Tensor9 Targets AI Privacy Bottlenecks to End Two-Product Dilemma

Solving AI's Data Privacy Dilemma: How to Consume AI Software Without Moving Your Data With Tensor9 — Ai4

Ai4

Deploying unified infrastructure-as-code directly into customer-owned environments eliminates the hundredfold operational cost penalty of building separate on-premises AI software variants.

AI-assisted summary of Solving AI's Data Privacy Dilemma: How to Consume AI Software Without Moving Your Data With Tensor9 — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 24, 2025, this presentation features Michael Ten-Pow, Founder and CEO of Tensor9. The video outlines how enterprises struggle to adopt AI due to the security risks and costs of data movement, exploring software consumption models and introducing an approach that enables AI utilization without data egress.

## TL;DR Enterprises frequently block AI adoption because security and compliance rules prohibit sensitive data from leaving internal company boundaries. While vendors can offer alternative hosting models—such as single-tenant SaaS, Cloud Premise/BYOC, on-premises, and air-gapped environments—supporting customer-hosted variants traditionally introduces an unsustainable 'two-product penalty' that inflates maintenance costs up to 100x. Tensor9 solves this impasse by allowing AI vendors to define software once using Infrastructure as Code (IaC) and deploy it seamlessly into any customer perimeter. This architecture guarantees strict data locality for enterprise buyers while maintaining a unified, continuous-delivery SaaS experience for vendors.

## Atmosphere & style The presentation carries a pragmatic, engineering-focused tone grounded in practical cloud architecture and enterprise sales realities. Ten-Pow speaks with the steady, authoritative cadence of a seasoned AWS and Amazon infrastructure veteran diagnosing familiar industry bottlenecks. The delivery feels crisp and analytical, pivoting methodically from the friction of vendor trust hurdles to a clear architectural taxonomy and a clean, unifying technical solution.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) The primary barrier preventing enterprises from adopting cutting-edge AI software is the refusal of security and legal teams to let sensitive data leave their internal perimeter. - [00:45](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=45s) Similar to clean room technologies developed for ad-tech data sharing, the enterprise AI dilemma fundamentally stems from a lack of mutual trust between vendors and customers. - [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) Multi-tenant SaaS creates maximum compliance friction and security audit overhead because proprietary enterprise data is transmitted to shared vendor cloud accounts. - [02:50](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=170s) While single-tenant SaaS provides workload isolation, heavily regulated industries like banking and healthcare still reject it because data leaves customer-controlled environments. - [03:40](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=220s) Customer-hosted architectures—such as Cloud Premise (BYOC), On-Premises bare metal, and Air-Gapped installations—guarantee data privacy by keeping compute inside the customer's security boundary. - [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) Supporting customer-hosted deployments traditionally imposes a 'two-product penalty' on AI vendors, forcing them to maintain a SaaS version alongside disconnected on-prem codebases. - [06:40](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=400s) Disconnected customer-premise deployments drastically increase operational friction, inflating maintenance and support costs by as much as 100x for both vendors and enterprises. - [07:15](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=435s) Tensor9 enables vendors to write software once using standard Infrastructure as Code (IaC) and deploy seamlessly across AWS, Azure, GCP, bare metal, and air-gapped environments. - [07:40](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=460s) By abstracting multi-environment deployment, vendors retain a unified product pipeline while customers receive full data locality with a SaaS-like operational experience.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) Overcoming the Trust Hurdle in Enterprise AI

Michael Ten-Pow, founder and CEO of Tensor9, opens by framing the central dilemma facing modern enterprise AI: organizations want to adopt cutting-edge AI tools to process sensitive enterprise data, but IT, legal, and security teams refuse to let that data leave their "four walls" to reside with a third-party vendor. Drawing on his engineering career, Ten-Pow explains that he spent 13 years at Amazon, including six years at AWS during its foundational era between 2007 and 2013, followed by a tenure in Amazon’s retail division working at the intersection of advertising and data privacy.

While at Amazon Retail, Ten-Pow repeatedly encountered roadblocks when attempting to collaborate with major ad platforms like Facebook and Google. Both parties held crucial datasets needed to measure advertising efficacy, yet neither side could share or expose their data because they lacked mutual trust. To overcome those friction points, engineering teams developed privacy-preserving technologies such as clean rooms. Ten-Pow argues that the enterprise AI privacy bottleneck is identical at its core: it boils down to an absence of trust between an AI software vendor and its prospective enterprise customer.

[01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) The Enterprise Dilemma and the Spectrum of Software Delivery

When enterprise leaders walk vendor exhibit halls and spot compelling tools—such as advanced enterprise search platforms—they inevitably hit a wall when consulting their security and legal departments. For organizations dealing with sensitive financial, legal, or healthcare data, sending proprietary information to a multi-tenant cloud is often deemed an absolute non-starter. Navigating these compliance risks and security audits introduces heavy friction and high overhead for both the vendor attempting to earn trust and the enterprise trying to verify it.

```mermaid flowchart LR accTitle: Enterprise AI Software Deployment Spectrum subgraph SaaS["Vendor Hosted"] direction TB A["Multi-Tenant SaaS<br/>(Shared Cloud Account)"] B["Single-Tenant SaaS<br/>(Dedicated Cloud Instance)"] end subgraph Hosted["Customer Hosted"] direction TB C["Cloud Prem / BYOC<br/>(Customer AWS/Azure/GCP)"] D["On-Premises<br/>(Bare Metal / VMs)"] E["Air-Gapped<br/>(Zero External Connectivity)"] end A --> B --> C --> D --> E ```

To solve this, vendors evaluate a spectrum of hosting models that trade vendor convenience for enterprise isolation:

- **Multi-Tenant SaaS:** The vendor runs a single shared environment in their own AWS account. The enterprise must transmit all data outside its perimeter, creating maximum compliance and security friction. - **Single-Tenant SaaS:** The vendor provisions dedicated infrastructure per customer (for example, isolating Capital One's workload entirely from JPMorgan Chase's). While straightforward for vendors using Infrastructure as Code (IaC), data still leaves customer premises—meaning highly regulated institutions will still reject it. - **Cloud Premise / Bring Your Own Cloud (BYOC):** The vendor deploys software directly into the customer's cloud tenant (AWS, Azure, or GCP), ensuring data stays inside the customer's boundary. - **On-Premises:** Deployments installed directly onto bare metal or virtual machines in private data centers, a common requirement for law firms and legacy banking infrastructure. - **Air-Gapped:** Strictly isolated, zero-connectivity environments typical in military, defense, and high-security government installations.

[06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) Eliminating the Two-Product Penalty with Tensor9

Under the status quo, supporting customer-hosted deployments forces AI software vendors to essentially rebuild and split their product into two distinct offerings. Vendors maintain the central SaaS product they love—one they can observe in real time and continuously deploy to ten times a day—alongside a customer-hosted or air-gapped variant that quickly turns into an "albatross around their neck." Updating and maintaining these disconnected customer-premise installations inflates operational and support costs by as much as 100x for both vendors and buyers.

Tensor9 resolves this tradeoff by allowing vendors to write their software once using standard Infrastructure as Code (IaC) and deploy it seamlessly into any target environment, including AWS, Google Cloud, Microsoft Azure, bare-metal on-prem, and air-gapped infrastructure. Vendors retain a single unified product pipeline without separate codebases, while enterprise customers get guaranteed data locality inside their own perimeter without needing to hire dedicated operational staff to decipher thick software manuals. Both sides achieve a frictionless, "SaaS-like experience" without compromising data privacy.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) organizations want to adopt cutting-edge AI tools to process sensitive enterprise data, but IT, legal, and security teams refuse to let that data leave their "four walls" to reside with a third-party vendor. — Michael Ten-Pow

> [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) the enterprise AI privacy bottleneck is identical at its core: it boils down to an absence of trust between an AI software vendor and its prospective enterprise customer. — Michael Ten-Pow

> [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) For organizations dealing with sensitive financial, legal, or healthcare data, sending proprietary information to a multi-tenant cloud is often deemed an absolute non-starter. — Michael Ten-Pow

> [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) The vendor runs a single shared environment in their own AWS account. The enterprise must transmit all data outside its perimeter, creating maximum compliance and security friction. — Michael Ten-Pow

> [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) While straightforward for vendors using Infrastructure as Code (IaC), data still leaves customer premises—meaning highly regulated institutions will still reject it. — Michael Ten-Pow

> [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) Vendors maintain the central SaaS product they love—one they can observe in real time and continuously deploy to ten times a day—alongside a customer-hosted or air-gapped variant that quickly turns into an "albatross around their neck." — Michael Ten-Pow

> [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) Updating and maintaining these disconnected customer-premise installations inflates operational and support costs by as much as 100x for both vendors and buyers. — Michael Ten-Pow

> [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) Tensor9 resolves this tradeoff by allowing vendors to write their software once using standard Infrastructure as Code (IaC) and deploy it seamlessly into any target environment — Michael Ten-Pow

> [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) Both sides achieve a frictionless, "SaaS-like experience" without compromising data privacy. — Michael Ten-Pow

## Mentioned Michael Ten-Pow, Tensor9, Amazon, AWS, Amazon Retail, Facebook, Google, Capital One, JPMorgan Chase, Microsoft Azure, Google Cloud

## Video description Michael Ten-Pow, Founder / CEO, Tensor9

Abstracts

Many companies are struggling to realize value from AI when data movement is either a security risk or cost prohibitive. This talk will explore the various options available to enterprise customers to consume AI software solutions, and will touch on why it’s so difficult for AI vendors to fulfill enterprise customer needs for data privacy. We’ll present a new approach that unlocks new revenue streams for AI vendors, while enabling enterprises to leverage AI without data egress.

[Source](https://www.youtube.com/watch?v=zI_xfbmfs-0)

中文简报

Tensor9瞄准AI隐私瓶颈，终结“双产品”困境 将统一的基础设施即代码直接部署至客户自有环境中，消除了构建独立本地化AI软件版本带来的百倍运维成本代价。 由于安全与合规规定禁止敏感数据流出企业内部边界，企业往往会阻碍AI的落地应用。尽管供应商可以提供单租户SaaS、Cloud Premise/BYOC、本地部署和气隙隔离环境等替代托管模式，但传统上支持客户自托管版本会带来难以承受的“双产品代价”，导致维护成本激增高达100倍。Tensor9通过让AI供应商只需利用基础设施即代码（IaC）定义一次软件，即可将其无缝部署到任何客户边界内，化解了这一僵局。该架构在为企业采购方确保严格数据本地化的同时，也为供应商维持了统一且持续交付的SaaS体验。 由于安全和法务部门严格禁止敏感数据离开企业内部边界，供应商与客户之间的信任缺失成为了企业采纳前沿 AI 软件的核心阻碍。 (0:04) 多租户及单租户 SaaS 模式均涉及数据外流，金融和医疗等强监管行业往往要求采用自带云（BYOC）、本地裸机或完全物理隔离（Air-Gapped）的部署方案以确保数据隐私。 (1:51) 传统客户托管部署会带来“双产品惩罚”，迫使 AI 供应商维护独立的代码库，并可能将双方的运维和技术支持成本推高多达 100 倍。 (6:03) Tensor9 支持供应商使用标准基础设施即代码（IaC）编写单套软件，无缝部署到 AWS、Azure、GCP、本地机房及物理隔离环境中，在保证数据不出域的同时消除代码分化问题。 (6:03)

Reading summary

Tensor9 founder and CEO Michael Ten-Pow argues that enterprise AI adoption is constrained by a lack of trust between vendors and enterprise buyers unwilling to let proprietary data leave their security perimeters. Overcoming this friction requires deployment architectures that run software within customer-controlled environments without imposing severe operational overhead on software vendors. The Enterprise Trust Deficit and Hosting Spectrum (0:04) Drawing on his background at Amazon Retail—where collaboration with platforms like Meta and Google stalled because neither party could safely share proprietary ad data—Ten-Pow frames the enterprise AI bottleneck as an issue of mutual trust. Organizations in healthcare, finance, and legal sectors face strict compliance barriers that prohibit transmitting sensitive information to external, vendor-hosted clouds. Software delivery models reflect varying degrees of isolation and operational friction. Vendor-hosted options like multi-tenant and single-tenant SaaS keep infrastructure under vendor management but fail enterprise compliance rules because data leaves customer walls. Conversely, customer-hosted models—including Cloud Premise / Bring Your Own Cloud (BYOC) on AWS, Azure, or GCP, private on-premises virtual machines, and air-gapped systems—keep data localized but introduce complex verification and configuration requirements. Eliminating the Two-Product Maintenance Penalty (6:03) Accommodating customer-hosted requirements traditionally imposes what Ten-Pow terms a 'two-product penalty.' Software vendors are forced to split their offerings into a preferred, continuously deployed SaaS platform alongside isolated, customer-premise versions that can inflate operational and maintenance costs by as much as 100x for buyers and vendors alike. Ten-Pow states that Tensor9 resolves this division by enabling vendors to write software once using standard Infrastructure as Code (IaC) and deploy it uniformly across multi-cloud tenants, bare-metal hardware, and disconnected air-gapped networks. This model is designed to preserve continuous SaaS-style product maintenance for vendors while guaranteeing enterprise buyers full data locality inside their own perimeters. Ten-Pow concludes that solving enterprise AI adoption requires reconciling vendor deployment efficiency with customer-controlled security, asserting that unified Infrastructure as Code can eliminate the traditional cost and maintenance trade-offs of disconnected customer-hosted software.

精炼阅读摘要

Tensor9 创始人兼首席执行官 Michael Ten-Pow 指出，企业采用前沿人工智能的核心障碍在于软件供应商与企业客户之间的信任缺失；通过统一的交付技术，供应商可以在不转移企业数据的前提下提供类似 SaaS 的体验。 企业AI面临的信任鸿沟与隐私瓶颈 (0:04) 企业 IT、法务与安全团队通常严格限制敏感数据流出自身的安全边界，这导致企业即便迫切需要先进的 AI 工具，也难以采购主流第三方的服务。Ten-Pow 结合其在亚马逊 13 年的工程经历（包括 2007 年至 2013 年在 AWS 基础设施部门的 6 年，以及随后在零售广告业务中的数据隐私实践）指出，零售部门过去在与 Facebook 和 Google 等大型广告平台合作以衡量投放效果时，也曾因缺乏互信而遭遇严峻的数据共享障碍。当时工程团队通过构建 Clean Rooms 等隐私保护技术解决了分歧；Ten-Pow 认为，当今企业级 AI 软件在数据隐私上的停滞局面，本质上同样是买卖双方信任缺失所致。 软件交付模式的光谱与合规隔离权衡 (1:51) 当企业在展会等场合接触到诸如企业级搜索等高效工具时，敏感的金融、法律或医疗机构往往会在内部安全审计环节否决多租户云方案，因为将专有数据上传到多租户环境存在极高的合规风险。为了应对不同程度的隔离需求，软件交付形成了一条在供应商便利度与客户隔离度之间权衡的光谱：多租户 SaaS（Multi-Tenant SaaS）由供应商在其 AWS 账户中统一维护，但需要客户将数据完全外发，摩擦最大；单租户 SaaS（Single-Tenant SaaS）虽通过基础设施即代码（IaC）为客户（如将 Capital One 与 JPMorgan Chase）分别配置独立实例，但数据依然留在供应商端，受严格监管的企业依然无法接受。 光谱的另一端则是客户托管方案，包括将软件直接部署至企业自身云租户（AWS、Azure 或 GCP）的云原生私有化部署（BYOC / Cloud Premise）；部署在私有数据中心裸机或虚拟机上的传统本地部署（On-Premises，常用于律所及传统银行系统）；以及完全切断外部网络连接的物理隔离环境（Air-Gapped，主要用于军工、国防与高密级政府机构）。这些客户托管模式虽消除了数据出域隐患，却给交付带来了巨大阻碍。 打破双产品惩罚与百倍运维成本困境 (6:03) 按照既有技术路径，AI 供应商若要支持客户托管部署，通常被迫将软件拆分为两个完全独立的产品线：一边是团队熟悉、可实时监控且能实现每日部署多达 10 次的集中式 SaaS；另一边则是更新迟缓、宛如沉重负担的客户本地或物理隔离版本。维护与更新这些脱节的客户本地安装，可能会将供应商与采购企业的运营及技术支持成本推高至多达 100 倍。 Tensor9 提出的方案主张消除这种“双产品惩罚”（Two-Product Penalty）。该方案允许供应商仅使用标准的基础设施即代码（IaC）编写一次软件，即可跨 AWS、Google Cloud、Microsoft Azure、裸机物理服务器以及物理隔离环境进行无缝部署。供应商得以保留单一代码库与统一的产品交付流水线，而企业客户则能在将数据严格保留在自身边界内的同时，无需配备专门运维人员去研读繁杂的软件操作手册，从而在兼顾数据隐私的同时实现无摩擦的交付体验。 Ten-Pow 强调，借助消除代码库割裂并适配多云与隔离环境的底层交付平台，企业无需再在数据绝对控制与先进 AI 功能之间进行妥协，然而如何在高度受限的无网环境下持续维持与云端一致的操作体验仍是长期工程挑战。

Detailed notes — zh

## Background 该视频由 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台，由 Tensor9 创始人兼首席执行官 Michael Ten-Pow 主讲。视频简介指出，许多企业因数据外泄的安全风险或高昂成本而难以释放 AI 的价值，本次演讲探讨了企业消费 AI 方案的各种路径、供应商满足隐私合规的难点，以及 Tensor9 如何通过免数据出境的新途径化解两难困境。 ## TL;DR 演讲指出，当前企业引入 AI 应用的核心阻碍在于 IT 与法务安全部门严禁敏感数据流出企业边界，而供应商若针对客户环境单独定制私有化交付，则会陷入代码分支分裂和高达百倍的运维成本沼泽。通过梳理从多租户 SaaS 到物理隔离（Air-gapped）的交付光谱，主讲人阐明了数据信任鸿沟带来的两败俱伤局面。针对此痛点，Tensor9 提出了一种全新交付架构，让软件供应商仅凭一套声明式 IaC 代码即可将产品一键分发并运行在客户自有的各类云端或本地隔离环境中，在保证企业数据绝不出域的同时兼顾开箱即用的 SaaS 级敏捷体验。 ## Atmosphere & style 整场演讲风格利落务实且切中行业要害，主讲人依托深厚的 AWS 核心工程背景与数据隐私实战经历，以生动幽默的行业洞察（如调侃在展厅看中产品却无法通过公司安全审查）迅速引发听众共鸣。演讲运用了“套在脖子上的信天翁”等极具画面感的比喻，层层剖析技术与商业死结，最终给出干脆利落的工程解法，传递出理性、自信且极具说服力的极客气质。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) 企业落地 AI 的最大阻碍是敏感数据出境风险，IT、法务与安全合规部门通常坚决禁止将核心数据传输至第三方云端。 - [00:35](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=35s) 生成式 AI 时代的隐私困境本质上是软件供应商与企业客户之间的信任鸿沟，与早期互联网广告跨机构数据合作受阻如出一辙。 - [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) 金融、医疗及法律等强监管行业面临极高的合规壁垒，在采购第三方 AI 工具时往往将“数据绝不出域”作为不可协商的红线。 - [03:15](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=195s) 软件交付架构存在从多租户 SaaS 到离网物理隔离的光谱，数据安全性要求越高，软件的交付与运维复杂度越呈指数级上升。 - [03:50](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=230s) 单租户 SaaS（Single-tenant SaaS）仅在基础设施层做到客户隔离，数据依旧需上传至供应商云端，仍无法被严格合规的企业接受。 - [04:30](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=270s) 自带云（BYOC）、本地物理机（On-Prem）及物理隔离（Air-gapped）虽满足了数据驻留，但给供应商技术架构施加了巨大的适配负担。 - [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) 为迎合企业私有部署需求，传统供应商往往被迫分裂出割裂的公有 SaaS 与私有交付双版本代码库，严重拖慢迭代速度。 - [06:25](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=385s) 传统的本地化与私有云交付模式如同“套在脖子上的信天翁”，发版困难，其综合支持与运营成本相比公有云高出 100 倍。 - [06:53](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=413s) Tensor9 提出自动化交付范式，让供应商通过同一套基础设施即代码（IaC）将单一产品无缝投递至客户的各类公有云、裸金属或离网环境中。 - [07:20](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=440s) 新型交付模式既保障了企业核心数据零外泄，又让客户享受到了无须自建繁琐运维团队的 SaaS 级轻量使用体验。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) 个人背景与企业级 AI 落地面临的“数据信任”困境 Tensor9 创始人兼首席执行官 Michael Ten-Pow 登台向现场观众打招呼，并直接切入当下企业在拥抱 AI 过程中普遍遭遇的痛点：市场上充斥着令人心动的 AI 产品，企业迫切希望引入它们来处理自身的核心业务；然而，一旦涉及敏感数据，这条路就会立刻受阻。原因很简单——企业的 IT、法务以及安全部门绝不允许将敏感数据传输到公司“四堵墙”（Four walls）之外的第三方供应商手中。 Michael 介绍了自己的技术背景与从业经历。他是一名资深工程师，曾在亚马逊工作了 13 年之久。在 2007 年至 2013 年云计算刚起步的早期阶段，他在 AWS 核心团队度过了 6 年；随后转入亚马逊零售业务部门，专注于广告系统与数据隐私的交汇领域。在那里，他主导推动亚马逊与 Facebook、Google 等广告巨头的外部合作，但双方频繁撞上一堵无法逾越的高墙：两家公司都极其渴望通过对方的数据来衡量广告投放效果（例如评估亚马逊广告在 Facebook 上的转化成效），但因为彼此无法在数据层面上建立完全信任，合作屡屡搁浅。正是为了打破这种不信任，他们当年参与研发了“数据清洁室”（Data Clean Rooms）等隐私计算技术。Michael 指出，当今生成式 AI 领域的隐私困境如出一辙，其核心本质就是**拥有 AI 软件的供应商与其企业客户之间缺乏信任机制**。 ### [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) 安全审查死局：从严苛合规到“绝不出门”的红线 Michael 调侃道，大家刚刚在楼下的展厅里参观时，肯定看到了大量令人眼花缭乱的 AI 展商，心中也许盘算着挑选一两款引入自己的企业。但紧接着每个人心里都会打鼓：一旦回到公司把方案递交给 IT、法务和安全部门（例如想采购一套企业级知识搜索工具），会发生什么？最好的情况是迎来一场漫长而严酷的安全审查流程；但在很多时候，安全团队会直截了当地宣判“没门”（A non-starter）——“我们绝对不可能把敏感数据发送给任何第三方供应商，软件必须部署在我们自己的私有环境内，没有任何商量余地”。 在涉及金融、医疗健康或法律等强监管行业时，合规性与监管门槛更为致命。这种固有的信任鸿沟带来了极其高昂的商业与协作成本：软件供应商为了达到不同大型企业繁琐的安全要求疲于奔命；而企业内部为了评估、验证第三方供应商并建立信任，也消耗了大量人力与财力。 ### [03:15](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=195s) 软件交付模式光谱：从多租户 SaaS 到极端隔离环境 为了更直观地说明当下的选择困难，Michael 梳理了当前主流的软件交付形态光谱。从最左侧开始是传统的**多租户 SaaS（Multi-tenant SaaS）**模式，供应商在自己的云环境（如 AWS 账号）中集中运行产品，企业必须将内部数据发送出境；随着模式向右推进，企业侧的数据安全性逐步提升，数据能够保留在本地，但供应商的交付与运维难度却呈指数级飙升。 ```mermaid flowchart LR accTitle: 软件交付架构演进光谱 subgraph S1["供应商托管 (Vendor Hosted)"] A["多租户 SaaS<br/>(数据出境/成本低)"] --> B["单租户 SaaS<br/>(独立隔离/仍需出境)"] end subgraph S2["客户本地托管 (Customer Hosted)"] C["自带云 (BYOC / Cloud Prem)<br/>(部署至客户 AWS/GCP/Azure)"] --> D["本地私有化 (On-Prem)<br/>(物理机/虚拟机/无云)"] D --> E["物理隔离 (Air-gapped)<br/>(完全切断外网/军工政务)"] end ``` - **单租户 SaaS（Single-tenant SaaS）**：供应商为每个大客户独立分配专用的基础设施实例。若同时服务 Capital One 与摩根大通（JPMorgan），两家银行的数据在基础设施层实现隔离，互不混杂。供应商借助基础设施即代码（IaC）相对容易复制一套环境，但这仍未能解决根本问题——企业依然要把敏感数据上传到供应商的云端。对于像摩根大通这样体量的银行，若想采购 Glean 这样的企业级搜索产品，单租户 SaaS 方案仍然会被安全部门直接否决。 - **自带云 / 云上私有化（BYOC / Cloud Premise）**：客户要求软件必须直接部署在企业自有的云账号内（如客户自己的 AWS、Azure 或 GCP 租户），供应商必须设法把整套软件打包装入客户的云基础设施中。 - **本地私有化（On-Prem）**：脱离任何公有云，直接运行在客户的数据中心、裸金属物理机（Bare Metal）或内部虚拟机上。许多传统律所和商业银行虽然在往云端迁移，但核心系统依然以这种极为受限的环境为主。 - **物理隔离（Air-gapped）**：光谱最右端的极端严苛场景，常见于政府或军工国防机构。整个环境与外部互联网彻底切断连接，没有任何公网出入口。 向右移动的每一步，对供应商的技术架构和运营运维能力都是一次沉重的打击。企业想要掌控数据的初衷合情合理，但这给软件供应商施加了无法承受的包袱。 ### [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) 供应商的噩梦：“套在脖子上的信天翁”与百倍成本跃升 Michael 直言，面对企业客户“数据不出域”的要求，当今技术现状下供应商的唯一做法就是**重构系统**。这不可避免地导致供应商分裂出两套完全割裂的产品形态：一套是他们最心爱、最引以为傲的公有 SaaS 版本，研发团队可以每天持续发布更新部署十多次，线上指标透明清晰，出现任何 Bug 都能在几分钟内定位并修复；另一套则是为了迎合大客户而东拼西凑出来的 Cloud Prem、On-prem 或 Air-gapped 私有交付版本。 Michael 用了一个生动的隐喻——这个私有化部署版本就像是“套在研发团队脖子上的信天翁”（An albatross around their neck，喻指甩不掉的沉重诅咒与负担）。它无法快速迭代更新，每次发版都需要经历极度痛苦的手动流程，后期的支持与维护成本高得惊人。结果导致买卖双方两败俱伤：供应商与企业为了在更安全的环境下运行系统，付出了高出公有云 **100 倍** 的综合支持与运营成本。 ### [06:53](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=413s) Tensor9 的解法：以 SaaS 体验在任意客户环境免数据迁移落地 为了彻底打破“数据安全”与“运维成本”之间的对抗妥协，Tensor9 提出了一种全新的架构范式：让软件供应商能够使用**同一套核心代码**，平滑部署到企业客户指定的任何计算环境中。 ```mermaid flowchart TD accTitle: Tensor9 跨环境交付架构 subgraph Dev["供应商侧 (单一产品视角)"] Code["基础设施即代码 (IaC)<br/>定义单一核心产品"] end subgraph Engine["Tensor9 自动化运行时与分发"] T9["Tensor9 引擎"] end subgraph Env["企业本地环境 (SaaS 级体验 / 数据零外泄)"] AWS["企业自建 AWS"] GCP["企业自建 GCP"] Azure["企业自建 Azure"] Bare["本地 Bare Metal 物理机"] Air["物理隔离网络 (Air-gapped)"] end Code --> T9 T9 --> AWS T9 --> GCP T9 --> Azure T9 --> Bare T9 --> Air ``` 只要供应商的产品架构能够通过现代基础设施即代码（IaC）进行声明式定义，Tensor9 就可以将其无缝下沉投递到客户的 AWS、Google Cloud、Azure，甚至是本地裸金属服务器以及彻底离网的 Air-gapped 隔离网络中。 - **对供应商而言**：研发团队眼中只有“一款产品”，彻底告别为了满足特定合规要求而分化维护多套代码分支的历史包袱。 - **对企业客户而言**：获得如同开箱即用 SaaS 般的流畅体验。企业不再需要接收一个塞满复杂组件的压缩包和几百页的手册，也不必专门雇佣一批专业运维工程师来照料这套软件。数据自始至终留在企业内部不动，双方却同时获得了敏捷、低成本、如 SaaS 般轻量易用的交付成果。 演讲最后，Michael 总结道，打破数据孤岛并不意味着必须妥协安全性，并热情邀请会后受此困扰的从业者在场下进一步交流探讨。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) 企业的 IT、法务以及安全部门绝不允许将敏感数据传输到公司“四堵墙”（Four walls）之外的第三方供应商手中。 — Michael Ten-Pow > [00:04](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=4s) 当今生成式 AI 领域的隐私困境如出一辙，其核心本质就是拥有 AI 软件的供应商与其企业客户之间缺乏信任机制。 — Michael Ten-Pow > [01:51](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=111s) 我们绝对不可能把敏感数据发送给任何第三方供应商，软件必须部署在我们自己的私有环境内，没有任何商量余地。 — Michael Ten-Pow > [03:15](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=195s) 企业想要掌控数据的初衷合情合理，但这给软件供应商施加了无法承受的包袱。 — Michael Ten-Pow > [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) 这个私有化部署版本就像是“套在研发团队脖子上的信天翁”（An albatross around their neck，喻指甩不掉的沉重诅咒与负担）。 — Michael Ten-Pow > [06:03](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=363s) 供应商与企业为了在更安全的环境下运行系统，付出了高出公有云 100 倍 的综合支持与运营成本。 — Michael Ten-Pow > [06:53](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=413s) 只要供应商的产品架构能够通过现代基础设施即代码（IaC）进行声明式定义，Tensor9 就可以将其无缝下沉投递到客户的 AWS、Google Cloud、Azure，甚至是本地裸金属服务器以及彻底离网的 Air-gapped 隔离网络中。 — Michael Ten-Pow > [06:53](https://www.youtube.com/watch?v=zI_xfbmfs-0&t=413s) 企业不再需要接收一个塞满复杂组件的压缩包和几百页的手册，也不必专门雇佣一批专业运维工程师来照料这套软件。数据自始至终留在企业内部不动，双方却同时获得了敏捷、低成本、如 SaaS 般轻量易用的交付成果。 — Michael Ten-Pow ## Mentioned Michael Ten-Pow, Tensor9, AWS, 亚马逊 (Amazon), Facebook, Google, Capital One, 摩根大通 (JPMorgan), Glean, Azure, Google Cloud ## Video description Michael Ten-Pow, Founder / CEO, Tensor9 Abstracts Many companies are struggling to realize value from AI when data movement is either a security risk or cost prohibitive. This talk will explore the various options available to enterprise customers to consume AI software solutions, and will touch on why it’s so difficult for AI vendors to fulfill enterprise customer needs for data privacy. We’ll present a new approach that unlocks new revenue streams for AI vendors, while enabling enterprises to leverage AI without data egress. [Source](https://www.youtube.com/watch?v=zI_xfbmfs-0)

Source quotes

[0:04] Michael Ten-Pow: 企业的 IT、法务以及安全部门绝不允许将敏感数据传输到公司“四堵墙”（Four walls）之外的第三方供应商手中。 [0:04] Michael Ten-Pow: 当今生成式 AI 领域的隐私困境如出一辙，其核心本质就是拥有 AI 软件的供应商与其企业客户之间缺乏信任机制。 [1:51] Michael Ten-Pow: 我们绝对不可能把敏感数据发送给任何第三方供应商，软件必须部署在我们自己的私有环境内，没有任何商量余地。 [3:15] Michael Ten-Pow: 企业想要掌控数据的初衷合情合理，但这给软件供应商施加了无法承受的包袱。 [6:03] Michael Ten-Pow: 这个私有化部署版本就像是“套在研发团队脖子上的信天翁”（An albatross around their neck，喻指甩不掉的沉重诅咒与负担）。 [6:03] Michael Ten-Pow: 供应商与企业为了在更安全的环境下运行系统，付出了高出公有云 100 倍 的综合支持与运营成本。 [6:53] Michael Ten-Pow: 只要供应商的产品架构能够通过现代基础设施即代码（IaC）进行声明式定义，Tensor9 就可以将其无缝下沉投递到客户的 AWS、Google Cloud、Azure，甚至是本地裸金属服务器以及彻底离网的 Air-gapped 隔离网络中。 [6:53] Michael Ten-Pow: 企业不再需要接收一个塞满复杂组件的压缩包和几百页的手册，也不必专门雇佣一批专业运维工程师来照料这套软件。数据自始至终留在企业内部不动，双方却同时获得了敏捷、低成本、如 SaaS 般轻量易用的交付成果。

Source quotes

[0:04] Michael Ten-Pow: organizations want to adopt cutting-edge AI tools to process sensitive enterprise data, but IT, legal, and security teams refuse to let that data leave their "four walls" to reside with a third-party vendor. [0:04] Michael Ten-Pow: the enterprise AI privacy bottleneck is identical at its core: it boils down to an absence of trust between an AI software vendor and its prospective enterprise customer. [1:51] Michael Ten-Pow: For organizations dealing with sensitive financial, legal, or healthcare data, sending proprietary information to a multi-tenant cloud is often deemed an absolute non-starter. [1:51] Michael Ten-Pow: The vendor runs a single shared environment in their own AWS account. The enterprise must transmit all data outside its perimeter, creating maximum compliance and security friction. [1:51] Michael Ten-Pow: While straightforward for vendors using Infrastructure as Code (IaC), data still leaves customer premises—meaning highly regulated institutions will still reject it. [6:03] Michael Ten-Pow: Vendors maintain the central SaaS product they love—one they can observe in real time and continuously deploy to ten times a day—alongside a customer-hosted or air-gapped variant that quickly turns into an "albatross around their neck." [6:03] Michael Ten-Pow: Updating and maintaining these disconnected customer-premise installations inflates operational and support costs by as much as 100x for both vendors and buyers. [6:03] Michael Ten-Pow: Tensor9 resolves this tradeoff by allowing vendors to write their software once using standard Infrastructure as Code (IaC) and deploy it seamlessly into any target environment [6:03] Michael Ten-Pow: Both sides achieve a frictionless, "SaaS-like experience" without compromising data privacy.

Original transcript

[0:04] Hello everybody. My name is Michael [0:05] Tenpow. I'm the CEO of a company called [0:08] Tensor. helps solve this uh this problem [0:13] this problem that we have in AI where [0:15] there are AI products that we'd like to [0:17] use in the enterprise dealing with [0:20] sensitive data and we are unable to do [0:21] so because [0:24] the IT staff legal security staff at our [0:27] enterprise unwilling to send sensitive [0:30] data outside our four walls to a third [0:32] party. I'll tell you just a bit about [0:34] myself and then how I got here and uh [0:36] then talk about the problem a bit and [0:38] how we intend to solve it. Um so I'm an [0:41] engineer. I worked at Amazon for a long [0:43] time for 13 years. I spent six years [0:46] working in AWS in the early days between [0:48] 2007 2013 and uh and then after that I [0:52] worked in Amazon retail working at the [0:54] inter intersection of advertising and [0:56] privacy and one of the challenges that I [0:58] encountered there was I was working on [1:01] helping Amazon [1:03] collaborate with advertisers like [1:06] Facebook and Google and we kept running [1:08] into this problem where we the two [1:10] companies wanted to collaborate but were [1:12] unable to do so. we had these data sets [1:15] that were very important for that uh to [1:17] measure for example the efficacy of of [1:20] uh our ads on Facebook and we couldn't [1:23] collaborate and because we didn't trust [1:25] each other with that data and so we [1:27] invented technology like clean rooms to [1:30] be able to um overcome those trust [1:33] hurdles. So fundamentally this is what [1:35] this is about. This AI data privacy [1:37] problem is about overcoming is about the [1:40] lack of trust between a software vendor [1:43] who has an AI product and their [1:45] customer. [1:51] So this is this is the this is the [1:53] dilemma here. We have this promise of [1:54] AI. You know, you guys all walked the [1:57] exhibit uh hall downstairs and saw a [2:00] number of AI companies maybe that you [2:02] were interested in bringing into your [2:04] enterprise and no doubt you thought [2:06] about what is it going to what's going [2:07] to happen when I go talk to my IT, my [2:09] legal, my security team about bringing [2:11] this let's say enterprise search product [2:15] into into my company. Um they're likely [2:19] going to say, "Well, we're going to need [2:20] a very rigorous security review." Um, [2:23] but they may say this is a non-starter. [2:25] They may say there is no way, there's [2:27] absolutely no way we're going to send [2:29] our data to a thirdparty vendor. Um, [2:33] this needs to be on prem in our [2:35] environment and there's no uh other way [2:37] around it. So there there are difficult [2:40] security risks here. There are [2:41] compliance and regulatory hurdles if [2:43] you're dealing with financial or [2:44] healthcare data or legal data. And [2:47] because of all of this um these trust [2:50] hurdles, there is a high cost. There's a [2:52] high cost to the vendor to be able to [2:55] meet uh the enterprises security [2:58] requirements and there's a high cost to [3:00] the enterprise to be able to uh [3:03] basically earn you know earn the trust [3:06] uh you know build trust with uh that [3:08] vendor in order to over the trust [3:10] hurdles and and work and collaborate and [3:12] work together. [3:15] So what happens here are the options [3:19] on the left we have the multi-tenant SAS [3:22] model. This is where a vendor has their [3:26] product running in their cloud [3:28] environment in their AWS [3:30] and in order to work with an enterprise [3:32] the enterprise needs to send their data [3:34] to that AWS account outside their four [3:36] walls. And as you move to the right you [3:39] have options that are more secure for [3:41] the enterprise. [3:43] they keep the data localized inside the [3:45] enterprises environment. However, uh [3:48] they create lots of challenges for the [3:50] vendor in order to be able to um fit [3:54] that model. And so if you move one right [3:56] over from the multi-tenant SAS case, [3:58] there's a single tenant SAS case. This [4:00] is where a vendor has a dedicated [4:03] environment just for that just for the [4:05] enterprise customer. And so there's no [4:07] co-mingling of data. So that vendor, if [4:09] they're working with Capital One and [4:11] they're working with JP Morgan, Capital [4:12] One and JP Morgan's data will be [4:14] isolated from each other. There will be [4:16] no co-mingling. That's fairly easy for a [4:18] vendor to do. They need to be able to [4:21] take their existing stack and maybe if [4:23] they're using like infrastructure as [4:25] code, they can make another version of [4:27] it um that's just for that enterprise [4:29] customer. The problem though there is [4:32] that the enterprise still has to be able [4:34] to trust that vendor with their [4:36] sensitive data. And for many cases, [4:38] that's a non-starter. If you think about [4:40] a company like Gleam that's doing [4:42] enterprise search, if they're, you know, [4:44] have they're trying to sell to, let's [4:46] say, JP Morgan, there is no way that JP [4:49] Morgan is going to be comfortable with a [4:51] single tenant SAS solution. They need [4:53] one of these other solutions to the [4:55] right, which are labeled customer [4:57] hosted. There are three of them. One is [4:59] cloud premise known as BYOC or bring [5:02] your own cloud. This is where the [5:04] enterprise has a cloud account. They're [5:06] in AWS. They're in Azure, they're in [5:07] GCP. Um, and the vendor figures out how [5:10] to deploy their software into this [5:12] environment. As we move to the right, we [5:14] have on-prem. This is can be bare metal. [5:17] It could be virtual machines. This is [5:19] not in the cloud. This is typically far [5:22] more constrained environments. A lot of [5:24] law firms operate this way. Banks [5:26] operate this way, though they are moving [5:28] to the cloud. And on the far right is [5:30] airgapped. These are highly constrained [5:33] environments. often it's government or [5:34] military where uh there is no [5:37] connectivity to the outside world. As [5:39] you move to the right, it becomes [5:40] increasingly difficult for the vendor to [5:43] be able to support support that [5:44] environment. And so while enterprises [5:46] may like the idea of an on-prem or cloud [5:48] prem or a gap thing, what they're asking [5:51] of their vendor um is a lot of burden to [5:54] be able to deploy their product into [5:56] this challenging environment and to be [5:58] able to operate it. [6:03] So this is really hard. The [6:04] state-of-the-art today is that a vendor [6:06] effectively has to rebuild their product [6:08] to be able to fit these environments. [6:11] Um, and frankly what ends up happening [6:13] is they have two products. They have [6:16] their SAS product that they love that [6:18] they deploy to 10 times a day that [6:20] they're able to observe issues very [6:23] quickly. Um and then they have their [6:26] cloud prem onrem or airgapped product [6:29] and it's an albatross around their neck. [6:32] Um it uh they can't update it quickly. [6:35] Um and the support cost is extremely [6:38] high. The end result is that both the [6:40] vendor and the enterprise have these [6:42] really high costs 100x in some cases uh [6:46] in order to support these more secure [6:48] environments. [6:53] So we seek to um to break this trade-off [6:58] um and make it so that the vendor [7:01] uh can uh take their existing product [7:04] and deploy it into any enterprise [7:06] environment. So here we see this is how [7:07] our product works. We'll take an [7:09] existing product um as long as it's [7:12] defined in infrastructure as code and we [7:14] can deploy to various customer [7:16] environments into AWS into Google cloud [7:19] into Azure onrem bare metal and air [7:21] gapped environments. The vendor sees one [7:24] product they don't have multiple [7:26] products and the enterprise [7:29] um also uh has a SAS-like experience. [7:32] They don't have to get a package of [7:34] software with a manual that they have to [7:36] hire folks in order to operate. This [7:39] keeps the cost down. It gives both [7:41] parties an enterprise uh sorry a [7:44] SAS-like uh experience. [7:49] And that's it. Thank you all. If you're [7:51] interested, please uh please find me [7:52] later. Thank you.