VAST Data CEO：AI 中间层的隐形成功与爆发需求VAST Data CEO: The Hidden $30B Layer Feeding GPUs
The Takeaway：AI 工厂的真正瓶颈不在模型，而在喂养 GPU 的软件基础设施层，VAST Data 正以“共享一切”架构成为这个被忽视的操作系统，估值已达 300 亿美元。
Renen Hallak 是 VAST Data 创始人兼 CEO。公司为 xAI、CoreWeave、Nebius、Mistral 等顶级 AI 玩家提供存储、数据库与模型管理，却鲜为人知。他用 Jensen 的五层蛋糕比喻：最底层是电力与硬件，中间是软件基础设施，上层是模型与应用。VAST 就坐在中间，把旧栈的分片共享无架构彻底推翻，改成 disaggregated shared everything，让成千上万节点像直接连着一样共享所有数据。
他最令人印象深刻的观察是需求失控加速：“我们有个 AI 云客户三个月前说未来三年需要 500 PB，上周回来改口说还要额外加 2 EB。”训练相对简单，推理与 agent 时代则复杂得多，需要模型路由、KV cache、RAG、agent 记忆与细粒度权限。为解决企业数据与模型权重互不信任的问题，VAST 刚发布 Data Enclave：借助 NVIDIA 加密内存，让推理在企业侧跑，两边都看不到对方的敏感信息。
Hallak 认为软件基础设施层会像当年的云与移动 OS 一样沉淀大部分价值，而模型层可能与应用层合并。他从 xAI 学到的速度哲学是“找到限制因素就干掉它”，并坚持“坏事要大声说、经常说”。十年后如果 AI 超越人类智能并接入物理世界，“未来十年的变化将超过过去一千年”。
The Takeaway: The real bottleneck in AI factories is not models but the software infrastructure that feeds GPUs; VAST Data has become the overlooked operating system for this layer and is now valued at $30B.
Renen Hallak is founder and CEO of VAST Data, the company quietly powering xAI, CoreWeave, Nebius, Mistral and other top AI players with storage, database and model-management software. Using Jensen’s five-layer cake, he places VAST squarely in the middle software-infrastructure layer. The firm discarded the old shared-nothing sharding model in favor of a disaggregated shared-everything architecture so tens of thousands of nodes can see every byte as if it were local.
Demand is accelerating faster than anyone planned: “One of these AI clouds told us three months ago they would need about 500 petabytes over the next three years. Last week they came back and said we’re gonna need an extra two exabytes on top of that.” Training is relatively simple; inference and agents introduce model routing, KV caches, RAG, agent memory and fine-grained identity. To solve the trust gap between enterprise data and model weights, VAST just launched Data Enclave, using NVIDIA encrypted memory so inference runs inside the customer’s environment while both sides remain blind to each other’s secrets.
Hallak expects the software-infrastructure layer to capture the bulk of value the way cloud and mobile operating systems did, while models may eventually merge with applications. From xAI he learned “find the limiting factor and get rid of it,” and he runs the company on the rule “bad things should be stated loudly and often.” If AI surpasses human intelligence and gains physical access, “in the next ten years we’ll see more difference than we did in the last thousand years.”
查看原文 →查看原文 →
Renen Hallak 是 VAST Data 创始人兼 CEO。公司为 xAI、CoreWeave、Nebius、Mistral 等顶级 AI 玩家提供存储、数据库与模型管理，却鲜为人知。他用 Jensen 的五层蛋糕比喻：最底层是电力与硬件，中间是软件基础设施，上层是模型与应用。VAST 就坐在中间，把旧栈的分片共享无架构彻底推翻，改成 disaggregated shared everything，让成千上万节点像直接连着一样共享所有数据。
他最令人印象深刻的观察是需求失控加速：“我们有个 AI 云客户三个月前说未来三年需要 500 PB，上周回来改口说还要额外加 2 EB。”训练相对简单，推理与 agent 时代则复杂得多，需要模型路由、KV cache、RAG、agent 记忆与细粒度权限。为解决企业数据与模型权重互不信任的问题，VAST 刚发布 Data Enclave：借助 NVIDIA 加密内存，让推理在企业侧跑，两边都看不到对方的敏感信息。
Hallak 认为软件基础设施层会像当年的云与移动 OS 一样沉淀大部分价值，而模型层可能与应用层合并。他从 xAI 学到的速度哲学是“找到限制因素就干掉它”，并坚持“坏事要大声说、经常说”。十年后如果 AI 超越人类智能并接入物理世界，“未来十年的变化将超过过去一千年”。
The Takeaway: The real bottleneck in AI factories is not models but the software infrastructure that feeds GPUs; VAST Data has become the overlooked operating system for this layer and is now valued at $30B.
Renen Hallak is founder and CEO of VAST Data, the company quietly powering xAI, CoreWeave, Nebius, Mistral and other top AI players with storage, database and model-management software. Using Jensen’s five-layer cake, he places VAST squarely in the middle software-infrastructure layer. The firm discarded the old shared-nothing sharding model in favor of a disaggregated shared-everything architecture so tens of thousands of nodes can see every byte as if it were local.
Demand is accelerating faster than anyone planned: “One of these AI clouds told us three months ago they would need about 500 petabytes over the next three years. Last week they came back and said we’re gonna need an extra two exabytes on top of that.” Training is relatively simple; inference and agents introduce model routing, KV caches, RAG, agent memory and fine-grained identity. To solve the trust gap between enterprise data and model weights, VAST just launched Data Enclave, using NVIDIA encrypted memory so inference runs inside the customer’s environment while both sides remain blind to each other’s secrets.
Hallak expects the software-infrastructure layer to capture the bulk of value the way cloud and mobile operating systems did, while models may eventually merge with applications. From xAI he learned “find the limiting factor and get rid of it,” and he runs the company on the rule “bad things should be stated loudly and often.” If AI surpasses human intelligence and gains physical access, “in the next ten years we’ll see more difference than we did in the last thousand years.”