Flexera Expands FinOps Platform to Curb Runaway Enterprise AI Spending

Flexera扩展FinOps平台以遏制失控的企业AI支出 Flexera产品经理Scott Bulloch警告称，受补贴的企业AI模型掩盖了实际算力成本，若不通过近乎实时的追踪进行管理，将对企业预算构成威胁。 Scott Bulloch探讨了AI技术、单点解决方案和自主智能体的快速普及如何导致企业面临前所未有的“曲棍球棒式”成本激增及严重的账单冲击。他重点指出从模型训练向推理阶段的转变，解释了复杂的定价模式（即“Token经济学”）和供应商补贴的订阅费用是如何掩盖企业真实支出的。依托Flexera在IT资产管理和FinOps原则方面的基础，Bulloch提出了实现实时可见性、模型选择和算力优化的策略。最后，他强调企业管理者正在超越单纯的成本担忧，转而聚焦于明确的投资回报率（ROI）和跨团队协同。 随着自主 AI Agent 与工作负载的爆发式采用，企业面临前所未有的指数级成本飙升，而各单点解决方案的拼凑也导致了严重的数据滞后与工作流碎片化问题。 (0:04) 企业 AI 的成本驱动主体已由人类用户延伸至智能体、自动化项目及算力资源；Flexera 借助其 IT 资产管理与云成本管理经验，为应对 AI 治理提供关键的资产与算力利用率洞察。 (7:10) 企业正面临严峻的 AI 账单冲击（如 Uber 在一个季度内耗尽全年 AI 预算），支出重心已全面转向推理与执行步骤，且现有服务商的巨额算力补贴意味着未来企业实际成本将进一步攀升。 (10:56) Flexera 正将 Flexera One 拓展至 AI 成本管理领域，提供数小时内的近实时支出可见性与模型选型优化，推动企业管理层将重心从评估成本转向衡量清晰的投资回报率（ROI）。 (14:45)

Scott Bulloch, Senior Product Manager at Flexera, argues that rapid enterprise adoption of AI agents and inferencing is driving an unprecedented cost crisis, requiring organizations to apply IT asset management and FinOps principles to control spending and prove return on investment. Anatomy of the Enterprise AI Cost Crisis (10:56) Enterprise AI adoption is triggering severe end-of-month invoice shocks, with Bulloch observing that "AI just doesn't innovate, it invoices." For example, Uber exhausted an entire annual AI budget within a single quarter. Commercial spending has pivoted almost entirely from training models to ongoing inferencing, prompt design, and data I/O, while custom model training is now largely restricted to specialized fields like medicine and data science. Bulloch warns that current software-as-a-service AI pricing is heavily subsidized by providers: for a $200 subscription, Anthropic pays roughly $800 in compute expenses, and OpenAI pays nearly $1,400, foreshadowing sharp enterprise price hikes. Furthermore, enterprise "tokconomics" encompasses not just raw token volume, but provider charges for execution steps, workflow actions, and media rendering. In response, some CIOs and CTOs are reassessing weighted on-premise infrastructure to ensure predictable baseline costs, reversing cloud-only trends by reserving cloud compute primarily for bursting. Fragmented Tooling and Non-Human Cost Drivers (3:10) Organizations frequently depend on disconnected point solutions for individual tasks, burdening technology leaders with stitching disparate tools together and managing fragmented workflows and stale data. At the same time, customer expectations are shifting rapidly; Bulloch notes that one banking client moved from banning AI over data privacy concerns to demanding autonomous agents within a single twelve-month window. Simultaneously, enterprise cost drivers have expanded beyond individual human users to include automated workloads, autonomous agents, and specialized models. Flexera applies its IT Asset Management (ITAM) roots to help organizations audit and identify unused, costly GPUs and CPUs. Addressing labor trends, Bulloch characterizes AI as an executive "scapegoat" for corporate layoffs, contending that human employees become more valuable when augmented by AI rather than displaced by it. Operationalizing FinOps and Real-Time ROI Governance (14:45) To manage AI estates, Bulloch frames governance through Albert Camus's observation that revolution demands clarity, unity, and order—translating in an enterprise context to visibility, convergence, and control. Flexera is extending its Flexera One platform and Technopedia repository to deliver cost visibility within hours rather than at month-end, enabling teams to right-size workloads by selecting budget models over expensive frontier models when high complexity is unnecessary. Ultimately, Bulloch highlights that enterprise leadership discussions are shifting from "what AI costs" to "what it returns." Executive teams demonstrate a willingness to maintain high compute spending, provided cross-team alignment and normalized, business-enriched data can clearly demonstrate return on investment. While rapid advancements like Llama 4 reduce the need for custom model training, Bulloch emphasizes that organizations must replace fragmented point solutions with continuous, real-time cost governance to survive the transition from subsidized AI experimentation to sustainable enterprise deployment.

Flexera高级产品经理Scott Bulloch指出，企业AI应用的激增正推动支出呈现失控的“曲棍球棒”式增长，行业亟须借助IT资产管理与FinOps框架应对成本危机。他认为，技术领导者必须建立近实时的算力可视化与模型配置策略，将管理重心从单纯被动承受账单转向衡量实际的投资回报率。 账单冲击与高额补贴下的成本失控 (10:56) 企业由于未能实时监控AI支出，普遍在月末面临严重的“账单冲击”。Bulloch直言“AI不仅带来创新，更带来账单”，并以Uber为例指出，该公司在短短一个季度内就耗尽了全年的AI预算。当前的AI经济结构已从模型训练全面转向推理、提示词设计以及输入输出数据管理，传统模型训练已基本退缩至医学和数据科学等特定领域。同时，Bulloch将“代币经济学”称为一个令人头痛的概念，强调服务商的实际收费已不仅限于Token数量，还扩展至工作流执行步骤、系统操作及媒体渲染等多重维度。 更为严峻的是，当前终端用户的订阅费用背后存在云服务商的高额补贴。Bulloch披露，对于一份200美元的订阅服务，Anthropic需要支付约800美元的算力成本，而OpenAI的计算支出更是接近1400美元，这意味着企业未来的AI采购成本必将大幅上涨。为建立可预测的财务模型，部分CIO和CTO正开始评估加权本地部署模式以锁定固定开销，仅将云端用于应对突发算力需求，这一举措正在扭转近年来完全上云的趋势。 单点工具泛滥与治理维度的转变 (3:10) 目前企业大量引入解决特定任务的单点解决方案，导致技术负责人不得不艰难拼凑互不连通的异构系统，进而引发数据陈旧与工作流割裂。与此同时，客户需求在极短时间内发生了剧烈分化与演进。Bulloch列举了一家大型银行客户的案例：该机构在12个月前还因数据隐私风险坚决抵制任何AI集成，而如今却主动要求部署自主AI智能体，以期彻底替代传统员工培训流程。此外，诸如Llama 4的技术进步与提示词工程的成熟，使得许多数据提取项目不再需要耗费精力的定制模型训练。 Bulloch指出，AI革命的发展节奏已经全面超越了昔日的互联网、云计算及FinOps演进周期。他援引哲学家阿尔贝·加缪关于“革命的本质是对清晰、统一与秩序的渴望”这一观点，阐述AI治理的对应逻辑：在企业AI生态中，“清晰”代表着全面的资产可见性，“统一”代表着异构系统与工具的收敛融合，而“秩序”则对应着控制机制与治理规范，这与经典的FinOps管理框架高度一致。 依托ITAM根基识别算力与硬件闲置 (7:10) Flexera依托在IT资产管理（ITAM）、SaaS软件资产管理以及云成本控制领域的积累，主张通过规范化与业务上下文丰富的数据集来治理AI资产，强调唯有高质量的数据输入才能防止产生“垃圾进，垃圾出”的失控局面。在硬件治理层面，企业正频繁借助成熟的ITAM工具排查闲置且昂贵的高性能GPU和CPU，以准确评估现有计算硬件的真实利用率。 当前的成本驱动主体已不再局限于传统的人类员工账号，各类后台工作负载、自主AI智能体、垂直专业模型以及自动化脚本正在消耗庞大的企业级算力。针对伴随AI技术推进而出现的企业裁员现象，Bulloch明确表示，将裁员归咎于AI往往只是一种“替罪羊”说辞；他认为，当员工通过AI实现能力增强而非被直接替代时，人力资源所能释放的商业价值反而会显著提高。 践行FinOps闭环与ROI驱动管理 (14:45) 为消除月末账单滞后的被动局面，Flexera正将其核心平台Flexera One拓展至AI成本管理领域，并借助其技术情报库Technopedia，在数小时内为企业提供近实时的支出可见性。企业能够依照FinOps生命周期追踪应用系统与智能体的底层消耗，制定精细化的模型适配策略，在经济型基础模型与昂贵的前沿模型之间进行权衡，避免在低复杂度任务上滥用高成本前沿算力。 Bulloch强调，高管层的关注焦点正在经历根本性转移，从单纯询问“AI消耗了多少预算”转向审视“AI创造了多少商业回报”。只要能够展现明确的投资回报率（ROI），管理层依然愿意持续注资。因此，建立跨团队的紧密协同、保持技术领导力并实现精确的成本追踪，是企业化解AI工具盲目扩张风险、保障投资实效的关键路径。 Bulloch认为应对AI成本危机需要结合近实时可见性、严谨的FinOps控制以及将AI作为员工赋能工具的战略认知；然而，随着基础模型提供商对算力补贴的逐步退坡，企业未来能否在算力成本持续走高的压力下有效证明AI项目的真实ROI，仍是亟待解决的治理挑战。

## Background 该内容由频道 Ai4 于 2026 年 8 月 20 日在 YouTube 平台发布，由 Flexera AI 产品管理高级经理 Scott Bulloch 主讲。视频简介指出，当前企业在加速采用 AI 的过程中面临预算超支、Token 消耗失控及巨额账单等成本危机，演讲重点探讨了企业如何建立财务问责制以有效控制 AI 支出，并确保技术投资带来可衡量的商业价值。 ## TL;DR 演讲指出，当前企业 AI 应用的核心重心已从模型训练全面转向推理调用与自主 Agent，但伴随而来的“曲棍球棒式”算力成本暴增与单点工具泛滥正引发严重的“账单休克”。面对 GPU 闲置、模型高配低用以及供应商服务成本倒挂等现实困境，企业必须抛弃花哨的“代币经济学”概念，回归对真实资源消耗的穿透治理。Flexera 提倡将其在 IT 资产管理（ITAM）、SaaS 及云成本治理（FinOps）领域的深厚沉淀迁移至 AI 领域，通过“跟踪、落地、右规化优化”三步闭环实现小时级异常预警。最终，企业 AI 治理的核心诉求应从单纯审视“花费金额”升维至衡量确凿的“业务投资回报率（ROI）”与组织领导力。 ## Atmosphere & style 演讲风格务实、犀利且极具警示性，主讲人通过引入 Uber 预算暴雷与厂商成本严重倒挂等真实行业案例，直击光鲜 AI 创新背后的财务黑洞。叙事逻辑冷静而干练，在直斥“代币经济学”等虚浮术语的同时，展现出深耕资产管理领域老牌团队的沉稳与洞察。整场分享节奏紧凑、层次清晰，既传达出应对 AI 成本海啸的紧迫感，又为深陷账单焦虑的企业决策者提供了条理分明的秩序感与破局路径。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=4s) AI 时代的唯一恒量是剧烈变革，昨日有效的方法明日将失效，从业者需依托社区协作与多重视角共同应对挑战。 - [02:08](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=128s) AI 应用的普及带来了指数级的曲棍球棒式成本激增，而市面上零散单点工具的泛滥与拼凑进一步加剧了企业数据割裂与全局可见性缺失。 - [04:06](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=246s) 企业需求发生戏剧性反转，从严禁数据接触 AI 迅速转向全面拥抱自动化 Agent，落地重心也由昂贵的模型微调转向优质的提示工程（Prompting）。 - [06:03](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=363s) 应对 AI 混乱局面的三支柱模型为：通过可见性实现清晰、通过聚合实现统一、通过控制实现秩序，与传统 FinOps 闭环理念一致。 - [07:10](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=430s) Flexera 将深耕数十年的 IT 资产管理（ITAM）、SaaS 治理与云成本优化经验及上下文标准化数据，全面迁移并应用于 AI 资产治理。 - [08:45](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=525s) AI 资产消耗主体已从单一员工扩展至并发 Agent、模型与工作流，企业需高度警惕 GPU 闲置这一昂贵的财务漏洞。 - [08:45](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=525s) AI 的本质是赋能与增强人类员工而非单纯替代；企业面临的真正商业危机在于失控的边际运行成本而非人力优化。 - [10:56](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=656s) 频繁发生的账单休克促使企业决策回摆，许多高管考虑在本地部署加权模型确保基准成本可预测，仅在算力不足时弹性爆发到云端。 - [12:21](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=741s) AI 经济学重心已从模型训练全面转向模型推理，企业治理应抛弃过度包装的代币经济学概念，直击算力与工作流的真实资源消耗。 - [13:18](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=798s) 当前模型订阅存在严重的交付成本倒挂现象（如 200 美元订阅对应高达 800 至 1400 美元的实际服务成本），预示未来企业 AI 账单面临剧烈上涨压力。 - [14:19](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=859s) AI 成本管理需依托基准测试与模型分级，通过小时级异常监控和模型使用右规化（Right-Sizing），避免拿旗舰大模型低效处理简单任务。 - [16:20](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=980s) 单点工具无法解决复合型 AI 支出难题，必须在统一平台（如 Flexera One）上联动打通 GPU 硬件、软件许可、云配额及 Token 调用等多维指标。 - [17:33](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=1053s) AI 治理的核心不仅是关注花了多少钱，而是向管理层证明投资回报率（ROI），依托敏捷领导力实现小时级响应与跨部门成本共识。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=4s) AI 时代的唯一恒量：变革、重塑与社区视角 Scott Bulloch 以 Flexera AI 高级产品经理的身份开场，直奔当今科技领域最核心的现实——变革是唯一的恒量。在当前的 AI 浪潮下，整个行业正经历前所未有的剧烈震荡，从业人员被迫从传统岗位走向全新角色，不仅需要重新学习技能，更要跨越不同层次的认知门槛。这种变革正在从底层彻底重塑“创新”的定义：昨天行之有效的方法，明天就会彻底失效；过去习以为常的工作流程，如今全部必须推倒重来。 面对这种颠覆，Scott 强调技术创新绝非零和博弈，不是单纯的“你对我”或“某家公司对另一家公司”的竞争，而是需要合作伙伴、客户以及整个技术社区共同协作、互相扶持。在加入 Flexera 之前，Scott 曾在 Snow Software 工作（后被 Flexera 收购），他特别提到了 Snow Software 当年的座右铭——“视角即力量”（"Perspective is power"）。这一理念始终贯穿于他对 AI 领域的思考：身处日常挑战之中的从业者，唯有借助技术社区的交流与多元视角，才能在瞬息万变的生态中找准方向。 - 行业正处于全面转型期，员工被迫学习新技能以适应全新岗位。 - 创新逻辑被重写，“昨天有效的方法在明天将不再适用”。 - 强调社区协作与跨企业交流的重要性，秉持“视角即力量”的认知框架。 ### [02:08](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=128s) 创新的曲棍球棒成本曲线与单点工具泛滥困境 随着大语言模型、智能体（Agents）以及工作负载的爆发式普及，创新的迭代速率达到了历史最高峰。然而，伴随业务高速增长而来的，是令人心惊肉跳的“曲棍球棒式”成本激增曲线（Hockey Stick Growth Costs）——企业在进入平稳期后，算力与使用成本突然呈指数级飙升。无论是 Flexera 自身的内部 AI 实践，还是其庞大的企业客户群体，每天都在真切经历这种失控的成本压力。 讽刺的是，推动这一轮技术变革的各种工具，本身也变成了企业内部的新麻烦。当前市场上充斥着大量“单点解决方案”（Point Solutions），只要在技术展厅走上一圈，就能看到无数只专注于解决某一个细分问题的独立工具。企业与技术团队被迫像“缝合怪”一样将几十个零散工具拼接在一起。这种拼凑不仅带来了数据陈旧、集成脆弱等一系列问题，而且严重缺乏全局可见性，但令人无奈的是，这却成了当下企业的普遍现状（Status Quo）。 ```mermaid flowchart TD accTitle: 单点工具泛滥导致的集成与成本困境 subgraph S1["工具层：单点工具泛滥"] A["细分工具 A"] B["细分工具 B"] C["细分工具 C"] end subgraph S2["执行层：企业技术人员手工拼接"] D["数据割裂与陈旧"] E["集成脆弱与维护负担"] end subgraph S3["结果层：业务失控"] F["曲棍球棒式成本激增"] G["缺乏统一可见性"] end A --> D B --> D C --> E D --> F E --> G ``` ### [04:06](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=246s) 需求巨变：从“绝不碰 AI”到全自动 Agent 与免训练提示工程 企业对 AI 的态度在过去短短 12 个月内发生了戏剧性的 180 度大转弯。Scott 分享了一个大型银行客户的真实案例：大约一年前，Flexera 团队向该银行推介利用 AI 自动提取发票和文档数据的软件方案时，对方态度极为坚决地回绝：“绝不使用 AI，连想都别想，我们的数据绝对不能碰 AI。”然而仅仅一年之后，这家客户的态度彻底反转，急迫地要求：“赶紧给我们一个 Agent，让它替我搞定一切，我根本不想花时间培训员工，直接让 AI 自动完成！” 这种剧变不仅体现在客户心理上，也体现在底层技术路径的颠覆上。以数据提取项目为例，过去团队曾认为必须自己收集数据、从头微调和训练专有模型；但随着 Llama 等新一代强大模型的问世，团队惊奇地发现根本无需进行任何额外模型训练，仅凭优质的提示工程（Prompting）就能完美满足业务需求。这场 AI 革命的推进速度和颠覆力度，已经远远超越了当年的互联网革命、云计算兴起（AWS、Azure）以及早期 FinOps 的演进。 - 某大型银行客户在 12 个月内从严禁 AI 接触数据，转变为全面拥抱自动化 AI Agent。 - 业务落地重心从昂贵的模型从头训练，快速转向依赖先进基础模型（如 Llama 系列）的提示工程（Prompting）。 - 变革节奏超越了以往互联网与云计算浪潮，成为时代性的技术跃迁。 ### [06:03](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=363s) 变革三要素：可见性带来清晰，融合带来统一，治理带来秩序 借用阿尔贝·加缪（Albert Camus）关于革命与反抗的经典论述，Scott 提出 AI 变革的核心诉求同样是三件事：“对清晰的渴望、对统一的追求、对秩序的向往”。企业在应对 AI 混乱局面时，必须建立对应的三层能力支柱： ```mermaid flowchart LR accTitle: AI 变革的三大支柱模型 subgraph C1["清晰 (Clarity)"] A["可见性 (Visibility)"] end subgraph C2["统一 (Unity)"] B["融合聚合 (Convergence)"] end subgraph C3["秩序 (Order)"] C["控制与治理 (Governance)"] end A --> B --> C ``` 1. **通过可见性实现清晰（Clarity through Visibility）**：全面掌握企业内所有 AI 资产、用量与成本流向。 2. **通过聚合实现统一（Unity through Convergence）**：打破单点工具的割裂，将散落的数据与系统融为一体。 3. **通过控制实现秩序（Order through Control）**：建立严密的治理规则与预算管控机制。 这一演进逻辑与 FinOps 基金会（FinOps Foundation）所提倡的“可见、优化、运营”飞轮如出一辙。企业目前面对的是完全相同的管理本质，只是应用场景切换到了 AI 领域，需要使用升级的工具链重新走一遍治理之路。 ### [07:10](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=430s) Flexera 核心根基：从 ITAM、SaaS 到云成本治理的经验迁移 在应对 AI 成本危机上，Flexera 拥有深厚的历史底蕴。Flexera 在 IT 资产管理（ITAM）领域深耕数十年，长期帮助企业处理复杂的软件许可合规、发票数据抽取与海量资产盘点。Scott 强调，“垃圾进，垃圾出”（"Garbage in, garbage out"）是数据领域的铁律，企业若想构建在未来真正有商业价值的 AI 系统，核心在于拥有经过上下文丰富化、标准化的真实最新数据（Contextual Normalized Data）。 Flexera 目前已经构建了完整的三维管理矩阵：覆盖算力层面的云成本管理（Cloud Cost Management / FinOps）、覆盖按席位订阅的 SaaS 许可管理，以及传统的 IT 资产管理（ITAM）。这三大支柱共同构成了对企业整体 IT 资产状态（Estate）的全局洞察，涵盖成本、消耗量、许可授权以及实际使用情况。如今，Flexera 正是将这一套历经验证的成熟资产治理框架与标准化数据能力，全方位迁移并应用到爆发式增长的 AI 资产管理之中。 - **IT 资产管理（ITAM）**：沉淀数十年的资产拓扑与许可管理经验。 - **云成本管理（FinOps）**：深度的基础架构与云资源开销优化能力。 - **SaaS 管理**：针对标准席位制软件与订阅服务的全生命周期管控。 - **上下文标准化数据**：结合业务上下文对底层消耗进行切片分析与成本回溯（Showback / Chargeback）。 ### [08:45](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=525s) 算力黑洞与复合消耗：GPU 闲置、Agent 普及与裁员真相 随着 AI 应用的纵深推进，企业的资产消耗维度发生了结构性剧变：消耗主体不再仅仅是“人类用户”，而是大量并发运行的 AI 智能体（Agents）、基础模型、自动化项目以及后台工作流。与此相伴的是对底层 GPU 和 CPU 算力的巨额消耗。由于 GPU 采购与租赁费用极其昂贵，大量客户频繁调出 Flexera 的 ITAM 工具，首要目的就是排查企业内部到底有哪些高价值 GPU 处于闲置状态，以及它们的利用率是否达标。 Scott 还针对近期频频见诸报端的科技巨头（如某知名社交媒体公司）因 AI 裁员的新闻发表了独到见解。他指出，将裁员单纯归咎于“AI 取代人类工作”往往只是管理层实现既定精简目标的借口（Scapegoat）。在现实中，AI 本质上是对人类员工的赋能与增强（Augmentation），一个善用 AI 的员工创造的商业价值远高于单打独斗的个体。企业面临的真正危机不是人力替代，而是**失控的 AI 运行与扩张成本**，如果无法在规模化运行时有效遏制成本溢出，反噬企业的财务健康，业务创新便难以为继。 - AI 资产消耗单元多元化：从单纯的员工席位扩展至 Agents、Models、Projects 与 Workflows。 - GPU 闲置成为企业重度关注的财务漏洞，资产可见性成为优化算力的第一步。 - 强调当前 AI 的核心角色是“人类增强工具”，失控的边际运行成本才是真正的商业阻碍。 ### [10:56](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=656s) 账单休克危机：“AI 不仅在创新，更在寄账单” 在当下的 AI 狂欢中，“成本危机”已经成为各家企业 CIO 与 CFO 的噩梦。Scott 列举了行业内广为人知的案例：例如 Uber 在 AI 上的开销严重超标，仅用了一个季度就彻底烧光了原本规划一整年的 AI 预算。许多企业在月底收到云厂商与模型供应商的天价账单时，陷入极度的“账单休克”（Invoice Shock），由于前期缺乏实时追踪机制，财务团队只能在事后痛苦地逆向倒推排查超支原因。 > “AI 并不只是在搞创新，它还在不断给你寄账单（AI just doesn't innovate, it invoices）。” 为了规避这种不可预测的财务黑洞，企业高管层的决策逻辑开始发生回摆：许多 CIO 和 CTO 正在研究如何在本地（On-Prem）部署加权模型，试图重拾传统 IT 时代固定成本的可预测性，同时借鉴成熟的 FinOps 模式，在本地算力不足时才弹性爆发（Burst）到云端。这种从最初依赖 VMware 本地虚拟化到云端溢出的轮回，正在 AI 算力领域原样重演。 - **Uber 预算失控案例**：单季度耗尽全年的 AI 预算规划。 - **账单休克（Invoice Shock）**：由于缺乏实时监控，企业在月末面临不可承受的巨额超支账单。 - **混合架构回摆**：企业高管尝试通过本地部署核心模型确保基准成本可预测，辅以云端弹性突发以控制总开销。 ### [12:21](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=741s) 新 AI 经济学：从模型训练走向推理，抛弃伪概念关注真实消耗 当前的 AI 经济形态已经发生了根本性的范式转移：整个产业重心已经全面从“模型训练（Training）”转向“模型推理（Inferencing）”。除数据科学实验室或医疗等高度专业的极少数垂直领域仍需自建训练外，绝大多数企业已不再讨论训练专属大模型。当下的核心竞争力完全聚焦于优质的上下文数据、输入输出控制以及高效的提示词设计。 对于业界热炒的“AI Tokenomics（代币经济学）”，Scott 表达了明确的保留与反感态度。他直言行业过分沉迷于“Token”这一概念，但本质上这只是一种计费度量单位。对于企业财务治理而言，核心本质永远是**真实的资源消耗（Consumption）**。企业如果不能穿透花哨的概念直击实际的算力与用量消耗，就无法真正驾驭规模化 AI 带来的财务挑战。 ### [13:18](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=798s) “代币经济学”的幻象与暴涨的服务端实际成本 当前的计费模式早已不再仅局限于单纯的 Token 计算。各大服务商开始按执行次数（executions）、工作流步骤（workflow steps）以及视频渲染等各种维度进行综合计价。Scott Bulloch 直言不讳地表示，对他个人而言，所谓的“代币经济学”（tokenomics）现在听起来几乎成了一个“骂人词”（swear word）——因为这些本质上全都是基于资源消耗的综合计费，即便其中包含 AI 计算成分，也不应被简单包装。更令人警惕的是当前的定价倒挂现象。用户购买一份 200 美元的订阅服务，而背后的实际交付成本却远超想象：据他引用的最新指标，Anthropic 为支撑这套 200 美元的订阅，背后实际需要耗费大约 800 美元的服务成本；OpenAI 的情况更加夸张，其真实服务成本据估算高达 1,400 美元左右。 这种严重的补贴和倒挂预示着未来企业 AI 成本不可避免的剧烈上涨。随着供应商逐步收紧补贴，未来的 AI 账单势必“直冲云霄”（go through the roof）。正因如此，在本地部署私有 GPU、在本地运行带权重的开源模型以规避公有云高昂持续成本的方案，正在演变成不可忽视的现实路径。 - **订阅与服务成本倒挂**：用户支付 200 美元月费，Anthropic 实际交付成本约 800 美元，OpenAI 实际交付成本高达约 1,400 美元。 - **多元化计量维度**：计费点已从单纯的 Token 扩展到工作流执行步数、模型执行次数以及多媒体/视频渲染等综合消耗指标。 - **本地部署回归**：高企的长期推理开销推动企业认真考虑本地私有化 GPU 及模型权重部署方案。 --- ### [14:19](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=859s) 模型分级基准与 Flexera 向 AI 成本管理的自然延伸 面对市面上琳琅满目的模型生态，企业绝不能盲目依赖单一选项。尖端的“前沿模型”（frontier model）与经济实惠的“预算模型”（budget model）之间存在巨大的成本与性能鸿沟，企业必须清楚在何种业务场景下该选用何种模型。因此，针对实际业务数据开展模型基准测试（benchmarking），明确知晓潜在的替代方案与最低门槛需求，将成为控制支出的核心能力。 为了在整套 IT 资产中实现全局可视化，Flexera 正将其核心能力全面拓展至 AI 成本管理领域。Scott Bulloch 强调，这种延伸是水到渠成的：Flexera 长期深耕云成本管理（Cloud Cost Management）、SaaS 治理和 IT 资产管理（ITAM），对软件许可、计量指标（licensing & metrics）了如指掌。将这些积累迁移到 AI 领域是必然选择，因为在企业构建 AI 应用的过程中，透明度至关重要，而且这种洞察绝不能等到“月底看账单”才迟钝地发现。如果在几天甚至数周后才获知异常，企业早已被失控的超支账单灼伤；管理者必须在“数小时之内”感知突发峰值与异常消耗。 ```mermaid flowchart TD accTitle: AI Cost Management Lifecycle subgraph Track["跟踪 (Track)"] T1["全面监控业务应用"] T2["追踪自主 Agent 消耗"] end subgraph Implement["执行 (Implement)"] I1["制定内部治理策略"] I2["模型选型 (避免高配低用)"] end subgraph Optimize["优化 (Optimize)"] O1["模型规格右规化 (Right-Sizing)"] O2["降配至高性价比替代方案"] end Track --> Implement --> Optimize --> Track ``` --- ### [15:43](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=943s) FinOps 闭环实操：跟踪、落地与模型右规化 将传统 FinOps 轮环（FinOps wheel）迁移至 AI 治理，主要依托三个关键阶段：“跟踪（Track）、落地（Implement）、优化（Optimize）”。在跟踪层面，企业需要全面盘点并监控内部所有的应用程序与自主 Agent 运行状态；在落地实施层面，需要确立明确的采购与接入规范，确保团队选用正确的模型类型。Scott Bulloch 指出，团队在现实中屡见不鲜的问题是“拿旗舰级模型去跑所有简单任务”，这种过度配置造成了惊人的浪费。 优化阶段的核心理念是“模型使用右规化”（right sizing for model usage）。就像云原生资源根据负载动态升配或降配一样，如果一项任务正在调用昂贵的前沿模型，但其实际输出并不需要如此高的算力，企业就应果断调整，将其迁移至轻量级模型或专用预算模型，把每分钱花在刀刃上。 - **小时级异常预警**：告别以月为周期的滞后账单，将消耗监控粒度细化至小时级，防止失控的调用循环在一夜之间烧毁预算。 - **模型使用右规化（Right-Sizing）**：严禁过度依赖昂贵前沿模型执行低复杂度任务，根据任务难度动态匹配高性价比替代模型。 --- ### [16:20](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=980s) Flexera One 平台集成与跨领域多维可见性 在技术架构上，AI 成本管理并非凭空搭建，而是深度嵌入在统一的 Flexera One 平台之上。其底座汇集了 Flexera 拥有行业领先地位的知识层——包括囊括海量软件元数据与资产目录的 Technopedia 数据库。Flexera 在软件许可、公有云资源治理方面已经为客户挽回了数以百万计的资金损失，将这套沉淀复制到 AI 算力与模型治理上，能够释放出极为可观的商业回报。 Scott Bulloch 强调，AI 成本管理绝不能依靠某个单一维度的点状工具（point tool）来解决。它是一个复合型难题：既涵盖底层 GPU 物理算力、中间层模型推理调用，又穿插着上层 SaaS 授权和传统云基础设施配额。市场上没有哪一款单点产品能够跨越所有孤岛提供正确的全局视角。唯有打通硬件利用率、软件许可证、云端消费以及 Token 调用的多维度指标联动，才能看清企业 AI 支出的全貌。 - **底层沉淀复用**：基于 Flexera One 与 Technopedia 丰富的数据资产，直接扩展支持 AI 基础设施与模型成本核算。 - **打破单点工具局限**：单一工具无法解决全局问题，必须整合 GPU 硬件监控、云支出与 SaaS 许可等多维度指标，提供端到端穿透视图。 --- ### [17:33](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=1053s) 商业价值归宿：从关注“花费”到衡量“投资回报” 讨论最终回归到企业高层的核心诉求上。Scott Bulloch 犀利地指出，真正的问题从来不是“AI 花了多少钱”，而是“AI 带来了什么回报”以及“它最终值多少钱”。当向首席执行官（CEO）、首席信息官（CIO）或业务副总裁汇报时，话题的核心永远是投资回报率（ROI）。管理层并不排斥为高昂的 AI 成本买单，前提是他们必须清晰获知这笔巨资换回了什么确凿的商业成效。Scott 直言，他日常与自己公司的 CEO 交流时，这也始终是需要优先解答的核心命题。 在业务节奏全面提速的当下，企业响应与决策的周期已从“数周、数天”被极限压缩到了“数小时”以内。在这种飞速迭代中，最关键的制胜要素是“领导力”（leadership）。伴随着 AI 技术在组织各业务线的大规模普及（proliferation），跨部门的信息共享、精准的成本归属追踪以及全团队目标的统一步调，显得比以往任何时候都更加迫切和致命。Scott Bulloch 以此作结，呼吁企业必须依托统一的治理框架与敏捷的管理共识，共同应对正在到来的 AI 成本风暴。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=4s) 视角即力量 — Scott Bulloch > [00:04](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=4s) 昨天有效的方法，明天就会彻底失效；过去习以为常的工作流程，如今全部必须推倒重来。 — Scott Bulloch > [04:06](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=246s) 绝不使用 AI，连想都别想，我们的数据绝对不能碰 AI。 — 大型银行客户 > [04:06](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=246s) 赶紧给我们一个 Agent，让它替我搞定一切，我根本不想花时间培训员工，直接让 AI 自动完成！ — 大型银行客户 > [06:03](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=363s) 对清晰的渴望、对统一的追求、对秩序的向往 — 阿尔贝·加缪（Albert Camus） > [07:10](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=430s) 垃圾进，垃圾出（Garbage in, garbage out） — Scott Bulloch > [10:56](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=656s) AI 并不只是在搞创新，它还在不断给你寄账单（AI just doesn't innovate, it invoices）。 — Scott Bulloch > [13:18](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=798s) 所谓的“代币经济学”（tokenomics）现在听起来几乎成了一个“骂人词”（swear word） — Scott Bulloch > [15:43](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=943s) 拿旗舰级模型去跑所有简单任务 — Scott Bulloch > [17:33](https://www.youtube.com/watch?v=ab0ndfJWnaM&t=1053s) 真正的问题从来不是“AI 花了多少钱”，而是“AI 带来了什么回报”以及“它最终值多少钱”。 — Scott Bulloch ## Mentioned Scott Bulloch, Flexera, Snow Software, Llama, 阿尔贝·加缪（Albert Camus）, FinOps 基金会（FinOps Foundation）, AWS, Azure, VMware, Uber, Anthropic, OpenAI, Flexera One, Technopedia ## Video description Scott Bulloch - Senior Manager, Product Management - AI, Flexera Recent technology news has been dominated by the growing concern of organizations losing control of AI spend, with companies reporting budget blowouts in months, runaway token usage, and nine-figure bills tied to unmanaged access. As organizations accelerate AI adoption, they are discovering that the biggest barrier to scale isn't innovation, it's cost. Join Flexera to explore the drivers behind today's AI cost crisis and learn how leading organizations are enabling financial accountability, controlling costs, and affordably running AI at scale while delivering measurable business value from their investments. [Source](https://www.youtube.com/watch?v=ab0ndfJWnaM)

[0:04] Scott Bulloch: 视角即力量 [0:04] Scott Bulloch: 昨天有效的方法，明天就会彻底失效；过去习以为常的工作流程，如今全部必须推倒重来。 [4:06] 大型银行客户: 绝不使用 AI，连想都别想，我们的数据绝对不能碰 AI。 [4:06] 大型银行客户: 赶紧给我们一个 Agent，让它替我搞定一切，我根本不想花时间培训员工，直接让 AI 自动完成！ [6:03] 阿尔贝·加缪（Albert Camus）: 对清晰的渴望、对统一的追求、对秩序的向往 [7:10] Scott Bulloch: 垃圾进，垃圾出（Garbage in, garbage out） [10:56] Scott Bulloch: AI 并不只是在搞创新，它还在不断给你寄账单（AI just doesn't innovate, it invoices）。 [13:18] Scott Bulloch: 所谓的“代币经济学”（tokenomics）现在听起来几乎成了一个“骂人词”（swear word） [15:43] Scott Bulloch: 拿旗舰级模型去跑所有简单任务 [17:33] Scott Bulloch: 真正的问题从来不是“AI 花了多少钱”，而是“AI 带来了什么回报”以及“它最终值多少钱”。

[0:04] Scott Bulloch: perspective is power [3:10] Albert Camus: the act of revolution is demand for clarity, unity, and aspiration for order. [10:56] Scott Bulloch: AI just doesn't innovate, it invoices.

[0:04] Hey everyone, like I said, my name is Scott  Bulock, senior product manager at Flexera. And [0:09] uh as part of this conversation today, I'm  going to talk about cost visibility and Flex [0:13] Air as a whole as in the AI world. I think  everybody here would probably agree change [0:20] is constant. Everything that we're seeing,  everything that we're doing, you're here, [0:25] you're at an AI conference, you're living through  this today. And the only constant in this industry [0:30] right now is change. People are being forced from  standard jobs to new jobs. They have to learn new [0:37] things and skill up in different levels as  well. It's also redefining innovation. The [0:44] things that worked yesterday aren't going  to work tomorrow. The things that you're [0:49] used to doing all have to be reworked from the  ground up. And the one thing about innovation, [0:59] we all have to help each other get to the right  place in innovation. It's not just you versus me [1:05] or somebody else versus somebody else, right? For  our partners, customers, for ourselves, and most [1:13] most importantly each other. The community  around the technology space is key for us. [1:22] With that, we believe that every person  in this room is facing daily challenges [1:28] when it comes to the AI space. And  because of this change, technology, [1:35] community events like this are key. I don't think  anybody would disagree with that. They're huge for [1:41] learning conversations. Even as you're sitting  having breakfast or lunch, you can learn a lot [1:46] from the person sitting beside you. You know, I  come from Snow Software before I came to Flexer, [1:52] before we were acquired. And the the slogan  for Snow Software was perspective is power. [2:00] That's something I take with me all of the  time when I think about AI in this space. [2:08] And I don't think any of us would disagree. The  pace of innovation is accelerating. It's faster [2:16] than it's ever been before. And like I said,  the things that we used to do yesterday are [2:21] not going to be the same things that are going  to work tomorrow. As the technology changes, [2:28] adoption changes, models changes, adoption and  explosion, workloads, agents, all of that type of [2:34] stuff. Companies are actually experiencing faster  growth costs. The hockey stick in this scenario, [2:41] if you guys know what I'm talking about, it's  real, right? you get to a steady state and [2:47] at the very end your costs are going through the  roof. So we see this day in and day out across our [2:54] customer base internally as we use AI ourselves.  We're experiencing this and unfortunately for many [3:03] the same technology that has kind of facilitated  this change has also become a bit of a problem [3:10] within the organizations today. This has resulted  in a proliferation of point solutions. And if [3:19] you walk around the floor, if you're down  in the booth looking at all the vendors, [3:24] you can see this. It's it's basically it's  point blank. It's really really hard to not [3:29] see these specific point solutions that do  one in particular job really, really well. [3:35] And as you look at that, what that means is  us as technologists, customers, consumers, [3:43] we have to stitch all of those pieces together.  And that in itself can be difficult. It can be [3:49] stale data. There's a whole bunch of reasons  why that could be bad in a bunch of different [3:55] scenarios. Unfortunately, this is status  quo. This is the way this works for us today. [4:06] with that change is demand you know as we go  through this challenge of new demand for change [4:15] and I'll I'll give an example of this one we had  a customer big banking customer about a year ago [4:22] you know we were like AI agents we're going to put  AI in our thing we had software we were extracting [4:27] invoices all that type of stuff making it easier  it was like the easy button all AI they're like [4:33] no No AI, none. Don't even think about it. My data  is not touching it. We're having none of it today. [4:41] That same customer, give me an agent, make the  agent do the thing for me. I don't want to train [4:46] my people. I want AI to do it for me. Like the  way things have shifted in the past 12 months is [4:54] absolutely crazy. And I don't think anybody here  would disagree with that. when that in particular [5:01] project when we were dealing with extraction of  data, we were looking at training models and doing [5:07] all of that type of stuff like we had to look  at that training. That was the only way we were [5:12] going to get the usage out of those models. And  then Llama 4 came out. Maverick done. We didn't [5:20] have to train anything. Prompting took care of  everything that we needed to do in that space. [5:26] just to give you an idea idea of how fast this  is changing as every model comes out. So part of [5:34] this change and this demand we would say is an AI  revolution. This is changing faster than anything [5:41] we've ever seen. You think about the internet, you  think about the cloud, you think about, you know, [5:46] AWS, Azure and all of that, the PHOPS space.  This is way bigger and way more transformative [5:53] than those have ever been at a rate we'll  probably never see again in our lifetime. [6:03] I don't think anybody would disagree with this.  The act of revolution is demand for clarity, [6:09] unity and aspiration for order. And  this is really really interesting as [6:16] you think about this. You know, Albert  Camas when he wrote this, you know, [6:22] was thinking about the revolution and the  rebellion and all of that type of stuff, [6:26] but it's perfectly fitting. Clarity, unity, and  order. And let's think about those for one second. [6:36] Clarity through visibility, kind of  common. Unity through convergence, [6:44] bringing everything together. And order through  control. Think of your governance. Think of all [6:48] of that type of stuff. This is exactly  what you're going through, businesses are [6:52] going through. This is the step. Think about  the um the PinOps foundation kind of wheel, [6:59] you know, the optimize visibility, that type  of thing. Exact same thing. Same processes, [7:04] different tooling, and we're  doing this all over again. [7:10] with that flex era as a whole for our ourselves  our customers we have a platform we've been in [7:18] the ITAM space for years and in the ITM space  we help customers with licensing and all that [7:24] type of stuff like I mentioned invoices and  extracting tons of data but the key to that [7:30] is contextual normalized data everybody knows  garbage in garbage out we've heard it I don't [7:37] know how many thousands of times it's a real  central data, real data, up-to-date data is [7:43] the key to building anything in the AI space  that's going to be worthy for you tomorrow. So, [7:50] enriching that with business data and everything  that allows you to break up that data, report it [7:56] back to the business, that's key. So, for us,  it's beyond kind of an idea. We've been doing [8:01] this for years. We're just going to now apply  this to AI. So with that we already have cloud [8:11] cost management that's a flex era thing today  that's PHOps we have SAS management standard [8:18] seatbased licenses those types of things and ITM  like I mentioned we've been in this ITM space for [8:25] decades now and every one of these has a view  on an estate whether that be cost consumption [8:36] licenses, you know, order usage, all the  common things that we're talking about here. [8:45] It's accelerating. Like I said, AI is  compounding this problem. And it's not [8:51] just about users anymore. It's about agents.  It's about models. It's about projects, [8:57] workflows. All of these types of things  are now consuming AI within your business. [9:04] GPUs and CPUs. I can't tell you how many  times people pull up our ITAM tool because [9:10] they want to know where their unused GPUs  are today because of the cost of them, [9:14] the sheer cost. They want to know if they're  actually being utilized properly or not. [9:22] So, this is pretty common. I think you guys  hear about this all the time. I definitely do. [9:29] Um, one of our customers happens to be um, a  social media company that deals with this and [9:36] has been in the news lately. And this is a this  is a real thing, but I think it's actually the [9:42] wrong reason. I'm not sure that these happen  because AI is taking over people's jobs and [9:49] things of that nature. I think it's actually  more about just a scapegoat to get to a place [9:54] that they were going to try and get to anyway.  AI is an augmentation of users today. All of [10:02] you as users are more valuable using AI than you  are standalone as a user. That's just a reality. [10:10] Everybody has a number. They have a value. And  then the real problem, it's the cost. And that's [10:16] where all of this stems from. It's controlling  those costs, running those AI things at scale, [10:24] right? At the end of the day, the runoff costs on  this is the real problem. Counter effects on those [10:31] come back to the business in different ways.  So, we've been basically forced, like I said, [10:39] a year ago, people were like, "Nope, no AI."  Now, they're like, "AI go faster. Use AI. Do [10:47] it as fast as you possibly can. Make sure  it's valid. But it all comes at a cost. [10:56] You know, pretty common in the news to see  this. I see this on LinkedIn all the time. [11:01] It's a cost crisis. Uber, everybody knows that  they got some crazy crazy overages and they blew [11:07] through their budgets in I think it was a  quarter what they were expecting to spend [11:11] in a year. Pretty wild. But this is real.  And people are getting shocked. You know, [11:18] invoice shock at the end of every month.  They're not tracking any of this stuff. [11:22] And this is why we say AI just doesn't innovate.  It invoices. You are actually getting build. [11:29] You're not tracking these things. Companies  don't necessarily know what's going on in the [11:33] AI space until they get the bill at the end of  the month and then they have to work backwards. [11:43] Back to every AI decision has a price tag.  everything that you build. You know, CIOS, CTO's [11:50] are now trying to figure out how they do weighted  models on prem. We're going backwards. They [11:56] want to do this, but they want the spend to be  predictable, kind of like things were back in the [12:02] day. You kind of knew your fixed costs, all that  type of stuff. And they want to burst to the cloud [12:07] like FinOps does. You know, VMware had it all on  prem and then they started bursting to the cloud [12:12] because people couldn't get capacity anymore.  Same thing, same problem, different thing. This [12:21] is a new AI economy. Let's let's be honest about  that. With that, AI has moved from training to [12:29] inferencing. Like I mentioned, we were looking at  training models and now you don't even have to do [12:35] any of that stuff anymore. They've just gotten so  good that you don't even need to do that anymore. [12:41] It's all about prompts. It's all about skill,  inputs, outputs, having the right data set behind [12:46] those. I haven't heard anybody talk about training  models except for in data science, medical, [12:53] those types of scenarios because they're very  very specialty business. It just doesn't happen [12:57] anymore. AI tokconomics, I I kind of hate this  word to be honest. Um AI itself to me isn't just [13:06] about tokens, it's about consumption. Everybody  says tokens this, tokens that. At the end of [13:13] the day, it's about consumption. And as we see,  if you walk around the floor, it's actually not [13:18] just about a token anymore. They're charging you  for executions, workflow steps, any of those types [13:27] of things, video render, all of those things.  That is just a form of consumption, even though [13:32] there's an AI portion to those. So tokconomics to  me is a little bit of a swear word, personally. [13:40] This is where we're expected to be. This  is crazy, right? If you think about what's [13:45] happening in Claude in OpenAI, you buy a $200  subscription. I think um the last metric I saw, [13:53] Enthropic, the $200 subscription was about $800  for them to actually service it to you. Open AAI [14:00] was actually even higher. It was about $1,400  for them to service that to you. So, just think [14:04] about what's going to happen to costs in this  space over time. they're going to go through the [14:09] roof. That GPU on prem, weighted models on prem  to kind of get around this, that's a real thing. [14:19] And the difference between these models,  the budget model versus the frontier model, [14:24] when to use these models. This is all key in  this vast amount of models for different things. [14:32] benchmarking on this data is going to be very  very real and you're going to want to know what [14:36] your alternatives are in this space and then  just what's required here. So when you think [14:45] about complete visibility in your estate, Flexer  is actually talking about AI cost management based [14:54] on what we do today as a business and what we have  in our customer base. We've extended into this [15:01] space which makes a ton of sense. Like I said,  cloud cost management, SAS, and ITM. This is flex [15:07] era's [clears throat] usual. Everything that we  know in this space, licensing, metrics, all of [15:12] that type of stuff, we're going to extend it to  AI cost management because as you build on this, [15:22] you need to see exactly what's happening. And it  can't be at the end of the month. You need to know [15:28] within hours that something's happening. You're  going to have a spike or you're going to be out of [15:34] control by the end of the month instead of getting  burnt on a bill. So going back to this track, [15:43] implement optimize back to the PHOPS wheel. What  does that look like? tracking all of your apps, [15:52] tracking your agents, implementing some type  of a strategy within your business, right? [15:58] Models. Make sure you're selecting the right one.  Don't use Fable for everything as an example. We [16:05] see this all the time. And optimize. If you're  using something and it's the incorrect thing, [16:13] obviously you're going to want to bump this up.  It's almost like right sizing for model usage. [16:20] We already have the platform. So we're  implementing that on top of the platform [16:24] of flexer1 with that the AI cost management stack.  This is basically high level what our stack looks [16:32] like today, right? Everything from knowledge  layer, our technopedia, which is our library of [16:39] all of our content data. We're basically leaders  in all of these already. And now we're just going [16:47] to be doing the extension of that. But when you  think about what's happening across the customer [16:51] base from a company like Flexera, giving you ROI,  saving customers millions of dollars and taking [16:58] what we know in this space and doing it in the  AI space is going to be very very powerful. This [17:06] segment across is key. We look at this as multiple  things. I don't think it's one product. We talk [17:13] about AI cost management, but given where we have  our visibility across our tools like GPUs, things [17:19] of that nature, you need visibility from multiple  metrics to get this. I don't think any one tool in [17:26] anyone's space is going to give you the view  that you're looking for at the right level. [17:33] This is an interesting question. It's not what AI  costs, it's what it returns. And it's ultimately [17:45] what it's worth. If you're talking to CEOs, VPs,  CIOS, it's going to come back to ROI. They don't [17:54] mind paying for it as long as they're going  to get something back for it. They just need [17:59] to know what that return is on investment.  That is point blank. Like I talk to our CEO [18:05] about this all of the time. It's it's a problem  that needs to be solved. And then what matters [18:17] helping customers, helping ourselves as we  do things in hours versus days versus weeks, [18:26] how you align internally, this is much faster than  you've ever done it before, but those things are [18:32] more critical than they've ever been. So  the key right now is actually leadership. [18:40] As you're going through your AI proliferation  and you're doing this across your business, [18:46] this is key to everything that you're going to do  in this space. Like sharing within your business, [18:54] tracking all of your costs properly,  and making sure you're in alignment [18:59] with other people on your team is more  important now than it's ever been before. [19:09] That's it from Flexera. Thank  you very much. [applause]