Nvidia, HCLTech Executives Push Custom AI Factories Over Model Renting

From Renting AI to Owning Intelligence: The Case for Purpose-Built AI — Ai4

Ai4

Enterprises must transition from renting third-party models to owning tailored AI factories to protect proprietary data and control long-term operational costs.

AI-assisted summary of From Renting AI to Owning Intelligence: The Case for Purpose-Built AI — original source. Check the original for full context.

Source published: 2026-08-22

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 22, 2026, this session features Fast Company's Harry McCracken, HCLTech's Kalyan Kumar (KK), and Nvidia's Kaushik Shirhatti. According to the show notes, the panel explores how organizations are transitioning from renting general-purpose AI models to owning purpose-built intelligence tailored to enterprise data, compliance, and long-term control.

## TL;DR The panel discusses the strategic evolution from consuming third-party frontier AI models to constructing owned, enterprise-grade AI capabilities. Rather than choosing strictly between renting or owning, the speakers advocate for a hybrid approach that integrates open-weight models, small language models, and private data-first architectures into an 'AI Factory.' By analyzing the economics of 'tokconomics' and the rise of agentic AI, they demonstrate how custom infrastructure allows enterprises to secure proprietary context, optimize long-term operational costs, and deliver tangible business outcomes.

## Atmosphere & style The session carries an engaging, highly authoritative, and pragmatic tone suited for a major tech conference. The dynamic between moderator Harry McCracken and speakers KK and Shirhatti feels collaborative and structured, balancing sharp industry insights with lighthearted moments like jokes about the intense Las Vegas heat. It offers an intellectually rich, forward-looking discussion that demystifies complex enterprise AI strategies into approachable conceptual frameworks like the 'five-layer cake' and 'AI in a box.'

## Key takeaways - [00:04](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=4s) Enterprise AI is shifting from an initial phase focused on access and general-purpose model consumption to a phase centered on owning customized, purpose-built intelligence. - [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) Renting and owning AI are not mutually exclusive; enterprises benefit most from a hybrid approach that combines third-party frontier models with proprietary internal context. - [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) AI paradigms are evolving from simple prompt-and-response Generative AI toward continuous, integrated Agentic AI and physical AI systems. - [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) An AI Factory transforms raw data and electrical energy into digital tokens, actionable insights, and measurable business outcomes across a five-layer technology stack. - [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) To unlock differentiation, organizations must adopt a 'data-first' architecture that decouples enterprise data from legacy applications to fine-tune specialized models. - [09:38](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=578s) Open-weight models are rapidly gaining traction over proprietary frontier models by offering greater data sovereignty, cost control, and deep customization using internal IP. - [13:14](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=794s) Corporate AI budgeting is shifting toward 'tokconomics,' evaluating hardware and infrastructure investments based on predictable token throughput versus business value. - [16:51](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1011s) AI factories are scalable to any size—ranging from desktop workstation clusters to multi-GPU arrays—making local AI execution accessible to mid-sized organizations. - [16:51](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1011s) The primary operational hurdle for smaller entities in owning AI capabilities is talent acquisition, which is increasingly addressed through managed service providers and systems integrators. - [19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) Future enterprise AI deployments will depend heavily on modular, outcome-based designs ('AI in a box') driven by an engineering mindset to deliver concrete business results.

## Detailed walkthrough Here is the detailed walkthrough of Part 1 of the panel discussion:

[00:04](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=4s) Introduction: Renting vs. Owning AI in the Enterprise

Harry McCracken, Global Technology Editor at *Fast Company*, introduces the session by addressing a central dilemma facing modern organizations: as artificial intelligence transitions into a core operational capability, should companies rely solely on third-party AI models, or lies the true competitive advantage in constructing custom, purpose-built intelligence?

To unpack this shift, McCracken welcomes two industry leaders to the AI4 stage: * **Kalyan Kumar (KK)**, President of Software Products at HCLTech (leading go-to-market strategies, product engineering, cloud operations, and shared enterprise functions). * **Kaushik Shirhatti**, Vice President of the AI Factory at Nvidia (leading strategic initiatives to help organizations deploy value-driven AI factories).

The session opens with a brief moment of lighthearted banter as KK jokes about the Las Vegas heat, suggesting that the current temperature of 109°F should ideally reach at least 115°F to really warm things up. Transitioning to the core topic, McCracken acknowledges that renting AI is appealing because it offers access to cutting-edge models while delegating heavy technical lifting to third parties. He asks KK why enterprises facing complex challenges should consider taking on direct responsibility for their own AI capabilities.

[01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) The Coexistence of Renting and Owning: Hybrid AI Strategies

KK emphasizes that the debate over renting versus owning AI is not a binary choice, but a complementary "and" strategy where both models must coexist. While third-party frontier models provide extraordinary general-purpose capabilities, an enterprise’s true differentiation resides inside its proprietary context—the data, IT systems, workflows, and business processes built and refined over the last four decades. Organizations must establish routing mechanisms to combine broad, general-purpose intelligence with customized, internal systems of differentiation.

Kaushik Shirhatti builds on KK's points, highlighting that enterprises are settling into a dual hybrid model that balances both cloud and on-premises environments, as well as frontier and open-weights models. Reflecting on how the landscape has evolved over the past year, Shirhatti illustrates the shift in AI paradigms: * **Generative AI** acts like a "vending machine": press a button, and it outputs a single answer. * **Agentic AI** operates like hiring a new employee: it requires proper onboarding, contextual training, ongoing feedback, and collaboration with other employees. Agents are persistent, continuous, self-evolving, and deeply integrated into core source code. * **Physical AI** represents the next upcoming frontier as AI continues to expand into non-digital domains.

[04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) Defining the AI Factory and Its Core Components

When McCracken asks Shirhatti to define an "AI Factory"—a concept popularized by Nvidia CEO Jensen Huang—Shirhatti begins by clarifying who the actual customer of an AI factory is. Rather than viewing giant institutions like Bank of America, Kaiser Permanente, or Google as the end customers, he frame them as partners. The true "customers" of an AI factory are the end users (e.g., bankers, clinicians), AI agents, and machines that rely on fast, reliable, and useful intelligence.

Comparing traditional manufacturing to artificial intelligence, Shirhatti outlines the core concept:

``` +-------------------+ +---------------------+ +-------------------+ | RAW INPUTS | ---> | AI FACTORY | ---> | OUTPUT VALUE | | Data & Electrical | | (5-Layer Tech Stack)| | Tokens, Insights, | | Energy | | Energy, Chips, Infra| | & Business Impact | +-------------------+ | Models, App Layer | +-------------------+ +---------------------+ ```

- **Traditional Factory Input/Output**: Raw materials and physical energy enter the factory to produce finished manufactured goods. - **AI Factory Input/Output**: Raw enterprise data and electrical power enter the system to generate digital tokens, actionable intelligence, and concrete business outcomes.

AI factories do not have to be massive infrastructure projects; an AI factory can be as compact as a single employee using an Nvidia DGX Station under her desk to process local tokens. Across all deployment scales, Shirhatti describes the underlying system as a "five-layer cake" consisting of energy, compute chips, infrastructure, models, and application software. Managing this stack functions like a relay race where every hardware and software handover must be seamlessly executed.

[06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) Data-First Architecture and Unlocking Enterprise Context

McCracken asks whether proprietary enterprise data provides a competitive edge regardless of whether a company rents cloud AI services or builds its own infrastructure. KK responds by cautioning against confusing AI with cloud computing, noting that they operate at fundamentally different layers. Historically, enterprise progress evolved through the mercantile, industrial, and information eras into today's age of intelligence, with process and data remaining the two constant drivers of differentiation.

However, KK highlights a major historical flaw in corporate IT design: organizations traditionally built processes, functions, and applications first, leaving data trapped behind siloed software architectures. AI forces a structural shift toward a "data-first" strategy: * **Data Decoupling**: Disentangling enterprise data from legacy applications and giving data its own identity. * **Data Management**: Defining clear metadata, data catalogs, data contracts, and dedicated data products. * **Model Customization**: Leveraging organized data to fine-tune Small Language Models (SLMs) or train open-weight models to capture domain-specific context. * **Traditional Machine Learning**: Retaining standard ML techniques for structured metrics data rather than forcing all processing tasks into generative AI models.

- Effective orchestration requires combining a data-intelligence strategy with automated process orchestration before routing requests to appropriate local or cloud-based models.

[09:38](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=578s) Open-Weight Models vs. Proprietary Frontier Models

McCracken notes that while leading frontier models have historically been proprietary, open-weight models have rapidly closed the performance gap. He asks if organizations lose strategic advantages when building on open-weight models instead of consuming turnkey AI services. Shirhatti explains that the rise of agentic AI is driving this re-evaluation due to growing operational costs and expanding workloads.

While end users simply want useful tools, infrastructure, security, and executive teams (such as CISOs and CFOs) must carefully balance cost, access control, security, and fine-tuning.

* **Proprietary Frontier Models**: Offer cutting-edge performance, rapid deployment out-of-the-box, and low setup friction, though with constrained deep customization. * **Open-Weight Models**: Can be downloaded directly into private infrastructure and fine-tuned using 30 to 40 years of internal corporate knowledge and institutional memory, unlocking high value from proprietary intellectual property.

KK adds that organizations must carefully distinguish between open-source software, open-weight models, and closed proprietary operating models, choosing their granular control based on specific operational needs.

[13:14](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=794s) The Economics of AI: Renting, Owning, and "Tokconomics"

McCracken turns to economic considerations, pointing out that paying per token can accumulate quickly ("token maxing"), whereas building private infrastructure incurs ongoing capital and operational management costs. KK reflects on how corporate technology budgeting has shifted from traditional CapEx to cloud-based OpEx, and is now evolving into a nuanced personal finance model of renting, leasing, or buying depending on usage patterns.

Small AI factories can be established using modest setups—such as a 16-GPU cluster—capable of running continuously to deliver predictable output, low latency, and continuous throughput.

Shirhatti explains that restricting innovative employees with strict token budgets hinders productivity and business outcomes. Instead, CFOs and CEOs must evaluate their investments through the lens of "tokconomics"—converting hardware into efficient token factories where token volume directly correlates with business value. Nvidia’s goal is to drive infrastructure efficiency so that for a static spend (e.g., $1 million), an enterprise can generate tenfold more intelligence year-over-year.

[16:51](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1011s) Democratizing AI Factories and Future Outlook

Addressing smaller and mid-sized enterprises, McCracken asks whether owning AI infrastructure is feasible without overextending internal resources. Shirhatti reassures that building an AI factory does not require a $50 million data center buildout. Smaller organizations, hospitals, local banks, and university research labs frequently deploy decentralized clusters or DGX workstation setups to generate substantial token volumes.

The primary operational constraint for smaller entities is not infrastructure cost, but attracting technical talent. To bridge this gap, organizations leverage managed AI service providers, co-location partners, and systems integrators like HCLTech to manage the full five-layer stack.

Looking ahead two to three years, McCracken asks if enterprise adoption will lean more heavily toward ownership. KK predicts that as data sovereignty and context differentiation become paramount, enterprises will build holistic, purpose-built AI strategies. Organizations will rely on a balanced architecture: utilizing local edge compute and on-premises infrastructure for core proprietary operations, while connecting to third-party frontier models via APIs for general-purpose requests.

[19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) Flexibility and the Future of the AI Factory

As the session wraps up, Kalyan Kumar (KK) emphasizes the paramount importance of flexibility when implementing purpose-built AI solutions in the enterprise. Organizations need the ability to structure their deployments around outcome-based models—taking specific business use cases and packaging them into an "AI in a box" format designed explicitly to deliver tangible business value. To successfully connect infrastructure, models, and real-world applications into a production-ready system, adopting an engineering mindset is essential.

Looking ahead, KK expresses strong optimism about the direction of the industry, predicting a significant increase in the adoption of the "AI factory" concept over the next two years as companies transition from merely renting third-party models to owning their intelligence. Moderator Harry McCracken concludes the session by thanking the panelists and audience for a high-value discussion and encouraging everyone to enjoy the remainder of the Ai4 conference.

- **Key Takeaways:** - AI implementations must prioritize outcome-based value delivery and modularity ("AI in a box"). - Enterprise success relies heavily on an engineering mindset to integrate infrastructure and software platforms. - Widespread adoption of dedicated AI factories is expected to accelerate significantly over the next two years.

For more details and full context, watch the video on YouTube: [From Renting AI to Owning Intelligence: The Case for Purpose-Built AI](https://www.youtube.com/watch?v=j8AwTGKeyJU).

## Notable quotes > [00:04](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=4s) Renting AI is appealing because it offers access to cutting-edge models while delegating heavy technical lifting to third parties. — Harry McCracken

> [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) The debate over renting versus owning AI is not a binary choice, but a complementary 'and' strategy where both models must coexist. — Kalyan Kumar (KK)

> [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) Generative AI acts like a 'vending machine': press a button, and it outputs a single answer. — Kaushik Shirhatti

> [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) Agentic AI operates like hiring a new employee: it requires proper onboarding, contextual training, ongoing feedback, and collaboration with other employees. — Kaushik Shirhatti

> [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) The true 'customers' of an AI factory are the end users (e.g., bankers, clinicians), AI agents, and machines that rely on fast, reliable, and useful intelligence. — Kaushik Shirhatti

> [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) Raw enterprise data and electrical power enter the system to generate digital tokens, actionable intelligence, and concrete business outcomes. — Kaushik Shirhatti

> [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) AI forces a structural shift toward a 'data-first' strategy — Kalyan Kumar (KK)

> [09:38](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=578s) Open-Weight Models: Can be downloaded directly into private infrastructure and fine-tuned using 30 to 40 years of internal corporate knowledge and institutional memory, unlocking high value from proprietary intellectual property. — Kaushik Shirhatti

> [13:14](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=794s) CFOs and CEOs must evaluate their investments through the lens of 'tokconomics'—converting hardware into efficient token factories where token volume directly correlates with business value. — Kaushik Shirhatti

> [16:51](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1011s) Building an AI factory does not require a $50 million data center buildout. — Kaushik Shirhatti

> [16:51](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1011s) As data sovereignty and context differentiation become paramount, enterprises will build holistic, purpose-built AI strategies. — Kalyan Kumar (KK)

> [19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) Organizations need the ability to structure their deployments around outcome-based models—taking specific business use cases and packaging them into an 'AI in a box' format designed explicitly to deliver tangible business value. — Kalyan Kumar (KK)

## Mentioned Harry McCracken, Kalyan Kumar (KK), Kaushik Shirhatti, Fast Company, HCLTech, Nvidia, Ai4, Jensen Huang, Bank of America, Kaiser Permanente, Google, Nvidia DGX Station

## Video description Harry McCracken - Global Technology Editor, Fast Company Kalyan Kumar (KK) - President - Software Products, HCLTech Kaushik Shirhatti - VP, AI Factory, Nvidia

For many enterprises, the first phase of AI was about access, and experimenting with powerful general-purpose models. The next phase is about ownership: building AI capabilities that are tailored to the business, grounded in enterprise data and deployed with the right levels of control, compliance and performance. This session explores how organizations can evolve from renting AI to owning intelligence by AI programs that connect infrastructure, platforms, models and use cases into a production-ready system for growth.

What the audience will take away: - Why enterprises are shifting from generic AI consumption to owned, purpose-built intelligence. - How ownership of AI can improve control, economics, compliance and long-term competitive advantage.- - Why sovereign AI is moving from niche requirement to board-level priority. - How control over infrastructure, data and models affects risk and value creation. - What it takes to deploy outcome-based AI implementations - Success stories

[Source](https://www.youtube.com/watch?v=j8AwTGKeyJU)

中文简报

英伟达与HCLTech高管力推定制AI工厂而非租用模型企业必须从租用第三方模型转向拥有定制的AI工厂，以保护专有数据并控制长期运营成本。专家组探讨了从使用第三方前沿AI模型向构建自主企业级AI能力转型的战略演进。发言嘉宾并未在“租用”与“自建”之间做绝对的选择，而是倡导采取混合模式，将开放权重模型、小语言模型以及私有数据优先架构整合为“AI工厂”。通过分析“Token经济学”（tokconomics）与智能体AI（Agentic AI）的崛起，他们阐述了定制化基础设施如何帮助企业保障专有上下文安全、优化长期运营成本并取得切实可见的业务成果。企业 AI 正在从单纯“租用”通用前沿模型迈向“拥有”定制化智能阶段；专家建议采用兼顾外部前沿模型与企业内部专有流程的混合架构，构建持久的差异化壁垒。 (1:47) 英伟达将“AI 工厂”定义为涵盖能源、芯片、基础设施、模型与应用的五层技术栈体系，能将原始业务数据和电能高效转化为驱动业务成果的数字 Token 与洞察。 (4:35) 企业构建自主智能需推行“数据优先”架构，将数据从旧应用中解耦，并借助具备完整数据主权和可结合企业知识产权微调的开放权重模型（Open-weight models）实现深度定制。 (6:49) 企业预算正转向衡量硬件吞吐量与业务回报挂钩的“Token 经济学（Tokconomics）”；AI 工厂具备高度可扩展性，从桌面级工作站到小型集群均可运行，使中小机构无需巨额机房投入也能实现本地化部署。 (13:14)

Reading summary

Kalyan Kumar of HCLTech and Kaushik Shirhatti of Nvidia argue that enterprise AI strategy must evolve from solely renting third-party cloud models into a hybrid approach centered on purpose-built, on-premises AI factories. They contend that direct infrastructure ownership and open-weight models are necessary to leverage proprietary corporate data and manage long-term token economics. Hybrid Strategies for Evolving AI Workloads (1:47) Kumar emphasizes that renting cloud models and building internal capacity is not a binary choice, but a complementary coexistence. While third-party frontier models provide strong general-purpose intelligence, corporate differentiation lies in proprietary data, workflows, and enterprise systems developed over decades. Organizations require routing mechanisms to combine external baseline models with customized internal environments. Shirhatti links this hybrid necessity to shifting AI paradigms. While standard generative AI functions like a transactional vending machine providing one-off outputs, emerging agentic AI behaves like an employee—requiring onboarding, contextual training, ongoing feedback, and deep integration into production codebases. This continuous workload increases operational reliance on stable, dedicated infrastructure alongside future physical AI deployments. The Five-Layer AI Factory and Data Architecture (4:35) Shirhatti defines an AI factory as a five-layer stack—spanning energy, compute chips, infrastructure, models, and application software—that takes raw data and electricity as inputs to produce digital tokens, insights, and business outcomes. Rather than viewing enterprise institutions as the end customer, Shirhatti identifies the true consumers as internal professionals, software agents, and connected machines that require dependable, low-latency intelligence. Kumar warns that maximizing this architecture requires rectifying a fundamental legacy IT flaw: applications were historically built first, trapping critical data inside functional silos. A viable AI strategy requires a data-first pivot that decouples data from legacy software, establishes formal data contracts and catalogs, and uses curated internal context to fine-tune small language models while preserving traditional machine learning for structured metrics. Open-Weight Models and Corporate Data Sovereignty (9:38) Although proprietary frontier models offer fast deployment and high baseline output, Shirhatti observes that enterprise leadership, including CFOs and CISOs, must balance out-of-the-box convenience against long-term cost, access controls, and security. Open-weight models deployed on private infrastructure allow organizations to fine-tune systems using 30 to 40 years of accumulated institutional knowledge without transmitting sensitive intellectual property to third parties. Kumar adds that enterprises must establish clear governance boundaries between open-source software, open-weight models, and proprietary commercial offerings, selecting model tiers according to specific operational needs and regulatory sovereignty constraints. Tokconomics and Accessible Infrastructure Scaling (13:14) Addressing the risk of per-token API cost inflation, Shirhatti introduces the concept of "tokconomics," arguing that placing strict token budgets on employees suppresses productivity and innovation. Instead, leadership should treat hardware as an amortized production engine designed to maximize token yield per dollar spent over time, allowing organizations to scale intelligence output under relatively static IT budgets. Shirhatti notes that building an AI factory does not necessitate a $50 million data center buildout. Smaller organizations, regional banks, and research labs can achieve continuous, low-latency throughput using setups as compact as a 16-GPU cluster or a single desktop workstation. Both speakers identify access to specialized technical talent, rather than capital hardware expense, as the primary barrier for smaller firms, necessitating partnerships with systems integrators and managed service providers. Kumar forecasts that enterprise adoption of dedicated AI factories will accelerate significantly over the next two years, driven by modular "AI in a box" architectures aligned with outcome-based business cases. However, the panelists note that realizing this shift depends on overcoming technical talent constraints and applying rigorous systems engineering to bridge local infrastructure with third-party cloud APIs.

精炼阅读摘要

在Ai4大会上，HCLTech的Kalyan Kumar与Nvidia的Kaushik Shirhatti指出，企业拥抱人工智能并非在租用第三方前沿模型与自建专用智能之间进行二选一，而是需要构建融合两者的混合AI战略；企业的长期核心壁垒在于将内部积累的专有数据和业务流程转化为自主可控的“AI工厂”。混合AI架构与智能体的协同演化 (1:47) Kalyan Kumar（KK）指出，租用第三方通用模型与自建专有智能是一种互补的共存策略。第三方前沿模型虽然具备强大的通用理解能力，但企业真正的差异化竞争优势存在于其在过去四十年间沉淀的专有背景之中，包括私有数据、IT系统、工作流及业务流程。因此，企业需要建立精细的路由机制，将广泛的通用智能与内部特有的业务系统高效融合。 Kaushik Shirhatti补充强调，企业正在走向平衡云端与本地部署、兼顾前沿模型与开源权重模型的双重混合架构。他指出AI范式正在经历从生成式AI到智能体AI（Agentic AI）的转变：生成式AI如同“自动售货机”，仅能响应单次按键输出；智能体AI则像“入职新员工”，需要长期的业务背景培训、持续的反馈机制以及与人类同事的协作，具备常驻运行、持续进化并深度嵌入核心代码的特征；而物理AI则是下一步向非数字实体领域延伸的关键前沿。 AI工厂概念与五层技术栈分解 (4:35) 针对Nvidia首席执行官Jensen Huang提出的“AI工厂”概念，Shirhatti明确了其服务定位：AI工厂的最终客户并非大型机构本身，而是一线的业务人员（如银行家、临床医生）、AI智能体以及依靠高效智能运转的终端机器。类比于传统工业制造输入原料与物理能源以生产工业品，AI工厂的本质是输入企业原始数据与电力，持续生产数字Token、决策见解与实际业务成果。 Shirhatti将AI工厂的技术底座拆解为“五层架构”，涵盖能源、算力芯片、基础设施、模型以及应用软件层，其高效协同犹如一场每个软硬件环节都必须精准交接的接力赛。他指出，AI工厂不仅限于超大型数据中心，其形态极具弹性，既可以扩展至大规模集群，也可以小型化至员工办公桌下运行的一台Nvidia DGX工作站。数据优先策略与开源权重模型优势 (6:49) 关于企业私有数据的战略价值，KK提示不能将人工智能与云计算混为一谈，两者处于不同的架构层级。历史上企业IT通常先构建业务流程和应用程序，导致关键数据被锁死在孤立的软件烟囱中。AI的普及倒逼企业推行“数据优先”策略，包括将数据与传统应用解耦并赋予其独立身份、建立完备的元数据管理、数据目录、数据契约与数据产品，并结合小语言模型（SLM）及传统机器学习处理特定任务。随着开源权重模型与专有前沿模型性能差距的快速收窄，Shirhatti与KK分析认为，闭源前沿模型具有低部署门槛和开箱即用的优势，但在深度定制上存在局限；开源权重模型则可直接下载至企业私有设施，允许组织运用自身沉淀数十年的内部知识库进行微调。面对智能体带来的持续调用需求，这一模式能够为首席信息安全官（CISO）和首席财务官（CFO）提供更高的数据主权掌控力、安全性与定制自由度。 Token经济学与中小企业落地路径 (13:14) 在IT经济学维度，KK谈到企业技术预算正从早期的资本支出（CapEx）和云端运营支出（OpEx），转变为根据使用模式灵活选择租用、租赁或直接购买的复合模式。企业通过部署一个由16块GPU构成的小型本地集群，即可在维持较低资本投入的同时，保障全天候、可预测的稳定吞吐量与低时延响应。 Shirhatti进一步提出了“Token经济学”（Tokconomics）的概念，反对通过严苛的Token预算限制员工的业务创新，建议管理层将底层硬件视为高效的Token生产工厂，促使Token产出与实际商业价值直接锚定。Nvidia的目标是在维持固定投入（例如100万美元）的情况下，推动企业每年获得的智能输出实现十倍级增长。针对中小机构的落地顾虑，Shirhatti认为硬件成本已非核心瓶颈，技术人才匮乏才是主要挑战，这促使企业转向托管AI服务提供商、托管数据中心及系统集成商寻求全栈工程支持。 KK预测，未来两到三年内，由于企业对数据主权与差异化核心竞争力的重视，以业务成果为导向的模块化“开箱即用式AI”（AI in a box）将加速普及；企业将普遍采用由本地边缘算力承载核心业务、由API接入第三方前沿模型处理通用请求的兼顾架构，而工程化整合能力将成为决定AI工厂落地成败的关键。

Detailed notes — zh

## Background 本视频由 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。视频记录了 Fast Company 全球科技编辑 Harry McCracken、HCLTech 软件产品总裁 Kalyan Kumar (KK) 以及 NVIDIA AI 工厂副总裁 Kaushik Shirhatti 在大会现场的研讨，主要探讨企业如何从租用通用 AI 模型转向自建专属智能，打通基础设施、平台、模型与业务场景。 ## TL;DR 随着人工智能深入企业核心业务，仅依赖第三方 API 的通用 AI 已无法满足对核心竞争力和数据安全的需求。嘉宾们指出，企业正在迈向“双重混合”与“并行共存”的新阶段，通过建立涵盖能源、芯片、基础设施、模型和应用的五层架构“AI 工厂”，将私有数据与开源权重模型或专用模型相结合。通过引入智能体 AI（Agentic AI）与结果导向的工程思维，企业能够精细化掌控成本与风险，真正拥有高度定制且可持续进化的专属智能体系。 ## Atmosphere & style 研讨会开场气氛轻松融洽，嘉宾们借拉斯维加斯的酷热天气幽默打趣，随即迅速切入硬核的技术与商业战略话题。对话节奏紧凑且富有洞察力，主持人提问切中要害，两位行业领袖则配合默契，将复杂的 AI 基础设施概念转化为“自动售货机 vs 招聘新员工”和“五层架构蛋糕”等生动比喻。整场讨论充满了前瞻性的技术工程思维与理性的商业经济考量，呈现出高端科技峰会上专业、务实且启发性极强的交流氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=4s) 企业 AI 正在经历从“租赁”通用模型向“拥有”业务专属智能（Purpose-built Intelligence）的范式转移。 - [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) 企业的 AI 战略并非非此即彼，而是采用“并行共存（AND, not OR）”与“双重混合（Dual Hybrid）”模式，有机结合云端/本地以及闭源前沿模型/开源权重模型。 - [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) AI 交互模式正从单次按需响应的生成式 AI（像“自动售货机”）演变为具备长期驻留、深入业务流程与源码的智能体 AI（像“招聘新员工”）。 - [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) “AI 工厂”（AI Factory）输入数据与算力能源，持续输出 Token 与业务成果；其服务对象是终端用户、智能体与机器设备。 - [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) 完整的 AI 工厂遵循“五层蛋糕”架构：包含能源、芯片、基础设施与网络、模型、应用与生态。 - [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) 企业的核心竞争壁垒在于业务流程与私有数据，AI 的落地倒逼企业实施“数据优先”战略，实现数据与应用的解耦。 - [09:38](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=578s) 开源权重模型（Open-weight Models）允许下载到本地私有环境并结合企业 IP 进行微调，在成本、安全和深度定制上具备独特优势。 - [13:14](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=794s) AI 工厂的经济学建立在明确的代币经济学（Tokenomics）、硬件 7x24 小时满载运转以及确定性产出的基础之上。 - [19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) 企业落地 AI 工厂需要具备工程化思维，采用基于结果导向（Outcome-based）的部署策略与开箱即用的套件。 - [19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) 预测未来两年内，企业对专属 AI 工厂和自主智能体系的采纳度将迎来爆发式增长。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=4s) 企业的 AI 抉择：租赁还是自建？在 AI4 大会现场，主持人 Harry McCracken（《Fast Company》全球科技编辑）拉开了研讨的序幕。随着人工智能逐渐成为企业运营的核心组件，各大机构正在面临一个根本性的战略抉择：仅仅依赖第三方的通用 AI 模型是否足够？抑或是将真正的竞争优势建立在“为自身业务量身定制”的专用智能（Purpose-built Intelligence）之上？为了探讨这一重大转变，主办方邀请了两位重磅嘉宾：HCLTech 软件产品总裁 Kalyan Kumar（简称 KK）以及 NVIDIA AI 工厂副总裁 Kaushik Shirhatti。在登台引出话题时，嘉宾们还轻松地开起了现场天气的玩笑，提到拉斯维加斯当天高达 109 华氏度（约 42.8 摄氏度）的高温，笑称“这还不够热，至少得升到 115 华氏度才算够味”。随后，话题迅速转向企业 AI 的基础设施布局。主持人指出，直接“租赁 AI”的优势显而易见——企业可以即时获取顶尖的前沿模型，且复杂的底层运维和“重活”都由云厂商代劳；那么企业在面对重大挑战时，为何还要考虑承担更多责任，去自建专属的 AI 基础设施呢？ - 演讲嘉宾包括 Fast Company 的 Harry McCracken、HCLTech 的 Kalyan Kumar (KK) 以及 NVIDIA 的 Kaushik Shirhatti。 - 讨论的核心议题是企业从通用 API 租赁向自建与专属 AI 基础设施的演进。 ### [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) 专用智能与混合架构：从自动售货机到智能体针对自建与租赁的博弈，KK 明确指出，这绝非非此即彼的单选题，而是一种“两者并存（AND, not OR）”的融合共生模式。企业固然可以继续利用前沿大模型（Frontier Models）提供的通用智能，但企业真正的核心竞争力在于内部沉淀了几十年的独特上下文、业务流程与私有数据。为了将这些企业独有的优势转化为智能化能力，必须建立内部的差异化系统，并通过智能路由机制将通用模型与自建的专用模型有机结合起来。 Kaushik 补充强调，企业正加速迈向云端与本地混合部署（Dual Hybrid Strategy）的新时代。与去年大热的生成式 AI（GenAI）不同，当前的 AI 演进正在经历向智能体 AI（Agentic AI）乃至物理 AI（Physical AI）的跨越。Kaushik 做出了一个形象的比喻：早期的生成式 AI 就像一台“自动售货机”，用户按下按钮（输入 Prompt），系统就吐出一个答案；而智能体 AI 则更像是“招聘一位新员工”——你需要对它进行入职培训（Onboarding），提供充足的业务上下文，在其出错时进行纠正，并安排它与其他员工协同工作。由于智能体具备持续性、持久性、自进化能力，甚至可以直接触及企业核心源代码，这种深度植入业务流程的特性极大地推动了企业对数据安全、自主可控以及本地化/开放模型（Open-weight Models）的刚性需求。 ``` 传统 GenAI 模式 (自动售货机) [ 用户 Prompt ] ---> ( 按下按钮) ---> [ 吐出单次回答 ] 智能体 AI 模式 (招聘新员工) [ 业务上下文 ] ---> ( 深度培训/校正) ---> [ 持续进化/协同执行 ] ---> [ 触及源码与核心流程 ] ``` ### [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) 什么是“AI 工厂”？五层蛋糕架构解析针对 NVIDIA 首席执行官黄仁勋（Jensen Huang）近年来大力推行的“AI 工厂”（AI Factory）概念，Kaushik 进行了详细的拆解。他首先重新定义了 AI 工厂的“客户”身份：真正的客户并不是名义上的企业实体（如美国银行或谷歌），而是那些实际使用 AI 的终端用户、AI 智能体以及各类机器设备。对于这些用户（如银行交易员、临床医生）而言，他们并不关心工厂内部复杂的“装配线”，他们只在乎 AI 是否足够快、足够实用。为了满足这种体验，产业界必须像接力赛一样紧密合作。传统工厂消耗原料和能源，产出物理实体商品；而 AI 工厂则是输入数据（Data）与算力能源（Energy），产出 Token、情报以及具体的业务结果（Business Outcomes）。构建一座完整的 AI 工厂包含所谓的“五层蛋糕”（Five-layer Cake）架构，需要各环节完美交接。 <Image alt="概念示意图：展示 AI 工厂从数据输入到 Token 和业务成果输出的过程" caption="AI 工厂：从数据到 Token 的转换流程" src="image_agent_tag_5086037000220319900"/> - **五层蛋糕架构组件**： 1. 能源（Energy） 2. 芯片（Chips） 3. 基础设施与网络（Infrastructure & Network） 4. 模型（Models） 5. 应用与生态（Applications & Ecosystem） - **形态的多样性**：AI 工厂既可以是大型数据中心，也可以是员工桌下配备的一台 NVIDIA DGX Station 桌面工作站。 <Image alt="NVIDIA DGX 架构基础设施示例" caption="NVIDIA DGX 系列 AI 基础设施示例" src="image_agent_tag_5086037000220322253"/> ### [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) 数据优先战略与模型分类讨论深入到企业数据的核心价值时，KK 指出“绝对不能将 AI 等同于云”，两者处于完全不同的层级。纵观从商埠时代、工业时代、信息时代到如今的智能时代，企业的差异化竞争力始终建立在两大基石之上：**业务流程（Process）**与**私有数据（Data）**。智能体恰恰就是流程与数据的交汇点。然而，许多企业历史上犯过一个错误——缺乏“数据优先”（Data-first）的顶层设计，导致数据被封锁在各个孤立的应用程序之后。为了迎接 AI 时代，企业必须重新解耦数据与应用，赋予数据独立的身份，建立数据目录、数据契约与数据产品。企业需要清晰划分数据资产的走向：哪些用于训练开源权重模型（Open-weight Model），哪些用于微调小语言模型（SLM），哪些用于构建企业专属模型。此外，KK 提醒大家不要盲目将所有任务都推给生成式 AI，传统的机器学习（ML）在处理指标和结构化数据方面依然高效。 - **开放权重 vs. 前沿闭源**： - **前沿模型（Frontier/Proprietary）**：开箱即用、性能极高、更新迅速，适合通用任务与无需深度定制的场景。 - **开放权重模型（Open-weight）**：允许企业将其下载到本地私有环境中，结合自身积累数十年的核心数据与知识产权（IP）进行微调，能将企业私有数据的价值发挥到极致。 - **架构重构核心**：企业需建立“数据智能策略### 引言与嘉宾介绍：从租赁 AI 到拥有智能的范式转变随着人工智能迅速演变为企业核心业务体系不可或缺的一部分，全球组织正面临着一项根本性的战略决策：究竟是继续依赖第三方通用 AI 模型服务，还是通过构建专门针对自身业务场景（Purpose-Built）的专属智能体系来获取核心竞争优势？针对这一行业痛点与技术转型路径，本次研讨会邀请到了两档顶级科技领袖：HCLTech 软件产品总裁 Kalyan Kumar（业内常称为 KK），以及 NVIDIA（英伟达）AI Factory 副总裁 Kaushik Shirhatti。KK 在 HCLTech 掌管着市场拓展、产品工程、云运营及所有共享企业职能，而 Kaushik 则在 NVIDIA 领导战略规划，旨在帮助各类企业搭建能够持续输出可衡量业务价值的“AI 工厂”（AI Factory）。现场主持人 Fast Company 全球科技主编 Harry McCracken 开启了对话，他打趣道“希望拉斯维加斯的天气能从现在的 109 华氏度上升到 115 华氏度”，随后迅速将焦点拉回主题。主持人提到，企业用户对“租赁 AI”（Renting AI）的好处早已了然于胸——既能即时获取最顶尖的前沿模型（Frontier Models），又可以将庞大的底层重活累活全部外包给云厂商。基于此，他向两位嘉宾抛出了核心议题：当企业面临日益复杂的业务挑战时，为什么应当开始考虑承担更多自主责任，去亲自构建属于自己的 AI 基础设施与专属智能？ --- ### [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) 专用智能与混合架构：为什么企业必须走向“兼收并蓄” KK 明确指出，企业的 AI 战略绝非非黑即白的“二选一”单行道（Not an either-or conversation），而是“并行共存”（And and）的混合演进。企业之所以需要着手“自主构建”（Build），核心动力来自于**专用智能（Purpose Intelligence）**的需求。过去四十年中，企业倾注了无数资源来梳理 IT 系统、业务工作流与数据资产，而这些沉淀在客户环境内部的独有上下文（Context），才是企业真正的核心差异化壁垒。前沿通用大模型固然拥有惊人的基础能力，但企业必须建立一套科学的路由机制（Routing Mechanism），将前沿通用能力与企业内部独特的上下文、系统差异化优势完美结合。 Kaushik 对此高度赞同，并补充道，未来绝大多数企业最终都会落脚于**双重混合战略（Dual Hybrid Strategy）**。传统的“云端 vs 线上/本地（Cloud vs On-Prem）”争议已持续了二十年，而在 AI 时代，企业将根据具体负载性质，同时采用云端与本地部署，并混合使用前沿闭源模型与开源权重模型（Open-Weight Models）。相较于去年的探讨，如今推动这一巨变的核心驱动力正是从生成式 AI（GenAI）向**智能体 AI（Agentic AI）**乃至物理 AI（Physical AI）的演进。为了形象阐述这一概念转变，Kaushik 提到了著名的“自动售货机与新员工”比喻： - **生成式 AI（GenAI）**：类似于一台**自动售货机**——按下一个按钮，掉出一个答案；再按一次，再给一个结果，缺乏长久的上下文延续。 - **智能体 AI（Agentic AI）**：类似于**招聘一名新员工**。你不可能只给新员工指令，你必须对他进行入职培训（Onboarding），输入海量业务上下文；如果发现偏差，还要不断纠偏；此外，该智能体还需与其他“员工”协作，具备长期驻留、持续进化并直接触碰底层源代码的能力。正是智能体这种高度持久化、自我演进且深度融入业务系统的特性，使得关于安全、数据隐私以及“本地 vs 云端”的讨论变得极其迫切。 ``` +-----------------------------------------------------------------+ | 企业双重混合 AI 战略 (Dual Hybrid) | +-----------------------------------------------------------------+ | +------------------------+------------------------+ | | v v +------------------------+ +------------------------+ | 生成式 AI (GenAI) | | 智能体 AI (Agentic) | | 交互模式: 自动售货机 | | 交互模式: 招聘新员工 | | 单次按需响应, 上下文弱 | | 深入源码, 持续学习协作 | +------------------------+ +------------------------+ | | +------------------------+------------------------+ | v +-----------------------------------------------------------------+ | 底层的算力与模型选型 | | 前沿闭源 (Frontier) + 开源权重 (Open-Weight) | 云端 + 本地 (On-Prem) | +-----------------------------------------------------------------+ ``` --- ### [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) 解密“AI 工厂”与五层架构蛋糕（Five-Layer Cake）当话题转向 NVIDIA 首席执行官黄仁勋（Jensen Huang）频繁提及的“AI 工厂”（AI Factory）概念时，Kaushik 重新定义了 AI 工厂的终极服务对象。他指出，真正的客户并非通常理解的美国银行（Bank of America）、凯aiser 医疗（Kaiser Permanente）或谷歌，这些机构都是生态构建的合作伙伴；**AI 工厂真正的“客户”，是那些正在驾驭 AI 能力的终端用户、AI 智能体（Agents）以及自动化机器**。例如，坐在银行柜台里的打理业务的交易员，或是医院里诊断病情的临床医生，他们根本不关心工厂流水线具体的工程细节，他们只要求 AI 响应够快、切实有用。从底层物理逻辑来看，传统工厂输入的是原材料与能源，输出的是工业成品；而 **AI 工厂则是输入数据与“智能能源”（算力），持续输出 Token、智能与最终的业务成果**。AI 工厂既可以是规模宏大的超大规模数据中心，也可以是员工办公桌下运行的一台 NVIDIA DGX Spark 工作站。无论形态大小，所有 AI 工厂都遵循极其严密的**“五层架构蛋糕”（Five-Layer Cake）**结构： - **第一层：能源（Energy）**—— 底层电力与物理能源供给。 - **第二层：芯片（Chips）**—— GPU/TPU 等核心算力硬件。 - **第三层：基础设施（Infrastructure）**—— 网络架构、服务器与存储集群。 - **第四层：模型（Models）**—— 基础大模型、开源权重模型与微调模型。 - **第五层：应用（Applications）**—— 贴合实际场景的业务应用与智能体服务。 Kaushik 强调，构建 AI 工厂就像是一场接力赛（Relay Race），硬件、软件、网络与生态系统必须完美配合，确保每一次“接力棒的传递”零失误，才能高效率地生成 Token 并转化成商业价值。 --- ### [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) 数据优先战略：重新打通企业内部的解耦与重构 KK 深入剖析了企业数据与 AI 的深层关系。他警示行业：绝不能将“AI”与“云”画等号，因为两者处于完全不同的层级。通用智能是建立在海量公共数据集与互联网信息的蒸馏基础上的，而企业的核心竞争壁垒——无论是在重商主义时代、工业时代、信息时代，还是如今的**智能时代（Age of Intelligence）**——始终由两大要素决定：**流程（Process）与数据（Data）**。智能体（Agents）正是业务流程与数据的交叉点。然而，绝大多数企业在历史上都有一个致命缺陷：**从来没有采用过“数据优先”（Data-First Approach）的建构思路**。过去的做法是先定义业务功能、开发应用程序，最后把数据死死锁在孤立的软件系统之后。AI 的爆发正倒逼企业重新思考：如何实现数据与应用的解耦（Delinking data and app）？如何赋予数据独立的身份认同？为了打造生产级的 AI 工厂，企业必须重塑自身的数据治理架构，包括严格界定： - **数据元数据（Metadata）与数据目录（Data Catalog）** - **数据契约（Data Contracts）与数据产品（Data Products）** - **选择性微调与蒸馏**：明确哪些数据用于微调小语言模型（SLM），哪些用于训练开源权重模型，从而打造企业专属的上下文。 KK 还特别提到了被许多人忽视的一点：**并非所有数据都需要推给生成式 AI**。对于大量的指标与结构化数据（Metrics Data），传统的机器学习（Machine Learning）依然极为高效。只有建立起“数据智能策略”与“流程编排”两大基石，企业才能真正驾驭 AI 工厂。 --- ### [09:38](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=578s) 开源权重 vs 前沿闭源模型：成本、安全与微调的博弈面对开源权重模型（Open-Weight Models）快速追赶前沿闭源模型的现状，主持人询问：如果企业选择基于开源权重构建自己的 AI 体系，是否会失去性能优势？Kaushik 分析指出，这一变化的核心拉动力依然是智能体 AI 的落地。随着 AI 变得越来越管用，企业部署的业务负载成倍增加，这就把三大核心议题推到了决策层的面前：**成本（Cost）、安全性（Security）与微调能力（Fine-Tuning）**。虽然前沿闭源模型速度极快、性能卓绝且开箱即用，但开源权重模型的战略价值在于“权重可下载”： - 企业可以**下载模型权重并在私有环境中部署**； - 结合企业过去 30 年、40 年积累的独有知识产权（IP）与机构知识（Institutional Knowledge）进行**深度微调（Fine-Tuning）**； - 一旦私有数据与开源权重融合，该模型产生的商业价值将呈现爆发式增长。 Kaushik 总结道，如果在一年前，答案可能是前沿模型单方面碾压；但今天，客户正在变得非常聪明，他们会进行精细化的负载划分——在需要极致创新与不需定制的场景使用前沿闭源模型；在涉及核心数据资产、需要高度定制与成本可控的场景，全面拥抱开源权重模型。KK 补充提醒，行业内必须清晰区分开源代码（Open Source）、开源权重（Open Weight）与闭源（Closed/Proprietary）这三种完全不同的运营模式与粒度控制。 --- ### [13:14](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=794s) 经济学账本：Token 计费、硬件折旧与“代币经济学” 在讨论财务与经济效益时，KK 指出企业 IT 的预算模式正在经历又一次重大轮回：企业最初高度依赖资本支出（CapEx），随后全面转向云计算的运营支出（OpEx）；而现在，企业就像在个人财务中衡量“租用、租赁还是购买”（Rent, Lease, or Own）一样，重新审视算力资产。建立 AI 工厂的经济学极其明确： - 以最小的 AI 工厂为例，仅需 16 块 GPU 即可起步； - 硬件基础设施可以 **7x24 小时满载运转**，吞吐量（Throughput）与产出上限完全可预测； - 能够根据速度、时间、成本与性能均衡计算整体投资回报（ROI）； - 结合边缘计算（Edge）与本地 PC 的算力调度，通过路由机制实现成本的最优化平衡。 Kaushik 从 CEO 与 CFO 的角度切 ### [19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) 从出租 AI 到拥有智能：AI 工厂的灵活性与工程思维讨论在对 AI 工厂（AI Factory）未来发展的展望中走向尾声。演讲者特别强调，未来企业应用 AI 的核心在于**灵活性**与**工程化思维**。企业可以根据具体的业务场景，采用基于结果导向（outcome-based）的模型构建方式，例如将 AI 打造成类似“开箱即用”（AI in a box）的套件，直接为企业输出明确的业务价值。要将这些复杂的系统真正落地并运转起来，具备坚实的工程思维（Engineering mindset）至关重要。演讲者对未来的行业趋势持非常乐观的预测，认为在未来两年内，市场对于 AI 工厂这种模式的采纳度（uptake）将会迎来爆发式增长。企业正在从最初简单“租用”通用 AI 能力，全面转向深度构建与“拥有”专属智能体系。最后，主持人对全体嘉宾的高价值讨论表示感谢，并祝愿现场观众在接下来的 Ai4 大会活动中收获满满。 - **核心落地模式**：采用基于结果（Outcome-based）的部署策略，将特定业务场景打包为“开箱即用 AI”（AI in a box）以直接释放业务价值。 - **关键能力支撑**：工程化思维（Engineering mindset）与系统搭建能力是决定 AI 工厂能否成功的核心要素。 - **行业趋势预测**：未来两年内，定制化、专属化的 AI 工厂将在企业级市场中得到更广泛的采纳与普及。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=4s) 这还不够热，至少得升到 115 华氏度才算够味 — Harry McCracken > [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) 这绝非非此即彼的单选题，而是一种“两者并存（AND, not OR）”的融合共生模式 — Kalyan Kumar > [01:47](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=107s) 早期的生成式 AI 就像一台“自动售货机”，用户按下按钮（输入 Prompt），系统就吐出一个答案；而智能体 AI 则更像是“招聘一位新员工” — Kaushik Shirhatti > [04:35](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=275s) 真正的客户并不是名义上的企业实体，而是那些实际使用 AI 的终端用户、AI 智能体以及各类机器设备 — Kaushik Shirhatti > [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) 绝对不能将 AI 等同于云，两者处于完全不同的层级 — Kalyan Kumar > [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) 过去的做法是先定义业务功能、开发应用程序，最后把数据死死锁在孤立的软件系统之后 — Kalyan Kumar > [06:49](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=409s) 并不是所有数据都需要推给生成式 AI，对于大量的指标与结构化数据，传统的机器学习依然极为高效 — Kalyan Kumar > [09:38](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=578s) 企业可以下载模型权重并在私有环境中部署，结合企业过去 30 年、40 年积累的独有知识产权与机构知识进行深度微调 — Kaushik Shirhatti > [13:14](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=794s) 像在个人财务中衡量“租用、租赁还是购买”一样，重新审视算力资产 — Kalyan Kumar > [19:56](https://www.youtube.com/watch?v=j8AwTGKeyJU&t=1196s) 未来企业应用 AI 的核心在于灵活性与工程化思维 — Kaushik Shirhatti ## Mentioned Harry McCracken, Fast Company, Kalyan Kumar (KK), HCLTech, Kaushik Shirhatti, NVIDIA, Jensen Huang, Ai4, NVIDIA DGX Station, NVIDIA DGX Spark, Bank of America, Kaiser Permanente, Google ## Video description Harry McCracken - Global Technology Editor, Fast Company Kalyan Kumar (KK) - President - Software Products, HCLTech Kaushik Shirhatti - VP, AI Factory, Nvidia For many enterprises, the first phase of AI was about access, and experimenting with powerful general-purpose models. The next phase is about ownership: building AI capabilities that are tailored to the business, grounded in enterprise data and deployed with the right levels of control, compliance and performance. This session explores how organizations can evolve from renting AI to owning intelligence by AI programs that connect infrastructure, platforms, models and use cases into a production-ready system for growth. What the audience will take away: - Why enterprises are shifting from generic AI consumption to owned, purpose-built intelligence. - How ownership of AI can improve control, economics, compliance and long-term competitive advantage.- - Why sovereign AI is moving from niche requirement to board-level priority. - How control over infrastructure, data and models affects risk and value creation. - What it takes to deploy outcome-based AI implementations - Success stories [Source](https://www.youtube.com/watch?v=j8AwTGKeyJU)

Source quotes

[0:04] Harry McCracken: 这还不够热，至少得升到 115 华氏度才算够味 [1:47] Kalyan Kumar: 这绝非非此即彼的单选题，而是一种“两者并存（AND, not OR）”的融合共生模式 [1:47] Kaushik Shirhatti: 早期的生成式 AI 就像一台“自动售货机”，用户按下按钮（输入 Prompt），系统就吐出一个答案；而智能体 AI 则更像是“招聘一位新员工” [4:35] Kaushik Shirhatti: 真正的客户并不是名义上的企业实体，而是那些实际使用 AI 的终端用户、AI 智能体以及各类机器设备 [6:49] Kalyan Kumar: 绝对不能将 AI 等同于云，两者处于完全不同的层级 [6:49] Kalyan Kumar: 过去的做法是先定义业务功能、开发应用程序，最后把数据死死锁在孤立的软件系统之后 [6:49] Kalyan Kumar: 并不是所有数据都需要推给生成式 AI，对于大量的指标与结构化数据，传统的机器学习依然极为高效 [9:38] Kaushik Shirhatti: 企业可以下载模型权重并在私有环境中部署，结合企业过去 30 年、40 年积累的独有知识产权与机构知识进行深度微调 [13:14] Kalyan Kumar: 像在个人财务中衡量“租用、租赁还是购买”一样，重新审视算力资产 [19:56] Kaushik Shirhatti: 未来企业应用 AI 的核心在于灵活性与工程化思维

Source quotes

[0:04] Harry McCracken: Renting AI is appealing because it offers access to cutting-edge models while delegating heavy technical lifting to third parties. [1:47] Kalyan Kumar (KK): The debate over renting versus owning AI is not a binary choice, but a complementary 'and' strategy where both models must coexist. [1:47] Kaushik Shirhatti: Generative AI acts like a 'vending machine': press a button, and it outputs a single answer. [1:47] Kaushik Shirhatti: Agentic AI operates like hiring a new employee: it requires proper onboarding, contextual training, ongoing feedback, and collaboration with other employees. [4:35] Kaushik Shirhatti: The true 'customers' of an AI factory are the end users (e.g., bankers, clinicians), AI agents, and machines that rely on fast, reliable, and useful intelligence. [4:35] Kaushik Shirhatti: Raw enterprise data and electrical power enter the system to generate digital tokens, actionable intelligence, and concrete business outcomes. [6:49] Kalyan Kumar (KK): AI forces a structural shift toward a 'data-first' strategy [9:38] Kaushik Shirhatti: Open-Weight Models: Can be downloaded directly into private infrastructure and fine-tuned using 30 to 40 years of internal corporate knowledge and institutional memory, unlocking high value from proprietary intellectual property. [13:14] Kaushik Shirhatti: CFOs and CEOs must evaluate their investments through the lens of 'tokconomics'—converting hardware into efficient token factories where token volume directly correlates with business value. [16:51] Kaushik Shirhatti: Building an AI factory does not require a $50 million data center buildout. [16:51] Kalyan Kumar (KK): As data sovereignty and context differentiation become paramount, enterprises will build holistic, purpose-built AI strategies. [19:56] Kalyan Kumar (KK): Organizations need the ability to structure their deployments around outcome-based models—taking specific business use cases and packaging them into an 'AI in a box' format designed explicitly to deliver tangible business value.

Original transcript

[0:04] As AI becomes a core part of the enterprise, organizations are facing a fundamental question. [0:11] Is it enough to rely on thirdparty AI models or is the real advantage in building intelligence [0:17] that is purpose-built for your business? Helping organizations navigate that shift is HCL software, [0:24] a company focused on building the next generation of enterprise software. Kalyan Kumar KK serves as [0:33] president of software products for HCL Tech where he leads go to market product engineering cloud [0:40] operations and all shared enterprise functions. Joining him on stage is Koshek Sherati who is [0:48] the vice president of AI factory at NVIDIA where he leads strategic initiatives to help [0:53] organizations build AI factories that deliver measurable business value. Please welcome to [1:00] the AI4 stage Kan Kumar and Kashak Sharate KK and Kachek. Thanks so much for joining us today. It's [1:08] a pleasure. Thank you for the invite. Good morning everybody. Good morning. We need the weather in [1:14] Vegas to be a little warmer. I saw it's 109. We need it to be 115 at least. KK, I don't think the [1:22] people in this room need a lot of explanation for why renting AI is useful in a lot of cases. you [1:28] you get access to the best models and somebody else takes on the heavy lifting. So let's talk [1:34] about building your own AI infrastructure and as enterprises are facing their biggest challenges [1:41] why might they want to at least consider taking on more responsibility for their own AI. Thanks. [1:47] Thanks Harry and good morning. So it's not an and or conversation. It's always together. [1:54] You would need both to coexist. The reason you start to look at why you need to start to build is [2:02] primarily around purpose intelligence. Enterprises have a lot of context and those information which [2:08] exist within the customers environment. It's taken like last four tickets for information systems, [2:14] IT systems, enterprise workflows, processes, data to be streamlined. But as you start to move into [2:20] the intelligence layer, a lot of the enterprise differentiation is in the data. And hence if you [2:27] want to start to build an intelligence model, build your information context. It's extremely [2:32] important to decide what's general purpose intelligence, extremely great capability [2:37] available from the frontier models. But how do you then bring all of them back together and [2:41] say what's enterprise differentiation, systems of differentiation, what's my uniqueness? and [2:48] then how do I blend both of them by some kind of a routing mechanism. So I think it's you [2:53] need both but you need to start to look at built differentiation with your own systems. I think [2:58] it's a very good point right which is it's it's a very subtle thing. It's not, you know, open [3:06] versus or versus frontier. It's and and really if you kind of take a step back for a second, [3:12] what's driving this is most of the enterprises we believe will end up in a dual hybrid strategy. [3:21] um the the cloud versus on-prem discussions been going on for 20 years but now there will be cloud [3:27] and on-prem based on the workloads and there will be frontier models and openwave models [3:33] and fundamentally what's driving that and what's different between last year when we were all here [3:38] was geni aentic AI and physical AI is around the corner because when you think about genai it was [3:47] more like a vending machine I think about it like a vending machine you know you Press a button, [3:51] you get an answer. You press a button, you get an answer. Aentic AI, that's like hiring a new [3:56] employee. And when you hire a new employee, what's the first thing you do? You onboard [4:00] that employee. You give them as much context. Then you realize that you haven't onboarded the [4:05] employee correctly. So you have to correct it. And then that employee has to work with other [4:10] employees. And so agents are now deeply rooted. They're persistent. They're continuous. They're [4:16] self-evolving. They touch your source code. And so all of that as agents have become more and [4:22] more useful is driving that discussion of on-prim versus cloud, open versus frontier. Uh and I think [4:30] we believe that it's going to be an end. Kashak your CEO Jensen started talking about the concept [4:35] of the AI factory and when Jensen starts talking about something everybody starts talking about [4:41] it. Step us through exactly what an AI factory is in the context of this conversation. Yeah, [4:47] that's a great question. Um, let me start with who the customer is for an AI factory. When a lot of [4:54] times we think about customers, we think of Bank of America as a customer or Kaiser Permanente or [5:01] Google and all of these companies. None of these are the customers. We're all partners in this. [5:06] We're all doing this together. The real customers are the users, the agents, and the machines that [5:12] are harnessing the power of AI. So now when you think about the user, you think about the banker [5:18] working at a bank, you think about the clinician, what do they need? What do they care about? They [5:24] don't care about the assembly line in the factory. They care about the experience. They want AI to be [5:28] fast. They want AI to be useful. Now to make that happen, all of us have to come together [5:35] and do the hard things. And so when we think about an AI factory, a lot of times people think about, [5:40] you know, let's look at the basic definition. you know a traditional factory you have raw materials [5:45] going in you've got energy going in and you get a a finished good a finished product and similarly [5:51] when we think about an AI factory you got data going in you got intelligent uh you know energy [5:57] going in and you get tokens intelligence business outcomes coming out and these are not just like [6:02] gigantic huge factories right there of course we build huge factories but it could be an employee [6:08] with a DJX Spark under her desk who's generating tokens And then for enterprises you have you know [6:15] amazing choices. You can rent it from the cloud or you can build your own factory. But the basic [6:21] components of the factory remain the same across what we call as the five layer kick which is the [6:27] energy, the chips, the infrastructure, the models and applications. And it's like a relay race. [6:32] You take compute, you take network, you take software and you take all of the ecosystem. And [6:38] just like a relay race, you have to make sure you are perfecting the handover. And when all [6:44] of that comes together, you generate tokens, you generate outcomes. And that's what we call as an [6:49] AI factory. KK, you you touched on the fact that the real competitive advantage in AI is is not in [6:56] the models you use, it's in the unique proprietary data your enterprise has. Um, is that equally true [7:04] whether you're using AI as a cloud service or choosing to build your own or do do people who do [7:11] take on more of this responsibility get even more opportunity to take advantage of their data? So we [7:18] should not equate AI with cloud because they sit effectively at two different layers. Intelligence [7:26] is built upon a lot of information and context. General purpose intelligence has always been built [7:33] using public data sets distillation of whatever different content and there's always a big debate [7:39] saying that how it's being built and and the whole process. When you come back to the enterprise, [7:44] go back from the modern renaissance era, from the merkantile age to the industrial age to the age of [7:51] information and now the age of intelligence. Two things have consistently been common for [7:57] enterprise differentiation, process and data. And that's where enterprises need to really start to [8:03] look at what's the inter like koshik talked about agentic use cases. Agents are the intersection of [8:08] process and data or a function human executing activities. You treat agent as onboarding an [8:14] agent just like you onboard an employee or are they a co-orker or they the primary lead or an [8:20] assist but enterprises have one big flaw and based on our experiences we never built with [8:27] data first approach historically the enterprises we we looked at process we created functions we [8:31] built applications and systems the data was locked behind them. AI is forcing companies to reimagine [8:38] saying that how do you dealinking data an app? Can data has its own identity? How do I define [8:44] the definition of data metadata data catalog data contracts data products? Then you decide [8:51] what goes into training an openweight model fine-tuning an S lm to make it an enterprise [8:57] context build my own small language model. And one very important thing people have forgotten that [9:04] there is something called machine learning for metrics data. You don't need to shove everything [9:07] to genai. You still can do a lot of those things back. So you classify your data. You decide this [9:12] piece. This effectively means that companies need to rethink their entire approach. And we believe [9:18] that a data intelligence strategy along with a process orchestration will be the foundation stone [9:25] for building enterprise AI factory intelligence. Then then you decide based on routing what goes [9:31] to the cloud and what goes to your own models. So regardless of how you go about it being data first [9:38] is going to give you an advantage. Absolutely. How um historically the leading frontier models have [9:44] come from a handful of companies and have been proprietary. Uh we've seen openweight models do [9:52] a tremendous amount of catching up just recently. But if you uh build your own AI and you build it [9:58] on openw weight models, are you giving up anything versus if if you used AI as a service? Yeah, [10:04] I think again that's a great question and again I'll go back to what's driving all of this up. [10:09] It's you know agentic AI is driving a huge amount of discussion now and the fundamental reason why [10:18] it's driving this discussion is agents are useful AI is useful people use it more people use it more [10:27] you're getting more business outcomes enterprises are realizing new and new workloads that they [10:32] can put in once you start doing that there are few things that come in the forefront first of [10:39] of all cost, right? We all talk about cost, so I'm not going to talk too much about the cost, [10:43] right? It's it's obvious like people have to think about the cost and the budget and the tokens and [10:48] all of that stuff. And that's what the CFO thinks. But then there's the huge infrastructure team. [10:54] The user shouldn't care. Like I said, to the user, it shouldn't matter what model. As long [10:58] as AI is useful, they can do amazing work. But the infrastructure teams, the security teams, [11:03] the people who are making decisions on what models, the CESOs, what connectors they can use, [11:10] these guys are really thinking about the balance. And what's driving that is cost, security as well [11:19] as finetuning, right? So when you think about the proprietary models, they are super useful. [11:26] We have some of the best proprietary models today, right? You can be very fast, extremely performant [11:34] and you can have some level of customization. But the reason open models become very very important [11:42] and when we say open weights, what do we mean by that? What you mean by weights is you can [11:47] actually download it in your own enterprise and then you can fine-tune that with your [11:53] own data with your own intellectual property. All the institutional knowledge that a bank [12:00] has collected over the last 30 years, 40 years, once you're able to fine-tune that with the data, [12:07] suddenly the value of that model is tremendous. And so what I expect to see is if you had asked [12:15] me this question a year and a half ago, the answer was clearly no. But today, what I expect to see [12:22] that it's a nuanced answer. Customers are going to be very smart. They will look at the workloads. [12:29] Some they will use the openweight models and then where you need good performance and you want you [12:36] don't want customization uh you want cutting edge you're going to use frontier models. I [12:41] just want to add there's also this whole confusion which is industry around open source versus open [12:47] weight versus closed. I think it's again how much granularity you want to really have in building [12:56] and engineering and and sometimes you want to take it down to the physical layer correct in the [13:00] previous panel they talked about 80% of the world is still physical right so I think as you start to [13:04] take that up understanding those three as koshik mentioned becomes very very important open source [13:09] open weight versus frontier closed or whatever the operating model is let's talk a little bit more [13:14] about the economics um paying by the token can add up. If you're token maxing, it can really add up. [13:22] But building your own infrastructure and paying people to manage it also costs money over time. [13:29] Is one of these approaches more cost effective than the other? Think economics for enterprise. [13:39] There are a couple of things which is changing our IT or technology budgeting cycles have anyway [13:44] fundamentally completely shifted. We were once upon a time completely capex heavy to pivoting [13:50] towards cloud completely moving to opex and now you starting to realize that it's goes back to [13:56] even on a personal finance you rent lease own you make those decisions it's like power I think the [14:04] economic model and as koshik pointed out on tokens if you build an AI factory and the smallest AI [14:10] factory I think koshik is what 16 GPUs is that the smallest one you can start with you know [14:15] what you can get with that you know the sized up infrastructure it can run 24 by7 you know what's [14:21] the max throughput output going back and then you then start to decide what benefit I'm getting out [14:26] of this what's the speed time cost performance value and then you connect it to the model so [14:33] I think the economics is all about you want it fast speed there is a different cost model you [14:39] want it to run for longer term it's a different cost model and that's where orchestration running [14:46] edge has latency needs localization running stuff on your local PC and and running local actions [14:53] and then you're able to route across local your own infrastructure or to the frontier model and [14:59] a mix of all of them. So I think economics has a lot of other aspects to play in but long-term if [15:04] you believe that you can depreciate and extract the value of an AI infrastructure then you should [15:09] look at that from that lens. Any thoughts? Yeah, I mean if any of you guys if you are the CEO of your [15:16] company and you have some amazing employees who are the power users and who are doing incredible [15:24] body of work and bringing innovation to your processes, they're reimagining processes. They're [15:30] changing the arc of your company. Do you want to get in a meeting with them and tell no sorry your [15:36] token budget is up to this many? You don't. Right? And so that's where it starts. If you're a CFO, [15:43] if you're a CEO, what you want to be able to do is you want to make sure that the tokens, [15:48] not all to tokens are equal, that the tokens equal to usefulness, tokens equal to innovation, [15:57] take tokens equal to business outcomes. And then you have to look at the infrastructure investment [16:03] I make. How do I convert an AI factory to an token factory and what the tokconomics looks like? And [16:09] in that you're going to have a rent versus buy decision. You'll have a frontier model versus an [16:16] open decision. But ultimately what you don't want to do is and this is something that Nvidia takes [16:24] very very seriously which is we believe that if today you let's say spend a million dollars and a [16:31] company is generating this much intelligence next year for the same million dollars everybody should [16:38] be able to generate like 10x more intelligence for the same dog right and it's not just Nvidia [16:45] it's not just HCL It's everybody all the decisions that the customers make, infrastructure decisions [16:51] make and that's where token maxing comes in. Large companies always have the option of taking things [16:58] on premises. Um but 16 GPUs doesn't sound all that intimidating. Um is it possible [17:05] for mediumsiz and smaller companies also to go in this direction without taking on more than [17:10] they can handle? 100%. You know I started with AI factories. When people think about AI factories [17:17] a lot of time people thinks about this like huge gigawatt scale factories. And similarly when when [17:22] I am talking about onprem I think most people would think about like a big bang a $50 million [17:29] data center build out but that's not the case you know even before you go to 16 GPUs like I said you [17:35] can start with a DJX Parker station and you can build your own factory but you can you know we [17:42] have hundreds of customers hospitals small banks you know small labs universities where you can [17:48] start with a very decentralized his cluster um and you know start generating huge amount of tokens, [17:56] start generating great workloads. The bottleneck really in that is not the cost. The bottleneck [18:02] really is the talent right is that do these companies have the talent uh to go you know think [18:09] through all the workloads and bring innovation there and this is where somebody like HCL and [18:15] a lot of other providers you know colo providers where they can actually do completely managed AI [18:21] solutions where really you are now not renting it out but you have a partner through the five layer [18:28] cake who can actually go be a your kind extended team to go generate that. KK, do you expect if [18:35] we have this conversation again in two or three years that more of the people in this room will [18:40] have taken on more responsibility for their AI or or will the balance between renting and owning [18:45] remain fairly consistent over time? Yep. Just to chime this 30 seconds on what Koshi said. When you [18:51] look at when you're taking this decision, renting is consuming like an API with a client, correct? [18:59] Yes. start to look at partners who can look at the full stack AI solution thinking engineering le uh [19:07] you have to start thinking AI will exist in every nook and corner of the enterprise in the edge in [19:13] the inferencing in the intelligence in physical AI in your PC as Kashik talked about there's [19:18] enough compute intelligence local and then into that so think of a partner who can look at this [19:22] whole thing now coming back in two years I'm quite convinced that the way enterprise context and data [19:31] and data differentiation within an enterprise is companies will want to build up the AI strategy [19:39] which is holistic for purpose-built intelligence. So they will still use frontier connect to consume [19:46] and then they'll partner with the right providers who can help them build this infrastructure with [19:51] various different consumption model. It could be you could build it in a leasing kind of a [19:56] model. You could build it in an outcome based. You could take business use cases, think of like AI in [20:01] a box, put it in to give me a outcome value. So flexibility is going to be extremely important. [20:06] Engineering mindset to be able to put that thing is going to be important. So I'm quite wondering [20:09] in two years you're going to see more uptake of the AI factory. This has been a very high [20:14] value discussion. Thank you so much for taking the time. Enjoy the rest of your AO4 everybody.