在哈里·斯特宾斯（Harry Stebbings）的20VC节目的一次广泛采访中，OpenRouter 的联合创始人兼首席执行官亚历克斯·阿塔拉（Alex Atala）探讨了人工智能的未来、他的公司作为大型语言模型（LLMs）入口的作用以及更广阔的生态系统。OpenRouter 致力于促进对各种大型语言模型的访问，据报道估值超过15亿美元，并一直有传言称 Stripe 将以100亿美元收购该公司，但阿塔拉对此不予置评。 阿塔拉从他在 OpenSea 工作期间的经验中汲取教训，强调了强大基础设施和可扩展性的重要性。他将这一经验应用到 OpenRouter 上，确保了高可用性，并能抵御人工智能领域常见的不可预测的流量峰值。 OpenRouter 的核心理念围绕着人工智能的“多模型未来”和“神经多样性”。阿塔拉认为，公司不会只依赖单一模型，而是会整合各种模型来完成不同的任务，并激发创造力。他强调了专用推理提供商（如 Fireworks 和 Together AI）的出现，这些提供商在托管开源模型方面优于超大规模提供商，并指出它们在速度、处理边缘情况的创新以及实时性能方面的优势。他不同意推理层容易商品化的观点，理由是英伟达（NVIDIA）对多元化客户群的兴趣以及提供商在令牌效率方面的不断创新。OpenRouter 根据实时质量、速度和成本动态路由流量，确保用户始终能获得最符合其需求的最佳性能模型。 关于 OpenRouter 的商业模式，阿塔拉解释了他们从5.5%的抽成率（take rate）演变为基于承诺消费的企业方案，对于自带推理的用户则不收取任何费用。他预计未来的收入将主要来自帮助企业和初创公司管理计划外推理能力，尤其是在市场持续低估其不断增长的人工智能需求的情况下。他引用了“杰文斯悖论”（Jevons paradox），即 OpenRouter 上 GPT 5.6 Luna 等模型价格下降10倍，导致使用量增加13倍，这说明了令牌价格下降仍然能使他们的业务受益。 阿塔拉承认，OpenRouter 的令牌量可能低估了前沿模型，因为他们的平台迎合的是寻求多模型解决方案的用户。他指出，美国公司在使用前沿模型时常常更加紧张，更多是出于对数据政策和缺乏控制的担忧，而非针对中国模型。他还讨论了模型开发的巨大速度，OpenRouter 在一个月内就推出了70个模型，并预测代理实验室（agent labs）将越来越多地开发自己的模型。 在竞争格局方面，阿塔拉表示担忧，认为美国在开源模型方面“远远落后”于中国，中国得益于集中的国家支持以及针对外部使用模型较少的内部监管障碍。他建议鼓励对中国模型进行蒸馏（distillation），并利用美国硬件公司向美国新兴实验室提供计算资源，以培育一个更具竞争力的生态系统。他辩护说，蒸馏是一种常见的、道德的模型构建技术，即使是闭源实验室也在使用。 他还谈到了开发者忠诚度，指出尽管 OpenRouter 的目标是零转换成本，但开发者通常会因为稳定性、成本效益以及对输出的信任而坚持使用某些模型。他认为“记忆”是一个关键的保留机制，应用程序可能在存储重要的上下文信息方面发挥关键作用。他将“harnesses”（集成框架/工具）与传统应用程序区分开来，将其描述为更具可组合性、确定性和用户友好性，为开发者提供了更大的灵活性。他还认为 Meta 的 Muse 有潜力，一旦确定了明确的利基市场，就可能成为人工智能领域的一个强大挑战者。 在快速问答环节中，阿塔拉将 Poolside 列为一个被低估的模型。他不同意70%的新兴实验室会失败的预测，认为更接近50%，其中包括整合。他赞赏 Anthropic 的“偏执”，认为这是人工智能安全讨论中不可或缺的声音。他“最疯狂”的观察是人工智能时代员工成本的动态变化，即个人在使用昂贵模型与廉价模型时的效率直接影响其被感知的价值。最后，他对人工智能加速罕见病研究的潜力以及众包城市和乡村生活质量改善解决方案的潜力表示兴奋。

In a wide-ranging interview on 20VC with Harry Stebbings, Alex Atala, co-founder and CEO of OpenRouter, discussed the future of AI, his company's role as a gateway to LLMs, and the broader ecosystem. OpenRouter, which facilitates access to various large language models, is reportedly valued at over $1.5 billion and has been subject to rumors of a $10 billion acquisition by Stripe, which Atala declined to comment on. Drawing lessons from his time at OpenSea, Atala emphasized the importance of robust infrastructure and scalability. He applied this experience to OpenRouter, ensuring high uptime and resilience against unpredictable traffic spikes common in the AI space. OpenRouter's core thesis revolves around a "multi-model future" and "neurodiversity" in AI. Atala believes that companies will not rely on a single model but will integrate various models for different tasks and to foster creativity. He highlighted the emergence of dedicated inference providers (like Fireworks and Together AI) as superior to hyperscalers for hosting open-weight models, noting their speed, innovation in handling edge cases, and real-time performance. He rejected the idea that this inference layer is easily commoditizable, citing NVIDIA's interest in a diverse customer base and providers' constant innovation in token efficiency. OpenRouter dynamically routes traffic based on real-time quality, speed, and cost, ensuring users always get the best performing model for their needs. Regarding OpenRouter's business model, Atala explained their evolution from a 5.5% take rate to enterprise plans based on committed spend, with no fees for users bringing their own inference. He anticipates future revenue will largely come from helping enterprises and startups manage unplanned inference capacity, especially as the market continues to underestimate its growing AI needs. He cited the "Jevons paradox" where a 10x price drop for models like GPT 5.6 Luna on OpenRouter led to a 13x increase in usage, illustrating how falling token prices can still benefit their business. Atala acknowledged that OpenRouter's token volumes might undercount frontier models, as their platform caters to users seeking multi-model solutions. He noted that US companies are often more nervous about using frontier models due to concerns over data policy and lack of control, rather than Chinese models. He also discussed the immense rate of model development, with OpenRouter launching 70 models in a month, and predicted agent labs will increasingly develop their own models. On the competitive landscape, Atala expressed concern that America is "very, very behind" in open-weight models compared to China, which benefits from concentrated state support and fewer internal regulatory hurdles for models designed for external use. He proposed encouraging the distillation of Chinese models and leveraging US hardware companies to provide compute resources to American neo-labs to foster a more competitive ecosystem. He defended distillation as a common, ethical technique for building models, even used by closed-source labs. He also touched on developer loyalty, noting that while OpenRouter aims for zero switching costs, developers often stick to models due to stability, cost-effectiveness, and trust in outputs. He believes "memory" is a key retention mechanism, with apps likely playing a crucial role in holding important contextual information. He differentiated "harnesses" from traditional applications, describing them as more composable, deterministic, and user-friendly, offering developers greater flexibility. He also sees potential in Meta's Muse to become a serious challenger in the AI space once they define a clear niche. In a quick-fire round, Atala named Poolside as an underrated model. He disagreed with the prediction that 70% of neo-labs would fail, suggesting closer to 50% including consolidation. He appreciated Anthropic's "paranoia" as a necessary voice in AI safety discussions. His "craziest" observation was the shift towards dynamic employee costs in the AI era, where an individual's efficiency in using expensive versus cheap models directly impacts their perceived value. Finally, he expressed excitement about AI's potential to accelerate rare disease research and to crowdsource solutions for urban and rural quality-of-life improvements.