曹卿云 took the idea further, suggesting that differences among labs may come from people with better taste. Ion’s response was that if this were really the case, rankings would not change every two months; several labs are currently close in capability, capital can still attract talent, and talent itself also moves between labs.

5. Open Contributions and Real-Human Evaluation Determine the Pace of Model Iteration

• Ion believes the structural advantage of open source, when it works well, is that it lets innovation diffuse beyond lab boundaries. Using Linux as a reference point, he wants AI to open not only the models but, as far as possible, the data, the software—including everything used in training—and the evaluations, so that “people around the world can contribute.”

• Arena’s vision is to help people build AI that is “more accurate, more reliable, more predictable, and of course more cost-effective.” To answer which model is best suited to a given task, evaluation is not an ancillary feature but a necessary selection mechanism.

• Ion recognizes the value of static benchmarks but does not treat them as the final yardstick. The best evaluation method is to have real users run real tasks and solve their own problems, then assess the model or agent through explicit or implicit preferences: “the ultimate measure is humans themselves.”