VAST Data：AI时代的操作系统，价值300亿美元的隐藏层VAST Data: The $30B Hidden Layer Feeding AI's GPUs
The Takeaway：AI工厂的真正瓶颈不在模型或算力，而在中间那层软件基础设施，它决定了数据如何高效、安全地喂给GPU，并让企业在不泄露数据的情况下使用外部模型。
VAST Data创始人兼CEO Renen Hallak打造了这层“新纪元的操作系统”。公司最新估值300亿美元，客户包括xAI和全球最大的AI neo-clouds。Hallak用NVIDIA的五层蛋糕类比：电力、硬件、软件基础设施、模型、应用。VAST就坐在中间，负责存储、网络、数据库、训练/推理编排，以及未来的模型管理和agent记忆。
他指出，AI工厂最终会成为每个企业和国家的标配，就像电力一样。企业需要把IP蒸馏进自己的权重里，通过强化学习让agent不断进化，而不是把敏感数据交给OpenAI或Anthropic。为此VAST刚发布了模型管理与机密计算（enclave）能力，利用NVIDIA加密内存，让推理可以在企业本地运行，模型权重与用户数据双向隔离。
Hallak最震撼的观察是需求爆炸：一家AI云客户三个月前说未来三年要500PB，上周直接加码到额外2EB。他判断，未来十年的变化将超过过去一千年：“I think in the next ten years, we'll see more difference than we did in the last thousand years.”
在商业模式上，VAST卖软件、毛利高、零主动流失、客户每年数据量翻2-3倍，在AI供应链中罕见地实现了盈利。Hallak认为软件基础设施层将像微软、亚马逊一样长期积累巨大价值，而模型层最终可能与应用层融合。
VAST Data创始人兼CEO Renen Hallak打造了这层“新纪元的操作系统”。公司最新估值300亿美元，客户包括xAI和全球最大的AI neo-clouds。Hallak用NVIDIA的五层蛋糕类比：电力、硬件、软件基础设施、模型、应用。VAST就坐在中间，负责存储、网络、数据库、训练/推理编排，以及未来的模型管理和agent记忆。
他指出，AI工厂最终会成为每个企业和国家的标配，就像电力一样。企业需要把IP蒸馏进自己的权重里，通过强化学习让agent不断进化，而不是把敏感数据交给OpenAI或Anthropic。为此VAST刚发布了模型管理与机密计算（enclave）能力，利用NVIDIA加密内存，让推理可以在企业本地运行，模型权重与用户数据双向隔离。
Hallak最震撼的观察是需求爆炸：一家AI云客户三个月前说未来三年要500PB，上周直接加码到额外2EB。他判断，未来十年的变化将超过过去一千年：“I think in the next ten years, we'll see more difference than we did in the last thousand years.”
在商业模式上，VAST卖软件、毛利高、零主动流失、客户每年数据量翻2-3倍，在AI供应链中罕见地实现了盈利。Hallak认为软件基础设施层将像微软、亚马逊一样长期积累巨大价值，而模型层最终可能与应用层融合。
The Takeaway: The real bottleneck in the AI factory is not models or GPUs, but the middle software-infrastructure layer that feeds data at massive scale, keeps both enterprise data and model weights confidential, and turns agents into continuously learning systems.
Renen Hallak, founder and CEO of VAST Data (valued at $30B), sits exactly in that middle layer. The company powers xAI and many of the largest AI neo-clouds. Using Jensen’s five-layer cake analogy—power, hardware, software infrastructure, models, applications—VAST owns the operating system that manages storage, networking, databases, training/inference orchestration, and now model management plus agent memory.
Hallak argues every enterprise and nation will eventually need its own AI factory. Organizations must distill their IP into private weights via reinforcement learning so agents keep improving without handing proprietary data to the big model labs. VAST’s latest announcement adds confidential computing enclaves: inference runs inside the customer’s environment while model weights stay encrypted end-to-end using NVIDIA hardware, protecting both sides.
Demand is already outrunning supply. One AI-cloud customer told VAST three months ago it would need 500 PB over three years; last week it came back asking for an extra 2 EB on top. Hallak’s longer view is stark: “I think in the next ten years, we’ll see more difference than we did in the last thousand years.”
Because VAST sells software with high gross margins, it is already profitable, has zero active churn, and sees customers double or triple data every year—rare in the AI stack. He believes the software-infrastructure layer will capture the same enduring value Microsoft and Amazon captured in previous platform shifts, while the model layer may ultimately merge with applications.
查看原文 →
Renen Hallak, founder and CEO of VAST Data (valued at $30B), sits exactly in that middle layer. The company powers xAI and many of the largest AI neo-clouds. Using Jensen’s five-layer cake analogy—power, hardware, software infrastructure, models, applications—VAST owns the operating system that manages storage, networking, databases, training/inference orchestration, and now model management plus agent memory.
Hallak argues every enterprise and nation will eventually need its own AI factory. Organizations must distill their IP into private weights via reinforcement learning so agents keep improving without handing proprietary data to the big model labs. VAST’s latest announcement adds confidential computing enclaves: inference runs inside the customer’s environment while model weights stay encrypted end-to-end using NVIDIA hardware, protecting both sides.
Demand is already outrunning supply. One AI-cloud customer told VAST three months ago it would need 500 PB over three years; last week it came back asking for an extra 2 EB on top. Hallak’s longer view is stark: “I think in the next ten years, we’ll see more difference than we did in the last thousand years.”
Because VAST sells software with high gross margins, it is already profitable, has zero active churn, and sees customers double or triple data every year—rare in the AI stack. He believes the software-infrastructure layer will capture the same enduring value Microsoft and Amazon captured in previous platform shifts, while the model layer may ultimately merge with applications.